1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34d6606683SLinus Torvalds 3563551ae0SDavid Gibson #include <asm/page.h> 36ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3724669e58SAneesh Kumar K.V #include <asm/tlb.h> 3863551ae0SDavid Gibson 3924669e58SAneesh Kumar K.V #include <linux/io.h> 4063551ae0SDavid Gibson #include <linux/hugetlb.h> 419dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 429a305230SLee Schermerhorn #include <linux/node.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 45f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 461da177e4SLinus Torvalds 47c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 48e5ff2159SAndi Kleen unsigned int default_hstate_idx; 49e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 50cf11e85fSRoman Gushchin 51dbda8feaSBarry Song #ifdef CONFIG_CMA 52cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5338e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 54a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 55a01f4390SMike Kravetz { 56a01f4390SMike Kravetz return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, 57a01f4390SMike Kravetz 1 << order); 58a01f4390SMike Kravetz } 59a01f4390SMike Kravetz #else 60a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 61a01f4390SMike Kravetz { 62a01f4390SMike Kravetz return false; 63a01f4390SMike Kravetz } 64dbda8feaSBarry Song #endif 65dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 66cf11e85fSRoman Gushchin 67641844f5SNaoya Horiguchi /* 68641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 69641844f5SNaoya Horiguchi * at boot time. 70641844f5SNaoya Horiguchi */ 71641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 72e5ff2159SAndi Kleen 7353ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7453ba51d2SJon Tollefson 75e5ff2159SAndi Kleen /* for command line parsing */ 76e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 77e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 789fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 79282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 80b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 81e5ff2159SAndi Kleen 823935baa9SDavid Gibson /* 8331caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8431caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 853935baa9SDavid Gibson */ 86c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 870bd0f9fbSEric Paris 888382d914SDavidlohr Bueso /* 898382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 908382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 918382d914SDavidlohr Bueso */ 928382d914SDavidlohr Bueso static int num_fault_mutexes; 93c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 948382d914SDavidlohr Bueso 957ca02d0aSMike Kravetz /* Forward declaration */ 967ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 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 { 2640db9d74eSMina Almasry struct file_region *nrg = NULL; 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 { 346a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 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 3722103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *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); 3812103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 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; 4042103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 405d75c6af9SMina Almasry 4060db9d74eSMina Almasry if (regions_needed) 4070db9d74eSMina Almasry *regions_needed = 0; 408d75c6af9SMina Almasry 4090db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 4100db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 4110db9d74eSMina Almasry * bounds checking. 4120db9d74eSMina Almasry */ 4130db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4140db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4150db9d74eSMina Almasry if (rg->from < f) { 4160db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4170db9d74eSMina Almasry * then we need to update last_accounted_offset. 4180db9d74eSMina Almasry */ 4190db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4200db9d74eSMina Almasry last_accounted_offset = rg->to; 4210db9d74eSMina Almasry continue; 4220db9d74eSMina Almasry } 423d75c6af9SMina Almasry 4240db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4250db9d74eSMina Almasry * finished. 4260db9d74eSMina Almasry */ 427ca7e0457SPeter Xu if (rg->from >= t) 428d75c6af9SMina Almasry break; 429d75c6af9SMina Almasry 4300db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4310db9d74eSMina Almasry * update last_accounted_offset. 432d75c6af9SMina Almasry */ 4332103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4342103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4352103cf9cSPeter Xu last_accounted_offset, 4362103cf9cSPeter Xu rg->from, h, h_cg, 4372103cf9cSPeter Xu regions_needed); 438d75c6af9SMina Almasry 4390db9d74eSMina Almasry last_accounted_offset = rg->to; 4400db9d74eSMina Almasry } 4410db9d74eSMina Almasry 4420db9d74eSMina Almasry /* Handle the case where our range extends beyond 4430db9d74eSMina Almasry * last_accounted_offset. 4440db9d74eSMina Almasry */ 4452103cf9cSPeter Xu if (last_accounted_offset < t) 4462103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4472103cf9cSPeter Xu t, h, h_cg, regions_needed); 4480db9d74eSMina Almasry 4490db9d74eSMina Almasry return add; 4500db9d74eSMina Almasry } 4510db9d74eSMina Almasry 4520db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4530db9d74eSMina Almasry */ 4540db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4550db9d74eSMina Almasry int regions_needed) 4560db9d74eSMina Almasry __must_hold(&resv->lock) 4570db9d74eSMina Almasry { 4580db9d74eSMina Almasry struct list_head allocated_regions; 4590db9d74eSMina Almasry int to_allocate = 0, i = 0; 4600db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4610db9d74eSMina Almasry 4620db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4630db9d74eSMina Almasry 4640db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4650db9d74eSMina Almasry 4660db9d74eSMina Almasry /* 4670db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4680db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4690db9d74eSMina Almasry * 4700db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4710db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4720db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4730db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4740db9d74eSMina Almasry */ 4750db9d74eSMina Almasry while (resv->region_cache_count < 4760db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4770db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4780db9d74eSMina Almasry resv->region_cache_count; 4790db9d74eSMina Almasry 4800db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 481f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4820db9d74eSMina Almasry * needing to allocate for regions_needed. 4830db9d74eSMina Almasry */ 4840db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4850db9d74eSMina Almasry 4860db9d74eSMina Almasry spin_unlock(&resv->lock); 4870db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4880db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4890db9d74eSMina Almasry if (!trg) 4900db9d74eSMina Almasry goto out_of_memory; 4910db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4920db9d74eSMina Almasry } 4930db9d74eSMina Almasry 4940db9d74eSMina Almasry spin_lock(&resv->lock); 4950db9d74eSMina Almasry 496d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 497d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4980db9d74eSMina Almasry } 4990db9d74eSMina Almasry 5000db9d74eSMina Almasry return 0; 5010db9d74eSMina Almasry 5020db9d74eSMina Almasry out_of_memory: 5030db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 504d75c6af9SMina Almasry list_del(&rg->link); 505d75c6af9SMina Almasry kfree(rg); 506d75c6af9SMina Almasry } 5070db9d74eSMina Almasry return -ENOMEM; 508d75c6af9SMina Almasry } 509d75c6af9SMina Almasry 5101dd308a7SMike Kravetz /* 5111dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5120db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5130db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5140db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5150db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5160db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 517cf3ad20bSMike Kravetz * 5180db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5190db9d74eSMina Almasry * 5200db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5210db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5227c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5230db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5240db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5250db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5261dd308a7SMike Kravetz */ 5270db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 528075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 529075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 53096822904SAndy Whitcroft { 5310db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53296822904SAndy Whitcroft 5337b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5340db9d74eSMina Almasry retry: 5350db9d74eSMina Almasry 5360db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 537972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 538972a3da3SWei Yang &actual_regions_needed); 53996822904SAndy Whitcroft 5405e911373SMike Kravetz /* 5410db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5420db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5430db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5440db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5450db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5460db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5470db9d74eSMina Almasry * operation. 5485e911373SMike Kravetz */ 5490db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5500db9d74eSMina Almasry resv->region_cache_count < 5510db9d74eSMina Almasry resv->adds_in_progress + 5520db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5530db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5540db9d74eSMina Almasry * allocate file_region entries. 5550db9d74eSMina Almasry */ 5560db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5575e911373SMike Kravetz 5580db9d74eSMina Almasry if (allocate_file_region_entries( 5590db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5600db9d74eSMina Almasry return -ENOMEM; 5615e911373SMike Kravetz } 5625e911373SMike Kravetz 5630db9d74eSMina Almasry goto retry; 5640db9d74eSMina Almasry } 565cf3ad20bSMike Kravetz 566972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5670db9d74eSMina Almasry 5680db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5690db9d74eSMina Almasry 5707b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 571cf3ad20bSMike Kravetz return add; 57296822904SAndy Whitcroft } 57396822904SAndy Whitcroft 5741dd308a7SMike Kravetz /* 5751dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5761dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5771dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5781dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5791dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5801dd308a7SMike Kravetz * not change the number of huge pages represented by the 5810db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5820db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5830db9d74eSMina Almasry * file_region structure is added. 5840db9d74eSMina Almasry * 5850db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5860db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5870db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5885e911373SMike Kravetz * 5895e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5905e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5915e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5925e911373SMike Kravetz * is needed and can not be allocated. 5931dd308a7SMike Kravetz */ 5940db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5950db9d74eSMina Almasry long *out_regions_needed) 59696822904SAndy Whitcroft { 59796822904SAndy Whitcroft long chg = 0; 59896822904SAndy Whitcroft 5997b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6005e911373SMike Kravetz 601972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 602075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 603972a3da3SWei Yang out_regions_needed); 6045e911373SMike Kravetz 6050db9d74eSMina Almasry if (*out_regions_needed == 0) 6060db9d74eSMina Almasry *out_regions_needed = 1; 6075e911373SMike Kravetz 6080db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6095e911373SMike Kravetz return -ENOMEM; 6105e911373SMike Kravetz 6110db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61296822904SAndy Whitcroft 6137b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61496822904SAndy Whitcroft return chg; 61596822904SAndy Whitcroft } 61696822904SAndy Whitcroft 6171dd308a7SMike Kravetz /* 6185e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6195e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6205e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6215e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6220db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6230db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6240db9d74eSMina Almasry * the adds_in_progress counter. 6255e911373SMike Kravetz * 6265e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6275e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6285e911373SMike Kravetz * arguments will match the associated region_chg call. 6295e911373SMike Kravetz */ 6300db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6310db9d74eSMina Almasry long regions_needed) 6325e911373SMike Kravetz { 6335e911373SMike Kravetz spin_lock(&resv->lock); 6345e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6350db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6365e911373SMike Kravetz spin_unlock(&resv->lock); 6375e911373SMike Kravetz } 6385e911373SMike Kravetz 6395e911373SMike Kravetz /* 640feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 641feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 642feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 643feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 644feba16e2SMike Kravetz * 645feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 646feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 647feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 648feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 649feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 650feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 651feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6521dd308a7SMike Kravetz */ 653feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65496822904SAndy Whitcroft { 6551406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65696822904SAndy Whitcroft struct file_region *rg, *trg; 657feba16e2SMike Kravetz struct file_region *nrg = NULL; 658feba16e2SMike Kravetz long del = 0; 65996822904SAndy Whitcroft 660feba16e2SMike Kravetz retry: 6617b24d861SDavidlohr Bueso spin_lock(&resv->lock); 662feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 663dbe409e4SMike Kravetz /* 664dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 665dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 666dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 667dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 668dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 669dbe409e4SMike Kravetz */ 670dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 671feba16e2SMike Kravetz continue; 672dbe409e4SMike Kravetz 673feba16e2SMike Kravetz if (rg->from >= t) 67496822904SAndy Whitcroft break; 67596822904SAndy Whitcroft 676feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 677feba16e2SMike Kravetz /* 678feba16e2SMike Kravetz * Check for an entry in the cache before dropping 679feba16e2SMike Kravetz * lock and attempting allocation. 680feba16e2SMike Kravetz */ 681feba16e2SMike Kravetz if (!nrg && 682feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 683feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 684feba16e2SMike Kravetz struct file_region, 685feba16e2SMike Kravetz link); 686feba16e2SMike Kravetz list_del(&nrg->link); 687feba16e2SMike Kravetz resv->region_cache_count--; 68896822904SAndy Whitcroft } 68996822904SAndy Whitcroft 690feba16e2SMike Kravetz if (!nrg) { 691feba16e2SMike Kravetz spin_unlock(&resv->lock); 692feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 693feba16e2SMike Kravetz if (!nrg) 694feba16e2SMike Kravetz return -ENOMEM; 695feba16e2SMike Kravetz goto retry; 696feba16e2SMike Kravetz } 697feba16e2SMike Kravetz 698feba16e2SMike Kravetz del += t - f; 69979aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 700d85aecf2SMiaohe Lin resv, rg, t - f, false); 701feba16e2SMike Kravetz 702feba16e2SMike Kravetz /* New entry for end of split region */ 703feba16e2SMike Kravetz nrg->from = t; 704feba16e2SMike Kravetz nrg->to = rg->to; 705075a61d0SMina Almasry 706075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 707075a61d0SMina Almasry 708feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 709feba16e2SMike Kravetz 710feba16e2SMike Kravetz /* Original entry is trimmed */ 711feba16e2SMike Kravetz rg->to = f; 712feba16e2SMike Kravetz 713feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 714feba16e2SMike Kravetz nrg = NULL; 71596822904SAndy Whitcroft break; 716feba16e2SMike Kravetz } 717feba16e2SMike Kravetz 718feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 719feba16e2SMike Kravetz del += rg->to - rg->from; 720075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 721d85aecf2SMiaohe Lin rg->to - rg->from, true); 72296822904SAndy Whitcroft list_del(&rg->link); 72396822904SAndy Whitcroft kfree(rg); 724feba16e2SMike Kravetz continue; 72596822904SAndy Whitcroft } 7267b24d861SDavidlohr Bueso 727feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 728075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 729d85aecf2SMiaohe Lin t - rg->from, false); 730075a61d0SMina Almasry 73179aa925bSMike Kravetz del += t - rg->from; 73279aa925bSMike Kravetz rg->from = t; 73379aa925bSMike Kravetz } else { /* Trim end of region */ 734075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 735d85aecf2SMiaohe Lin rg->to - f, false); 73679aa925bSMike Kravetz 73779aa925bSMike Kravetz del += rg->to - f; 73879aa925bSMike Kravetz rg->to = f; 739feba16e2SMike Kravetz } 740feba16e2SMike Kravetz } 741feba16e2SMike Kravetz 7427b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 743feba16e2SMike Kravetz kfree(nrg); 744feba16e2SMike Kravetz return del; 74596822904SAndy Whitcroft } 74696822904SAndy Whitcroft 7471dd308a7SMike Kravetz /* 748b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 749b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 750b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 751b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 752b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 753b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 754b5cec28dSMike Kravetz * counts. 755b5cec28dSMike Kravetz */ 75672e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 757b5cec28dSMike Kravetz { 758b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 759b5cec28dSMike Kravetz long rsv_adjust; 760da56388cSMiaohe Lin bool reserved = false; 761b5cec28dSMike Kravetz 762b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 763da56388cSMiaohe Lin if (rsv_adjust > 0) { 764b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 765b5cec28dSMike Kravetz 766da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 767da56388cSMiaohe Lin reserved = true; 768da56388cSMiaohe Lin } else if (!rsv_adjust) { 769da56388cSMiaohe Lin reserved = true; 770b5cec28dSMike Kravetz } 771da56388cSMiaohe Lin 772da56388cSMiaohe Lin if (!reserved) 773da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 774b5cec28dSMike Kravetz } 775b5cec28dSMike Kravetz 776b5cec28dSMike Kravetz /* 7771dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7781dd308a7SMike Kravetz * that intersect with the range [f, t). 7791dd308a7SMike Kravetz */ 7801406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78184afd99bSAndy Whitcroft { 7821406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78384afd99bSAndy Whitcroft struct file_region *rg; 78484afd99bSAndy Whitcroft long chg = 0; 78584afd99bSAndy Whitcroft 7867b24d861SDavidlohr Bueso spin_lock(&resv->lock); 78784afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 78884afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 789f2135a4aSWang Sheng-Hui long seg_from; 790f2135a4aSWang Sheng-Hui long seg_to; 79184afd99bSAndy Whitcroft 79284afd99bSAndy Whitcroft if (rg->to <= f) 79384afd99bSAndy Whitcroft continue; 79484afd99bSAndy Whitcroft if (rg->from >= t) 79584afd99bSAndy Whitcroft break; 79684afd99bSAndy Whitcroft 79784afd99bSAndy Whitcroft seg_from = max(rg->from, f); 79884afd99bSAndy Whitcroft seg_to = min(rg->to, t); 79984afd99bSAndy Whitcroft 80084afd99bSAndy Whitcroft chg += seg_to - seg_from; 80184afd99bSAndy Whitcroft } 8027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80384afd99bSAndy Whitcroft 80484afd99bSAndy Whitcroft return chg; 80584afd99bSAndy Whitcroft } 80684afd99bSAndy Whitcroft 80796822904SAndy Whitcroft /* 808e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 809e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 810e7c4b0bfSAndy Whitcroft */ 811a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 812a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 813e7c4b0bfSAndy Whitcroft { 814a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 815a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 816e7c4b0bfSAndy Whitcroft } 817e7c4b0bfSAndy Whitcroft 8180fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8190fe6e20bSNaoya Horiguchi unsigned long address) 8200fe6e20bSNaoya Horiguchi { 8210fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8220fe6e20bSNaoya Horiguchi } 823dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8240fe6e20bSNaoya Horiguchi 82584afd99bSAndy Whitcroft /* 82608fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 82708fba699SMel Gorman * cases this will be same size as used by the page table entries. 82808fba699SMel Gorman */ 82908fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 83008fba699SMel Gorman { 83105ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83205ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83308fba699SMel Gorman return PAGE_SIZE; 83408fba699SMel Gorman } 835f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83608fba699SMel Gorman 83708fba699SMel Gorman /* 8383340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8393340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 84009135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84109135cc5SDan Williams * version of this symbol is required. 8423340289dSMel Gorman */ 84309135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8443340289dSMel Gorman { 8453340289dSMel Gorman return vma_kernel_pagesize(vma); 8463340289dSMel Gorman } 8473340289dSMel Gorman 8483340289dSMel Gorman /* 84984afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 85084afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85184afd99bSAndy Whitcroft * alignment. 85284afd99bSAndy Whitcroft */ 85384afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85484afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85504f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85684afd99bSAndy Whitcroft 857a1e78772SMel Gorman /* 858a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 859a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 860a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 861a1e78772SMel Gorman * 862a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 863a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 864a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 865a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86684afd99bSAndy Whitcroft * 86784afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 86884afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 86984afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 87084afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87184afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87284afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87384afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87484afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 875a1e78772SMel Gorman */ 876e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 877e7c4b0bfSAndy Whitcroft { 878e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 879e7c4b0bfSAndy Whitcroft } 880e7c4b0bfSAndy Whitcroft 881e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 882e7c4b0bfSAndy Whitcroft unsigned long value) 883e7c4b0bfSAndy Whitcroft { 884e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 885e7c4b0bfSAndy Whitcroft } 886e7c4b0bfSAndy Whitcroft 887e9fe92aeSMina Almasry static void 888e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 889e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 890e9fe92aeSMina Almasry struct hstate *h) 891e9fe92aeSMina Almasry { 892e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 893e9fe92aeSMina Almasry if (!h_cg || !h) { 894e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 895e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 896e9fe92aeSMina Almasry resv_map->css = NULL; 897e9fe92aeSMina Almasry } else { 898e9fe92aeSMina Almasry resv_map->reservation_counter = 899e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 900e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 901e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 902e9fe92aeSMina Almasry } 903e9fe92aeSMina Almasry #endif 904e9fe92aeSMina Almasry } 905e9fe92aeSMina Almasry 9069119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 90784afd99bSAndy Whitcroft { 90884afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9095e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9105e911373SMike Kravetz 9115e911373SMike Kravetz if (!resv_map || !rg) { 9125e911373SMike Kravetz kfree(resv_map); 9135e911373SMike Kravetz kfree(rg); 91484afd99bSAndy Whitcroft return NULL; 9155e911373SMike Kravetz } 91684afd99bSAndy Whitcroft 91784afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9187b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 91984afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 92084afd99bSAndy Whitcroft 9215e911373SMike Kravetz resv_map->adds_in_progress = 0; 922e9fe92aeSMina Almasry /* 923e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 924e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 925e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 926e9fe92aeSMina Almasry * reservations are to be un-charged from here. 927e9fe92aeSMina Almasry */ 928e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9295e911373SMike Kravetz 9305e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9315e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9325e911373SMike Kravetz resv_map->region_cache_count = 1; 9335e911373SMike Kravetz 93484afd99bSAndy Whitcroft return resv_map; 93584afd99bSAndy Whitcroft } 93684afd99bSAndy Whitcroft 9379119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 93884afd99bSAndy Whitcroft { 93984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9405e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9415e911373SMike Kravetz struct file_region *rg, *trg; 94284afd99bSAndy Whitcroft 94384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 944feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9455e911373SMike Kravetz 9465e911373SMike Kravetz /* ... and any entries left in the cache */ 9475e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9485e911373SMike Kravetz list_del(&rg->link); 9495e911373SMike Kravetz kfree(rg); 9505e911373SMike Kravetz } 9515e911373SMike Kravetz 9525e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9535e911373SMike Kravetz 95484afd99bSAndy Whitcroft kfree(resv_map); 95584afd99bSAndy Whitcroft } 95684afd99bSAndy Whitcroft 9574e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9584e35f483SJoonsoo Kim { 959f27a5136SMike Kravetz /* 960f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 961f27a5136SMike Kravetz * address space within the inode. This original address space 962f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 963f27a5136SMike Kravetz * address space embedded within the inode. 964f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 965f27a5136SMike Kravetz * this may not be true for device special inodes. 966f27a5136SMike Kravetz */ 967f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9684e35f483SJoonsoo Kim } 9694e35f483SJoonsoo Kim 97084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 971a1e78772SMel Gorman { 97281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9734e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9744e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9754e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9764e35f483SJoonsoo Kim 9774e35f483SJoonsoo Kim return inode_resv_map(inode); 9784e35f483SJoonsoo Kim 9794e35f483SJoonsoo Kim } else { 98084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9824e35f483SJoonsoo Kim } 983a1e78772SMel Gorman } 984a1e78772SMel Gorman 98584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 986a1e78772SMel Gorman { 98781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 989a1e78772SMel Gorman 99084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99204f2cbe3SMel Gorman } 99304f2cbe3SMel Gorman 99404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99504f2cbe3SMel Gorman { 99681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 998e7c4b0bfSAndy Whitcroft 999e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 100004f2cbe3SMel Gorman } 100104f2cbe3SMel Gorman 100204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100304f2cbe3SMel Gorman { 100481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1005e7c4b0bfSAndy Whitcroft 1006e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1007a1e78772SMel Gorman } 1008a1e78772SMel Gorman 100904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 1010a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1011a1e78772SMel Gorman { 101281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1013f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1014a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1015a1e78772SMel Gorman } 1016a1e78772SMel Gorman 1017550a7d60SMina Almasry /* 1018550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1019550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1020550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1021550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1022550a7d60SMina Almasry * reservation. 1023550a7d60SMina Almasry */ 1024550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1025550a7d60SMina Almasry { 1026550a7d60SMina Almasry /* 1027550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1028550a7d60SMina Almasry * It has already been transferred to new_vma. 1029550a7d60SMina Almasry * 1030550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1031550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1032550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1033550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1034550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1035550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1036550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1037550a7d60SMina Almasry */ 1038550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1039550a7d60SMina Almasry 1040afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1041afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1042550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1043afe041c2SBui Quang Minh } 1044550a7d60SMina Almasry 1045550a7d60SMina Almasry reset_vma_resv_huge_pages(vma); 1046550a7d60SMina Almasry } 1047550a7d60SMina Almasry 1048a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1049559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1050a1e78772SMel Gorman { 1051af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1052af0ed73eSJoonsoo Kim /* 1053af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1054af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1055af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1056af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1057af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1058af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1059af0ed73eSJoonsoo Kim * properly, so add work-around here. 1060af0ed73eSJoonsoo Kim */ 1061af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1062559ec2f8SNicholas Krause return true; 1063af0ed73eSJoonsoo Kim else 1064559ec2f8SNicholas Krause return false; 1065af0ed73eSJoonsoo Kim } 1066a63884e9SJoonsoo Kim 1067a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10681fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10691fb1b0e9SMike Kravetz /* 10701fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10711fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10721fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10737c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10741fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10751fb1b0e9SMike Kravetz */ 10761fb1b0e9SMike Kravetz if (chg) 10771fb1b0e9SMike Kravetz return false; 10781fb1b0e9SMike Kravetz else 1079559ec2f8SNicholas Krause return true; 10801fb1b0e9SMike Kravetz } 1081a63884e9SJoonsoo Kim 1082a63884e9SJoonsoo Kim /* 1083a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1084a63884e9SJoonsoo Kim * private mappings. 1085a63884e9SJoonsoo Kim */ 108667961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 108767961f9dSMike Kravetz /* 108867961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 108967961f9dSMike Kravetz * could have been performed on the private mapping. 109067961f9dSMike Kravetz * Examine the value of chg to determine if reserves 109167961f9dSMike Kravetz * actually exist or were previously consumed. 109267961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 109367961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 109467961f9dSMike Kravetz * private mappings has different (opposite) semantics 109567961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 109667961f9dSMike Kravetz * has already taken this difference in semantics into 109767961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 109867961f9dSMike Kravetz * as in the shared case above. Code could easily be 109967961f9dSMike Kravetz * combined, but keeping it separate draws attention to 110067961f9dSMike Kravetz * subtle differences. 110167961f9dSMike Kravetz */ 110267961f9dSMike Kravetz if (chg) 110367961f9dSMike Kravetz return false; 110467961f9dSMike Kravetz else 1105559ec2f8SNicholas Krause return true; 110667961f9dSMike Kravetz } 1107a63884e9SJoonsoo Kim 1108559ec2f8SNicholas Krause return false; 1109a1e78772SMel Gorman } 1110a1e78772SMel Gorman 1111a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11121da177e4SLinus Torvalds { 11131da177e4SLinus Torvalds int nid = page_to_nid(page); 11149487ca60SMike Kravetz 11159487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1116b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1117b65a4edaSMike Kravetz 11180edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1119a5516438SAndi Kleen h->free_huge_pages++; 1120a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11216c037149SMike Kravetz SetHPageFreed(page); 11221da177e4SLinus Torvalds } 11231da177e4SLinus Torvalds 112494310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1125bf50bab2SNaoya Horiguchi { 1126bf50bab2SNaoya Horiguchi struct page *page; 11271a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1128bf50bab2SNaoya Horiguchi 11299487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1130bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11318e3560d9SPavel Tatashin if (pin && !is_pinnable_page(page)) 1132bbe88753SJoonsoo Kim continue; 1133bbe88753SJoonsoo Kim 11346664bfc8SWei Yang if (PageHWPoison(page)) 11356664bfc8SWei Yang continue; 1136bbe88753SJoonsoo Kim 11370edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1138a9869b83SNaoya Horiguchi set_page_refcounted(page); 11396c037149SMike Kravetz ClearHPageFreed(page); 1140bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1141bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1142bf50bab2SNaoya Horiguchi return page; 1143bf50bab2SNaoya Horiguchi } 1144bf50bab2SNaoya Horiguchi 11456664bfc8SWei Yang return NULL; 11466664bfc8SWei Yang } 11476664bfc8SWei Yang 11483e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11493e59fcb0SMichal Hocko nodemask_t *nmask) 115094310cbcSAnshuman Khandual { 11513e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11523e59fcb0SMichal Hocko struct zonelist *zonelist; 11533e59fcb0SMichal Hocko struct zone *zone; 11543e59fcb0SMichal Hocko struct zoneref *z; 115598fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11563e59fcb0SMichal Hocko 11573e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11583e59fcb0SMichal Hocko 11593e59fcb0SMichal Hocko retry_cpuset: 11603e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11613e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 116294310cbcSAnshuman Khandual struct page *page; 116394310cbcSAnshuman Khandual 11643e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11653e59fcb0SMichal Hocko continue; 11663e59fcb0SMichal Hocko /* 11673e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11683e59fcb0SMichal Hocko * zone aware 11693e59fcb0SMichal Hocko */ 11703e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11713e59fcb0SMichal Hocko continue; 11723e59fcb0SMichal Hocko node = zone_to_nid(zone); 117394310cbcSAnshuman Khandual 117494310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 117594310cbcSAnshuman Khandual if (page) 117694310cbcSAnshuman Khandual return page; 117794310cbcSAnshuman Khandual } 11783e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11793e59fcb0SMichal Hocko goto retry_cpuset; 11803e59fcb0SMichal Hocko 118194310cbcSAnshuman Khandual return NULL; 118294310cbcSAnshuman Khandual } 118394310cbcSAnshuman Khandual 1184a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1185a5516438SAndi Kleen struct vm_area_struct *vma, 1186af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1187af0ed73eSJoonsoo Kim long chg) 11881da177e4SLinus Torvalds { 1189cfcaa66fSBen Widawsky struct page *page = NULL; 1190480eccf9SLee Schermerhorn struct mempolicy *mpol; 119104ec6264SVlastimil Babka gfp_t gfp_mask; 11923e59fcb0SMichal Hocko nodemask_t *nodemask; 119304ec6264SVlastimil Babka int nid; 11941da177e4SLinus Torvalds 1195a1e78772SMel Gorman /* 1196a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1197a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1198a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1199a1e78772SMel Gorman */ 1200af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1201a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1202c0ff7453SMiao Xie goto err; 1203a1e78772SMel Gorman 120404f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1205a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 12066eab04a8SJustin P. Mattock goto err; 120704f2cbe3SMel Gorman 120804ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 120904ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1210cfcaa66fSBen Widawsky 1211cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12123e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1213cfcaa66fSBen Widawsky 1214cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1215cfcaa66fSBen Widawsky nodemask = NULL; 1216cfcaa66fSBen Widawsky } 1217cfcaa66fSBen Widawsky 1218cfcaa66fSBen Widawsky if (!page) 1219cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1220cfcaa66fSBen Widawsky 12213e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1222d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1223a63884e9SJoonsoo Kim h->resv_huge_pages--; 1224bf50bab2SNaoya Horiguchi } 1225cc9a6c87SMel Gorman 1226cc9a6c87SMel Gorman mpol_cond_put(mpol); 1227cc9a6c87SMel Gorman return page; 1228cc9a6c87SMel Gorman 1229c0ff7453SMiao Xie err: 1230cc9a6c87SMel Gorman return NULL; 12311da177e4SLinus Torvalds } 12321da177e4SLinus Torvalds 12331cac6f2cSLuiz Capitulino /* 12341cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12351cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12361cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12371cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12381cac6f2cSLuiz Capitulino * node for alloc or free. 12391cac6f2cSLuiz Capitulino */ 12401cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12411cac6f2cSLuiz Capitulino { 12420edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12431cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12441cac6f2cSLuiz Capitulino 12451cac6f2cSLuiz Capitulino return nid; 12461cac6f2cSLuiz Capitulino } 12471cac6f2cSLuiz Capitulino 12481cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12491cac6f2cSLuiz Capitulino { 12501cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12511cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12521cac6f2cSLuiz Capitulino return nid; 12531cac6f2cSLuiz Capitulino } 12541cac6f2cSLuiz Capitulino 12551cac6f2cSLuiz Capitulino /* 12561cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12571cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12581cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12591cac6f2cSLuiz Capitulino * mask. 12601cac6f2cSLuiz Capitulino */ 12611cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12621cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12631cac6f2cSLuiz Capitulino { 12641cac6f2cSLuiz Capitulino int nid; 12651cac6f2cSLuiz Capitulino 12661cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12671cac6f2cSLuiz Capitulino 12681cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12691cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12701cac6f2cSLuiz Capitulino 12711cac6f2cSLuiz Capitulino return nid; 12721cac6f2cSLuiz Capitulino } 12731cac6f2cSLuiz Capitulino 12741cac6f2cSLuiz Capitulino /* 127510c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12761cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12771cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12781cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12791cac6f2cSLuiz Capitulino */ 12801cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12811cac6f2cSLuiz Capitulino { 12821cac6f2cSLuiz Capitulino int nid; 12831cac6f2cSLuiz Capitulino 12841cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12851cac6f2cSLuiz Capitulino 12861cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12871cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12881cac6f2cSLuiz Capitulino 12891cac6f2cSLuiz Capitulino return nid; 12901cac6f2cSLuiz Capitulino } 12911cac6f2cSLuiz Capitulino 12921cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12931cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12941cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12951cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12961cac6f2cSLuiz Capitulino nr_nodes--) 12971cac6f2cSLuiz Capitulino 12981cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12991cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13001cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13011cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13021cac6f2cSLuiz Capitulino nr_nodes--) 13031cac6f2cSLuiz Capitulino 13048531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 130534d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 130634d9e35bSMike Kravetz unsigned int order, bool demote) 1307944d9fecSLuiz Capitulino { 1308944d9fecSLuiz Capitulino int i; 1309944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1310944d9fecSLuiz Capitulino struct page *p = page + 1; 1311944d9fecSLuiz Capitulino 1312c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 131347e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 131447e29d32SJohn Hubbard 1315944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1316a01f4390SMike Kravetz p->mapping = NULL; 13171d798ca3SKirill A. Shutemov clear_compound_head(p); 131834d9e35bSMike Kravetz if (!demote) 1319944d9fecSLuiz Capitulino set_page_refcounted(p); 1320944d9fecSLuiz Capitulino } 1321944d9fecSLuiz Capitulino 1322944d9fecSLuiz Capitulino set_compound_order(page, 0); 1323ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1324944d9fecSLuiz Capitulino __ClearPageHead(page); 1325944d9fecSLuiz Capitulino } 1326944d9fecSLuiz Capitulino 13278531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13288531fc6fSMike Kravetz unsigned int order) 13298531fc6fSMike Kravetz { 13308531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13318531fc6fSMike Kravetz } 13328531fc6fSMike Kravetz 13338531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 133434d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 133534d9e35bSMike Kravetz unsigned int order) 133634d9e35bSMike Kravetz { 133734d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 133834d9e35bSMike Kravetz } 133934d9e35bSMike Kravetz 1340d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1341944d9fecSLuiz Capitulino { 1342cf11e85fSRoman Gushchin /* 1343cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1344cf11e85fSRoman Gushchin * cma_release() returns false. 1345cf11e85fSRoman Gushchin */ 1346dbda8feaSBarry Song #ifdef CONFIG_CMA 1347dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1348cf11e85fSRoman Gushchin return; 1349dbda8feaSBarry Song #endif 1350cf11e85fSRoman Gushchin 1351944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1352944d9fecSLuiz Capitulino } 1353944d9fecSLuiz Capitulino 13544eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1355d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1356d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1357944d9fecSLuiz Capitulino { 135804adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1359953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1360953f064aSLi Xinhai nid = numa_mem_id(); 1361944d9fecSLuiz Capitulino 1362dbda8feaSBarry Song #ifdef CONFIG_CMA 1363dbda8feaSBarry Song { 1364cf11e85fSRoman Gushchin struct page *page; 1365cf11e85fSRoman Gushchin int node; 1366cf11e85fSRoman Gushchin 1367953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1368953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1369953f064aSLi Xinhai huge_page_order(h), true); 1370953f064aSLi Xinhai if (page) 1371953f064aSLi Xinhai return page; 1372953f064aSLi Xinhai } 1373953f064aSLi Xinhai 1374953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1375cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1376953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1377cf11e85fSRoman Gushchin continue; 1378cf11e85fSRoman Gushchin 1379cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1380cf11e85fSRoman Gushchin huge_page_order(h), true); 1381cf11e85fSRoman Gushchin if (page) 1382cf11e85fSRoman Gushchin return page; 1383cf11e85fSRoman Gushchin } 1384cf11e85fSRoman Gushchin } 1385953f064aSLi Xinhai } 1386dbda8feaSBarry Song #endif 1387cf11e85fSRoman Gushchin 13885e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1389944d9fecSLuiz Capitulino } 1390944d9fecSLuiz Capitulino 13914eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13924eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13934eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13944eb0716eSAlexandre Ghiti { 13954eb0716eSAlexandre Ghiti return NULL; 13964eb0716eSAlexandre Ghiti } 13974eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1398944d9fecSLuiz Capitulino 1399e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1400d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14014eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14024eb0716eSAlexandre Ghiti { 14034eb0716eSAlexandre Ghiti return NULL; 14044eb0716eSAlexandre Ghiti } 1405d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1406944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1407d00181b9SKirill A. Shutemov unsigned int order) { } 1408944d9fecSLuiz Capitulino #endif 1409944d9fecSLuiz Capitulino 14106eb4e88aSMike Kravetz /* 14116eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 141234d9e35bSMike Kravetz * as just a compound page. 141334d9e35bSMike Kravetz * 141434d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14156eb4e88aSMike Kravetz * 14166eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14176eb4e88aSMike Kravetz */ 141834d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 141934d9e35bSMike Kravetz bool adjust_surplus, 142034d9e35bSMike Kravetz bool demote) 14216eb4e88aSMike Kravetz { 14226eb4e88aSMike Kravetz int nid = page_to_nid(page); 14236eb4e88aSMike Kravetz 14246eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14256eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14266eb4e88aSMike Kravetz 14279487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14286eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14296eb4e88aSMike Kravetz return; 14306eb4e88aSMike Kravetz 14316eb4e88aSMike Kravetz list_del(&page->lru); 14326eb4e88aSMike Kravetz 14336eb4e88aSMike Kravetz if (HPageFreed(page)) { 14346eb4e88aSMike Kravetz h->free_huge_pages--; 14356eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14366eb4e88aSMike Kravetz } 14376eb4e88aSMike Kravetz if (adjust_surplus) { 14386eb4e88aSMike Kravetz h->surplus_huge_pages--; 14396eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14406eb4e88aSMike Kravetz } 14416eb4e88aSMike Kravetz 1442e32d20c0SMike Kravetz /* 1443e32d20c0SMike Kravetz * Very subtle 1444e32d20c0SMike Kravetz * 1445e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1446e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1447e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1448e32d20c0SMike Kravetz * their reference. 1449e32d20c0SMike Kravetz * 1450e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1451e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1452e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1453e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1454e32d20c0SMike Kravetz * apply. 1455e32d20c0SMike Kravetz * 1456e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1457e32d20c0SMike Kravetz * after update_and_free_page is called. 145834d9e35bSMike Kravetz * 145934d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 146034d9e35bSMike Kravetz * be turned into a page of smaller size. 1461e32d20c0SMike Kravetz */ 146234d9e35bSMike Kravetz if (!demote) 14636eb4e88aSMike Kravetz set_page_refcounted(page); 1464e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14656eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1466e32d20c0SMike Kravetz else 1467e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14686eb4e88aSMike Kravetz 14696eb4e88aSMike Kravetz h->nr_huge_pages--; 14706eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14716eb4e88aSMike Kravetz } 14726eb4e88aSMike Kravetz 147334d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 147434d9e35bSMike Kravetz bool adjust_surplus) 147534d9e35bSMike Kravetz { 147634d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 147734d9e35bSMike Kravetz } 147834d9e35bSMike Kravetz 14798531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14808531fc6fSMike Kravetz bool adjust_surplus) 14818531fc6fSMike Kravetz { 14828531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 14838531fc6fSMike Kravetz } 14848531fc6fSMike Kravetz 1485ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1486ad2fa371SMuchun Song bool adjust_surplus) 1487ad2fa371SMuchun Song { 1488ad2fa371SMuchun Song int zeroed; 1489ad2fa371SMuchun Song int nid = page_to_nid(page); 1490ad2fa371SMuchun Song 1491ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1492ad2fa371SMuchun Song 1493ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1494ad2fa371SMuchun Song 1495ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1496ad2fa371SMuchun Song h->nr_huge_pages++; 1497ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1498ad2fa371SMuchun Song 1499ad2fa371SMuchun Song if (adjust_surplus) { 1500ad2fa371SMuchun Song h->surplus_huge_pages++; 1501ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1502ad2fa371SMuchun Song } 1503ad2fa371SMuchun Song 1504ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1505ad2fa371SMuchun Song set_page_private(page, 0); 1506ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1507ad2fa371SMuchun Song 1508ad2fa371SMuchun Song /* 1509b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1510b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1511b65a4edaSMike Kravetz * just in case. 1512ad2fa371SMuchun Song */ 1513ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1514b65a4edaSMike Kravetz if (!zeroed) 1515b65a4edaSMike Kravetz /* 1516b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1517b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1518b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1519b65a4edaSMike Kravetz * when this other ref is dropped. 1520b65a4edaSMike Kravetz */ 1521b65a4edaSMike Kravetz return; 1522b65a4edaSMike Kravetz 1523ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1524ad2fa371SMuchun Song enqueue_huge_page(h, page); 1525ad2fa371SMuchun Song } 1526ad2fa371SMuchun Song 1527b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15286af2acb6SAdam Litke { 15296af2acb6SAdam Litke int i; 1530dbfee5aeSMike Kravetz struct page *subpage = page; 1531a5516438SAndi Kleen 15324eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1533944d9fecSLuiz Capitulino return; 153418229df5SAndy Whitcroft 1535ad2fa371SMuchun Song if (alloc_huge_page_vmemmap(h, page)) { 1536ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1537ad2fa371SMuchun Song /* 1538ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1539ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1540ad2fa371SMuchun Song * as a surplus page. 1541ad2fa371SMuchun Song */ 1542ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1543ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1544ad2fa371SMuchun Song return; 1545ad2fa371SMuchun Song } 1546ad2fa371SMuchun Song 1547dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1548dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1549dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 155032f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1551a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1552a7407a27SLuiz Capitulino 1 << PG_writeback); 15536af2acb6SAdam Litke } 1554a01f4390SMike Kravetz 1555a01f4390SMike Kravetz /* 1556a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1557a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1558a01f4390SMike Kravetz */ 1559a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1560a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1561944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1562944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1563944d9fecSLuiz Capitulino } else { 1564a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 15656af2acb6SAdam Litke } 1566944d9fecSLuiz Capitulino } 15676af2acb6SAdam Litke 1568b65d4adbSMuchun Song /* 1569b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1570b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1571b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1572b65d4adbSMuchun Song * the vmemmap pages. 1573b65d4adbSMuchun Song * 1574b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1575b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1576b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1577b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1578b65d4adbSMuchun Song */ 1579b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1580b65d4adbSMuchun Song 1581b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1582b65d4adbSMuchun Song { 1583b65d4adbSMuchun Song struct llist_node *node; 1584b65d4adbSMuchun Song 1585b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1586b65d4adbSMuchun Song 1587b65d4adbSMuchun Song while (node) { 1588b65d4adbSMuchun Song struct page *page; 1589b65d4adbSMuchun Song struct hstate *h; 1590b65d4adbSMuchun Song 1591b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1592b65d4adbSMuchun Song struct page, mapping); 1593b65d4adbSMuchun Song node = node->next; 1594b65d4adbSMuchun Song page->mapping = NULL; 1595b65d4adbSMuchun Song /* 1596b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1597b65d4adbSMuchun Song * is going to trigger because a previous call to 1598b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1599b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1600b65d4adbSMuchun Song */ 1601b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1602b65d4adbSMuchun Song 1603b65d4adbSMuchun Song __update_and_free_page(h, page); 1604b65d4adbSMuchun Song 1605b65d4adbSMuchun Song cond_resched(); 1606b65d4adbSMuchun Song } 1607b65d4adbSMuchun Song } 1608b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1609b65d4adbSMuchun Song 1610b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1611b65d4adbSMuchun Song { 1612b65d4adbSMuchun Song if (free_vmemmap_pages_per_hpage(h)) 1613b65d4adbSMuchun Song flush_work(&free_hpage_work); 1614b65d4adbSMuchun Song } 1615b65d4adbSMuchun Song 1616b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1617b65d4adbSMuchun Song bool atomic) 1618b65d4adbSMuchun Song { 1619ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1620b65d4adbSMuchun Song __update_and_free_page(h, page); 1621b65d4adbSMuchun Song return; 1622b65d4adbSMuchun Song } 1623b65d4adbSMuchun Song 1624b65d4adbSMuchun Song /* 1625b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1626b65d4adbSMuchun Song * 1627b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1628b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1629b65d4adbSMuchun Song * hasn't retrieved the list yet. 1630b65d4adbSMuchun Song */ 1631b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1632b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1633b65d4adbSMuchun Song } 1634b65d4adbSMuchun Song 163510c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 163610c6ec49SMike Kravetz { 163710c6ec49SMike Kravetz struct page *page, *t_page; 163810c6ec49SMike Kravetz 163910c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1640b65d4adbSMuchun Song update_and_free_page(h, page, false); 164110c6ec49SMike Kravetz cond_resched(); 164210c6ec49SMike Kravetz } 164310c6ec49SMike Kravetz } 164410c6ec49SMike Kravetz 1645e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1646e5ff2159SAndi Kleen { 1647e5ff2159SAndi Kleen struct hstate *h; 1648e5ff2159SAndi Kleen 1649e5ff2159SAndi Kleen for_each_hstate(h) { 1650e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1651e5ff2159SAndi Kleen return h; 1652e5ff2159SAndi Kleen } 1653e5ff2159SAndi Kleen return NULL; 1654e5ff2159SAndi Kleen } 1655e5ff2159SAndi Kleen 1656db71ef79SMike Kravetz void free_huge_page(struct page *page) 165727a85ef1SDavid Gibson { 1658a5516438SAndi Kleen /* 1659a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1660a5516438SAndi Kleen * compound page destructor. 1661a5516438SAndi Kleen */ 1662e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16637893d1d5SAdam Litke int nid = page_to_nid(page); 1664d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 166507443a85SJoonsoo Kim bool restore_reserve; 1666db71ef79SMike Kravetz unsigned long flags; 166727a85ef1SDavid Gibson 1668b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1669b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 16708ace22bcSYongkai Wu 1671d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 16728ace22bcSYongkai Wu page->mapping = NULL; 1673d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1674d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 167527a85ef1SDavid Gibson 16761c5ecae3SMike Kravetz /* 1677d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 16780919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 16790919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 16800919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 16816c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 16820919e1b6SMike Kravetz * remove the reserved page from the subpool. 16830919e1b6SMike Kravetz */ 16840919e1b6SMike Kravetz if (!restore_reserve) { 16850919e1b6SMike Kravetz /* 16860919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 16870919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 16880919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 16890919e1b6SMike Kravetz * operation. 16901c5ecae3SMike Kravetz */ 16911c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 16921c5ecae3SMike Kravetz restore_reserve = true; 16930919e1b6SMike Kravetz } 16941c5ecae3SMike Kravetz 1695db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 16968f251a3dSMike Kravetz ClearHPageMigratable(page); 16976d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 16986d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 169908cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 170008cf9fafSMina Almasry pages_per_huge_page(h), page); 170107443a85SJoonsoo Kim if (restore_reserve) 170207443a85SJoonsoo Kim h->resv_huge_pages++; 170307443a85SJoonsoo Kim 17049157c311SMike Kravetz if (HPageTemporary(page)) { 17056eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1706db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1707b65d4adbSMuchun Song update_and_free_page(h, page, true); 1708ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17090edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17106eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1711db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1712b65d4adbSMuchun Song update_and_free_page(h, page, true); 17137893d1d5SAdam Litke } else { 17145d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1715a5516438SAndi Kleen enqueue_huge_page(h, page); 1716db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 171727a85ef1SDavid Gibson } 17181121828aSMike Kravetz } 171927a85ef1SDavid Gibson 1720d3d99fccSOscar Salvador /* 1721d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1722d3d99fccSOscar Salvador */ 1723d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1724d3d99fccSOscar Salvador { 1725d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1726d3d99fccSOscar Salvador h->nr_huge_pages++; 1727d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1728d3d99fccSOscar Salvador } 1729d3d99fccSOscar Salvador 1730f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1731b7ba30c6SAndi Kleen { 1732f41f2ed4SMuchun Song free_huge_page_vmemmap(h, page); 17330edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1734f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1735ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17369dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17371adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1738d3d99fccSOscar Salvador } 1739d3d99fccSOscar Salvador 1740d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1741d3d99fccSOscar Salvador { 1742f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1743db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1744d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1745db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1746b7ba30c6SAndi Kleen } 1747b7ba30c6SAndi Kleen 174834d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 174934d9e35bSMike Kravetz bool demote) 175020a0307cSWu Fengguang { 17517118fc29SMike Kravetz int i, j; 175220a0307cSWu Fengguang int nr_pages = 1 << order; 175320a0307cSWu Fengguang struct page *p = page + 1; 175420a0307cSWu Fengguang 175520a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 175620a0307cSWu Fengguang set_compound_order(page, order); 1757ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1758de09d31dSKirill A. Shutemov __SetPageHead(page); 175920a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1760ef5a22beSAndrea Arcangeli /* 1761ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1762ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1763ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 17647c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1765ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1766ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1767ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1768ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1769ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1770ef5a22beSAndrea Arcangeli * after get_user_pages(). 1771ef5a22beSAndrea Arcangeli */ 1772ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 17737118fc29SMike Kravetz /* 17747118fc29SMike Kravetz * Subtle and very unlikely 17757118fc29SMike Kravetz * 17767118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 17777118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 17787118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 17797118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 17807118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 17817118fc29SMike Kravetz * We need to respect any increased ref count, and only set 17827118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1783416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1784416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1785416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1786416d85edSMike Kravetz * pages using the appropriate free interface. 178734d9e35bSMike Kravetz * 178834d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 17897118fc29SMike Kravetz */ 179034d9e35bSMike Kravetz if (!demote) { 17917118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1792416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 17937118fc29SMike Kravetz goto out_error; 17947118fc29SMike Kravetz } 179534d9e35bSMike Kravetz } else { 179634d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 179734d9e35bSMike Kravetz } 17981d798ca3SKirill A. Shutemov set_compound_head(p, page); 179920a0307cSWu Fengguang } 1800b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 180147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18027118fc29SMike Kravetz return true; 18037118fc29SMike Kravetz 18047118fc29SMike Kravetz out_error: 18057118fc29SMike Kravetz /* undo tail page modifications made above */ 18067118fc29SMike Kravetz p = page + 1; 18077118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 18087118fc29SMike Kravetz clear_compound_head(p); 18097118fc29SMike Kravetz set_page_refcounted(p); 18107118fc29SMike Kravetz } 18117118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 18127118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 18137118fc29SMike Kravetz __ClearPageReserved(p); 18147118fc29SMike Kravetz set_compound_order(page, 0); 18157118fc29SMike Kravetz page[1].compound_nr = 0; 18167118fc29SMike Kravetz __ClearPageHead(page); 18177118fc29SMike Kravetz return false; 181820a0307cSWu Fengguang } 181920a0307cSWu Fengguang 182034d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 182134d9e35bSMike Kravetz { 182234d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 182334d9e35bSMike Kravetz } 182434d9e35bSMike Kravetz 18258531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18268531fc6fSMike Kravetz unsigned int order) 18278531fc6fSMike Kravetz { 18288531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18298531fc6fSMike Kravetz } 18308531fc6fSMike Kravetz 18317795912cSAndrew Morton /* 18327795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18337795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18347795912cSAndrew Morton * details. 18357795912cSAndrew Morton */ 183620a0307cSWu Fengguang int PageHuge(struct page *page) 183720a0307cSWu Fengguang { 183820a0307cSWu Fengguang if (!PageCompound(page)) 183920a0307cSWu Fengguang return 0; 184020a0307cSWu Fengguang 184120a0307cSWu Fengguang page = compound_head(page); 1842f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 184320a0307cSWu Fengguang } 184443131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 184543131e14SNaoya Horiguchi 184627c73ae7SAndrea Arcangeli /* 184727c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 184827c73ae7SAndrea Arcangeli * normal or transparent huge pages. 184927c73ae7SAndrea Arcangeli */ 185027c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 185127c73ae7SAndrea Arcangeli { 185227c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 185327c73ae7SAndrea Arcangeli return 0; 185427c73ae7SAndrea Arcangeli 1855d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 185627c73ae7SAndrea Arcangeli } 185727c73ae7SAndrea Arcangeli 1858c0d0381aSMike Kravetz /* 1859c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1860c0d0381aSMike Kravetz * 1861336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1862336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1863336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1864c0d0381aSMike Kravetz */ 1865c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1866c0d0381aSMike Kravetz { 1867336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1868c0d0381aSMike Kravetz 1869c0d0381aSMike Kravetz if (!mapping) 1870c0d0381aSMike Kravetz return mapping; 1871c0d0381aSMike Kravetz 1872c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1873c0d0381aSMike Kravetz return mapping; 1874c0d0381aSMike Kravetz 1875c0d0381aSMike Kravetz return NULL; 1876c0d0381aSMike Kravetz } 1877c0d0381aSMike Kravetz 1878fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 187913d60f4bSZhang Yi { 188013d60f4bSZhang Yi struct page *page_head = compound_head(page); 188113d60f4bSZhang Yi pgoff_t index = page_index(page_head); 188213d60f4bSZhang Yi unsigned long compound_idx; 188313d60f4bSZhang Yi 188413d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 188513d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 188613d60f4bSZhang Yi else 188713d60f4bSZhang Yi compound_idx = page - page_head; 188813d60f4bSZhang Yi 188913d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 189013d60f4bSZhang Yi } 189113d60f4bSZhang Yi 18920c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1893f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1894f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 18951da177e4SLinus Torvalds { 1896af0fb9dfSMichal Hocko int order = huge_page_order(h); 18971da177e4SLinus Torvalds struct page *page; 1898f60858f9SMike Kravetz bool alloc_try_hard = true; 1899f96efd58SJoe Jin 1900f60858f9SMike Kravetz /* 1901f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1902f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1903f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1904f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1905f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1906f60858f9SMike Kravetz */ 1907f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1908f60858f9SMike Kravetz alloc_try_hard = false; 1909f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1910f60858f9SMike Kravetz if (alloc_try_hard) 1911f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1912af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1913af0fb9dfSMichal Hocko nid = numa_mem_id(); 191484172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1915af0fb9dfSMichal Hocko if (page) 1916af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1917af0fb9dfSMichal Hocko else 1918af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 191963b4613cSNishanth Aravamudan 1920f60858f9SMike Kravetz /* 1921f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1922f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1923f60858f9SMike Kravetz * normal 'try hard' allocations. 1924f60858f9SMike Kravetz */ 1925f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1926f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1927f60858f9SMike Kravetz 1928f60858f9SMike Kravetz /* 1929f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1930f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1931f60858f9SMike Kravetz * overall state change. 1932f60858f9SMike Kravetz */ 1933f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1934f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1935f60858f9SMike Kravetz 193663b4613cSNishanth Aravamudan return page; 193763b4613cSNishanth Aravamudan } 193863b4613cSNishanth Aravamudan 1939af0fb9dfSMichal Hocko /* 19400c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 19410c397daeSMichal Hocko * should use this function to get new hugetlb pages 19420c397daeSMichal Hocko */ 19430c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1944f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1945f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19460c397daeSMichal Hocko { 19470c397daeSMichal Hocko struct page *page; 19487118fc29SMike Kravetz bool retry = false; 19490c397daeSMichal Hocko 19507118fc29SMike Kravetz retry: 19510c397daeSMichal Hocko if (hstate_is_gigantic(h)) 19520c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 19530c397daeSMichal Hocko else 19540c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1955f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 19560c397daeSMichal Hocko if (!page) 19570c397daeSMichal Hocko return NULL; 19580c397daeSMichal Hocko 19597118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 19607118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 19617118fc29SMike Kravetz /* 19627118fc29SMike Kravetz * Rare failure to convert pages to compound page. 19637118fc29SMike Kravetz * Free pages and try again - ONCE! 19647118fc29SMike Kravetz */ 19657118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 19667118fc29SMike Kravetz if (!retry) { 19677118fc29SMike Kravetz retry = true; 19687118fc29SMike Kravetz goto retry; 19697118fc29SMike Kravetz } 19707118fc29SMike Kravetz return NULL; 19717118fc29SMike Kravetz } 19727118fc29SMike Kravetz } 19730c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 19740c397daeSMichal Hocko 19750c397daeSMichal Hocko return page; 19760c397daeSMichal Hocko } 19770c397daeSMichal Hocko 19780c397daeSMichal Hocko /* 1979af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1980af0fb9dfSMichal Hocko * manner. 1981af0fb9dfSMichal Hocko */ 1982f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1983f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1984b2261026SJoonsoo Kim { 1985b2261026SJoonsoo Kim struct page *page; 1986b2261026SJoonsoo Kim int nr_nodes, node; 1987af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1988b2261026SJoonsoo Kim 1989b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1990f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1991f60858f9SMike Kravetz node_alloc_noretry); 1992af0fb9dfSMichal Hocko if (page) 1993b2261026SJoonsoo Kim break; 1994b2261026SJoonsoo Kim } 1995b2261026SJoonsoo Kim 1996af0fb9dfSMichal Hocko if (!page) 1997af0fb9dfSMichal Hocko return 0; 1998b2261026SJoonsoo Kim 1999af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2000af0fb9dfSMichal Hocko 2001af0fb9dfSMichal Hocko return 1; 2002b2261026SJoonsoo Kim } 2003b2261026SJoonsoo Kim 2004e8c5c824SLee Schermerhorn /* 200510c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 200610c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 200710c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 200810c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2009e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2010e8c5c824SLee Schermerhorn */ 201110c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 201210c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20136ae11b27SLee Schermerhorn bool acct_surplus) 2014e8c5c824SLee Schermerhorn { 2015b2261026SJoonsoo Kim int nr_nodes, node; 201610c6ec49SMike Kravetz struct page *page = NULL; 2017e8c5c824SLee Schermerhorn 20189487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2019b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2020685f3457SLee Schermerhorn /* 2021685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2022685f3457SLee Schermerhorn * nodes with surplus pages. 2023685f3457SLee Schermerhorn */ 2024b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2025b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 202610c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2027e8c5c824SLee Schermerhorn struct page, lru); 20286eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20299a76db09SLee Schermerhorn break; 2030e8c5c824SLee Schermerhorn } 2031b2261026SJoonsoo Kim } 2032e8c5c824SLee Schermerhorn 203310c6ec49SMike Kravetz return page; 2034e8c5c824SLee Schermerhorn } 2035e8c5c824SLee Schermerhorn 2036c8721bbbSNaoya Horiguchi /* 2037c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2038faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2039faf53defSNaoya Horiguchi * This function returns values like below: 2040faf53defSNaoya Horiguchi * 2041ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2042ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2043ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2044ad2fa371SMuchun Song * is enabled. 2045faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2046faf53defSNaoya Horiguchi * (allocated or reserved.) 2047faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2048faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2049c8721bbbSNaoya Horiguchi */ 2050c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2051c8721bbbSNaoya Horiguchi { 20526bc9b564SNaoya Horiguchi int rc = -EBUSY; 2053082d5b6bSGerald Schaefer 20547ffddd49SMuchun Song retry: 2055faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2056faf53defSNaoya Horiguchi if (!PageHuge(page)) 2057faf53defSNaoya Horiguchi return 0; 2058faf53defSNaoya Horiguchi 2059db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2060faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2061faf53defSNaoya Horiguchi rc = 0; 2062faf53defSNaoya Horiguchi goto out; 2063faf53defSNaoya Horiguchi } 2064faf53defSNaoya Horiguchi 2065faf53defSNaoya Horiguchi if (!page_count(page)) { 20662247bb33SGerald Schaefer struct page *head = compound_head(page); 20672247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 20686bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 2069082d5b6bSGerald Schaefer goto out; 20707ffddd49SMuchun Song 20717ffddd49SMuchun Song /* 20727ffddd49SMuchun Song * We should make sure that the page is already on the free list 20737ffddd49SMuchun Song * when it is dissolved. 20747ffddd49SMuchun Song */ 20756c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2076db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20777ffddd49SMuchun Song cond_resched(); 20787ffddd49SMuchun Song 20797ffddd49SMuchun Song /* 20807ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 20817ffddd49SMuchun Song * encounter this race. In fact, we have a chance 20827ffddd49SMuchun Song * to successfully dissolve the page if we do a 20837ffddd49SMuchun Song * retry. Because the race window is quite small. 20847ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 20857ffddd49SMuchun Song * for increasing the success rate of dissolving page. 20867ffddd49SMuchun Song */ 20877ffddd49SMuchun Song goto retry; 20887ffddd49SMuchun Song } 20897ffddd49SMuchun Song 2090ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2091ad2fa371SMuchun Song h->max_huge_pages--; 2092ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2093ad2fa371SMuchun Song 2094c3114a84SAnshuman Khandual /* 2095ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2096ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2097ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2098ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2099ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2100ad2fa371SMuchun Song * appropriate action on failure. 2101ad2fa371SMuchun Song */ 2102ad2fa371SMuchun Song rc = alloc_huge_page_vmemmap(h, head); 2103ad2fa371SMuchun Song if (!rc) { 2104ad2fa371SMuchun Song /* 2105ad2fa371SMuchun Song * Move PageHWPoison flag from head page to the raw 2106ad2fa371SMuchun Song * error page, which makes any subpages rather than 2107ad2fa371SMuchun Song * the error page reusable. 2108c3114a84SAnshuman Khandual */ 2109c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 2110c3114a84SAnshuman Khandual SetPageHWPoison(page); 2111c3114a84SAnshuman Khandual ClearPageHWPoison(head); 2112c3114a84SAnshuman Khandual } 2113b65d4adbSMuchun Song update_and_free_page(h, head, false); 2114ad2fa371SMuchun Song } else { 2115ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2116ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2117ad2fa371SMuchun Song h->max_huge_pages++; 2118ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2119ad2fa371SMuchun Song } 2120ad2fa371SMuchun Song 2121ad2fa371SMuchun Song return rc; 2122c8721bbbSNaoya Horiguchi } 2123082d5b6bSGerald Schaefer out: 2124db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2125082d5b6bSGerald Schaefer return rc; 2126c8721bbbSNaoya Horiguchi } 2127c8721bbbSNaoya Horiguchi 2128c8721bbbSNaoya Horiguchi /* 2129c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2130c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21312247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21322247bb33SGerald Schaefer * part of it lies within the given range. 2133082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2134082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2135c8721bbbSNaoya Horiguchi */ 2136082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2137c8721bbbSNaoya Horiguchi { 2138c8721bbbSNaoya Horiguchi unsigned long pfn; 2139eb03aa00SGerald Schaefer struct page *page; 2140082d5b6bSGerald Schaefer int rc = 0; 2141c8721bbbSNaoya Horiguchi 2142d0177639SLi Zhong if (!hugepages_supported()) 2143082d5b6bSGerald Schaefer return rc; 2144d0177639SLi Zhong 2145eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 2146eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2147eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2148eb03aa00SGerald Schaefer if (rc) 2149082d5b6bSGerald Schaefer break; 2150eb03aa00SGerald Schaefer } 2151082d5b6bSGerald Schaefer 2152082d5b6bSGerald Schaefer return rc; 2153c8721bbbSNaoya Horiguchi } 2154c8721bbbSNaoya Horiguchi 2155ab5ac90aSMichal Hocko /* 2156ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2157ab5ac90aSMichal Hocko */ 21580c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2159b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 21607893d1d5SAdam Litke { 21619980d744SMichal Hocko struct page *page = NULL; 2162b65a4edaSMike Kravetz bool retry = false; 21637893d1d5SAdam Litke 2164bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2165aa888a74SAndi Kleen return NULL; 2166aa888a74SAndi Kleen 2167db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21689980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 21699980d744SMichal Hocko goto out_unlock; 2170db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2171d1c3fb1fSNishanth Aravamudan 2172b65a4edaSMike Kravetz retry: 2173f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 21749980d744SMichal Hocko if (!page) 21750c397daeSMichal Hocko return NULL; 2176d1c3fb1fSNishanth Aravamudan 2177db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21789980d744SMichal Hocko /* 21799980d744SMichal Hocko * We could have raced with the pool size change. 21809980d744SMichal Hocko * Double check that and simply deallocate the new page 21819980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 21829980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 21839980d744SMichal Hocko * codeflow 21849980d744SMichal Hocko */ 21859980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 21869157c311SMike Kravetz SetHPageTemporary(page); 2187db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21889980d744SMichal Hocko put_page(page); 21892bf753e6SKai Shen return NULL; 2190b65a4edaSMike Kravetz } 2191b65a4edaSMike Kravetz 2192b65a4edaSMike Kravetz if (zero_ref) { 2193b65a4edaSMike Kravetz /* 2194b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2195b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2196b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2197b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2198b65a4edaSMike Kravetz */ 2199b65a4edaSMike Kravetz SetHPageTemporary(page); 2200b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2201b65a4edaSMike Kravetz /* 2202b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2203b65a4edaSMike Kravetz * huge. Retry once. 2204b65a4edaSMike Kravetz */ 2205b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2206b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2207b65a4edaSMike Kravetz if (retry) 2208b65a4edaSMike Kravetz return NULL; 2209b65a4edaSMike Kravetz 2210b65a4edaSMike Kravetz retry = true; 2211b65a4edaSMike Kravetz goto retry; 2212b65a4edaSMike Kravetz } 2213b65a4edaSMike Kravetz ClearHPageTemporary(page); 2214b65a4edaSMike Kravetz } 2215b65a4edaSMike Kravetz 22169980d744SMichal Hocko h->surplus_huge_pages++; 22174704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22189980d744SMichal Hocko 22199980d744SMichal Hocko out_unlock: 2220db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22217893d1d5SAdam Litke 22227893d1d5SAdam Litke return page; 22237893d1d5SAdam Litke } 22247893d1d5SAdam Litke 2225bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2226ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2227ab5ac90aSMichal Hocko { 2228ab5ac90aSMichal Hocko struct page *page; 2229ab5ac90aSMichal Hocko 2230ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2231ab5ac90aSMichal Hocko return NULL; 2232ab5ac90aSMichal Hocko 2233f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2234ab5ac90aSMichal Hocko if (!page) 2235ab5ac90aSMichal Hocko return NULL; 2236ab5ac90aSMichal Hocko 2237ab5ac90aSMichal Hocko /* 2238ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2239ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2240ab5ac90aSMichal Hocko */ 22419157c311SMike Kravetz SetHPageTemporary(page); 2242ab5ac90aSMichal Hocko 2243ab5ac90aSMichal Hocko return page; 2244ab5ac90aSMichal Hocko } 2245ab5ac90aSMichal Hocko 2246e4e574b7SAdam Litke /* 2247099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2248099730d6SDave Hansen */ 2249e0ec90eeSDave Hansen static 22500c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2251099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2252099730d6SDave Hansen { 2253cfcaa66fSBen Widawsky struct page *page = NULL; 2254aaf14e40SMichal Hocko struct mempolicy *mpol; 2255aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2256aaf14e40SMichal Hocko int nid; 2257aaf14e40SMichal Hocko nodemask_t *nodemask; 2258aaf14e40SMichal Hocko 2259aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2260cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2261cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2262cfcaa66fSBen Widawsky 2263cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2264cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2265cfcaa66fSBen Widawsky 2266cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2267cfcaa66fSBen Widawsky nodemask = NULL; 2268cfcaa66fSBen Widawsky } 2269cfcaa66fSBen Widawsky 2270cfcaa66fSBen Widawsky if (!page) 2271b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2272aaf14e40SMichal Hocko mpol_cond_put(mpol); 2273aaf14e40SMichal Hocko return page; 2274099730d6SDave Hansen } 2275099730d6SDave Hansen 2276ab5ac90aSMichal Hocko /* page migration callback function */ 22773e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2278d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 22794db9b2efSMichal Hocko { 2280db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22814db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 22823e59fcb0SMichal Hocko struct page *page; 22833e59fcb0SMichal Hocko 22843e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 22853e59fcb0SMichal Hocko if (page) { 2286db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22873e59fcb0SMichal Hocko return page; 22884db9b2efSMichal Hocko } 22894db9b2efSMichal Hocko } 2290db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22914db9b2efSMichal Hocko 22920c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 22934db9b2efSMichal Hocko } 22944db9b2efSMichal Hocko 2295ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2296389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2297389c8178SMichal Hocko unsigned long address) 2298ebd63723SMichal Hocko { 2299ebd63723SMichal Hocko struct mempolicy *mpol; 2300ebd63723SMichal Hocko nodemask_t *nodemask; 2301ebd63723SMichal Hocko struct page *page; 2302ebd63723SMichal Hocko gfp_t gfp_mask; 2303ebd63723SMichal Hocko int node; 2304ebd63723SMichal Hocko 2305ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2306ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2307d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2308ebd63723SMichal Hocko mpol_cond_put(mpol); 2309ebd63723SMichal Hocko 2310ebd63723SMichal Hocko return page; 2311ebd63723SMichal Hocko } 2312ebd63723SMichal Hocko 2313bf50bab2SNaoya Horiguchi /* 231425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2315e4e574b7SAdam Litke * of size 'delta'. 2316e4e574b7SAdam Litke */ 23170a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23181b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2319e4e574b7SAdam Litke { 2320e4e574b7SAdam Litke struct list_head surplus_list; 2321e4e574b7SAdam Litke struct page *page, *tmp; 23220a4f3d1bSLiu Xiang int ret; 23230a4f3d1bSLiu Xiang long i; 23240a4f3d1bSLiu Xiang long needed, allocated; 232528073b02SHillf Danton bool alloc_ok = true; 2326e4e574b7SAdam Litke 23279487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2328a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2329ac09b3a1SAdam Litke if (needed <= 0) { 2330a5516438SAndi Kleen h->resv_huge_pages += delta; 2331e4e574b7SAdam Litke return 0; 2332ac09b3a1SAdam Litke } 2333e4e574b7SAdam Litke 2334e4e574b7SAdam Litke allocated = 0; 2335e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2336e4e574b7SAdam Litke 2337e4e574b7SAdam Litke ret = -ENOMEM; 2338e4e574b7SAdam Litke retry: 2339db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2340e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23410c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2342b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 234328073b02SHillf Danton if (!page) { 234428073b02SHillf Danton alloc_ok = false; 234528073b02SHillf Danton break; 234628073b02SHillf Danton } 2347e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 234869ed779aSDavid Rientjes cond_resched(); 2349e4e574b7SAdam Litke } 235028073b02SHillf Danton allocated += i; 2351e4e574b7SAdam Litke 2352e4e574b7SAdam Litke /* 2353e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2354e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2355e4e574b7SAdam Litke */ 2356db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2357a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2358a5516438SAndi Kleen (h->free_huge_pages + allocated); 235928073b02SHillf Danton if (needed > 0) { 236028073b02SHillf Danton if (alloc_ok) 2361e4e574b7SAdam Litke goto retry; 236228073b02SHillf Danton /* 236328073b02SHillf Danton * We were not able to allocate enough pages to 236428073b02SHillf Danton * satisfy the entire reservation so we free what 236528073b02SHillf Danton * we've allocated so far. 236628073b02SHillf Danton */ 236728073b02SHillf Danton goto free; 236828073b02SHillf Danton } 2369e4e574b7SAdam Litke /* 2370e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 237125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2372e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2373ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2374ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2375ac09b3a1SAdam Litke * before they are reserved. 2376e4e574b7SAdam Litke */ 2377e4e574b7SAdam Litke needed += allocated; 2378a5516438SAndi Kleen h->resv_huge_pages += delta; 2379e4e574b7SAdam Litke ret = 0; 2380a9869b83SNaoya Horiguchi 238119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 238219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 238319fc3f0aSAdam Litke if ((--needed) < 0) 238419fc3f0aSAdam Litke break; 2385b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2386a5516438SAndi Kleen enqueue_huge_page(h, page); 238719fc3f0aSAdam Litke } 238828073b02SHillf Danton free: 2389db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 239019fc3f0aSAdam Litke 2391b65a4edaSMike Kravetz /* 2392b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2393b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2394b65a4edaSMike Kravetz */ 2395c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2396b65a4edaSMike Kravetz free_huge_page(page); 2397db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2398e4e574b7SAdam Litke 2399e4e574b7SAdam Litke return ret; 2400e4e574b7SAdam Litke } 2401e4e574b7SAdam Litke 2402e4e574b7SAdam Litke /* 2403e5bbc8a6SMike Kravetz * This routine has two main purposes: 2404e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2405e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2406e5bbc8a6SMike Kravetz * to the associated reservation map. 2407e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2408e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2409e4e574b7SAdam Litke */ 2410a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2411a5516438SAndi Kleen unsigned long unused_resv_pages) 2412e4e574b7SAdam Litke { 2413e4e574b7SAdam Litke unsigned long nr_pages; 241410c6ec49SMike Kravetz struct page *page; 241510c6ec49SMike Kravetz LIST_HEAD(page_list); 241610c6ec49SMike Kravetz 24179487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 241810c6ec49SMike Kravetz /* Uncommit the reservation */ 241910c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2420e4e574b7SAdam Litke 2421aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2422bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2423e5bbc8a6SMike Kravetz goto out; 2424aa888a74SAndi Kleen 2425e5bbc8a6SMike Kravetz /* 2426e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2427e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2428e5bbc8a6SMike Kravetz */ 2429a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2430e4e574b7SAdam Litke 2431685f3457SLee Schermerhorn /* 2432685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24339b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24349b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24359b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 243610c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24379b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2438685f3457SLee Schermerhorn */ 2439685f3457SLee Schermerhorn while (nr_pages--) { 244010c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 244110c6ec49SMike Kravetz if (!page) 2442e5bbc8a6SMike Kravetz goto out; 244310c6ec49SMike Kravetz 244410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2445e4e574b7SAdam Litke } 2446e5bbc8a6SMike Kravetz 2447e5bbc8a6SMike Kravetz out: 2448db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 244910c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2450db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2451e4e574b7SAdam Litke } 2452e4e574b7SAdam Litke 24535e911373SMike Kravetz 2454c37f9fb1SAndy Whitcroft /* 2455feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24565e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2457cf3ad20bSMike Kravetz * 2458cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2459cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2460cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2461cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2462cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2463feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2464feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2465feba16e2SMike Kravetz * is called in such cases. 2466cf3ad20bSMike Kravetz * 2467cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2468cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2469cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2470cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2471cf3ad20bSMike Kravetz * take appropriate action. 247296b96a96SMike Kravetz * 247396b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 247496b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 247596b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 247696b96a96SMike Kravetz * reservation exists. 2477846be085SMike Kravetz * 2478846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2479846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2480846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2481846be085SMike Kravetz * exists. 2482c37f9fb1SAndy Whitcroft */ 24835e911373SMike Kravetz enum vma_resv_mode { 24845e911373SMike Kravetz VMA_NEEDS_RESV, 24855e911373SMike Kravetz VMA_COMMIT_RESV, 2486feba16e2SMike Kravetz VMA_END_RESV, 248796b96a96SMike Kravetz VMA_ADD_RESV, 2488846be085SMike Kravetz VMA_DEL_RESV, 24895e911373SMike Kravetz }; 2490cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2491cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 24925e911373SMike Kravetz enum vma_resv_mode mode) 2493c37f9fb1SAndy Whitcroft { 24944e35f483SJoonsoo Kim struct resv_map *resv; 24954e35f483SJoonsoo Kim pgoff_t idx; 2496cf3ad20bSMike Kravetz long ret; 24970db9d74eSMina Almasry long dummy_out_regions_needed; 2498c37f9fb1SAndy Whitcroft 24994e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25004e35f483SJoonsoo Kim if (!resv) 2501c37f9fb1SAndy Whitcroft return 1; 2502c37f9fb1SAndy Whitcroft 25034e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25045e911373SMike Kravetz switch (mode) { 25055e911373SMike Kravetz case VMA_NEEDS_RESV: 25060db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25070db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25080db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25090db9d74eSMina Almasry * ever require 1 region. 25100db9d74eSMina Almasry */ 25110db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25125e911373SMike Kravetz break; 25135e911373SMike Kravetz case VMA_COMMIT_RESV: 2514075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25150db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25160db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25175e911373SMike Kravetz break; 2518feba16e2SMike Kravetz case VMA_END_RESV: 25190db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25205e911373SMike Kravetz ret = 0; 25215e911373SMike Kravetz break; 252296b96a96SMike Kravetz case VMA_ADD_RESV: 25230db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2524075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25250db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25260db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25270db9d74eSMina Almasry } else { 25280db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 252996b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 253096b96a96SMike Kravetz } 253196b96a96SMike Kravetz break; 2532846be085SMike Kravetz case VMA_DEL_RESV: 2533846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2534846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2535846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2536846be085SMike Kravetz } else { 2537846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2538846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2539846be085SMike Kravetz VM_BUG_ON(ret < 0); 2540846be085SMike Kravetz } 2541846be085SMike Kravetz break; 25425e911373SMike Kravetz default: 25435e911373SMike Kravetz BUG(); 25445e911373SMike Kravetz } 254584afd99bSAndy Whitcroft 2546846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2547cf3ad20bSMike Kravetz return ret; 254867961f9dSMike Kravetz /* 2549bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2550bf3d12b9SMiaohe Lin * 255167961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 255267961f9dSMike Kravetz * However, a file associated with mapping could have been 255367961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 255467961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 255567961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 255667961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 255767961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 255867961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 255967961f9dSMike Kravetz * reservation has already been consumed. As a result, the 256067961f9dSMike Kravetz * return value of this routine is the opposite of the 256167961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 256267961f9dSMike Kravetz */ 2563bf3d12b9SMiaohe Lin if (ret > 0) 256467961f9dSMike Kravetz return 0; 2565bf3d12b9SMiaohe Lin if (ret == 0) 256667961f9dSMike Kravetz return 1; 2567bf3d12b9SMiaohe Lin return ret; 256884afd99bSAndy Whitcroft } 2569cf3ad20bSMike Kravetz 2570cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2571a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2572c37f9fb1SAndy Whitcroft { 25735e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2574cf3ad20bSMike Kravetz } 2575c37f9fb1SAndy Whitcroft 2576cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2577cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2578cf3ad20bSMike Kravetz { 25795e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 25805e911373SMike Kravetz } 25815e911373SMike Kravetz 2582feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 25835e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 25845e911373SMike Kravetz { 2585feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2586c37f9fb1SAndy Whitcroft } 2587c37f9fb1SAndy Whitcroft 258896b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 258996b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 259096b96a96SMike Kravetz { 259196b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 259296b96a96SMike Kravetz } 259396b96a96SMike Kravetz 2594846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2595846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 259696b96a96SMike Kravetz { 2597846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2598846be085SMike Kravetz } 2599846be085SMike Kravetz 2600846be085SMike Kravetz /* 2601846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2602846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2603846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2604846be085SMike Kravetz * 2605846be085SMike Kravetz * It handles two specific cases: 2606846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2607846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2608846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2609846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2610846be085SMike Kravetz * 2611846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2612846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2613846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2614846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2615846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2616846be085SMike Kravetz * reserve map indicates there is a reservation present. 2617846be085SMike Kravetz * 2618846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2619846be085SMike Kravetz */ 2620846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2621846be085SMike Kravetz unsigned long address, struct page *page) 2622846be085SMike Kravetz { 262396b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 262496b96a96SMike Kravetz 2625846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2626846be085SMike Kravetz if (unlikely(rc < 0)) 262796b96a96SMike Kravetz /* 262896b96a96SMike Kravetz * Rare out of memory condition in reserve map 2629d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 263096b96a96SMike Kravetz * global reserve count will not be incremented 263196b96a96SMike Kravetz * by free_huge_page. This will make it appear 263296b96a96SMike Kravetz * as though the reservation for this page was 263396b96a96SMike Kravetz * consumed. This may prevent the task from 263496b96a96SMike Kravetz * faulting in the page at a later time. This 263596b96a96SMike Kravetz * is better than inconsistent global huge page 263696b96a96SMike Kravetz * accounting of reserve counts. 263796b96a96SMike Kravetz */ 2638d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2639846be085SMike Kravetz else if (rc) 2640846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2641846be085SMike Kravetz else 2642846be085SMike Kravetz vma_end_reservation(h, vma, address); 2643846be085SMike Kravetz } else { 2644846be085SMike Kravetz if (!rc) { 264596b96a96SMike Kravetz /* 2646846be085SMike Kravetz * This indicates there is an entry in the reserve map 2647c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2648846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2649846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2650846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2651846be085SMike Kravetz * does not consume a reservation. 265296b96a96SMike Kravetz */ 2653846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2654846be085SMike Kravetz if (rc < 0) 2655846be085SMike Kravetz /* 2656846be085SMike Kravetz * VERY rare out of memory condition. Since 2657846be085SMike Kravetz * we can not delete the entry, set 2658846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2659846be085SMike Kravetz * count will be incremented when the page 2660846be085SMike Kravetz * is freed. This reserve will be consumed 2661846be085SMike Kravetz * on a subsequent allocation. 2662846be085SMike Kravetz */ 2663846be085SMike Kravetz SetHPageRestoreReserve(page); 2664846be085SMike Kravetz } else if (rc < 0) { 2665846be085SMike Kravetz /* 2666846be085SMike Kravetz * Rare out of memory condition from 2667846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2668846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2669846be085SMike Kravetz * this implies there is not an entry in the 2670846be085SMike Kravetz * reserve map. 2671846be085SMike Kravetz * 2672846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2673846be085SMike Kravetz * no reservation. We are done. 2674846be085SMike Kravetz */ 2675846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2676846be085SMike Kravetz /* 2677846be085SMike Kravetz * For private mappings, no entry indicates 2678846be085SMike Kravetz * a reservation is present. Since we can 2679846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2680846be085SMike Kravetz * on the page so reserve count will be 2681846be085SMike Kravetz * incremented when freed. This reserve will 2682846be085SMike Kravetz * be consumed on a subsequent allocation. 2683846be085SMike Kravetz */ 2684846be085SMike Kravetz SetHPageRestoreReserve(page); 268596b96a96SMike Kravetz } else 2686846be085SMike Kravetz /* 2687846be085SMike Kravetz * No reservation present, do nothing 2688846be085SMike Kravetz */ 268996b96a96SMike Kravetz vma_end_reservation(h, vma, address); 269096b96a96SMike Kravetz } 269196b96a96SMike Kravetz } 269296b96a96SMike Kravetz 2693369fa227SOscar Salvador /* 2694369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2695369fa227SOscar Salvador * @h: struct hstate old page belongs to 2696369fa227SOscar Salvador * @old_page: Old page to dissolve 2697ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2698369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2699369fa227SOscar Salvador */ 2700ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2701ae37c7ffSOscar Salvador struct list_head *list) 2702369fa227SOscar Salvador { 2703369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2704369fa227SOscar Salvador int nid = page_to_nid(old_page); 2705b65a4edaSMike Kravetz bool alloc_retry = false; 2706369fa227SOscar Salvador struct page *new_page; 2707369fa227SOscar Salvador int ret = 0; 2708369fa227SOscar Salvador 2709369fa227SOscar Salvador /* 2710369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2711f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2712f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2713f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2714f41f2ed4SMuchun Song * under the lock. 2715369fa227SOscar Salvador */ 2716b65a4edaSMike Kravetz alloc_retry: 2717369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2718369fa227SOscar Salvador if (!new_page) 2719369fa227SOscar Salvador return -ENOMEM; 2720b65a4edaSMike Kravetz /* 2721b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2722b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2723b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2724b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2725b65a4edaSMike Kravetz * 2726b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2727b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2728b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2729b65a4edaSMike Kravetz */ 2730b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2731b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2732b65a4edaSMike Kravetz if (alloc_retry) 2733b65a4edaSMike Kravetz return -EBUSY; 2734b65a4edaSMike Kravetz 2735b65a4edaSMike Kravetz alloc_retry = true; 2736b65a4edaSMike Kravetz goto alloc_retry; 2737b65a4edaSMike Kravetz } 2738b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2739b65a4edaSMike Kravetz 2740f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2741369fa227SOscar Salvador 2742369fa227SOscar Salvador retry: 2743369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2744369fa227SOscar Salvador if (!PageHuge(old_page)) { 2745369fa227SOscar Salvador /* 2746369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2747369fa227SOscar Salvador */ 2748369fa227SOscar Salvador goto free_new; 2749369fa227SOscar Salvador } else if (page_count(old_page)) { 2750369fa227SOscar Salvador /* 2751ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2752ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2753369fa227SOscar Salvador */ 2754ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2755ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2756369fa227SOscar Salvador ret = -EBUSY; 2757ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2758369fa227SOscar Salvador goto free_new; 2759369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2760369fa227SOscar Salvador /* 2761369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2762369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2763369fa227SOscar Salvador * we retry. 2764369fa227SOscar Salvador */ 2765369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2766369fa227SOscar Salvador cond_resched(); 2767369fa227SOscar Salvador goto retry; 2768369fa227SOscar Salvador } else { 2769369fa227SOscar Salvador /* 2770369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2771369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2772369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2773369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2774369fa227SOscar Salvador * stable since this happens under the lock. 2775369fa227SOscar Salvador */ 2776369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2777369fa227SOscar Salvador 2778369fa227SOscar Salvador /* 2779b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2780b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2781369fa227SOscar Salvador */ 2782369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2783369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2784369fa227SOscar Salvador 2785369fa227SOscar Salvador /* 2786369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2787369fa227SOscar Salvador */ 2788369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2789b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2790369fa227SOscar Salvador } 2791369fa227SOscar Salvador 2792369fa227SOscar Salvador return ret; 2793369fa227SOscar Salvador 2794369fa227SOscar Salvador free_new: 2795369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2796b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2797b65a4edaSMike Kravetz set_page_refcounted(new_page); 2798b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2799369fa227SOscar Salvador 2800369fa227SOscar Salvador return ret; 2801369fa227SOscar Salvador } 2802369fa227SOscar Salvador 2803ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2804369fa227SOscar Salvador { 2805369fa227SOscar Salvador struct hstate *h; 2806369fa227SOscar Salvador struct page *head; 2807ae37c7ffSOscar Salvador int ret = -EBUSY; 2808369fa227SOscar Salvador 2809369fa227SOscar Salvador /* 2810369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2811369fa227SOscar Salvador * to carefully check the state under the lock. 2812369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2813369fa227SOscar Salvador */ 2814369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2815369fa227SOscar Salvador if (PageHuge(page)) { 2816369fa227SOscar Salvador head = compound_head(page); 2817369fa227SOscar Salvador h = page_hstate(head); 2818369fa227SOscar Salvador } else { 2819369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2820369fa227SOscar Salvador return 0; 2821369fa227SOscar Salvador } 2822369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2823369fa227SOscar Salvador 2824369fa227SOscar Salvador /* 2825369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2826369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2827369fa227SOscar Salvador * of bailing out right away without further retrying. 2828369fa227SOscar Salvador */ 2829369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2830369fa227SOscar Salvador return -ENOMEM; 2831369fa227SOscar Salvador 2832ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2833ae37c7ffSOscar Salvador ret = 0; 2834ae37c7ffSOscar Salvador else if (!page_count(head)) 2835ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2836ae37c7ffSOscar Salvador 2837ae37c7ffSOscar Salvador return ret; 2838369fa227SOscar Salvador } 2839369fa227SOscar Salvador 284070c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 284104f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2842348ea204SAdam Litke { 284390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2844a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2845348ea204SAdam Litke struct page *page; 2846d85f69b0SMike Kravetz long map_chg, map_commit; 2847d85f69b0SMike Kravetz long gbl_chg; 28486d76dcf4SAneesh Kumar K.V int ret, idx; 28496d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 285008cf9fafSMina Almasry bool deferred_reserve; 28512fc39cecSAdam Litke 28526d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2853a1e78772SMel Gorman /* 2854d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2855d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2856d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2857a1e78772SMel Gorman */ 2858d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2859d85f69b0SMike Kravetz if (map_chg < 0) 286076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2861d85f69b0SMike Kravetz 2862d85f69b0SMike Kravetz /* 2863d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2864d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2865d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2866d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2867d85f69b0SMike Kravetz * checked against any subpool limit. 2868d85f69b0SMike Kravetz */ 2869d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2870d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2871d85f69b0SMike Kravetz if (gbl_chg < 0) { 2872feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 287376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 28745e911373SMike Kravetz } 287590d8b7e6SAdam Litke 2876d85f69b0SMike Kravetz /* 2877d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2878d85f69b0SMike Kravetz * map, there could be reservations associated with the 2879d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2880d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2881d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2882d85f69b0SMike Kravetz * the subpool reservations. 2883d85f69b0SMike Kravetz */ 2884d85f69b0SMike Kravetz if (avoid_reserve) 2885d85f69b0SMike Kravetz gbl_chg = 1; 2886d85f69b0SMike Kravetz } 2887d85f69b0SMike Kravetz 288808cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 288908cf9fafSMina Almasry */ 28906501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 289108cf9fafSMina Almasry if (deferred_reserve) { 289208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 289308cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 28948f34af6fSJianyu Zhan if (ret) 28958f34af6fSJianyu Zhan goto out_subpool_put; 289608cf9fafSMina Almasry } 289708cf9fafSMina Almasry 289808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 289908cf9fafSMina Almasry if (ret) 290008cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29018f34af6fSJianyu Zhan 2902db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2903d85f69b0SMike Kravetz /* 2904d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2905d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2906d85f69b0SMike Kravetz * a reservation exists for the allocation. 2907d85f69b0SMike Kravetz */ 2908d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 290981a6fcaeSJoonsoo Kim if (!page) { 2910db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29110c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29128f34af6fSJianyu Zhan if (!page) 29138f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2914a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2915d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2916a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2917a88c7695SNaoya Horiguchi } 2918db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 291915a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 292081a6fcaeSJoonsoo Kim /* Fall through */ 2921a1e78772SMel Gorman } 292281a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 292308cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 292408cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 292508cf9fafSMina Almasry */ 292608cf9fafSMina Almasry if (deferred_reserve) { 292708cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 292808cf9fafSMina Almasry h_cg, page); 292908cf9fafSMina Almasry } 293008cf9fafSMina Almasry 2931db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2932a1e78772SMel Gorman 2933d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2934a1e78772SMel Gorman 2935d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2936d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 293733039678SMike Kravetz /* 293833039678SMike Kravetz * The page was added to the reservation map between 293933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 294033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 294133039678SMike Kravetz * Adjust for the subpool count incremented above AND 294233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 294333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 294433039678SMike Kravetz * no longer applies. 294533039678SMike Kravetz */ 294633039678SMike Kravetz long rsv_adjust; 294733039678SMike Kravetz 294833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 294933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 295079aa925bSMike Kravetz if (deferred_reserve) 295179aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 295279aa925bSMike Kravetz pages_per_huge_page(h), page); 295333039678SMike Kravetz } 29547893d1d5SAdam Litke return page; 29558f34af6fSJianyu Zhan 29568f34af6fSJianyu Zhan out_uncharge_cgroup: 29578f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 295808cf9fafSMina Almasry out_uncharge_cgroup_reservation: 295908cf9fafSMina Almasry if (deferred_reserve) 296008cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 296108cf9fafSMina Almasry h_cg); 29628f34af6fSJianyu Zhan out_subpool_put: 2963d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29648f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2965feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29668f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2967b45b5bd6SDavid Gibson } 2968b45b5bd6SDavid Gibson 2969b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 2970e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2971b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 2972aa888a74SAndi Kleen { 2973b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 2974b2261026SJoonsoo Kim int nr_nodes, node; 2975aa888a74SAndi Kleen 29764178158eSZhenguo Yao if (nid != NUMA_NO_NODE && nid >= nr_online_nodes) 2977b5389086SZhenguo Yao return 0; 2978b5389086SZhenguo Yao /* do node specific alloc */ 2979b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 2980b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 2981b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 2982b5389086SZhenguo Yao if (!m) 2983b5389086SZhenguo Yao return 0; 2984b5389086SZhenguo Yao goto found; 2985b5389086SZhenguo Yao } 2986b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 2987b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2988b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 29898b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 299097ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2991aa888a74SAndi Kleen /* 2992aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2993aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2994aa888a74SAndi Kleen * puts them into the mem_map). 2995aa888a74SAndi Kleen */ 2996b5389086SZhenguo Yao if (!m) 2997b5389086SZhenguo Yao return 0; 2998aa888a74SAndi Kleen goto found; 2999aa888a74SAndi Kleen } 3000aa888a74SAndi Kleen 3001aa888a74SAndi Kleen found: 3002aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3003330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3004aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3005aa888a74SAndi Kleen m->hstate = h; 3006aa888a74SAndi Kleen return 1; 3007aa888a74SAndi Kleen } 3008aa888a74SAndi Kleen 300948b8d744SMike Kravetz /* 301048b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 301148b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 301248b8d744SMike Kravetz */ 3013aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3014aa888a74SAndi Kleen { 3015aa888a74SAndi Kleen struct huge_bootmem_page *m; 3016aa888a74SAndi Kleen 3017aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 301840d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3019aa888a74SAndi Kleen struct hstate *h = m->hstate; 3020ee8f248dSBecky Bruce 302148b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3022aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30237118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3024ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3025aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30267118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 30277118fc29SMike Kravetz } else { 3028416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30297118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30307118fc29SMike Kravetz } 3031af0fb9dfSMichal Hocko 3032b0320c7bSRafael Aquini /* 303348b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 303448b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 303548b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3036b0320c7bSRafael Aquini */ 3037c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3038520495feSCannon Matthews cond_resched(); 3039aa888a74SAndi Kleen } 3040aa888a74SAndi Kleen } 3041b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3042b5389086SZhenguo Yao { 3043b5389086SZhenguo Yao unsigned long i; 3044b5389086SZhenguo Yao char buf[32]; 3045b5389086SZhenguo Yao 3046b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3047b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3048b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3049b5389086SZhenguo Yao break; 3050b5389086SZhenguo Yao } else { 3051b5389086SZhenguo Yao struct page *page; 3052b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3053b5389086SZhenguo Yao 3054b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3055b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3056b5389086SZhenguo Yao if (!page) 3057b5389086SZhenguo Yao break; 3058b5389086SZhenguo Yao put_page(page); /* free it into the hugepage allocator */ 3059b5389086SZhenguo Yao } 3060b5389086SZhenguo Yao cond_resched(); 3061b5389086SZhenguo Yao } 3062b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3063b5389086SZhenguo Yao return; 3064b5389086SZhenguo Yao 3065b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3066b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3067b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3068b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3069b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3070b5389086SZhenguo Yao } 3071aa888a74SAndi Kleen 30728faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 30731da177e4SLinus Torvalds { 30741da177e4SLinus Torvalds unsigned long i; 3075f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3076b5389086SZhenguo Yao bool node_specific_alloc = false; 3077f60858f9SMike Kravetz 3078b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3079b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3080b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3081b5389086SZhenguo Yao return; 3082b5389086SZhenguo Yao } 3083b5389086SZhenguo Yao 3084b5389086SZhenguo Yao /* do node specific alloc */ 3085b5389086SZhenguo Yao for (i = 0; i < nr_online_nodes; i++) { 3086b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3087b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3088b5389086SZhenguo Yao node_specific_alloc = true; 3089b5389086SZhenguo Yao } 3090b5389086SZhenguo Yao } 3091b5389086SZhenguo Yao 3092b5389086SZhenguo Yao if (node_specific_alloc) 3093b5389086SZhenguo Yao return; 3094b5389086SZhenguo Yao 3095b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3096f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3097f60858f9SMike Kravetz /* 3098f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3099f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3100f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3101f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3102f60858f9SMike Kravetz */ 3103f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3104f60858f9SMike Kravetz GFP_KERNEL); 3105f60858f9SMike Kravetz } else { 3106f60858f9SMike Kravetz /* allocations done at boot time */ 3107f60858f9SMike Kravetz node_alloc_noretry = NULL; 3108f60858f9SMike Kravetz } 3109f60858f9SMike Kravetz 3110f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3111f60858f9SMike Kravetz if (node_alloc_noretry) 3112f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31131da177e4SLinus Torvalds 3114e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3115bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3116b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3117aa888a74SAndi Kleen break; 31180c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3119f60858f9SMike Kravetz &node_states[N_MEMORY], 3120f60858f9SMike Kravetz node_alloc_noretry)) 31211da177e4SLinus Torvalds break; 312269ed779aSDavid Rientjes cond_resched(); 31231da177e4SLinus Torvalds } 3124d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3125d715cf80SLiam R. Howlett char buf[32]; 3126d715cf80SLiam R. Howlett 3127c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3128d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3129d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31308faa8b07SAndi Kleen h->max_huge_pages = i; 3131e5ff2159SAndi Kleen } 3132f60858f9SMike Kravetz kfree(node_alloc_noretry); 3133d715cf80SLiam R. Howlett } 3134e5ff2159SAndi Kleen 3135e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3136e5ff2159SAndi Kleen { 313779dfc695SMike Kravetz struct hstate *h, *h2; 3138e5ff2159SAndi Kleen 3139e5ff2159SAndi Kleen for_each_hstate(h) { 3140641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 3141641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 3142641844f5SNaoya Horiguchi 31438faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3144bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31458faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 314679dfc695SMike Kravetz 314779dfc695SMike Kravetz /* 314879dfc695SMike Kravetz * Set demote order for each hstate. Note that 314979dfc695SMike Kravetz * h->demote_order is initially 0. 315079dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 315179dfc695SMike Kravetz * is not supported, so skip this. 3152a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3153a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 315479dfc695SMike Kravetz */ 315579dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 315679dfc695SMike Kravetz continue; 3157a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3158a01f4390SMike Kravetz continue; 315979dfc695SMike Kravetz for_each_hstate(h2) { 316079dfc695SMike Kravetz if (h2 == h) 316179dfc695SMike Kravetz continue; 316279dfc695SMike Kravetz if (h2->order < h->order && 316379dfc695SMike Kravetz h2->order > h->demote_order) 316479dfc695SMike Kravetz h->demote_order = h2->order; 316579dfc695SMike Kravetz } 3166e5ff2159SAndi Kleen } 3167641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 3168e5ff2159SAndi Kleen } 3169e5ff2159SAndi Kleen 3170e5ff2159SAndi Kleen static void __init report_hugepages(void) 3171e5ff2159SAndi Kleen { 3172e5ff2159SAndi Kleen struct hstate *h; 3173e5ff2159SAndi Kleen 3174e5ff2159SAndi Kleen for_each_hstate(h) { 31754abd32dbSAndi Kleen char buf[32]; 3176c6247f72SMatthew Wilcox 3177c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3178ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 3179c6247f72SMatthew Wilcox buf, h->free_huge_pages); 3180e5ff2159SAndi Kleen } 3181e5ff2159SAndi Kleen } 3182e5ff2159SAndi Kleen 31831da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 31846ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 31856ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31861da177e4SLinus Torvalds { 31874415cc8dSChristoph Lameter int i; 31881121828aSMike Kravetz LIST_HEAD(page_list); 31894415cc8dSChristoph Lameter 31909487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3191bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3192aa888a74SAndi Kleen return; 3193aa888a74SAndi Kleen 31941121828aSMike Kravetz /* 31951121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 31961121828aSMike Kravetz */ 31976ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 319810c6ec49SMike Kravetz struct page *page, *next; 3199a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3200a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3201a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32021121828aSMike Kravetz goto out; 32031da177e4SLinus Torvalds if (PageHighMem(page)) 32041da177e4SLinus Torvalds continue; 32056eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32061121828aSMike Kravetz list_add(&page->lru, &page_list); 32071121828aSMike Kravetz } 32081121828aSMike Kravetz } 32091121828aSMike Kravetz 32101121828aSMike Kravetz out: 3211db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 321210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3213db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32141da177e4SLinus Torvalds } 32151da177e4SLinus Torvalds #else 32166ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32176ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32181da177e4SLinus Torvalds { 32191da177e4SLinus Torvalds } 32201da177e4SLinus Torvalds #endif 32211da177e4SLinus Torvalds 322220a0307cSWu Fengguang /* 322320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 322420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 322520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 322620a0307cSWu Fengguang */ 32276ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32286ae11b27SLee Schermerhorn int delta) 322920a0307cSWu Fengguang { 3230b2261026SJoonsoo Kim int nr_nodes, node; 323120a0307cSWu Fengguang 32329487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 323320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 323420a0307cSWu Fengguang 3235e8c5c824SLee Schermerhorn if (delta < 0) { 3236b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3237b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3238b2261026SJoonsoo Kim goto found; 3239b2261026SJoonsoo Kim } 3240b2261026SJoonsoo Kim } else { 3241b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3242b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3243b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3244b2261026SJoonsoo Kim goto found; 3245e8c5c824SLee Schermerhorn } 32469a76db09SLee Schermerhorn } 3247b2261026SJoonsoo Kim return 0; 324820a0307cSWu Fengguang 3249b2261026SJoonsoo Kim found: 325020a0307cSWu Fengguang h->surplus_huge_pages += delta; 3251b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3252b2261026SJoonsoo Kim return 1; 325320a0307cSWu Fengguang } 325420a0307cSWu Fengguang 3255a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3256fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32576ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32581da177e4SLinus Torvalds { 32597893d1d5SAdam Litke unsigned long min_count, ret; 326010c6ec49SMike Kravetz struct page *page; 326110c6ec49SMike Kravetz LIST_HEAD(page_list); 3262f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3263f60858f9SMike Kravetz 3264f60858f9SMike Kravetz /* 3265f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3266f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3267f60858f9SMike Kravetz * the requested huge pages. 3268f60858f9SMike Kravetz */ 3269f60858f9SMike Kravetz if (node_alloc_noretry) 3270f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3271f60858f9SMike Kravetz else 3272f60858f9SMike Kravetz return -ENOMEM; 32731da177e4SLinus Torvalds 327429383967SMike Kravetz /* 327529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 327629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 327729383967SMike Kravetz */ 327829383967SMike Kravetz mutex_lock(&h->resize_lock); 3279b65d4adbSMuchun Song flush_free_hpage_work(h); 3280db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32814eb0716eSAlexandre Ghiti 32824eb0716eSAlexandre Ghiti /* 3283fd875dcaSMike Kravetz * Check for a node specific request. 3284fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3285fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3286fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3287fd875dcaSMike Kravetz */ 3288fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3289fd875dcaSMike Kravetz unsigned long old_count = count; 3290fd875dcaSMike Kravetz 3291fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3292fd875dcaSMike Kravetz /* 3293fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3294fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3295fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3296fd875dcaSMike Kravetz * largest possible value to align with their intention. 3297fd875dcaSMike Kravetz */ 3298fd875dcaSMike Kravetz if (count < old_count) 3299fd875dcaSMike Kravetz count = ULONG_MAX; 3300fd875dcaSMike Kravetz } 3301fd875dcaSMike Kravetz 3302fd875dcaSMike Kravetz /* 33034eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33044eb0716eSAlexandre Ghiti * page range allocation. 33054eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33064eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33074eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33084eb0716eSAlexandre Ghiti */ 33094eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33104eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3311db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 331229383967SMike Kravetz mutex_unlock(&h->resize_lock); 3313f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33144eb0716eSAlexandre Ghiti return -EINVAL; 33154eb0716eSAlexandre Ghiti } 33164eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33174eb0716eSAlexandre Ghiti } 3318aa888a74SAndi Kleen 33197893d1d5SAdam Litke /* 33207893d1d5SAdam Litke * Increase the pool size 33217893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33227893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3323d1c3fb1fSNishanth Aravamudan * 33240c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3325d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3326d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3327d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3328d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33297893d1d5SAdam Litke */ 3330a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33316ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33327893d1d5SAdam Litke break; 33337893d1d5SAdam Litke } 33347893d1d5SAdam Litke 3335a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33367893d1d5SAdam Litke /* 33377893d1d5SAdam Litke * If this allocation races such that we no longer need the 33387893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33397893d1d5SAdam Litke * and reducing the surplus. 33407893d1d5SAdam Litke */ 3341db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3342649920c6SJia He 3343649920c6SJia He /* yield cpu to avoid soft lockup */ 3344649920c6SJia He cond_resched(); 3345649920c6SJia He 3346f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3347f60858f9SMike Kravetz node_alloc_noretry); 3348db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33497893d1d5SAdam Litke if (!ret) 33507893d1d5SAdam Litke goto out; 33517893d1d5SAdam Litke 3352536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3353536240f2SMel Gorman if (signal_pending(current)) 3354536240f2SMel Gorman goto out; 33557893d1d5SAdam Litke } 33567893d1d5SAdam Litke 33577893d1d5SAdam Litke /* 33587893d1d5SAdam Litke * Decrease the pool size 33597893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33607893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33617893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33627893d1d5SAdam Litke * to the desired size as pages become free. 3363d1c3fb1fSNishanth Aravamudan * 3364d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3365d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3366d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33670c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3368d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3369d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3370d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33717893d1d5SAdam Litke */ 3372a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33736b0c880dSAdam Litke min_count = max(count, min_count); 33746ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 337510c6ec49SMike Kravetz 337610c6ec49SMike Kravetz /* 337710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 337810c6ec49SMike Kravetz */ 3379a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 338010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 338110c6ec49SMike Kravetz if (!page) 33821da177e4SLinus Torvalds break; 338310c6ec49SMike Kravetz 338410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 33851da177e4SLinus Torvalds } 338610c6ec49SMike Kravetz /* free the pages after dropping lock */ 3387db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 338810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3389b65d4adbSMuchun Song flush_free_hpage_work(h); 3390db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 339110c6ec49SMike Kravetz 3392a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 33936ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 33947893d1d5SAdam Litke break; 33957893d1d5SAdam Litke } 33967893d1d5SAdam Litke out: 33974eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3398db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 339929383967SMike Kravetz mutex_unlock(&h->resize_lock); 34004eb0716eSAlexandre Ghiti 3401f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3402f60858f9SMike Kravetz 34034eb0716eSAlexandre Ghiti return 0; 34041da177e4SLinus Torvalds } 34051da177e4SLinus Torvalds 34068531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34078531fc6fSMike Kravetz { 34088531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34098531fc6fSMike Kravetz struct hstate *target_hstate; 34108531fc6fSMike Kravetz int rc = 0; 34118531fc6fSMike Kravetz 34128531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34138531fc6fSMike Kravetz 34148531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34158531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34168531fc6fSMike Kravetz 34178531fc6fSMike Kravetz rc = alloc_huge_page_vmemmap(h, page); 34188531fc6fSMike Kravetz if (rc) { 34198531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34208531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34218531fc6fSMike Kravetz set_page_refcounted(page); 34228531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34238531fc6fSMike Kravetz return rc; 34248531fc6fSMike Kravetz } 34258531fc6fSMike Kravetz 34268531fc6fSMike Kravetz /* 34278531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34288531fc6fSMike Kravetz * sizes as it will not ref count pages. 34298531fc6fSMike Kravetz */ 34308531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34318531fc6fSMike Kravetz 34328531fc6fSMike Kravetz /* 34338531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34348531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34358531fc6fSMike Kravetz * as surplus. 34368531fc6fSMike Kravetz * 34378531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34388531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34398531fc6fSMike Kravetz */ 34408531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34418531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34428531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 34438531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 34448531fc6fSMike Kravetz prep_compound_gigantic_page_for_demote(page + i, 34458531fc6fSMike Kravetz target_hstate->order); 34468531fc6fSMike Kravetz else 34478531fc6fSMike Kravetz prep_compound_page(page + i, target_hstate->order); 34488531fc6fSMike Kravetz set_page_private(page + i, 0); 34498531fc6fSMike Kravetz set_page_refcounted(page + i); 34508531fc6fSMike Kravetz prep_new_huge_page(target_hstate, page + i, nid); 34518531fc6fSMike Kravetz put_page(page + i); 34528531fc6fSMike Kravetz } 34538531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34548531fc6fSMike Kravetz 34558531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34568531fc6fSMike Kravetz 34578531fc6fSMike Kravetz /* 34588531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34598531fc6fSMike Kravetz * based on pool changes for the demoted page. 34608531fc6fSMike Kravetz */ 34618531fc6fSMike Kravetz h->max_huge_pages--; 34628531fc6fSMike Kravetz target_hstate->max_huge_pages += pages_per_huge_page(h); 34638531fc6fSMike Kravetz 34648531fc6fSMike Kravetz return rc; 34658531fc6fSMike Kravetz } 34668531fc6fSMike Kravetz 346779dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 346879dfc695SMike Kravetz __must_hold(&hugetlb_lock) 346979dfc695SMike Kravetz { 34708531fc6fSMike Kravetz int nr_nodes, node; 34718531fc6fSMike Kravetz struct page *page; 347279dfc695SMike Kravetz int rc = 0; 347379dfc695SMike Kravetz 347479dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 347579dfc695SMike Kravetz 347679dfc695SMike Kravetz /* We should never get here if no demote order */ 347779dfc695SMike Kravetz if (!h->demote_order) { 347879dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 347979dfc695SMike Kravetz return -EINVAL; /* internal error */ 348079dfc695SMike Kravetz } 348179dfc695SMike Kravetz 34828531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 34838531fc6fSMike Kravetz if (!list_empty(&h->hugepage_freelists[node])) { 34848531fc6fSMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 34858531fc6fSMike Kravetz struct page, lru); 34868531fc6fSMike Kravetz rc = demote_free_huge_page(h, page); 34878531fc6fSMike Kravetz break; 34888531fc6fSMike Kravetz } 34898531fc6fSMike Kravetz } 34908531fc6fSMike Kravetz 349179dfc695SMike Kravetz return rc; 349279dfc695SMike Kravetz } 349379dfc695SMike Kravetz 3494a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3495a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3496a3437870SNishanth Aravamudan 349779dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 349879dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 349979dfc695SMike Kravetz 3500a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 3501a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 3502a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 3503a3437870SNishanth Aravamudan 3504a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3505a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3506a3437870SNishanth Aravamudan 35079a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35089a305230SLee Schermerhorn 35099a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3510a3437870SNishanth Aravamudan { 3511a3437870SNishanth Aravamudan int i; 35129a305230SLee Schermerhorn 3513a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35149a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35159a305230SLee Schermerhorn if (nidp) 35169a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3517a3437870SNishanth Aravamudan return &hstates[i]; 35189a305230SLee Schermerhorn } 35199a305230SLee Schermerhorn 35209a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3521a3437870SNishanth Aravamudan } 3522a3437870SNishanth Aravamudan 352306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3524a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3525a3437870SNishanth Aravamudan { 35269a305230SLee Schermerhorn struct hstate *h; 35279a305230SLee Schermerhorn unsigned long nr_huge_pages; 35289a305230SLee Schermerhorn int nid; 35299a305230SLee Schermerhorn 35309a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35319a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35329a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35339a305230SLee Schermerhorn else 35349a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35359a305230SLee Schermerhorn 3536ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3537a3437870SNishanth Aravamudan } 3538adbe8726SEric B Munson 3539238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3540238d3c13SDavid Rientjes struct hstate *h, int nid, 3541238d3c13SDavid Rientjes unsigned long count, size_t len) 3542a3437870SNishanth Aravamudan { 3543a3437870SNishanth Aravamudan int err; 35442d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3545a3437870SNishanth Aravamudan 35462d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35472d0adf7eSOscar Salvador return -EINVAL; 3548adbe8726SEric B Munson 35499a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35509a305230SLee Schermerhorn /* 35519a305230SLee Schermerhorn * global hstate attribute 35529a305230SLee Schermerhorn */ 35539a305230SLee Schermerhorn if (!(obey_mempolicy && 35542d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35552d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35562d0adf7eSOscar Salvador else 35572d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35582d0adf7eSOscar Salvador } else { 35599a305230SLee Schermerhorn /* 3560fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3561fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35629a305230SLee Schermerhorn */ 35632d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35642d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3565fd875dcaSMike Kravetz } 35669a305230SLee Schermerhorn 35672d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 356806808b08SLee Schermerhorn 35694eb0716eSAlexandre Ghiti return err ? err : len; 357006808b08SLee Schermerhorn } 357106808b08SLee Schermerhorn 3572238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3573238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3574238d3c13SDavid Rientjes size_t len) 3575238d3c13SDavid Rientjes { 3576238d3c13SDavid Rientjes struct hstate *h; 3577238d3c13SDavid Rientjes unsigned long count; 3578238d3c13SDavid Rientjes int nid; 3579238d3c13SDavid Rientjes int err; 3580238d3c13SDavid Rientjes 3581238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3582238d3c13SDavid Rientjes if (err) 3583238d3c13SDavid Rientjes return err; 3584238d3c13SDavid Rientjes 3585238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3586238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3587238d3c13SDavid Rientjes } 3588238d3c13SDavid Rientjes 358906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 359006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 359106808b08SLee Schermerhorn { 359206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 359306808b08SLee Schermerhorn } 359406808b08SLee Schermerhorn 359506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 359606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 359706808b08SLee Schermerhorn { 3598238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3599a3437870SNishanth Aravamudan } 3600a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3601a3437870SNishanth Aravamudan 360206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 360306808b08SLee Schermerhorn 360406808b08SLee Schermerhorn /* 360506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 360606808b08SLee Schermerhorn * huge page alloc/free. 360706808b08SLee Schermerhorn */ 360806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3609ae7a927dSJoe Perches struct kobj_attribute *attr, 3610ae7a927dSJoe Perches char *buf) 361106808b08SLee Schermerhorn { 361206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 361306808b08SLee Schermerhorn } 361406808b08SLee Schermerhorn 361506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 361606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 361706808b08SLee Schermerhorn { 3618238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 361906808b08SLee Schermerhorn } 362006808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 362106808b08SLee Schermerhorn #endif 362206808b08SLee Schermerhorn 362306808b08SLee Schermerhorn 3624a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3625a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3626a3437870SNishanth Aravamudan { 36279a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3628ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3629a3437870SNishanth Aravamudan } 3630adbe8726SEric B Munson 3631a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3632a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3633a3437870SNishanth Aravamudan { 3634a3437870SNishanth Aravamudan int err; 3635a3437870SNishanth Aravamudan unsigned long input; 36369a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3637a3437870SNishanth Aravamudan 3638bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3639adbe8726SEric B Munson return -EINVAL; 3640adbe8726SEric B Munson 36413dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3642a3437870SNishanth Aravamudan if (err) 364373ae31e5SEric B Munson return err; 3644a3437870SNishanth Aravamudan 3645db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3646a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3647db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3648a3437870SNishanth Aravamudan 3649a3437870SNishanth Aravamudan return count; 3650a3437870SNishanth Aravamudan } 3651a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3652a3437870SNishanth Aravamudan 3653a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3654a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3655a3437870SNishanth Aravamudan { 36569a305230SLee Schermerhorn struct hstate *h; 36579a305230SLee Schermerhorn unsigned long free_huge_pages; 36589a305230SLee Schermerhorn int nid; 36599a305230SLee Schermerhorn 36609a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36619a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36629a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36639a305230SLee Schermerhorn else 36649a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36659a305230SLee Schermerhorn 3666ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3667a3437870SNishanth Aravamudan } 3668a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3669a3437870SNishanth Aravamudan 3670a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3671a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3672a3437870SNishanth Aravamudan { 36739a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3674ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3675a3437870SNishanth Aravamudan } 3676a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3677a3437870SNishanth Aravamudan 3678a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3679a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3680a3437870SNishanth Aravamudan { 36819a305230SLee Schermerhorn struct hstate *h; 36829a305230SLee Schermerhorn unsigned long surplus_huge_pages; 36839a305230SLee Schermerhorn int nid; 36849a305230SLee Schermerhorn 36859a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36869a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36879a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 36889a305230SLee Schermerhorn else 36899a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 36909a305230SLee Schermerhorn 3691ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3692a3437870SNishanth Aravamudan } 3693a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3694a3437870SNishanth Aravamudan 369579dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 369679dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 369779dfc695SMike Kravetz { 369879dfc695SMike Kravetz unsigned long nr_demote; 369979dfc695SMike Kravetz unsigned long nr_available; 370079dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 370179dfc695SMike Kravetz struct hstate *h; 370279dfc695SMike Kravetz int err = 0; 370379dfc695SMike Kravetz int nid; 370479dfc695SMike Kravetz 370579dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 370679dfc695SMike Kravetz if (err) 370779dfc695SMike Kravetz return err; 370879dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 370979dfc695SMike Kravetz 371079dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 371179dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 371279dfc695SMike Kravetz n_mask = &nodes_allowed; 371379dfc695SMike Kravetz } else { 371479dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 371579dfc695SMike Kravetz } 371679dfc695SMike Kravetz 371779dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 371879dfc695SMike Kravetz mutex_lock(&h->resize_lock); 371979dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 372079dfc695SMike Kravetz 372179dfc695SMike Kravetz while (nr_demote) { 372279dfc695SMike Kravetz /* 372379dfc695SMike Kravetz * Check for available pages to demote each time thorough the 372479dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 372579dfc695SMike Kravetz */ 372679dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 372779dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 372879dfc695SMike Kravetz else 372979dfc695SMike Kravetz nr_available = h->free_huge_pages; 373079dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 373179dfc695SMike Kravetz if (!nr_available) 373279dfc695SMike Kravetz break; 373379dfc695SMike Kravetz 373479dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 373579dfc695SMike Kravetz if (err) 373679dfc695SMike Kravetz break; 373779dfc695SMike Kravetz 373879dfc695SMike Kravetz nr_demote--; 373979dfc695SMike Kravetz } 374079dfc695SMike Kravetz 374179dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 374279dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 374379dfc695SMike Kravetz 374479dfc695SMike Kravetz if (err) 374579dfc695SMike Kravetz return err; 374679dfc695SMike Kravetz return len; 374779dfc695SMike Kravetz } 374879dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 374979dfc695SMike Kravetz 375079dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 375179dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 375279dfc695SMike Kravetz { 375379dfc695SMike Kravetz int nid; 375479dfc695SMike Kravetz struct hstate *h = kobj_to_hstate(kobj, &nid); 375579dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 375679dfc695SMike Kravetz 375779dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 375879dfc695SMike Kravetz } 375979dfc695SMike Kravetz 376079dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 376179dfc695SMike Kravetz struct kobj_attribute *attr, 376279dfc695SMike Kravetz const char *buf, size_t count) 376379dfc695SMike Kravetz { 376479dfc695SMike Kravetz struct hstate *h, *demote_hstate; 376579dfc695SMike Kravetz unsigned long demote_size; 376679dfc695SMike Kravetz unsigned int demote_order; 376779dfc695SMike Kravetz int nid; 376879dfc695SMike Kravetz 376979dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 377079dfc695SMike Kravetz 377179dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 377279dfc695SMike Kravetz if (!demote_hstate) 377379dfc695SMike Kravetz return -EINVAL; 377479dfc695SMike Kravetz demote_order = demote_hstate->order; 3775a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3776a01f4390SMike Kravetz return -EINVAL; 377779dfc695SMike Kravetz 377879dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 377979dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 378079dfc695SMike Kravetz if (demote_order >= h->order) 378179dfc695SMike Kravetz return -EINVAL; 378279dfc695SMike Kravetz 378379dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 378479dfc695SMike Kravetz mutex_lock(&h->resize_lock); 378579dfc695SMike Kravetz h->demote_order = demote_order; 378679dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 378779dfc695SMike Kravetz 378879dfc695SMike Kravetz return count; 378979dfc695SMike Kravetz } 379079dfc695SMike Kravetz HSTATE_ATTR(demote_size); 379179dfc695SMike Kravetz 3792a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3793a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3794a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3795a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3796a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3797a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 379806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 379906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 380006808b08SLee Schermerhorn #endif 3801a3437870SNishanth Aravamudan NULL, 3802a3437870SNishanth Aravamudan }; 3803a3437870SNishanth Aravamudan 380467e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3805a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3806a3437870SNishanth Aravamudan }; 3807a3437870SNishanth Aravamudan 380879dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 380979dfc695SMike Kravetz &demote_size_attr.attr, 381079dfc695SMike Kravetz &demote_attr.attr, 381179dfc695SMike Kravetz NULL, 381279dfc695SMike Kravetz }; 381379dfc695SMike Kravetz 381479dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 381579dfc695SMike Kravetz .attrs = hstate_demote_attrs, 381679dfc695SMike Kravetz }; 381779dfc695SMike Kravetz 3818094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38199a305230SLee Schermerhorn struct kobject **hstate_kobjs, 382067e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3821a3437870SNishanth Aravamudan { 3822a3437870SNishanth Aravamudan int retval; 3823972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3824a3437870SNishanth Aravamudan 38259a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38269a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3827a3437870SNishanth Aravamudan return -ENOMEM; 3828a3437870SNishanth Aravamudan 38299a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3830cc2205a6SMiaohe Lin if (retval) { 38319a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3832cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3833cc2205a6SMiaohe Lin } 3834a3437870SNishanth Aravamudan 383579dfc695SMike Kravetz if (h->demote_order) { 383679dfc695SMike Kravetz if (sysfs_create_group(hstate_kobjs[hi], 383779dfc695SMike Kravetz &hstate_demote_attr_group)) 383879dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 383979dfc695SMike Kravetz } 384079dfc695SMike Kravetz 3841a3437870SNishanth Aravamudan return retval; 3842a3437870SNishanth Aravamudan } 3843a3437870SNishanth Aravamudan 3844a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3845a3437870SNishanth Aravamudan { 3846a3437870SNishanth Aravamudan struct hstate *h; 3847a3437870SNishanth Aravamudan int err; 3848a3437870SNishanth Aravamudan 3849a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3850a3437870SNishanth Aravamudan if (!hugepages_kobj) 3851a3437870SNishanth Aravamudan return; 3852a3437870SNishanth Aravamudan 3853a3437870SNishanth Aravamudan for_each_hstate(h) { 38549a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 38559a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3856a3437870SNishanth Aravamudan if (err) 3857282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3858a3437870SNishanth Aravamudan } 3859a3437870SNishanth Aravamudan } 3860a3437870SNishanth Aravamudan 38619a305230SLee Schermerhorn #ifdef CONFIG_NUMA 38629a305230SLee Schermerhorn 38639a305230SLee Schermerhorn /* 38649a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 386510fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 386610fbcf4cSKay Sievers * index of a node device or _hstate == node id. 386710fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38689a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38699a305230SLee Schermerhorn */ 38709a305230SLee Schermerhorn struct node_hstate { 38719a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38729a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38739a305230SLee Schermerhorn }; 3874b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38759a305230SLee Schermerhorn 38769a305230SLee Schermerhorn /* 387710fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 38789a305230SLee Schermerhorn */ 38799a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 38809a305230SLee Schermerhorn &nr_hugepages_attr.attr, 38819a305230SLee Schermerhorn &free_hugepages_attr.attr, 38829a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 38839a305230SLee Schermerhorn NULL, 38849a305230SLee Schermerhorn }; 38859a305230SLee Schermerhorn 388667e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 38879a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 38889a305230SLee Schermerhorn }; 38899a305230SLee Schermerhorn 38909a305230SLee Schermerhorn /* 389110fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 38929a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 38939a305230SLee Schermerhorn */ 38949a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 38959a305230SLee Schermerhorn { 38969a305230SLee Schermerhorn int nid; 38979a305230SLee Schermerhorn 38989a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 38999a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39009a305230SLee Schermerhorn int i; 39019a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39029a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39039a305230SLee Schermerhorn if (nidp) 39049a305230SLee Schermerhorn *nidp = nid; 39059a305230SLee Schermerhorn return &hstates[i]; 39069a305230SLee Schermerhorn } 39079a305230SLee Schermerhorn } 39089a305230SLee Schermerhorn 39099a305230SLee Schermerhorn BUG(); 39109a305230SLee Schermerhorn return NULL; 39119a305230SLee Schermerhorn } 39129a305230SLee Schermerhorn 39139a305230SLee Schermerhorn /* 391410fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39159a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39169a305230SLee Schermerhorn */ 39173cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 39189a305230SLee Schermerhorn { 39199a305230SLee Schermerhorn struct hstate *h; 392010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39219a305230SLee Schermerhorn 39229a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39239b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39249a305230SLee Schermerhorn 3925972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3926972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3927972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3928972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3929972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3930972dc4deSAneesh Kumar K.V } 39319a305230SLee Schermerhorn } 39329a305230SLee Schermerhorn 39339a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39349a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39359a305230SLee Schermerhorn } 39369a305230SLee Schermerhorn 39379a305230SLee Schermerhorn 39389a305230SLee Schermerhorn /* 393910fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39409a305230SLee Schermerhorn * No-op if attributes already registered. 39419a305230SLee Schermerhorn */ 39423cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 39439a305230SLee Schermerhorn { 39449a305230SLee Schermerhorn struct hstate *h; 394510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39469a305230SLee Schermerhorn int err; 39479a305230SLee Schermerhorn 39489a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39499a305230SLee Schermerhorn return; /* already allocated */ 39509a305230SLee Schermerhorn 39519a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 395210fbcf4cSKay Sievers &node->dev.kobj); 39539a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39549a305230SLee Schermerhorn return; 39559a305230SLee Schermerhorn 39569a305230SLee Schermerhorn for_each_hstate(h) { 39579a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39589a305230SLee Schermerhorn nhs->hstate_kobjs, 39599a305230SLee Schermerhorn &per_node_hstate_attr_group); 39609a305230SLee Schermerhorn if (err) { 3961282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 396210fbcf4cSKay Sievers h->name, node->dev.id); 39639a305230SLee Schermerhorn hugetlb_unregister_node(node); 39649a305230SLee Schermerhorn break; 39659a305230SLee Schermerhorn } 39669a305230SLee Schermerhorn } 39679a305230SLee Schermerhorn } 39689a305230SLee Schermerhorn 39699a305230SLee Schermerhorn /* 39709b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 397110fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 397210fbcf4cSKay Sievers * registered their associated device by this time. 39739a305230SLee Schermerhorn */ 39747d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 39759a305230SLee Schermerhorn { 39769a305230SLee Schermerhorn int nid; 39779a305230SLee Schermerhorn 39788cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 39798732794bSWen Congyang struct node *node = node_devices[nid]; 398010fbcf4cSKay Sievers if (node->dev.id == nid) 39819a305230SLee Schermerhorn hugetlb_register_node(node); 39829a305230SLee Schermerhorn } 39839a305230SLee Schermerhorn 39849a305230SLee Schermerhorn /* 398510fbcf4cSKay Sievers * Let the node device driver know we're here so it can 39869a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 39879a305230SLee Schermerhorn */ 39889a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 39899a305230SLee Schermerhorn hugetlb_unregister_node); 39909a305230SLee Schermerhorn } 39919a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 39929a305230SLee Schermerhorn 39939a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39949a305230SLee Schermerhorn { 39959a305230SLee Schermerhorn BUG(); 39969a305230SLee Schermerhorn if (nidp) 39979a305230SLee Schermerhorn *nidp = -1; 39989a305230SLee Schermerhorn return NULL; 39999a305230SLee Schermerhorn } 40009a305230SLee Schermerhorn 40019a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40029a305230SLee Schermerhorn 40039a305230SLee Schermerhorn #endif 40049a305230SLee Schermerhorn 4005a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4006a3437870SNishanth Aravamudan { 40078382d914SDavidlohr Bueso int i; 40088382d914SDavidlohr Bueso 4009d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4010d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4011d6995da3SMike Kravetz 4012c2833a5bSMike Kravetz if (!hugepages_supported()) { 4013c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4014c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40150ef89d25SBenjamin Herrenschmidt return 0; 4016d715cf80SLiam R. Howlett } 4017d715cf80SLiam R. Howlett 4018282f4214SMike Kravetz /* 4019282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4020282f4214SMike Kravetz * architectures depend on setup being done here. 4021282f4214SMike Kravetz */ 4022a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4023282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4024282f4214SMike Kravetz /* 4025282f4214SMike Kravetz * If we did not parse a default huge page size, set 4026282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4027282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4028282f4214SMike Kravetz * specified, set that here as well. 4029282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4030282f4214SMike Kravetz * setting. A warning will be printed in this case. 4031282f4214SMike Kravetz */ 4032282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4033f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4034282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4035282f4214SMike Kravetz char buf[32]; 4036282f4214SMike Kravetz 4037282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4038282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4039282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4040282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4041282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4042282f4214SMike Kravetz default_hstate_max_huge_pages); 4043282f4214SMike Kravetz } 4044282f4214SMike Kravetz default_hstate.max_huge_pages = 4045282f4214SMike Kravetz default_hstate_max_huge_pages; 4046b5389086SZhenguo Yao 4047b5389086SZhenguo Yao for (i = 0; i < nr_online_nodes; i++) 4048b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4049b5389086SZhenguo Yao default_hugepages_in_node[i]; 4050282f4214SMike Kravetz } 4051f8b74815SVaishali Thakkar } 4052a3437870SNishanth Aravamudan 4053cf11e85fSRoman Gushchin hugetlb_cma_check(); 4054a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4055aa888a74SAndi Kleen gather_bootmem_prealloc(); 4056a3437870SNishanth Aravamudan report_hugepages(); 4057a3437870SNishanth Aravamudan 4058a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 40599a305230SLee Schermerhorn hugetlb_register_all_nodes(); 40607179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 40619a305230SLee Schermerhorn 40628382d914SDavidlohr Bueso #ifdef CONFIG_SMP 40638382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 40648382d914SDavidlohr Bueso #else 40658382d914SDavidlohr Bueso num_fault_mutexes = 1; 40668382d914SDavidlohr Bueso #endif 4067c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 40686da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 40696da2ec56SKees Cook GFP_KERNEL); 4070c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 40718382d914SDavidlohr Bueso 40728382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4073c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4074a3437870SNishanth Aravamudan return 0; 4075a3437870SNishanth Aravamudan } 40763e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4077a3437870SNishanth Aravamudan 4078ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4079ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 40809fee021dSVaishali Thakkar { 4081ae94da89SMike Kravetz return size == HPAGE_SIZE; 40829fee021dSVaishali Thakkar } 40839fee021dSVaishali Thakkar 4084d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4085a3437870SNishanth Aravamudan { 4086a3437870SNishanth Aravamudan struct hstate *h; 40878faa8b07SAndi Kleen unsigned long i; 40888faa8b07SAndi Kleen 4089a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4090a3437870SNishanth Aravamudan return; 4091a3437870SNishanth Aravamudan } 409247d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4093a3437870SNishanth Aravamudan BUG_ON(order == 0); 409447d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 409529383967SMike Kravetz mutex_init(&h->resize_lock); 4096a3437870SNishanth Aravamudan h->order = order; 4097aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 40988faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 40998faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41000edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 410154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 410254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4103a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4104a3437870SNishanth Aravamudan huge_page_size(h)/1024); 410577490587SMuchun Song hugetlb_vmemmap_init(h); 41068faa8b07SAndi Kleen 4107a3437870SNishanth Aravamudan parsed_hstate = h; 4108a3437870SNishanth Aravamudan } 4109a3437870SNishanth Aravamudan 4110b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4111b5389086SZhenguo Yao { 4112b5389086SZhenguo Yao return true; 4113b5389086SZhenguo Yao } 4114282f4214SMike Kravetz /* 4115282f4214SMike Kravetz * hugepages command line processing 4116282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4117282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4118282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4119282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4120282f4214SMike Kravetz */ 4121282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4122a3437870SNishanth Aravamudan { 4123a3437870SNishanth Aravamudan unsigned long *mhp; 41248faa8b07SAndi Kleen static unsigned long *last_mhp; 4125b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4126b5389086SZhenguo Yao int count; 4127b5389086SZhenguo Yao unsigned long tmp; 4128b5389086SZhenguo Yao char *p = s; 4129a3437870SNishanth Aravamudan 41309fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4131282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41329fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4133282f4214SMike Kravetz return 0; 41349fee021dSVaishali Thakkar } 4135282f4214SMike Kravetz 4136a3437870SNishanth Aravamudan /* 4137282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4138282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4139282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4140282f4214SMike Kravetz * default_hugepagesz. 4141a3437870SNishanth Aravamudan */ 41429fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4143a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4144a3437870SNishanth Aravamudan else 4145a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4146a3437870SNishanth Aravamudan 41478faa8b07SAndi Kleen if (mhp == last_mhp) { 4148282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4149282f4214SMike Kravetz return 0; 41508faa8b07SAndi Kleen } 41518faa8b07SAndi Kleen 4152b5389086SZhenguo Yao while (*p) { 4153b5389086SZhenguo Yao count = 0; 4154b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4155b5389086SZhenguo Yao goto invalid; 4156b5389086SZhenguo Yao /* Parameter is node format */ 4157b5389086SZhenguo Yao if (p[count] == ':') { 4158b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4159b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4160b5389086SZhenguo Yao return 0; 4161b5389086SZhenguo Yao } 4162b5389086SZhenguo Yao node = tmp; 4163b5389086SZhenguo Yao p += count + 1; 4164b5389086SZhenguo Yao if (node < 0 || node >= nr_online_nodes) 4165b5389086SZhenguo Yao goto invalid; 4166b5389086SZhenguo Yao /* Parse hugepages */ 4167b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4168b5389086SZhenguo Yao goto invalid; 4169b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4170b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4171b5389086SZhenguo Yao else 4172b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4173b5389086SZhenguo Yao *mhp += tmp; 4174b5389086SZhenguo Yao /* Go to parse next node*/ 4175b5389086SZhenguo Yao if (p[count] == ',') 4176b5389086SZhenguo Yao p += count + 1; 4177b5389086SZhenguo Yao else 4178b5389086SZhenguo Yao break; 4179b5389086SZhenguo Yao } else { 4180b5389086SZhenguo Yao if (p != s) 4181b5389086SZhenguo Yao goto invalid; 4182b5389086SZhenguo Yao *mhp = tmp; 4183b5389086SZhenguo Yao break; 4184b5389086SZhenguo Yao } 4185b5389086SZhenguo Yao } 4186a3437870SNishanth Aravamudan 41878faa8b07SAndi Kleen /* 41888faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 418904adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 41908faa8b07SAndi Kleen * use the bootmem allocator. 41918faa8b07SAndi Kleen */ 419204adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 41938faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 41948faa8b07SAndi Kleen 41958faa8b07SAndi Kleen last_mhp = mhp; 41968faa8b07SAndi Kleen 4197a3437870SNishanth Aravamudan return 1; 4198b5389086SZhenguo Yao 4199b5389086SZhenguo Yao invalid: 4200b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4201b5389086SZhenguo Yao return 0; 4202a3437870SNishanth Aravamudan } 4203282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4204e11bfbfcSNick Piggin 4205282f4214SMike Kravetz /* 4206282f4214SMike Kravetz * hugepagesz command line processing 4207282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4208282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4209282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4210282f4214SMike Kravetz * hugepagesz argument was valid. 4211282f4214SMike Kravetz */ 4212359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4213e11bfbfcSNick Piggin { 4214359f2544SMike Kravetz unsigned long size; 4215282f4214SMike Kravetz struct hstate *h; 4216282f4214SMike Kravetz 4217282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4218359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4219359f2544SMike Kravetz 4220359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4221282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4222359f2544SMike Kravetz return 0; 4223359f2544SMike Kravetz } 4224359f2544SMike Kravetz 4225282f4214SMike Kravetz h = size_to_hstate(size); 4226282f4214SMike Kravetz if (h) { 4227282f4214SMike Kravetz /* 4228282f4214SMike Kravetz * hstate for this size already exists. This is normally 4229282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4230282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4231282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4232282f4214SMike Kravetz * previously specified. 4233282f4214SMike Kravetz */ 4234282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4235282f4214SMike Kravetz default_hstate.max_huge_pages) { 4236282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 423738237830SMike Kravetz return 0; 423838237830SMike Kravetz } 423938237830SMike Kravetz 4240282f4214SMike Kravetz /* 4241282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4242282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4243282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4244282f4214SMike Kravetz */ 4245282f4214SMike Kravetz parsed_hstate = h; 4246282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4247e11bfbfcSNick Piggin return 1; 4248e11bfbfcSNick Piggin } 4249282f4214SMike Kravetz 4250359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4251282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4252359f2544SMike Kravetz return 1; 4253359f2544SMike Kravetz } 4254359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4255359f2544SMike Kravetz 4256282f4214SMike Kravetz /* 4257282f4214SMike Kravetz * default_hugepagesz command line input 4258282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4259282f4214SMike Kravetz */ 4260ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4261e11bfbfcSNick Piggin { 4262ae94da89SMike Kravetz unsigned long size; 4263b5389086SZhenguo Yao int i; 4264ae94da89SMike Kravetz 4265282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4266282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4267282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4268282f4214SMike Kravetz return 0; 4269282f4214SMike Kravetz } 4270282f4214SMike Kravetz 4271282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4272282f4214SMike Kravetz 4273282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4274282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4275282f4214SMike Kravetz return 0; 4276282f4214SMike Kravetz } 4277282f4214SMike Kravetz 4278282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4279282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4280282f4214SMike Kravetz parsed_default_hugepagesz = true; 4281282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4282282f4214SMike Kravetz 4283282f4214SMike Kravetz /* 4284282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4285282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4286282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4287282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4288282f4214SMike Kravetz * allocated here from bootmem allocator. 4289282f4214SMike Kravetz */ 4290282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4291282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 4292b5389086SZhenguo Yao for (i = 0; i < nr_online_nodes; i++) 4293b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4294b5389086SZhenguo Yao default_hugepages_in_node[i]; 4295282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4296282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4297282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4298282f4214SMike Kravetz } 4299282f4214SMike Kravetz 4300e11bfbfcSNick Piggin return 1; 4301e11bfbfcSNick Piggin } 4302ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4303a3437870SNishanth Aravamudan 43048ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43058a213460SNishanth Aravamudan { 43068a213460SNishanth Aravamudan int node; 43078a213460SNishanth Aravamudan unsigned int nr = 0; 43088ca39e68SMuchun Song nodemask_t *mpol_allowed; 43098ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43108ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43118a213460SNishanth Aravamudan 43128ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 43138ca39e68SMuchun Song 43148ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4315c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 43168a213460SNishanth Aravamudan nr += array[node]; 43178ca39e68SMuchun Song } 43188a213460SNishanth Aravamudan 43198a213460SNishanth Aravamudan return nr; 43208a213460SNishanth Aravamudan } 43218a213460SNishanth Aravamudan 43228a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 432317743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 432417743798SMuchun Song void *buffer, size_t *length, 432517743798SMuchun Song loff_t *ppos, unsigned long *out) 432617743798SMuchun Song { 432717743798SMuchun Song struct ctl_table dup_table; 432817743798SMuchun Song 432917743798SMuchun Song /* 433017743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 433117743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 433217743798SMuchun Song */ 433317743798SMuchun Song dup_table = *table; 433417743798SMuchun Song dup_table.data = out; 433517743798SMuchun Song 433617743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 433717743798SMuchun Song } 433817743798SMuchun Song 433906808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 434006808b08SLee Schermerhorn struct ctl_table *table, int write, 434132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 43421da177e4SLinus Torvalds { 4343e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4344238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 434508d4a246SMichal Hocko int ret; 4346e5ff2159SAndi Kleen 4347457c1b27SNishanth Aravamudan if (!hugepages_supported()) 434886613628SJan Stancek return -EOPNOTSUPP; 4349457c1b27SNishanth Aravamudan 435017743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 435117743798SMuchun Song &tmp); 435208d4a246SMichal Hocko if (ret) 435308d4a246SMichal Hocko goto out; 4354e5ff2159SAndi Kleen 4355238d3c13SDavid Rientjes if (write) 4356238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4357238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 435808d4a246SMichal Hocko out: 435908d4a246SMichal Hocko return ret; 43601da177e4SLinus Torvalds } 4361396faf03SMel Gorman 436206808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 436332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 436406808b08SLee Schermerhorn { 436506808b08SLee Schermerhorn 436606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 436706808b08SLee Schermerhorn buffer, length, ppos); 436806808b08SLee Schermerhorn } 436906808b08SLee Schermerhorn 437006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 437106808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 437232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 437306808b08SLee Schermerhorn { 437406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 437506808b08SLee Schermerhorn buffer, length, ppos); 437606808b08SLee Schermerhorn } 437706808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 437806808b08SLee Schermerhorn 4379a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 438032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4381a3d0c6aaSNishanth Aravamudan { 4382a5516438SAndi Kleen struct hstate *h = &default_hstate; 4383e5ff2159SAndi Kleen unsigned long tmp; 438408d4a246SMichal Hocko int ret; 4385e5ff2159SAndi Kleen 4386457c1b27SNishanth Aravamudan if (!hugepages_supported()) 438786613628SJan Stancek return -EOPNOTSUPP; 4388457c1b27SNishanth Aravamudan 4389e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4390e5ff2159SAndi Kleen 4391bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4392adbe8726SEric B Munson return -EINVAL; 4393adbe8726SEric B Munson 439417743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 439517743798SMuchun Song &tmp); 439608d4a246SMichal Hocko if (ret) 439708d4a246SMichal Hocko goto out; 4398e5ff2159SAndi Kleen 4399e5ff2159SAndi Kleen if (write) { 4400db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4401e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4402db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4403e5ff2159SAndi Kleen } 440408d4a246SMichal Hocko out: 440508d4a246SMichal Hocko return ret; 4406a3d0c6aaSNishanth Aravamudan } 4407a3d0c6aaSNishanth Aravamudan 44081da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44091da177e4SLinus Torvalds 4410e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44111da177e4SLinus Torvalds { 4412fcb2b0c5SRoman Gushchin struct hstate *h; 4413fcb2b0c5SRoman Gushchin unsigned long total = 0; 4414fcb2b0c5SRoman Gushchin 4415457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4416457c1b27SNishanth Aravamudan return; 4417fcb2b0c5SRoman Gushchin 4418fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4419fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4420fcb2b0c5SRoman Gushchin 4421aca78307SMiaohe Lin total += huge_page_size(h) * count; 4422fcb2b0c5SRoman Gushchin 4423fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4424e1759c21SAlexey Dobriyan seq_printf(m, 44251da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 44261da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4427b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 44287893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 44294f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4430fcb2b0c5SRoman Gushchin count, 4431a5516438SAndi Kleen h->free_huge_pages, 4432a5516438SAndi Kleen h->resv_huge_pages, 4433a5516438SAndi Kleen h->surplus_huge_pages, 4434aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4435fcb2b0c5SRoman Gushchin } 4436fcb2b0c5SRoman Gushchin 4437aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 44381da177e4SLinus Torvalds } 44391da177e4SLinus Torvalds 44407981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 44411da177e4SLinus Torvalds { 4442a5516438SAndi Kleen struct hstate *h = &default_hstate; 44437981593bSJoe Perches 4444457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4445457c1b27SNishanth Aravamudan return 0; 44467981593bSJoe Perches 44477981593bSJoe Perches return sysfs_emit_at(buf, len, 44481da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4449a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4450a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4451a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4452a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4453a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 44541da177e4SLinus Torvalds } 44551da177e4SLinus Torvalds 4456949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 4457949f7ec5SDavid Rientjes { 4458949f7ec5SDavid Rientjes struct hstate *h; 4459949f7ec5SDavid Rientjes int nid; 4460949f7ec5SDavid Rientjes 4461457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4462457c1b27SNishanth Aravamudan return; 4463457c1b27SNishanth Aravamudan 4464949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 4465949f7ec5SDavid Rientjes for_each_hstate(h) 4466949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4467949f7ec5SDavid Rientjes nid, 4468949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4469949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4470949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4471aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4472949f7ec5SDavid Rientjes } 4473949f7ec5SDavid Rientjes 44745d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 44755d317b2bSNaoya Horiguchi { 44765d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 44775d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 44785d317b2bSNaoya Horiguchi } 44795d317b2bSNaoya Horiguchi 44801da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 44811da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 44821da177e4SLinus Torvalds { 4483d0028588SWanpeng Li struct hstate *h; 4484d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4485d0028588SWanpeng Li 4486d0028588SWanpeng Li for_each_hstate(h) 4487d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4488d0028588SWanpeng Li return nr_total_pages; 44891da177e4SLinus Torvalds } 44901da177e4SLinus Torvalds 4491a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4492fc1b8a73SMel Gorman { 4493fc1b8a73SMel Gorman int ret = -ENOMEM; 4494fc1b8a73SMel Gorman 44950aa7f354SMiaohe Lin if (!delta) 44960aa7f354SMiaohe Lin return 0; 44970aa7f354SMiaohe Lin 4498db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4499fc1b8a73SMel Gorman /* 4500fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4501fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4502fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4503fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4504fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4505fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4506fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4507fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4508fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4509fc1b8a73SMel Gorman * 4510fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4511fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4512fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4513fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4514fc1b8a73SMel Gorman * semantics that cpuset has. 45158ca39e68SMuchun Song * 45168ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45178ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45188ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45198ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45208ca39e68SMuchun Song * above. 4521fc1b8a73SMel Gorman */ 4522fc1b8a73SMel Gorman if (delta > 0) { 4523a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4524fc1b8a73SMel Gorman goto out; 4525fc1b8a73SMel Gorman 45268ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4527a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4528fc1b8a73SMel Gorman goto out; 4529fc1b8a73SMel Gorman } 4530fc1b8a73SMel Gorman } 4531fc1b8a73SMel Gorman 4532fc1b8a73SMel Gorman ret = 0; 4533fc1b8a73SMel Gorman if (delta < 0) 4534a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4535fc1b8a73SMel Gorman 4536fc1b8a73SMel Gorman out: 4537db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4538fc1b8a73SMel Gorman return ret; 4539fc1b8a73SMel Gorman } 4540fc1b8a73SMel Gorman 454184afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 454284afd99bSAndy Whitcroft { 4543f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 454484afd99bSAndy Whitcroft 454584afd99bSAndy Whitcroft /* 454684afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 454784afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 454884afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 454925985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 455084afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 455184afd99bSAndy Whitcroft * new reference here without additional locking. 455284afd99bSAndy Whitcroft */ 455309a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 455409a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4555f522c3acSJoonsoo Kim kref_get(&resv->refs); 455684afd99bSAndy Whitcroft } 455709a26e83SMike Kravetz } 455884afd99bSAndy Whitcroft 4559a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4560a1e78772SMel Gorman { 4561a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4562f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 456390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 45644e35f483SJoonsoo Kim unsigned long reserve, start, end; 45651c5ecae3SMike Kravetz long gbl_reserve; 456684afd99bSAndy Whitcroft 45674e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 45684e35f483SJoonsoo Kim return; 45694e35f483SJoonsoo Kim 4570a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4571a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 457284afd99bSAndy Whitcroft 45734e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4574e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 45757251ff78SAdam Litke if (reserve) { 45761c5ecae3SMike Kravetz /* 45771c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 45781c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 45791c5ecae3SMike Kravetz */ 45801c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 45811c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 45827251ff78SAdam Litke } 4583e9fe92aeSMina Almasry 4584e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4585a1e78772SMel Gorman } 4586a1e78772SMel Gorman 458731383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 458831383c68SDan Williams { 458931383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 459031383c68SDan Williams return -EINVAL; 459131383c68SDan Williams return 0; 459231383c68SDan Williams } 459331383c68SDan Williams 459405ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 459505ea8860SDan Williams { 4596aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 459705ea8860SDan Williams } 459805ea8860SDan Williams 45991da177e4SLinus Torvalds /* 46001da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46011da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46026c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46031da177e4SLinus Torvalds * this far. 46041da177e4SLinus Torvalds */ 4605b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 46061da177e4SLinus Torvalds { 46071da177e4SLinus Torvalds BUG(); 4608d0217ac0SNick Piggin return 0; 46091da177e4SLinus Torvalds } 46101da177e4SLinus Torvalds 4611eec3636aSJane Chu /* 4612eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4613eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4614eec3636aSJane Chu * This is because under System V memory model, mappings created via 4615eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4616eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4617eec3636aSJane Chu */ 4618f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4619d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 462084afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4621a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4622dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 462305ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 46241da177e4SLinus Torvalds }; 46251da177e4SLinus Torvalds 46261e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 46271e8f889bSDavid Gibson int writable) 462863551ae0SDavid Gibson { 462963551ae0SDavid Gibson pte_t entry; 463079c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 463163551ae0SDavid Gibson 46321e8f889bSDavid Gibson if (writable) { 4633106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4634106c992aSGerald Schaefer vma->vm_page_prot))); 463563551ae0SDavid Gibson } else { 4636106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4637106c992aSGerald Schaefer vma->vm_page_prot)); 463863551ae0SDavid Gibson } 463963551ae0SDavid Gibson entry = pte_mkyoung(entry); 464063551ae0SDavid Gibson entry = pte_mkhuge(entry); 464179c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 464263551ae0SDavid Gibson 464363551ae0SDavid Gibson return entry; 464463551ae0SDavid Gibson } 464563551ae0SDavid Gibson 46461e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 46471e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 46481e8f889bSDavid Gibson { 46491e8f889bSDavid Gibson pte_t entry; 46501e8f889bSDavid Gibson 4651106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 465232f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 46534b3073e1SRussell King update_mmu_cache(vma, address, ptep); 46541e8f889bSDavid Gibson } 46551e8f889bSDavid Gibson 4656d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 46574a705fefSNaoya Horiguchi { 46584a705fefSNaoya Horiguchi swp_entry_t swp; 46594a705fefSNaoya Horiguchi 46604a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4661d5ed7444SAneesh Kumar K.V return false; 46624a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4663d79d176aSBaoquan He if (is_migration_entry(swp)) 4664d5ed7444SAneesh Kumar K.V return true; 46654a705fefSNaoya Horiguchi else 4666d5ed7444SAneesh Kumar K.V return false; 46674a705fefSNaoya Horiguchi } 46684a705fefSNaoya Horiguchi 46693e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 46704a705fefSNaoya Horiguchi { 46714a705fefSNaoya Horiguchi swp_entry_t swp; 46724a705fefSNaoya Horiguchi 46734a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 46743e5c3600SBaoquan He return false; 46754a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4676d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 46773e5c3600SBaoquan He return true; 46784a705fefSNaoya Horiguchi else 46793e5c3600SBaoquan He return false; 46804a705fefSNaoya Horiguchi } 46811e8f889bSDavid Gibson 46824eae4efaSPeter Xu static void 46834eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 46844eae4efaSPeter Xu struct page *new_page) 46854eae4efaSPeter Xu { 46864eae4efaSPeter Xu __SetPageUptodate(new_page); 46874eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 46881eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 46894eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 46904eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 46914eae4efaSPeter Xu SetHPageMigratable(new_page); 46924eae4efaSPeter Xu } 46934eae4efaSPeter Xu 469463551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 469563551ae0SDavid Gibson struct vm_area_struct *vma) 469663551ae0SDavid Gibson { 46975e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 469863551ae0SDavid Gibson struct page *ptepage; 46991c59827dSHugh Dickins unsigned long addr; 4700ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 4701a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4702a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47034eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4704c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4705ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4706e8569dd2SAndreas Sandberg int ret = 0; 47071e8f889bSDavid Gibson 4708ac46d4f3SJérôme Glisse if (cow) { 47097269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 47106f4f13e8SJérôme Glisse vma->vm_start, 4711ac46d4f3SJérôme Glisse vma->vm_end); 4712ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4713c0d0381aSMike Kravetz } else { 4714c0d0381aSMike Kravetz /* 4715c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4716c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4717c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4718c0d0381aSMike Kravetz * huge_pmd_unshare. 4719c0d0381aSMike Kravetz */ 4720c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4721ac46d4f3SJérôme Glisse } 4722e8569dd2SAndreas Sandberg 4723a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 4724cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 47257868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4726c74df32cSHugh Dickins if (!src_pte) 4727c74df32cSHugh Dickins continue; 4728aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 4729e8569dd2SAndreas Sandberg if (!dst_pte) { 4730e8569dd2SAndreas Sandberg ret = -ENOMEM; 4731e8569dd2SAndreas Sandberg break; 4732e8569dd2SAndreas Sandberg } 4733c5c99429SLarry Woodman 47345e41540cSMike Kravetz /* 47355e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 47365e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 47375e41540cSMike Kravetz * 47385e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 47395e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 47405e41540cSMike Kravetz * Check here before taking page table lock, and once again 47415e41540cSMike Kravetz * after taking the lock below. 47425e41540cSMike Kravetz */ 47435e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 47445e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 4745c5c99429SLarry Woodman continue; 4746c5c99429SLarry Woodman 4747cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4748cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4749cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 47504a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 47515e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 47524eae4efaSPeter Xu again: 47535e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 47545e41540cSMike Kravetz /* 47555e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 47565e41540cSMike Kravetz * unlikely case dst entry !none as this implies 47575e41540cSMike Kravetz * sharing with another vma. 47585e41540cSMike Kravetz */ 47594a705fefSNaoya Horiguchi ; 47604a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 47614a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 47624a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 47634a705fefSNaoya Horiguchi 47644dd845b5SAlistair Popple if (is_writable_migration_entry(swp_entry) && cow) { 47654a705fefSNaoya Horiguchi /* 47664a705fefSNaoya Horiguchi * COW mappings require pages in both 47674a705fefSNaoya Horiguchi * parent and child to be set to read. 47684a705fefSNaoya Horiguchi */ 47694dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 47704dd845b5SAlistair Popple swp_offset(swp_entry)); 47714a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4772e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4773e5251fd4SPunit Agrawal entry, sz); 47744a705fefSNaoya Horiguchi } 4775e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 47764a705fefSNaoya Horiguchi } else { 47774eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 47784eae4efaSPeter Xu ptepage = pte_page(entry); 47794eae4efaSPeter Xu get_page(ptepage); 47804eae4efaSPeter Xu 47814eae4efaSPeter Xu /* 47824eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 47834eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 47844eae4efaSPeter Xu * COW earlier during fork. 47854eae4efaSPeter Xu * 47864eae4efaSPeter Xu * When pre-allocating the page or copying data, we 47874eae4efaSPeter Xu * need to be without the pgtable locks since we could 47884eae4efaSPeter Xu * sleep during the process. 47894eae4efaSPeter Xu */ 47904eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 47914eae4efaSPeter Xu pte_t src_pte_old = entry; 47924eae4efaSPeter Xu struct page *new; 47934eae4efaSPeter Xu 47944eae4efaSPeter Xu spin_unlock(src_ptl); 47954eae4efaSPeter Xu spin_unlock(dst_ptl); 47964eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 47974eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 47984eae4efaSPeter Xu if (IS_ERR(new)) { 47994eae4efaSPeter Xu put_page(ptepage); 48004eae4efaSPeter Xu ret = PTR_ERR(new); 48014eae4efaSPeter Xu break; 48024eae4efaSPeter Xu } 48034eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 48044eae4efaSPeter Xu npages); 48054eae4efaSPeter Xu put_page(ptepage); 48064eae4efaSPeter Xu 48074eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 48084eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 48094eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 48104eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48114eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48124eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4813846be085SMike Kravetz restore_reserve_on_error(h, vma, addr, 4814846be085SMike Kravetz new); 48154eae4efaSPeter Xu put_page(new); 48164eae4efaSPeter Xu /* dst_entry won't change as in child */ 48174eae4efaSPeter Xu goto again; 48184eae4efaSPeter Xu } 48194eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 48204eae4efaSPeter Xu spin_unlock(src_ptl); 48214eae4efaSPeter Xu spin_unlock(dst_ptl); 48224eae4efaSPeter Xu continue; 48234eae4efaSPeter Xu } 48244eae4efaSPeter Xu 482534ee645eSJoerg Roedel if (cow) { 48260f10851eSJérôme Glisse /* 48270f10851eSJérôme Glisse * No need to notify as we are downgrading page 48280f10851eSJérôme Glisse * table protection not changing it to point 48290f10851eSJérôme Glisse * to a new page. 48300f10851eSJérôme Glisse * 4831ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 48320f10851eSJérôme Glisse */ 48337f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 483484894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 483534ee645eSJoerg Roedel } 48364eae4efaSPeter Xu 483753f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 483863551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 48394eae4efaSPeter Xu hugetlb_count_add(npages, dst); 48401c59827dSHugh Dickins } 4841cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4842cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 484363551ae0SDavid Gibson } 484463551ae0SDavid Gibson 4845e8569dd2SAndreas Sandberg if (cow) 4846ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4847c0d0381aSMike Kravetz else 4848c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4849e8569dd2SAndreas Sandberg 4850e8569dd2SAndreas Sandberg return ret; 485163551ae0SDavid Gibson } 485263551ae0SDavid Gibson 4853550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4854*db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4855550a7d60SMina Almasry { 4856550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4857550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4858550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4859*db110a99SAneesh Kumar K.V pte_t pte; 4860550a7d60SMina Almasry 4861550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4862550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4863550a7d60SMina Almasry 4864550a7d60SMina Almasry /* 4865550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4866550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4867550a7d60SMina Almasry */ 4868550a7d60SMina Almasry if (src_ptl != dst_ptl) 4869550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4870550a7d60SMina Almasry 4871550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4872550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4873550a7d60SMina Almasry 4874550a7d60SMina Almasry if (src_ptl != dst_ptl) 4875550a7d60SMina Almasry spin_unlock(src_ptl); 4876550a7d60SMina Almasry spin_unlock(dst_ptl); 4877550a7d60SMina Almasry } 4878550a7d60SMina Almasry 4879550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4880550a7d60SMina Almasry struct vm_area_struct *new_vma, 4881550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4882550a7d60SMina Almasry unsigned long len) 4883550a7d60SMina Almasry { 4884550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4885550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4886550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4887550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4888550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4889550a7d60SMina Almasry unsigned long old_addr_copy; 4890550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4891550a7d60SMina Almasry struct mmu_notifier_range range; 4892550a7d60SMina Almasry 4893550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 4894550a7d60SMina Almasry old_end); 4895550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4896550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 4897550a7d60SMina Almasry /* Prevent race with file truncation */ 4898550a7d60SMina Almasry i_mmap_lock_write(mapping); 4899550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 4900550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 4901550a7d60SMina Almasry if (!src_pte) 4902550a7d60SMina Almasry continue; 4903550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 4904550a7d60SMina Almasry continue; 4905550a7d60SMina Almasry 4906550a7d60SMina Almasry /* old_addr arg to huge_pmd_unshare() is a pointer and so the 4907550a7d60SMina Almasry * arg may be modified. Pass a copy instead to preserve the 4908550a7d60SMina Almasry * value in old_addr. 4909550a7d60SMina Almasry */ 4910550a7d60SMina Almasry old_addr_copy = old_addr; 4911550a7d60SMina Almasry 4912550a7d60SMina Almasry if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) 4913550a7d60SMina Almasry continue; 4914550a7d60SMina Almasry 4915550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 4916550a7d60SMina Almasry if (!dst_pte) 4917550a7d60SMina Almasry break; 4918550a7d60SMina Almasry 4919*db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 4920550a7d60SMina Almasry } 4921550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 4922550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 492313e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 4924550a7d60SMina Almasry 4925550a7d60SMina Almasry return len + old_addr - old_end; 4926550a7d60SMina Almasry } 4927550a7d60SMina Almasry 492873c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 492924669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 493024669e58SAneesh Kumar K.V struct page *ref_page) 493163551ae0SDavid Gibson { 493263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 493363551ae0SDavid Gibson unsigned long address; 4934c7546f8fSDavid Gibson pte_t *ptep; 493563551ae0SDavid Gibson pte_t pte; 4936cb900f41SKirill A. Shutemov spinlock_t *ptl; 493763551ae0SDavid Gibson struct page *page; 4938a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4939a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4940ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4941a4a118f2SNadav Amit bool force_flush = false; 4942a5516438SAndi Kleen 494363551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4944a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4945a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 494663551ae0SDavid Gibson 494707e32661SAneesh Kumar K.V /* 494807e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 494907e32661SAneesh Kumar K.V * to huge page. 495007e32661SAneesh Kumar K.V */ 4951ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 495224669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4953dff11abeSMike Kravetz 4954dff11abeSMike Kravetz /* 4955dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4956dff11abeSMike Kravetz */ 49576f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 49586f4f13e8SJérôme Glisse end); 4959ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4960ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4961569f48b8SHillf Danton address = start; 4962569f48b8SHillf Danton for (; address < end; address += sz) { 49637868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4964c7546f8fSDavid Gibson if (!ptep) 4965c7546f8fSDavid Gibson continue; 4966c7546f8fSDavid Gibson 4967cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 496834ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 496931d49da5SAneesh Kumar K.V spin_unlock(ptl); 4970a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 4971a4a118f2SNadav Amit force_flush = true; 497231d49da5SAneesh Kumar K.V continue; 497331d49da5SAneesh Kumar K.V } 497439dde65cSChen, Kenneth W 49756629326bSHillf Danton pte = huge_ptep_get(ptep); 497631d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 497731d49da5SAneesh Kumar K.V spin_unlock(ptl); 497831d49da5SAneesh Kumar K.V continue; 497931d49da5SAneesh Kumar K.V } 49806629326bSHillf Danton 49816629326bSHillf Danton /* 49829fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 49839fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 49846629326bSHillf Danton */ 49859fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 49869386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 498731d49da5SAneesh Kumar K.V spin_unlock(ptl); 498831d49da5SAneesh Kumar K.V continue; 49898c4894c6SNaoya Horiguchi } 49906629326bSHillf Danton 49916629326bSHillf Danton page = pte_page(pte); 499204f2cbe3SMel Gorman /* 499304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 499404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 499504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 499604f2cbe3SMel Gorman */ 499704f2cbe3SMel Gorman if (ref_page) { 499831d49da5SAneesh Kumar K.V if (page != ref_page) { 499931d49da5SAneesh Kumar K.V spin_unlock(ptl); 500031d49da5SAneesh Kumar K.V continue; 500131d49da5SAneesh Kumar K.V } 500204f2cbe3SMel Gorman /* 500304f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 500404f2cbe3SMel Gorman * future faults in this VMA will fail rather than 500504f2cbe3SMel Gorman * looking like data was lost 500604f2cbe3SMel Gorman */ 500704f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 500804f2cbe3SMel Gorman } 500904f2cbe3SMel Gorman 5010c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5011b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5012106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 50136649a386SKen Chen set_page_dirty(page); 50149e81130bSHillf Danton 50155d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5016d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 501731d49da5SAneesh Kumar K.V 5018cb900f41SKirill A. Shutemov spin_unlock(ptl); 5019e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 502024669e58SAneesh Kumar K.V /* 502131d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 502224669e58SAneesh Kumar K.V */ 502331d49da5SAneesh Kumar K.V if (ref_page) 502431d49da5SAneesh Kumar K.V break; 5025fe1668aeSChen, Kenneth W } 5026ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 502724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5028a4a118f2SNadav Amit 5029a4a118f2SNadav Amit /* 5030a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5031a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5032a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5033a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5034a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5035a4a118f2SNadav Amit * dropped as well. 5036a4a118f2SNadav Amit * 5037a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5038a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5039a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5040a4a118f2SNadav Amit * Instead, do flush now. 5041a4a118f2SNadav Amit */ 5042a4a118f2SNadav Amit if (force_flush) 5043a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 50441da177e4SLinus Torvalds } 504563551ae0SDavid Gibson 5046d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5047d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 5048d833352aSMel Gorman unsigned long end, struct page *ref_page) 5049d833352aSMel Gorman { 5050d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 5051d833352aSMel Gorman 5052d833352aSMel Gorman /* 5053d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 5054d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 5055d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 5056d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 5057d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 5058c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 5059d833352aSMel Gorman * because in the context this is called, the VMA is about to be 5060c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 5061d833352aSMel Gorman */ 5062d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 5063d833352aSMel Gorman } 5064d833352aSMel Gorman 5065502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 506604f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 5067502717f4SChen, Kenneth W { 506824669e58SAneesh Kumar K.V struct mmu_gather tlb; 5069dff11abeSMike Kravetz 5070a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 507124669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 5072ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5073502717f4SChen, Kenneth W } 5074502717f4SChen, Kenneth W 507504f2cbe3SMel Gorman /* 507604f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5077578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 507804f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 507904f2cbe3SMel Gorman * same region. 508004f2cbe3SMel Gorman */ 50812f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 50822a4b3dedSHarvey Harrison struct page *page, unsigned long address) 508304f2cbe3SMel Gorman { 50847526674dSAdam Litke struct hstate *h = hstate_vma(vma); 508504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 508604f2cbe3SMel Gorman struct address_space *mapping; 508704f2cbe3SMel Gorman pgoff_t pgoff; 508804f2cbe3SMel Gorman 508904f2cbe3SMel Gorman /* 509004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 509104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 509204f2cbe3SMel Gorman */ 50937526674dSAdam Litke address = address & huge_page_mask(h); 509436e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 509536e4f20aSMichal Hocko vma->vm_pgoff; 509693c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 509704f2cbe3SMel Gorman 50984eb2b1dcSMel Gorman /* 50994eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 51004eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 51014eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 51024eb2b1dcSMel Gorman */ 510383cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 51046b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 510504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 510604f2cbe3SMel Gorman if (iter_vma == vma) 510704f2cbe3SMel Gorman continue; 510804f2cbe3SMel Gorman 510904f2cbe3SMel Gorman /* 51102f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 51112f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 51122f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 51132f84a899SMel Gorman */ 51142f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 51152f84a899SMel Gorman continue; 51162f84a899SMel Gorman 51172f84a899SMel Gorman /* 511804f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 511904f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 512004f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 512104f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 512204f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 512304f2cbe3SMel Gorman */ 512404f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 512524669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 512624669e58SAneesh Kumar K.V address + huge_page_size(h), page); 512704f2cbe3SMel Gorman } 512883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 512904f2cbe3SMel Gorman } 513004f2cbe3SMel Gorman 51310fe6e20bSNaoya Horiguchi /* 51320fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 5133aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5134ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5135ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 51360fe6e20bSNaoya Horiguchi */ 51372b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 5138974e6d66SHuang Ying unsigned long address, pte_t *ptep, 5139cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 51401e8f889bSDavid Gibson { 51413999f52eSAneesh Kumar K.V pte_t pte; 5142a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 51431e8f889bSDavid Gibson struct page *old_page, *new_page; 51442b740303SSouptick Joarder int outside_reserve = 0; 51452b740303SSouptick Joarder vm_fault_t ret = 0; 5146974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5147ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 51481e8f889bSDavid Gibson 51493999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 51501e8f889bSDavid Gibson old_page = pte_page(pte); 51511e8f889bSDavid Gibson 515204f2cbe3SMel Gorman retry_avoidcopy: 51531e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 51541e8f889bSDavid Gibson * and just make the page writable */ 515537a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 51565a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 51575b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 515883c54070SNick Piggin return 0; 51591e8f889bSDavid Gibson } 51601e8f889bSDavid Gibson 516104f2cbe3SMel Gorman /* 516204f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 516304f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 516404f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 516504f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 516604f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 516704f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 516804f2cbe3SMel Gorman * of the full address range. 516904f2cbe3SMel Gorman */ 51705944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 517104f2cbe3SMel Gorman old_page != pagecache_page) 517204f2cbe3SMel Gorman outside_reserve = 1; 517304f2cbe3SMel Gorman 517409cbfeafSKirill A. Shutemov get_page(old_page); 5175b76c8cfbSLarry Woodman 5176ad4404a2SDavidlohr Bueso /* 5177ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5178ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5179ad4404a2SDavidlohr Bueso */ 5180cb900f41SKirill A. Shutemov spin_unlock(ptl); 51815b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 51821e8f889bSDavid Gibson 51832fc39cecSAdam Litke if (IS_ERR(new_page)) { 518404f2cbe3SMel Gorman /* 518504f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 518604f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 518704f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 518804f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 518904f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 519004f2cbe3SMel Gorman */ 519104f2cbe3SMel Gorman if (outside_reserve) { 5192e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 5193e7dd91c4SMike Kravetz pgoff_t idx; 5194e7dd91c4SMike Kravetz u32 hash; 5195e7dd91c4SMike Kravetz 519609cbfeafSKirill A. Shutemov put_page(old_page); 519704f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 5198e7dd91c4SMike Kravetz /* 5199e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 5200e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 5201e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 5202e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 5203e7dd91c4SMike Kravetz * PMD sharing. 5204e7dd91c4SMike Kravetz * 5205e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 5206e7dd91c4SMike Kravetz */ 5207e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5208e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 5209e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5210e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 5211e7dd91c4SMike Kravetz 52125b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 5213e7dd91c4SMike Kravetz 5214e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 5215e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 5216cb900f41SKirill A. Shutemov spin_lock(ptl); 52175b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5218a9af0c5dSNaoya Horiguchi if (likely(ptep && 5219a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 522004f2cbe3SMel Gorman goto retry_avoidcopy; 5221a734bcc8SHillf Danton /* 5222cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5223cb900f41SKirill A. Shutemov * lock, and our job is done. 5224a734bcc8SHillf Danton */ 5225a734bcc8SHillf Danton return 0; 522604f2cbe3SMel Gorman } 522704f2cbe3SMel Gorman 52282b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5229ad4404a2SDavidlohr Bueso goto out_release_old; 52301e8f889bSDavid Gibson } 52311e8f889bSDavid Gibson 52320fe6e20bSNaoya Horiguchi /* 52330fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 52340fe6e20bSNaoya Horiguchi * anon_vma prepared. 52350fe6e20bSNaoya Horiguchi */ 523644e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5237ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5238ad4404a2SDavidlohr Bueso goto out_release_all; 523944e2aa93SDean Nelson } 52400fe6e20bSNaoya Horiguchi 5241974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 524247ad8475SAndrea Arcangeli pages_per_huge_page(h)); 52430ed361deSNick Piggin __SetPageUptodate(new_page); 52441e8f889bSDavid Gibson 52457269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 52466f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5247ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5248ad4404a2SDavidlohr Bueso 5249b76c8cfbSLarry Woodman /* 5250cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5251b76c8cfbSLarry Woodman * before the page tables are altered 5252b76c8cfbSLarry Woodman */ 5253cb900f41SKirill A. Shutemov spin_lock(ptl); 52545b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5255a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5256d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 525707443a85SJoonsoo Kim 52581e8f889bSDavid Gibson /* Break COW */ 52595b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5260ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5261d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 52625b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 52631eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 52641eba86c0SPasha Tatashin make_huge_pte(vma, new_page, 1)); 52658f251a3dSMike Kravetz SetHPageMigratable(new_page); 52661e8f889bSDavid Gibson /* Make the old page be freed below */ 52671e8f889bSDavid Gibson new_page = old_page; 52681e8f889bSDavid Gibson } 5269cb900f41SKirill A. Shutemov spin_unlock(ptl); 5270ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5271ad4404a2SDavidlohr Bueso out_release_all: 5272c7b1850dSMike Kravetz /* No restore in case of successful pagetable update (Break COW) */ 5273c7b1850dSMike Kravetz if (new_page != old_page) 52745b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 527509cbfeafSKirill A. Shutemov put_page(new_page); 5276ad4404a2SDavidlohr Bueso out_release_old: 527709cbfeafSKirill A. Shutemov put_page(old_page); 52788312034fSJoonsoo Kim 5279ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5280ad4404a2SDavidlohr Bueso return ret; 52811e8f889bSDavid Gibson } 52821e8f889bSDavid Gibson 528304f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 5284a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 5285a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 528604f2cbe3SMel Gorman { 528704f2cbe3SMel Gorman struct address_space *mapping; 5288e7c4b0bfSAndy Whitcroft pgoff_t idx; 528904f2cbe3SMel Gorman 529004f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 5291a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 529204f2cbe3SMel Gorman 529304f2cbe3SMel Gorman return find_lock_page(mapping, idx); 529404f2cbe3SMel Gorman } 529504f2cbe3SMel Gorman 52963ae77f43SHugh Dickins /* 52973ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 52983ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 52993ae77f43SHugh Dickins */ 53003ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 53012a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 53022a15efc9SHugh Dickins { 53032a15efc9SHugh Dickins struct address_space *mapping; 53042a15efc9SHugh Dickins pgoff_t idx; 53052a15efc9SHugh Dickins struct page *page; 53062a15efc9SHugh Dickins 53072a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 53082a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 53092a15efc9SHugh Dickins 53102a15efc9SHugh Dickins page = find_get_page(mapping, idx); 53112a15efc9SHugh Dickins if (page) 53122a15efc9SHugh Dickins put_page(page); 53132a15efc9SHugh Dickins return page != NULL; 53142a15efc9SHugh Dickins } 53152a15efc9SHugh Dickins 5316ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 5317ab76ad54SMike Kravetz pgoff_t idx) 5318ab76ad54SMike Kravetz { 5319ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5320ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5321ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 5322ab76ad54SMike Kravetz 5323ab76ad54SMike Kravetz if (err) 5324ab76ad54SMike Kravetz return err; 5325d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5326ab76ad54SMike Kravetz 532722146c3cSMike Kravetz /* 532822146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 532922146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 533022146c3cSMike Kravetz */ 533122146c3cSMike Kravetz set_page_dirty(page); 533222146c3cSMike Kravetz 5333ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5334ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5335ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5336ab76ad54SMike Kravetz return 0; 5337ab76ad54SMike Kravetz } 5338ab76ad54SMike Kravetz 53397677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 53407677f7fdSAxel Rasmussen struct address_space *mapping, 53417677f7fdSAxel Rasmussen pgoff_t idx, 53427677f7fdSAxel Rasmussen unsigned int flags, 53437677f7fdSAxel Rasmussen unsigned long haddr, 53447677f7fdSAxel Rasmussen unsigned long reason) 53457677f7fdSAxel Rasmussen { 53467677f7fdSAxel Rasmussen vm_fault_t ret; 53477677f7fdSAxel Rasmussen u32 hash; 53487677f7fdSAxel Rasmussen struct vm_fault vmf = { 53497677f7fdSAxel Rasmussen .vma = vma, 53507677f7fdSAxel Rasmussen .address = haddr, 53517677f7fdSAxel Rasmussen .flags = flags, 53527677f7fdSAxel Rasmussen 53537677f7fdSAxel Rasmussen /* 53547677f7fdSAxel Rasmussen * Hard to debug if it ends up being 53557677f7fdSAxel Rasmussen * used by a callee that assumes 53567677f7fdSAxel Rasmussen * something about the other 53577677f7fdSAxel Rasmussen * uninitialized fields... same as in 53587677f7fdSAxel Rasmussen * memory.c 53597677f7fdSAxel Rasmussen */ 53607677f7fdSAxel Rasmussen }; 53617677f7fdSAxel Rasmussen 53627677f7fdSAxel Rasmussen /* 53637677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 53647677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 53657677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 53667677f7fdSAxel Rasmussen */ 53677677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 53687677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 53697677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 53707677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 53717677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 53727677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 53737677f7fdSAxel Rasmussen 53747677f7fdSAxel Rasmussen return ret; 53757677f7fdSAxel Rasmussen } 53767677f7fdSAxel Rasmussen 53772b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 53782b740303SSouptick Joarder struct vm_area_struct *vma, 53798382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5380788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 5381ac9b9c66SHugh Dickins { 5382a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 53832b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5384409eb8c2SHillf Danton int anon_rmap = 0; 53854c887265SAdam Litke unsigned long size; 53864c887265SAdam Litke struct page *page; 53871e8f889bSDavid Gibson pte_t new_pte; 5388cb900f41SKirill A. Shutemov spinlock_t *ptl; 5389285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5390c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 53914c887265SAdam Litke 539204f2cbe3SMel Gorman /* 539304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 539404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 539525985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 539604f2cbe3SMel Gorman */ 539704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5398910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 539904f2cbe3SMel Gorman current->pid); 540004f2cbe3SMel Gorman return ret; 540104f2cbe3SMel Gorman } 540204f2cbe3SMel Gorman 54034c887265SAdam Litke /* 540487bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 540587bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 540687bf91d3SMike Kravetz * once for faults beyond end of file. 54074c887265SAdam Litke */ 5408a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 5409ebed4bfcSHugh Dickins if (idx >= size) 5410ebed4bfcSHugh Dickins goto out; 54111a1aad8aSMike Kravetz 541287bf91d3SMike Kravetz retry: 5413c7b1850dSMike Kravetz new_page = false; 541487bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 541587bf91d3SMike Kravetz if (!page) { 54167677f7fdSAxel Rasmussen /* Check for page in userfault range */ 54171a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 54187677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 54197677f7fdSAxel Rasmussen flags, haddr, 54207677f7fdSAxel Rasmussen VM_UFFD_MISSING); 54211a1aad8aSMike Kravetz goto out; 54221a1aad8aSMike Kravetz } 54231a1aad8aSMike Kravetz 5424285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 54252fc39cecSAdam Litke if (IS_ERR(page)) { 54264643d67eSMike Kravetz /* 54274643d67eSMike Kravetz * Returning error will result in faulting task being 54284643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 54294643d67eSMike Kravetz * tasks from racing to fault in the same page which 54304643d67eSMike Kravetz * could result in false unable to allocate errors. 54314643d67eSMike Kravetz * Page migration does not take the fault mutex, but 54324643d67eSMike Kravetz * does a clear then write of pte's under page table 54334643d67eSMike Kravetz * lock. Page fault code could race with migration, 54344643d67eSMike Kravetz * notice the clear pte and try to allocate a page 54354643d67eSMike Kravetz * here. Before returning error, get ptl and make 54364643d67eSMike Kravetz * sure there really is no pte entry. 54374643d67eSMike Kravetz */ 54384643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 54394643d67eSMike Kravetz ret = 0; 5440d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 54412b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5442d83e6c8aSMiaohe Lin spin_unlock(ptl); 54436bda666aSChristoph Lameter goto out; 54446bda666aSChristoph Lameter } 544547ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 54460ed361deSNick Piggin __SetPageUptodate(page); 5447cb6acd01SMike Kravetz new_page = true; 5448ac9b9c66SHugh Dickins 5449f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 5450ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 54516bda666aSChristoph Lameter if (err) { 54526bda666aSChristoph Lameter put_page(page); 54536bda666aSChristoph Lameter if (err == -EEXIST) 54546bda666aSChristoph Lameter goto retry; 54556bda666aSChristoph Lameter goto out; 54566bda666aSChristoph Lameter } 5457c7b1850dSMike Kravetz new_pagecache_page = true; 545823be7468SMel Gorman } else { 54596bda666aSChristoph Lameter lock_page(page); 54600fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 54610fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 54620fe6e20bSNaoya Horiguchi goto backout_unlocked; 546323be7468SMel Gorman } 5464409eb8c2SHillf Danton anon_rmap = 1; 54650fe6e20bSNaoya Horiguchi } 54660fe6e20bSNaoya Horiguchi } else { 546757303d80SAndy Whitcroft /* 5468998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5469998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5470998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5471fd6a03edSNaoya Horiguchi */ 5472fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 54730eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5474972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5475fd6a03edSNaoya Horiguchi goto backout_unlocked; 54766bda666aSChristoph Lameter } 54777677f7fdSAxel Rasmussen 54787677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 54797677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 54807677f7fdSAxel Rasmussen unlock_page(page); 54817677f7fdSAxel Rasmussen put_page(page); 54827677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 54837677f7fdSAxel Rasmussen flags, haddr, 54847677f7fdSAxel Rasmussen VM_UFFD_MINOR); 54857677f7fdSAxel Rasmussen goto out; 54867677f7fdSAxel Rasmussen } 5487998b4382SNaoya Horiguchi } 54881e8f889bSDavid Gibson 548957303d80SAndy Whitcroft /* 549057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 549157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 549257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 549357303d80SAndy Whitcroft * the spinlock. 549457303d80SAndy Whitcroft */ 54955e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5496285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 54972b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 54982b26736cSAndy Whitcroft goto backout_unlocked; 54992b26736cSAndy Whitcroft } 55005e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5501285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 55025e911373SMike Kravetz } 550357303d80SAndy Whitcroft 55048bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 550583c54070SNick Piggin ret = 0; 55067f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 55074c887265SAdam Litke goto backout; 55084c887265SAdam Litke 550907443a85SJoonsoo Kim if (anon_rmap) { 5510d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5511285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5512ac714904SChoi Gi-yong } else 551353f9263bSKirill A. Shutemov page_dup_rmap(page, true); 55141e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 55151e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5516285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 55171e8f889bSDavid Gibson 55185d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5519788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 55201e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5521974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 55221e8f889bSDavid Gibson } 55231e8f889bSDavid Gibson 5524cb900f41SKirill A. Shutemov spin_unlock(ptl); 5525cb6acd01SMike Kravetz 5526cb6acd01SMike Kravetz /* 55278f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 55288f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 55298f251a3dSMike Kravetz * been isolated for migration. 5530cb6acd01SMike Kravetz */ 5531cb6acd01SMike Kravetz if (new_page) 55328f251a3dSMike Kravetz SetHPageMigratable(page); 5533cb6acd01SMike Kravetz 55344c887265SAdam Litke unlock_page(page); 55354c887265SAdam Litke out: 5536ac9b9c66SHugh Dickins return ret; 55374c887265SAdam Litke 55384c887265SAdam Litke backout: 5539cb900f41SKirill A. Shutemov spin_unlock(ptl); 55402b26736cSAndy Whitcroft backout_unlocked: 55414c887265SAdam Litke unlock_page(page); 5542c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5543c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5544285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 55454c887265SAdam Litke put_page(page); 55464c887265SAdam Litke goto out; 5547ac9b9c66SHugh Dickins } 5548ac9b9c66SHugh Dickins 55498382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5550188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 55518382d914SDavidlohr Bueso { 55528382d914SDavidlohr Bueso unsigned long key[2]; 55538382d914SDavidlohr Bueso u32 hash; 55548382d914SDavidlohr Bueso 55558382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 55568382d914SDavidlohr Bueso key[1] = idx; 55578382d914SDavidlohr Bueso 555855254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 55598382d914SDavidlohr Bueso 55608382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 55618382d914SDavidlohr Bueso } 55628382d914SDavidlohr Bueso #else 55638382d914SDavidlohr Bueso /* 55646c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 55658382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 55668382d914SDavidlohr Bueso */ 5567188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 55688382d914SDavidlohr Bueso { 55698382d914SDavidlohr Bueso return 0; 55708382d914SDavidlohr Bueso } 55718382d914SDavidlohr Bueso #endif 55728382d914SDavidlohr Bueso 55732b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5574788c7df4SHugh Dickins unsigned long address, unsigned int flags) 557586e5216fSAdam Litke { 55768382d914SDavidlohr Bueso pte_t *ptep, entry; 5577cb900f41SKirill A. Shutemov spinlock_t *ptl; 55782b740303SSouptick Joarder vm_fault_t ret; 55798382d914SDavidlohr Bueso u32 hash; 55808382d914SDavidlohr Bueso pgoff_t idx; 55810fe6e20bSNaoya Horiguchi struct page *page = NULL; 558257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5583a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55848382d914SDavidlohr Bueso struct address_space *mapping; 55850f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5586285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 558786e5216fSAdam Litke 5588285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5589fd6a03edSNaoya Horiguchi if (ptep) { 5590c0d0381aSMike Kravetz /* 5591c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 5592c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 5593c0d0381aSMike Kravetz * not actually modifying content here. 5594c0d0381aSMike Kravetz */ 5595fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5596290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5597cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 5598290408d4SNaoya Horiguchi return 0; 5599290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5600aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5601972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5602b43a9990SMike Kravetz } 56038382d914SDavidlohr Bueso 5604c0d0381aSMike Kravetz /* 5605c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 560687bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 560787bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 560887bf91d3SMike Kravetz * and making the ptep no longer valid. 560987bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 5610c0d0381aSMike Kravetz * 5611c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5612c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5613c0d0381aSMike Kravetz * something has changed. 5614c0d0381aSMike Kravetz */ 5615ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5616c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5617aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5618c0d0381aSMike Kravetz if (!ptep) { 5619c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5620c0d0381aSMike Kravetz return VM_FAULT_OOM; 5621c0d0381aSMike Kravetz } 5622ddeaab32SMike Kravetz 56233935baa9SDavid Gibson /* 56243935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 56253935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 56263935baa9SDavid Gibson * the same page in the page cache. 56273935baa9SDavid Gibson */ 5628c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5629188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5630c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 56318382d914SDavidlohr Bueso 56327f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 56337f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 56348382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 5635b4d1d99fSDavid Gibson goto out_mutex; 56363935baa9SDavid Gibson } 563786e5216fSAdam Litke 563883c54070SNick Piggin ret = 0; 56391e8f889bSDavid Gibson 564057303d80SAndy Whitcroft /* 56410f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 56420f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 56437c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 56447c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 56457c8de358SEthon Paul * properly handle it. 56460f792cf9SNaoya Horiguchi */ 56470f792cf9SNaoya Horiguchi if (!pte_present(entry)) 56480f792cf9SNaoya Horiguchi goto out_mutex; 56490f792cf9SNaoya Horiguchi 56500f792cf9SNaoya Horiguchi /* 565157303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 565257303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 565357303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 565457303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 565557303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 565657303d80SAndy Whitcroft * consumed. 565757303d80SAndy Whitcroft */ 5658106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5659285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 56602b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5661b4d1d99fSDavid Gibson goto out_mutex; 56622b26736cSAndy Whitcroft } 56635e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5664285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 566557303d80SAndy Whitcroft 5666f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 566757303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 5668285b8dcaSHuang Ying vma, haddr); 566957303d80SAndy Whitcroft } 567057303d80SAndy Whitcroft 56710f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 56720fe6e20bSNaoya Horiguchi 56731e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 5674b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5675cb900f41SKirill A. Shutemov goto out_ptl; 5676b4d1d99fSDavid Gibson 56770f792cf9SNaoya Horiguchi /* 56780f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 56790f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 56800f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 56810f792cf9SNaoya Horiguchi */ 56820f792cf9SNaoya Horiguchi page = pte_page(entry); 56830f792cf9SNaoya Horiguchi if (page != pagecache_page) 56840f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 56850f792cf9SNaoya Horiguchi need_wait_lock = 1; 56860f792cf9SNaoya Horiguchi goto out_ptl; 56870f792cf9SNaoya Horiguchi } 56880f792cf9SNaoya Horiguchi 56890f792cf9SNaoya Horiguchi get_page(page); 5690b4d1d99fSDavid Gibson 5691788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 5692106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5693974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 5694cb900f41SKirill A. Shutemov pagecache_page, ptl); 56950f792cf9SNaoya Horiguchi goto out_put_page; 5696b4d1d99fSDavid Gibson } 5697106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5698b4d1d99fSDavid Gibson } 5699b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5700285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5701788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5702285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 57030f792cf9SNaoya Horiguchi out_put_page: 57040f792cf9SNaoya Horiguchi if (page != pagecache_page) 57050f792cf9SNaoya Horiguchi unlock_page(page); 57060f792cf9SNaoya Horiguchi put_page(page); 5707cb900f41SKirill A. Shutemov out_ptl: 5708cb900f41SKirill A. Shutemov spin_unlock(ptl); 570957303d80SAndy Whitcroft 571057303d80SAndy Whitcroft if (pagecache_page) { 571157303d80SAndy Whitcroft unlock_page(pagecache_page); 571257303d80SAndy Whitcroft put_page(pagecache_page); 571357303d80SAndy Whitcroft } 5714b4d1d99fSDavid Gibson out_mutex: 5715c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5716c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 57170f792cf9SNaoya Horiguchi /* 57180f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 57190f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 57200f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 57210f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 57220f792cf9SNaoya Horiguchi * here without taking refcount. 57230f792cf9SNaoya Horiguchi */ 57240f792cf9SNaoya Horiguchi if (need_wait_lock) 57250f792cf9SNaoya Horiguchi wait_on_page_locked(page); 57261e8f889bSDavid Gibson return ret; 572786e5216fSAdam Litke } 572886e5216fSAdam Litke 5729714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 57308fb5debcSMike Kravetz /* 57318fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 57328fb5debcSMike Kravetz * modifications for huge pages. 57338fb5debcSMike Kravetz */ 57348fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 57358fb5debcSMike Kravetz pte_t *dst_pte, 57368fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 57378fb5debcSMike Kravetz unsigned long dst_addr, 57388fb5debcSMike Kravetz unsigned long src_addr, 5739f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 57408fb5debcSMike Kravetz struct page **pagep) 57418fb5debcSMike Kravetz { 5742f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 57438cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 57448cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 57458cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 57461e392147SAndrea Arcangeli unsigned long size; 57471c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 57488fb5debcSMike Kravetz pte_t _dst_pte; 57498fb5debcSMike Kravetz spinlock_t *ptl; 57508cc5fcbbSMina Almasry int ret = -ENOMEM; 57518fb5debcSMike Kravetz struct page *page; 5752f6191471SAxel Rasmussen int writable; 5753cc30042dSMina Almasry bool page_in_pagecache = false; 57548fb5debcSMike Kravetz 5755f6191471SAxel Rasmussen if (is_continue) { 5756f6191471SAxel Rasmussen ret = -EFAULT; 5757f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5758f6191471SAxel Rasmussen if (!page) 5759f6191471SAxel Rasmussen goto out; 5760cc30042dSMina Almasry page_in_pagecache = true; 5761f6191471SAxel Rasmussen } else if (!*pagep) { 5762d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5763d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5764d84cf06eSMina Almasry */ 5765d84cf06eSMina Almasry if (vm_shared && 5766d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5767d84cf06eSMina Almasry ret = -EEXIST; 57688fb5debcSMike Kravetz goto out; 5769d84cf06eSMina Almasry } 5770d84cf06eSMina Almasry 5771d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5772d84cf06eSMina Almasry if (IS_ERR(page)) { 5773d84cf06eSMina Almasry ret = -ENOMEM; 5774d84cf06eSMina Almasry goto out; 5775d84cf06eSMina Almasry } 57768fb5debcSMike Kravetz 57778fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 57788fb5debcSMike Kravetz (const void __user *) src_addr, 5779810a56b9SMike Kravetz pages_per_huge_page(h), false); 57808fb5debcSMike Kravetz 5781c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 57828fb5debcSMike Kravetz if (unlikely(ret)) { 57839e368259SAndrea Arcangeli ret = -ENOENT; 57848cc5fcbbSMina Almasry /* Free the allocated page which may have 57858cc5fcbbSMina Almasry * consumed a reservation. 57868cc5fcbbSMina Almasry */ 57878cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 57888cc5fcbbSMina Almasry put_page(page); 57898cc5fcbbSMina Almasry 57908cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 57918cc5fcbbSMina Almasry * contents. 57928cc5fcbbSMina Almasry */ 57938cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 57948cc5fcbbSMina Almasry if (!page) { 57958cc5fcbbSMina Almasry ret = -ENOMEM; 57968cc5fcbbSMina Almasry goto out; 57978cc5fcbbSMina Almasry } 57988fb5debcSMike Kravetz *pagep = page; 57998cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 58008cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 58018cc5fcbbSMina Almasry * page. 58028cc5fcbbSMina Almasry */ 58038fb5debcSMike Kravetz goto out; 58048fb5debcSMike Kravetz } 58058fb5debcSMike Kravetz } else { 58068cc5fcbbSMina Almasry if (vm_shared && 58078cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 58088cc5fcbbSMina Almasry put_page(*pagep); 58098cc5fcbbSMina Almasry ret = -EEXIST; 58108cc5fcbbSMina Almasry *pagep = NULL; 58118cc5fcbbSMina Almasry goto out; 58128cc5fcbbSMina Almasry } 58138cc5fcbbSMina Almasry 58148cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 58158cc5fcbbSMina Almasry if (IS_ERR(page)) { 58168cc5fcbbSMina Almasry ret = -ENOMEM; 58178cc5fcbbSMina Almasry *pagep = NULL; 58188cc5fcbbSMina Almasry goto out; 58198cc5fcbbSMina Almasry } 5820715cbfd6SMatthew Wilcox (Oracle) folio_copy(page_folio(page), page_folio(*pagep)); 58218cc5fcbbSMina Almasry put_page(*pagep); 58228fb5debcSMike Kravetz *pagep = NULL; 58238fb5debcSMike Kravetz } 58248fb5debcSMike Kravetz 58258fb5debcSMike Kravetz /* 58268fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 58278fb5debcSMike Kravetz * preceding stores to the page contents become visible before 58288fb5debcSMike Kravetz * the set_pte_at() write. 58298fb5debcSMike Kravetz */ 58308fb5debcSMike Kravetz __SetPageUptodate(page); 58318fb5debcSMike Kravetz 5832f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 5833f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 58341e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 58351e392147SAndrea Arcangeli ret = -EFAULT; 58361e392147SAndrea Arcangeli if (idx >= size) 58371e392147SAndrea Arcangeli goto out_release_nounlock; 58381c9e8defSMike Kravetz 58391e392147SAndrea Arcangeli /* 58401e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 58411e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 58421e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 58431e392147SAndrea Arcangeli * the caller. 58441e392147SAndrea Arcangeli */ 58451c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 58461c9e8defSMike Kravetz if (ret) 58471c9e8defSMike Kravetz goto out_release_nounlock; 5848cc30042dSMina Almasry page_in_pagecache = true; 58491c9e8defSMike Kravetz } 58501c9e8defSMike Kravetz 58518fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 58528fb5debcSMike Kravetz spin_lock(ptl); 58538fb5debcSMike Kravetz 58541e392147SAndrea Arcangeli /* 58551e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 58561e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 58571e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 58581e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 58591e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 58601e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 58611e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 58621e392147SAndrea Arcangeli */ 58631e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 58641e392147SAndrea Arcangeli ret = -EFAULT; 58651e392147SAndrea Arcangeli if (idx >= size) 58661e392147SAndrea Arcangeli goto out_release_unlock; 58671e392147SAndrea Arcangeli 58688fb5debcSMike Kravetz ret = -EEXIST; 58698fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 58708fb5debcSMike Kravetz goto out_release_unlock; 58718fb5debcSMike Kravetz 58721c9e8defSMike Kravetz if (vm_shared) { 58731c9e8defSMike Kravetz page_dup_rmap(page, true); 58741c9e8defSMike Kravetz } else { 5875d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 58768fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 58771c9e8defSMike Kravetz } 58788fb5debcSMike Kravetz 5879f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 5880f6191471SAxel Rasmussen if (is_continue && !vm_shared) 5881f6191471SAxel Rasmussen writable = 0; 5882f6191471SAxel Rasmussen else 5883f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 5884f6191471SAxel Rasmussen 5885f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 5886f6191471SAxel Rasmussen if (writable) 58878fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 58888fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 58898fb5debcSMike Kravetz 58908fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 58918fb5debcSMike Kravetz 58928fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 58938fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 58948fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 58958fb5debcSMike Kravetz 58968fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 58978fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 58988fb5debcSMike Kravetz 58998fb5debcSMike Kravetz spin_unlock(ptl); 5900f6191471SAxel Rasmussen if (!is_continue) 59018f251a3dSMike Kravetz SetHPageMigratable(page); 5902f6191471SAxel Rasmussen if (vm_shared || is_continue) 59031c9e8defSMike Kravetz unlock_page(page); 59048fb5debcSMike Kravetz ret = 0; 59058fb5debcSMike Kravetz out: 59068fb5debcSMike Kravetz return ret; 59078fb5debcSMike Kravetz out_release_unlock: 59088fb5debcSMike Kravetz spin_unlock(ptl); 5909f6191471SAxel Rasmussen if (vm_shared || is_continue) 59101c9e8defSMike Kravetz unlock_page(page); 59115af10dfdSAndrea Arcangeli out_release_nounlock: 5912cc30042dSMina Almasry if (!page_in_pagecache) 5913846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 59148fb5debcSMike Kravetz put_page(page); 59158fb5debcSMike Kravetz goto out; 59168fb5debcSMike Kravetz } 5917714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 59188fb5debcSMike Kravetz 591982e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 592082e5d378SJoao Martins int refs, struct page **pages, 592182e5d378SJoao Martins struct vm_area_struct **vmas) 592282e5d378SJoao Martins { 592382e5d378SJoao Martins int nr; 592482e5d378SJoao Martins 592582e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 592682e5d378SJoao Martins if (likely(pages)) 592782e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 592882e5d378SJoao Martins if (vmas) 592982e5d378SJoao Martins vmas[nr] = vma; 593082e5d378SJoao Martins } 593182e5d378SJoao Martins } 593282e5d378SJoao Martins 593328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 593463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 593528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 59364f6da934SPeter Xu long i, unsigned int flags, int *locked) 593763551ae0SDavid Gibson { 5938d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5939d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 594028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5941a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 59420fa5bc40SJoao Martins int err = -EFAULT, refs; 594363551ae0SDavid Gibson 594463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 594563551ae0SDavid Gibson pte_t *pte; 5946cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 59472a15efc9SHugh Dickins int absent; 594863551ae0SDavid Gibson struct page *page; 594963551ae0SDavid Gibson 59504c887265SAdam Litke /* 595102057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 595202057967SDavid Rientjes * potentially allocating memory. 595302057967SDavid Rientjes */ 5954fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 595502057967SDavid Rientjes remainder = 0; 595602057967SDavid Rientjes break; 595702057967SDavid Rientjes } 595802057967SDavid Rientjes 595902057967SDavid Rientjes /* 59604c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 59612a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 59624c887265SAdam Litke * first, for the page indexing below to work. 5963cb900f41SKirill A. Shutemov * 5964cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 59654c887265SAdam Litke */ 59667868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 59677868a208SPunit Agrawal huge_page_size(h)); 5968cb900f41SKirill A. Shutemov if (pte) 5969cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 59702a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 597163551ae0SDavid Gibson 59722a15efc9SHugh Dickins /* 59732a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 59743ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 59753ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 59763ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 59773ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 59782a15efc9SHugh Dickins */ 59793ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 59803ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5981cb900f41SKirill A. Shutemov if (pte) 5982cb900f41SKirill A. Shutemov spin_unlock(ptl); 59832a15efc9SHugh Dickins remainder = 0; 59842a15efc9SHugh Dickins break; 59852a15efc9SHugh Dickins } 59862a15efc9SHugh Dickins 59879cc3a5bdSNaoya Horiguchi /* 59889cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 59899cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 59909cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 59919cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 59929cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 59939cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 59949cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 59959cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 59969cc3a5bdSNaoya Horiguchi */ 59979cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5998106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5999106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 60002b740303SSouptick Joarder vm_fault_t ret; 600187ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 60024c887265SAdam Litke 6003cb900f41SKirill A. Shutemov if (pte) 6004cb900f41SKirill A. Shutemov spin_unlock(ptl); 600587ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 600687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 60074f6da934SPeter Xu if (locked) 600871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 600971335f37SPeter Xu FAULT_FLAG_KILLABLE; 601087ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 601187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 601287ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 601387ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 60144426e945SPeter Xu /* 60154426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 60164426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 60174426e945SPeter Xu */ 601887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 601987ffc118SAndrea Arcangeli } 602087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 602187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 60222be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 60231c59827dSHugh Dickins remainder = 0; 60241c59827dSHugh Dickins break; 60251c59827dSHugh Dickins } 602687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 60274f6da934SPeter Xu if (locked && 60281ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 60294f6da934SPeter Xu *locked = 0; 603087ffc118SAndrea Arcangeli *nr_pages = 0; 603187ffc118SAndrea Arcangeli /* 603287ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 603387ffc118SAndrea Arcangeli * error, it will return zero 603487ffc118SAndrea Arcangeli * instead. 603587ffc118SAndrea Arcangeli * 603687ffc118SAndrea Arcangeli * No need to update "position" as the 603787ffc118SAndrea Arcangeli * caller will not check it after 603887ffc118SAndrea Arcangeli * *nr_pages is set to 0. 603987ffc118SAndrea Arcangeli */ 604087ffc118SAndrea Arcangeli return i; 604187ffc118SAndrea Arcangeli } 604287ffc118SAndrea Arcangeli continue; 604387ffc118SAndrea Arcangeli } 604463551ae0SDavid Gibson 6045a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 60467f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 60478fde12caSLinus Torvalds 60488fde12caSLinus Torvalds /* 6049acbfb087SZhigang Lu * If subpage information not requested, update counters 6050acbfb087SZhigang Lu * and skip the same_page loop below. 6051acbfb087SZhigang Lu */ 6052acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6053acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6054acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6055acbfb087SZhigang Lu vaddr += huge_page_size(h); 6056acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6057acbfb087SZhigang Lu i += pages_per_huge_page(h); 6058acbfb087SZhigang Lu spin_unlock(ptl); 6059acbfb087SZhigang Lu continue; 6060acbfb087SZhigang Lu } 6061acbfb087SZhigang Lu 6062d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6063d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6064d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 60650fa5bc40SJoao Martins 606682e5d378SJoao Martins if (pages || vmas) 606782e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 606882e5d378SJoao Martins vma, refs, 606982e5d378SJoao Martins likely(pages) ? pages + i : NULL, 607082e5d378SJoao Martins vmas ? vmas + i : NULL); 607163551ae0SDavid Gibson 607282e5d378SJoao Martins if (pages) { 60730fa5bc40SJoao Martins /* 60740fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 60750fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 60760fa5bc40SJoao Martins * checked that the huge page is present in the page 60770fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 60780fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 60790fa5bc40SJoao Martins * head page and tail pages from being rearranged in 60800fa5bc40SJoao Martins * any way. So this page must be available at this 60810fa5bc40SJoao Martins * point, unless the page refcount overflowed: 60820fa5bc40SJoao Martins */ 608382e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 60840fa5bc40SJoao Martins refs, 60850fa5bc40SJoao Martins flags))) { 60860fa5bc40SJoao Martins spin_unlock(ptl); 60870fa5bc40SJoao Martins remainder = 0; 60880fa5bc40SJoao Martins err = -ENOMEM; 60890fa5bc40SJoao Martins break; 60900fa5bc40SJoao Martins } 6091d5d4b0aaSChen, Kenneth W } 609282e5d378SJoao Martins 609382e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 609482e5d378SJoao Martins remainder -= refs; 609582e5d378SJoao Martins i += refs; 609682e5d378SJoao Martins 6097cb900f41SKirill A. Shutemov spin_unlock(ptl); 609863551ae0SDavid Gibson } 609928a35716SMichel Lespinasse *nr_pages = remainder; 610087ffc118SAndrea Arcangeli /* 610187ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 610287ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 610387ffc118SAndrea Arcangeli * branch. 610487ffc118SAndrea Arcangeli */ 610563551ae0SDavid Gibson *position = vaddr; 610663551ae0SDavid Gibson 61072be7cfedSDaniel Jordan return i ? i : err; 610863551ae0SDavid Gibson } 61098f860591SZhang, Yanmin 61107da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 61118f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 61128f860591SZhang, Yanmin { 61138f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 61148f860591SZhang, Yanmin unsigned long start = address; 61158f860591SZhang, Yanmin pte_t *ptep; 61168f860591SZhang, Yanmin pte_t pte; 6117a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 61187da4d641SPeter Zijlstra unsigned long pages = 0; 6119dff11abeSMike Kravetz bool shared_pmd = false; 6120ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6121dff11abeSMike Kravetz 6122dff11abeSMike Kravetz /* 6123dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6124ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6125dff11abeSMike Kravetz * range if PMD sharing is possible. 6126dff11abeSMike Kravetz */ 61277269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 61287269f999SJérôme Glisse 0, vma, mm, start, end); 6129ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 61308f860591SZhang, Yanmin 61318f860591SZhang, Yanmin BUG_ON(address >= end); 6132ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 61338f860591SZhang, Yanmin 6134ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 613583cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 6136a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 6137cb900f41SKirill A. Shutemov spinlock_t *ptl; 61387868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 61398f860591SZhang, Yanmin if (!ptep) 61408f860591SZhang, Yanmin continue; 6141cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 614234ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 61437da4d641SPeter Zijlstra pages++; 6144cb900f41SKirill A. Shutemov spin_unlock(ptl); 6145dff11abeSMike Kravetz shared_pmd = true; 614639dde65cSChen, Kenneth W continue; 61477da4d641SPeter Zijlstra } 6148a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6149a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6150a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6151a8bda28dSNaoya Horiguchi continue; 6152a8bda28dSNaoya Horiguchi } 6153a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6154a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 6155a8bda28dSNaoya Horiguchi 61564dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) { 6157a8bda28dSNaoya Horiguchi pte_t newpte; 6158a8bda28dSNaoya Horiguchi 61594dd845b5SAlistair Popple entry = make_readable_migration_entry( 61604dd845b5SAlistair Popple swp_offset(entry)); 6161a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 6162e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 6163e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 6164a8bda28dSNaoya Horiguchi pages++; 6165a8bda28dSNaoya Horiguchi } 6166a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6167a8bda28dSNaoya Horiguchi continue; 6168a8bda28dSNaoya Horiguchi } 6169a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6170023bdd00SAneesh Kumar K.V pte_t old_pte; 617179c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6172023bdd00SAneesh Kumar K.V 6173023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 6174023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 617579c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 6176023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 61777da4d641SPeter Zijlstra pages++; 61788f860591SZhang, Yanmin } 6179cb900f41SKirill A. Shutemov spin_unlock(ptl); 61808f860591SZhang, Yanmin } 6181d833352aSMel Gorman /* 6182c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6183d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6184c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6185dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6186dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6187d833352aSMel Gorman */ 6188dff11abeSMike Kravetz if (shared_pmd) 6189ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6190dff11abeSMike Kravetz else 61915491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 61920f10851eSJérôme Glisse /* 61930f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 61940f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 61950f10851eSJérôme Glisse * 6196ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 61970f10851eSJérôme Glisse */ 619883cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 6199ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 62007da4d641SPeter Zijlstra 62017da4d641SPeter Zijlstra return pages << h->order; 62028f860591SZhang, Yanmin } 62038f860591SZhang, Yanmin 620433b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 620533b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6206a1e78772SMel Gorman long from, long to, 62075a6fe125SMel Gorman struct vm_area_struct *vma, 6208ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6209e4e574b7SAdam Litke { 621033b8f84aSMike Kravetz long chg, add = -1; 6211a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 621290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 62139119a41eSJoonsoo Kim struct resv_map *resv_map; 6214075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 62150db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6216e4e574b7SAdam Litke 621763489f8eSMike Kravetz /* This should never happen */ 621863489f8eSMike Kravetz if (from > to) { 621963489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 622033b8f84aSMike Kravetz return false; 622163489f8eSMike Kravetz } 622263489f8eSMike Kravetz 6223a1e78772SMel Gorman /* 622417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 622517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 622690481622SDavid Gibson * without using reserves 622717c9d12eSMel Gorman */ 6228ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 622933b8f84aSMike Kravetz return true; 623017c9d12eSMel Gorman 623117c9d12eSMel Gorman /* 6232a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6233a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6234a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6235a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6236a1e78772SMel Gorman */ 62379119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6238f27a5136SMike Kravetz /* 6239f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6240f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6241f27a5136SMike Kravetz * hugetlbfs_get_inode). 6242f27a5136SMike Kravetz */ 62434e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 62449119a41eSJoonsoo Kim 62450db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 62469119a41eSJoonsoo Kim 62479119a41eSJoonsoo Kim } else { 6248e9fe92aeSMina Almasry /* Private mapping. */ 62499119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 62505a6fe125SMel Gorman if (!resv_map) 625133b8f84aSMike Kravetz return false; 62525a6fe125SMel Gorman 625317c9d12eSMel Gorman chg = to - from; 625417c9d12eSMel Gorman 62555a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 62565a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 62575a6fe125SMel Gorman } 62585a6fe125SMel Gorman 625933b8f84aSMike Kravetz if (chg < 0) 6260c50ac050SDave Hansen goto out_err; 626117c9d12eSMel Gorman 626233b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 626333b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6264075a61d0SMina Almasry goto out_err; 6265075a61d0SMina Almasry 6266075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6267075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6268075a61d0SMina Almasry * of the resv_map. 6269075a61d0SMina Almasry */ 6270075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6271075a61d0SMina Almasry } 6272075a61d0SMina Almasry 62731c5ecae3SMike Kravetz /* 62741c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 62751c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 62761c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 62771c5ecae3SMike Kravetz */ 62781c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 627933b8f84aSMike Kravetz if (gbl_reserve < 0) 6280075a61d0SMina Almasry goto out_uncharge_cgroup; 628117c9d12eSMel Gorman 628217c9d12eSMel Gorman /* 628317c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 628490481622SDavid Gibson * Hand the pages back to the subpool if there are not 628517c9d12eSMel Gorman */ 628633b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6287075a61d0SMina Almasry goto out_put_pages; 628817c9d12eSMel Gorman 628917c9d12eSMel Gorman /* 629017c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 629117c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 629217c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 629317c9d12eSMel Gorman * the reservation was and the page cache tells how much of 629417c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 629517c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 629617c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 629717c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 629817c9d12eSMel Gorman * else has to be done for private mappings here 629917c9d12eSMel Gorman */ 630033039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6301075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 630233039678SMike Kravetz 63030db9d74eSMina Almasry if (unlikely(add < 0)) { 63040db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6305075a61d0SMina Almasry goto out_put_pages; 63060db9d74eSMina Almasry } else if (unlikely(chg > add)) { 630733039678SMike Kravetz /* 630833039678SMike Kravetz * pages in this range were added to the reserve 630933039678SMike Kravetz * map between region_chg and region_add. This 631033039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 631133039678SMike Kravetz * the subpool and reserve counts modified above 631233039678SMike Kravetz * based on the difference. 631333039678SMike Kravetz */ 631433039678SMike Kravetz long rsv_adjust; 631533039678SMike Kravetz 6316d85aecf2SMiaohe Lin /* 6317d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6318d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6319d85aecf2SMiaohe Lin */ 6320075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6321075a61d0SMina Almasry hstate_index(h), 6322075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6323075a61d0SMina Almasry 632433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 632533039678SMike Kravetz chg - add); 632633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6327d85aecf2SMiaohe Lin } else if (h_cg) { 6328d85aecf2SMiaohe Lin /* 6329d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6330d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6331d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6332d85aecf2SMiaohe Lin * done. 6333d85aecf2SMiaohe Lin */ 6334d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 633533039678SMike Kravetz } 633633039678SMike Kravetz } 633733b8f84aSMike Kravetz return true; 633833b8f84aSMike Kravetz 6339075a61d0SMina Almasry out_put_pages: 6340075a61d0SMina Almasry /* put back original number of pages, chg */ 6341075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6342075a61d0SMina Almasry out_uncharge_cgroup: 6343075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6344075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6345c50ac050SDave Hansen out_err: 63465e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 63470db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 63480db9d74eSMina Almasry * region_add failed or didn't run. 63490db9d74eSMina Almasry */ 63500db9d74eSMina Almasry if (chg >= 0 && add < 0) 63510db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6352f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6353f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 635433b8f84aSMike Kravetz return false; 6355a43a8c39SChen, Kenneth W } 6356a43a8c39SChen, Kenneth W 6357b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6358b5cec28dSMike Kravetz long freed) 6359a43a8c39SChen, Kenneth W { 6360a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 63614e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 63629119a41eSJoonsoo Kim long chg = 0; 636390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 63641c5ecae3SMike Kravetz long gbl_reserve; 636545c682a6SKen Chen 6366f27a5136SMike Kravetz /* 6367f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6368f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6369f27a5136SMike Kravetz */ 6370b5cec28dSMike Kravetz if (resv_map) { 6371b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6372b5cec28dSMike Kravetz /* 6373b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6374b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6375b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6376b5cec28dSMike Kravetz */ 6377b5cec28dSMike Kravetz if (chg < 0) 6378b5cec28dSMike Kravetz return chg; 6379b5cec28dSMike Kravetz } 6380b5cec28dSMike Kravetz 638145c682a6SKen Chen spin_lock(&inode->i_lock); 6382e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 638345c682a6SKen Chen spin_unlock(&inode->i_lock); 638445c682a6SKen Chen 63851c5ecae3SMike Kravetz /* 63861c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 63871c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6388dddf31a4SMiaohe Lin * 6389dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6390dddf31a4SMiaohe Lin * won't go negative. 63911c5ecae3SMike Kravetz */ 63921c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 63931c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6394b5cec28dSMike Kravetz 6395b5cec28dSMike Kravetz return 0; 6396a43a8c39SChen, Kenneth W } 639793f70f90SNaoya Horiguchi 63983212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 63993212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 64003212b535SSteve Capper struct vm_area_struct *vma, 64013212b535SSteve Capper unsigned long addr, pgoff_t idx) 64023212b535SSteve Capper { 64033212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 64043212b535SSteve Capper svma->vm_start; 64053212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 64063212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 64073212b535SSteve Capper 64083212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6409de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6410de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 64113212b535SSteve Capper 64123212b535SSteve Capper /* 64133212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 64143212b535SSteve Capper * page table page. 64153212b535SSteve Capper */ 64163212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 64173212b535SSteve Capper vm_flags != svm_flags || 641807e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 64193212b535SSteve Capper return 0; 64203212b535SSteve Capper 64213212b535SSteve Capper return saddr; 64223212b535SSteve Capper } 64233212b535SSteve Capper 642431aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 64253212b535SSteve Capper { 64263212b535SSteve Capper unsigned long base = addr & PUD_MASK; 64273212b535SSteve Capper unsigned long end = base + PUD_SIZE; 64283212b535SSteve Capper 64293212b535SSteve Capper /* 64303212b535SSteve Capper * check on proper vm_flags and page table alignment 64313212b535SSteve Capper */ 6432017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 643331aafb45SNicholas Krause return true; 643431aafb45SNicholas Krause return false; 64353212b535SSteve Capper } 64363212b535SSteve Capper 6437c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6438c1991e07SPeter Xu { 6439c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 6440c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 6441c1991e07SPeter Xu return false; 6442c1991e07SPeter Xu #endif 6443c1991e07SPeter Xu return vma_shareable(vma, addr); 6444c1991e07SPeter Xu } 6445c1991e07SPeter Xu 64463212b535SSteve Capper /* 6447017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6448017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6449017b1660SMike Kravetz * shared pmd mappings. 6450017b1660SMike Kravetz */ 6451017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6452017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6453017b1660SMike Kravetz { 6454a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6455a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6456017b1660SMike Kravetz 6457a1ba9da8SLi Xinhai /* 6458f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6459f0953a1bSIngo Molnar * must be at least partially within in. 6460a1ba9da8SLi Xinhai */ 6461a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6462a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6463017b1660SMike Kravetz return; 6464017b1660SMike Kravetz 646575802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6466a1ba9da8SLi Xinhai if (*start > v_start) 6467a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6468017b1660SMike Kravetz 6469a1ba9da8SLi Xinhai if (*end < v_end) 6470a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6471017b1660SMike Kravetz } 6472017b1660SMike Kravetz 6473017b1660SMike Kravetz /* 64743212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 64753212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 64763212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 6477c0d0381aSMike Kravetz * code much cleaner. 6478c0d0381aSMike Kravetz * 64790bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 64800bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 64810bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 64820bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 64833212b535SSteve Capper */ 6484aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6485aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 64863212b535SSteve Capper { 64873212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 64883212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 64893212b535SSteve Capper vma->vm_pgoff; 64903212b535SSteve Capper struct vm_area_struct *svma; 64913212b535SSteve Capper unsigned long saddr; 64923212b535SSteve Capper pte_t *spte = NULL; 64933212b535SSteve Capper pte_t *pte; 6494cb900f41SKirill A. Shutemov spinlock_t *ptl; 64953212b535SSteve Capper 64960bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 64973212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 64983212b535SSteve Capper if (svma == vma) 64993212b535SSteve Capper continue; 65003212b535SSteve Capper 65013212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 65023212b535SSteve Capper if (saddr) { 65037868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 65047868a208SPunit Agrawal vma_mmu_pagesize(svma)); 65053212b535SSteve Capper if (spte) { 65063212b535SSteve Capper get_page(virt_to_page(spte)); 65073212b535SSteve Capper break; 65083212b535SSteve Capper } 65093212b535SSteve Capper } 65103212b535SSteve Capper } 65113212b535SSteve Capper 65123212b535SSteve Capper if (!spte) 65133212b535SSteve Capper goto out; 65143212b535SSteve Capper 65158bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6516dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 65173212b535SSteve Capper pud_populate(mm, pud, 65183212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6519c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6520dc6c9a35SKirill A. Shutemov } else { 65213212b535SSteve Capper put_page(virt_to_page(spte)); 6522dc6c9a35SKirill A. Shutemov } 6523cb900f41SKirill A. Shutemov spin_unlock(ptl); 65243212b535SSteve Capper out: 65253212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 65263212b535SSteve Capper return pte; 65273212b535SSteve Capper } 65283212b535SSteve Capper 65293212b535SSteve Capper /* 65303212b535SSteve Capper * unmap huge page backed by shared pte. 65313212b535SSteve Capper * 65323212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 65333212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 65343212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 65353212b535SSteve Capper * 6536c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 65373212b535SSteve Capper * 65383212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 65393212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 65403212b535SSteve Capper */ 654134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 654234ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 65433212b535SSteve Capper { 65443212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 6545c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 6546c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 65473212b535SSteve Capper 654834ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 65493212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 65503212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 65513212b535SSteve Capper return 0; 65523212b535SSteve Capper 65533212b535SSteve Capper pud_clear(pud); 65543212b535SSteve Capper put_page(virt_to_page(ptep)); 6555dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 65563212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 65573212b535SSteve Capper return 1; 65583212b535SSteve Capper } 6559c1991e07SPeter Xu 65609e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 6561aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6562aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 65639e5fc74cSSteve Capper { 65649e5fc74cSSteve Capper return NULL; 65659e5fc74cSSteve Capper } 6566e81f2d22SZhang Zhen 656734ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 656834ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 6569e81f2d22SZhang Zhen { 6570e81f2d22SZhang Zhen return 0; 6571e81f2d22SZhang Zhen } 6572017b1660SMike Kravetz 6573017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6574017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6575017b1660SMike Kravetz { 6576017b1660SMike Kravetz } 6577c1991e07SPeter Xu 6578c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6579c1991e07SPeter Xu { 6580c1991e07SPeter Xu return false; 6581c1991e07SPeter Xu } 65823212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 65833212b535SSteve Capper 65849e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 6585aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 65869e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 65879e5fc74cSSteve Capper { 65889e5fc74cSSteve Capper pgd_t *pgd; 6589c2febafcSKirill A. Shutemov p4d_t *p4d; 65909e5fc74cSSteve Capper pud_t *pud; 65919e5fc74cSSteve Capper pte_t *pte = NULL; 65929e5fc74cSSteve Capper 65939e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6594f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 6595f4f0a3d8SKirill A. Shutemov if (!p4d) 6596f4f0a3d8SKirill A. Shutemov return NULL; 6597c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 65989e5fc74cSSteve Capper if (pud) { 65999e5fc74cSSteve Capper if (sz == PUD_SIZE) { 66009e5fc74cSSteve Capper pte = (pte_t *)pud; 66019e5fc74cSSteve Capper } else { 66029e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 6603c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 6604aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 66059e5fc74cSSteve Capper else 66069e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 66079e5fc74cSSteve Capper } 66089e5fc74cSSteve Capper } 66094e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 66109e5fc74cSSteve Capper 66119e5fc74cSSteve Capper return pte; 66129e5fc74cSSteve Capper } 66139e5fc74cSSteve Capper 66149b19df29SPunit Agrawal /* 66159b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 66169b19df29SPunit Agrawal * entry at address @addr 66179b19df29SPunit Agrawal * 66188ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 66198ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 66209b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 66219b19df29SPunit Agrawal * table. 66229b19df29SPunit Agrawal */ 66237868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 66247868a208SPunit Agrawal unsigned long addr, unsigned long sz) 66259e5fc74cSSteve Capper { 66269e5fc74cSSteve Capper pgd_t *pgd; 6627c2febafcSKirill A. Shutemov p4d_t *p4d; 66288ac0b81aSLi Xinhai pud_t *pud; 66298ac0b81aSLi Xinhai pmd_t *pmd; 66309e5fc74cSSteve Capper 66319e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6632c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6633c2febafcSKirill A. Shutemov return NULL; 6634c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6635c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6636c2febafcSKirill A. Shutemov return NULL; 66379b19df29SPunit Agrawal 6638c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 66398ac0b81aSLi Xinhai if (sz == PUD_SIZE) 66408ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 66419e5fc74cSSteve Capper return (pte_t *)pud; 66428ac0b81aSLi Xinhai if (!pud_present(*pud)) 66438ac0b81aSLi Xinhai return NULL; 66448ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 66459b19df29SPunit Agrawal 66469e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 66478ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 66489e5fc74cSSteve Capper return (pte_t *)pmd; 66499e5fc74cSSteve Capper } 66509e5fc74cSSteve Capper 665161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 665261f77edaSNaoya Horiguchi 665361f77edaSNaoya Horiguchi /* 665461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 665561f77edaSNaoya Horiguchi * behavior. 665661f77edaSNaoya Horiguchi */ 665761f77edaSNaoya Horiguchi struct page * __weak 665861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 665961f77edaSNaoya Horiguchi int write) 666061f77edaSNaoya Horiguchi { 666161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 666261f77edaSNaoya Horiguchi } 666361f77edaSNaoya Horiguchi 666461f77edaSNaoya Horiguchi struct page * __weak 66654dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 66664dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 66674dc71451SAneesh Kumar K.V { 66684dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 66694dc71451SAneesh Kumar K.V return NULL; 66704dc71451SAneesh Kumar K.V } 66714dc71451SAneesh Kumar K.V 66724dc71451SAneesh Kumar K.V struct page * __weak 66739e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6674e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 66759e5fc74cSSteve Capper { 6676e66f17ffSNaoya Horiguchi struct page *page = NULL; 6677e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6678c9d398faSNaoya Horiguchi pte_t pte; 66793faa52c0SJohn Hubbard 66803faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 66813faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 66823faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 66833faa52c0SJohn Hubbard return NULL; 66843faa52c0SJohn Hubbard 6685e66f17ffSNaoya Horiguchi retry: 6686e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 6687e66f17ffSNaoya Horiguchi spin_lock(ptl); 6688e66f17ffSNaoya Horiguchi /* 6689e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 6690e66f17ffSNaoya Horiguchi * unmapped from other threads. 6691e66f17ffSNaoya Horiguchi */ 6692e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 6693e66f17ffSNaoya Horiguchi goto out; 6694c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 6695c9d398faSNaoya Horiguchi if (pte_present(pte)) { 669697534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 66973faa52c0SJohn Hubbard /* 66983faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 66993faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 67003faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 67013faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 67023faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 67033faa52c0SJohn Hubbard * unless the page refcount overflowed: 67043faa52c0SJohn Hubbard */ 67053faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 67063faa52c0SJohn Hubbard page = NULL; 67073faa52c0SJohn Hubbard goto out; 67083faa52c0SJohn Hubbard } 6709e66f17ffSNaoya Horiguchi } else { 6710c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 6711e66f17ffSNaoya Horiguchi spin_unlock(ptl); 6712e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 6713e66f17ffSNaoya Horiguchi goto retry; 6714e66f17ffSNaoya Horiguchi } 6715e66f17ffSNaoya Horiguchi /* 6716e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 6717e66f17ffSNaoya Horiguchi * follow_page_mask(). 6718e66f17ffSNaoya Horiguchi */ 6719e66f17ffSNaoya Horiguchi } 6720e66f17ffSNaoya Horiguchi out: 6721e66f17ffSNaoya Horiguchi spin_unlock(ptl); 67229e5fc74cSSteve Capper return page; 67239e5fc74cSSteve Capper } 67249e5fc74cSSteve Capper 672561f77edaSNaoya Horiguchi struct page * __weak 67269e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 6727e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 67289e5fc74cSSteve Capper { 67293faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6730e66f17ffSNaoya Horiguchi return NULL; 67319e5fc74cSSteve Capper 6732e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 67339e5fc74cSSteve Capper } 67349e5fc74cSSteve Capper 6735faaa5b62SAnshuman Khandual struct page * __weak 6736faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 6737faaa5b62SAnshuman Khandual { 67383faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6739faaa5b62SAnshuman Khandual return NULL; 6740faaa5b62SAnshuman Khandual 6741faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 6742faaa5b62SAnshuman Khandual } 6743faaa5b62SAnshuman Khandual 674431caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 674531caf665SNaoya Horiguchi { 6746bcc54222SNaoya Horiguchi bool ret = true; 6747bcc54222SNaoya Horiguchi 6748db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 67498f251a3dSMike Kravetz if (!PageHeadHuge(page) || 67508f251a3dSMike Kravetz !HPageMigratable(page) || 67510eb2df2bSMuchun Song !get_page_unless_zero(page)) { 6752bcc54222SNaoya Horiguchi ret = false; 6753bcc54222SNaoya Horiguchi goto unlock; 6754bcc54222SNaoya Horiguchi } 67558f251a3dSMike Kravetz ClearHPageMigratable(page); 675631caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 6757bcc54222SNaoya Horiguchi unlock: 6758db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6759bcc54222SNaoya Horiguchi return ret; 676031caf665SNaoya Horiguchi } 676131caf665SNaoya Horiguchi 676225182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 676325182f05SNaoya Horiguchi { 676425182f05SNaoya Horiguchi int ret = 0; 676525182f05SNaoya Horiguchi 676625182f05SNaoya Horiguchi *hugetlb = false; 676725182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 676825182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 676925182f05SNaoya Horiguchi *hugetlb = true; 677025182f05SNaoya Horiguchi if (HPageFreed(page) || HPageMigratable(page)) 677125182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 67720ed950d1SNaoya Horiguchi else 67730ed950d1SNaoya Horiguchi ret = -EBUSY; 677425182f05SNaoya Horiguchi } 677525182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 677625182f05SNaoya Horiguchi return ret; 677725182f05SNaoya Horiguchi } 677825182f05SNaoya Horiguchi 677931caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 678031caf665SNaoya Horiguchi { 6781db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 67828f251a3dSMike Kravetz SetHPageMigratable(page); 678331caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 6784db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 678531caf665SNaoya Horiguchi put_page(page); 678631caf665SNaoya Horiguchi } 6787ab5ac90aSMichal Hocko 6788ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 6789ab5ac90aSMichal Hocko { 6790ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 6791ab5ac90aSMichal Hocko 6792ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 6793ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 6794ab5ac90aSMichal Hocko 6795ab5ac90aSMichal Hocko /* 6796ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 6797ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 6798ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 6799ab5ac90aSMichal Hocko * the temporary status. 6800ab5ac90aSMichal Hocko * 6801ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 6802ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 6803ab5ac90aSMichal Hocko * the per-node's. 6804ab5ac90aSMichal Hocko */ 68059157c311SMike Kravetz if (HPageTemporary(newpage)) { 6806ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 6807ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 6808ab5ac90aSMichal Hocko 68099157c311SMike Kravetz SetHPageTemporary(oldpage); 68109157c311SMike Kravetz ClearHPageTemporary(newpage); 6811ab5ac90aSMichal Hocko 68125af1ab1dSMiaohe Lin /* 68135af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 68145af1ab1dSMiaohe Lin * when we do not cross the node. 68155af1ab1dSMiaohe Lin */ 68165af1ab1dSMiaohe Lin if (new_nid == old_nid) 68175af1ab1dSMiaohe Lin return; 6818db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 6819ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 6820ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 6821ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 6822ab5ac90aSMichal Hocko } 6823db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6824ab5ac90aSMichal Hocko } 6825ab5ac90aSMichal Hocko } 6826cf11e85fSRoman Gushchin 68276dfeaff9SPeter Xu /* 68286dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 68296dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 68306dfeaff9SPeter Xu */ 68316dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 68326dfeaff9SPeter Xu { 68336dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 68346dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 68356dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 68366dfeaff9SPeter Xu struct mmu_notifier_range range; 68376dfeaff9SPeter Xu unsigned long address, start, end; 68386dfeaff9SPeter Xu spinlock_t *ptl; 68396dfeaff9SPeter Xu pte_t *ptep; 68406dfeaff9SPeter Xu 68416dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 68426dfeaff9SPeter Xu return; 68436dfeaff9SPeter Xu 68446dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 68456dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 68466dfeaff9SPeter Xu 68476dfeaff9SPeter Xu if (start >= end) 68486dfeaff9SPeter Xu return; 68496dfeaff9SPeter Xu 68506dfeaff9SPeter Xu /* 68516dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 68526dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 68536dfeaff9SPeter Xu */ 68546dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 68556dfeaff9SPeter Xu start, end); 68566dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 68576dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 68586dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 68596dfeaff9SPeter Xu unsigned long tmp = address; 68606dfeaff9SPeter Xu 68616dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 68626dfeaff9SPeter Xu if (!ptep) 68636dfeaff9SPeter Xu continue; 68646dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 68656dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 68666dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 68676dfeaff9SPeter Xu spin_unlock(ptl); 68686dfeaff9SPeter Xu } 68696dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 68706dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 68716dfeaff9SPeter Xu /* 68726dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 68736dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 68746dfeaff9SPeter Xu */ 68756dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 68766dfeaff9SPeter Xu } 68776dfeaff9SPeter Xu 6878cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 6879cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 6880cf11e85fSRoman Gushchin 6881cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 6882cf11e85fSRoman Gushchin { 688338e719abSBaolin Wang int nid, count = 0; 688438e719abSBaolin Wang unsigned long tmp; 688538e719abSBaolin Wang char *s = p; 688638e719abSBaolin Wang 688738e719abSBaolin Wang while (*s) { 688838e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 688938e719abSBaolin Wang break; 689038e719abSBaolin Wang 689138e719abSBaolin Wang if (s[count] == ':') { 689238e719abSBaolin Wang nid = tmp; 689338e719abSBaolin Wang if (nid < 0 || nid >= MAX_NUMNODES) 689438e719abSBaolin Wang break; 689538e719abSBaolin Wang 689638e719abSBaolin Wang s += count + 1; 689738e719abSBaolin Wang tmp = memparse(s, &s); 689838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 689938e719abSBaolin Wang hugetlb_cma_size += tmp; 690038e719abSBaolin Wang 690138e719abSBaolin Wang /* 690238e719abSBaolin Wang * Skip the separator if have one, otherwise 690338e719abSBaolin Wang * break the parsing. 690438e719abSBaolin Wang */ 690538e719abSBaolin Wang if (*s == ',') 690638e719abSBaolin Wang s++; 690738e719abSBaolin Wang else 690838e719abSBaolin Wang break; 690938e719abSBaolin Wang } else { 6910cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 691138e719abSBaolin Wang break; 691238e719abSBaolin Wang } 691338e719abSBaolin Wang } 691438e719abSBaolin Wang 6915cf11e85fSRoman Gushchin return 0; 6916cf11e85fSRoman Gushchin } 6917cf11e85fSRoman Gushchin 6918cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 6919cf11e85fSRoman Gushchin 6920cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 6921cf11e85fSRoman Gushchin { 6922cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 692338e719abSBaolin Wang bool node_specific_cma_alloc = false; 6924cf11e85fSRoman Gushchin int nid; 6925cf11e85fSRoman Gushchin 6926cf11e85fSRoman Gushchin cma_reserve_called = true; 6927cf11e85fSRoman Gushchin 6928cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 6929cf11e85fSRoman Gushchin return; 6930cf11e85fSRoman Gushchin 693138e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 693238e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 693338e719abSBaolin Wang continue; 693438e719abSBaolin Wang 693538e719abSBaolin Wang if (!node_state(nid, N_ONLINE)) { 693638e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 693738e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 693838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 693938e719abSBaolin Wang continue; 694038e719abSBaolin Wang } 694138e719abSBaolin Wang 694238e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 694338e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 694438e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 694538e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 694638e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 694738e719abSBaolin Wang } else { 694838e719abSBaolin Wang node_specific_cma_alloc = true; 694938e719abSBaolin Wang } 695038e719abSBaolin Wang } 695138e719abSBaolin Wang 695238e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 695338e719abSBaolin Wang if (!hugetlb_cma_size) 695438e719abSBaolin Wang return; 695538e719abSBaolin Wang 6956cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 6957cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 6958cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 6959a01f4390SMike Kravetz hugetlb_cma_size = 0; 6960cf11e85fSRoman Gushchin return; 6961cf11e85fSRoman Gushchin } 6962cf11e85fSRoman Gushchin 696338e719abSBaolin Wang if (!node_specific_cma_alloc) { 6964cf11e85fSRoman Gushchin /* 6965cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 6966cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 6967cf11e85fSRoman Gushchin */ 6968cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 6969cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 6970cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 697138e719abSBaolin Wang } 6972cf11e85fSRoman Gushchin 6973cf11e85fSRoman Gushchin reserved = 0; 6974cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 6975cf11e85fSRoman Gushchin int res; 69762281f797SBarry Song char name[CMA_MAX_NAME]; 6977cf11e85fSRoman Gushchin 697838e719abSBaolin Wang if (node_specific_cma_alloc) { 697938e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 698038e719abSBaolin Wang continue; 698138e719abSBaolin Wang 698238e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 698338e719abSBaolin Wang } else { 6984cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 698538e719abSBaolin Wang } 698638e719abSBaolin Wang 6987cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 6988cf11e85fSRoman Gushchin 69892281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 6990a01f4390SMike Kravetz /* 6991a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 6992a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 6993a01f4390SMike Kravetz * huge page demotion. 6994a01f4390SMike Kravetz */ 6995a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 6996a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 699729d0f41dSBarry Song 0, false, name, 6998cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 6999cf11e85fSRoman Gushchin if (res) { 7000cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7001cf11e85fSRoman Gushchin res, nid); 7002cf11e85fSRoman Gushchin continue; 7003cf11e85fSRoman Gushchin } 7004cf11e85fSRoman Gushchin 7005cf11e85fSRoman Gushchin reserved += size; 7006cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7007cf11e85fSRoman Gushchin size / SZ_1M, nid); 7008cf11e85fSRoman Gushchin 7009cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7010cf11e85fSRoman Gushchin break; 7011cf11e85fSRoman Gushchin } 7012a01f4390SMike Kravetz 7013a01f4390SMike Kravetz if (!reserved) 7014a01f4390SMike Kravetz /* 7015a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7016a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7017a01f4390SMike Kravetz */ 7018a01f4390SMike Kravetz hugetlb_cma_size = 0; 7019cf11e85fSRoman Gushchin } 7020cf11e85fSRoman Gushchin 7021cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 7022cf11e85fSRoman Gushchin { 7023cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7024cf11e85fSRoman Gushchin return; 7025cf11e85fSRoman Gushchin 7026cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7027cf11e85fSRoman Gushchin } 7028cf11e85fSRoman Gushchin 7029cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7030