1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34f9317f77SMike Kravetz #include <linux/nospec.h> 35662ce1dcSYang Yang #include <linux/delayacct.h> 36d6606683SLinus Torvalds 3763551ae0SDavid Gibson #include <asm/page.h> 38ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3924669e58SAneesh Kumar K.V #include <asm/tlb.h> 4063551ae0SDavid Gibson 4124669e58SAneesh Kumar K.V #include <linux/io.h> 4263551ae0SDavid Gibson #include <linux/hugetlb.h> 439dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 449a305230SLee Schermerhorn #include <linux/node.h> 45ab5ac90aSMichal Hocko #include <linux/page_owner.h> 467835e98bSNick Piggin #include "internal.h" 47f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 481da177e4SLinus Torvalds 49c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 50e5ff2159SAndi Kleen unsigned int default_hstate_idx; 51e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 52cf11e85fSRoman Gushchin 53dbda8feaSBarry Song #ifdef CONFIG_CMA 54cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5538e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 56a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 57a01f4390SMike Kravetz { 58a01f4390SMike Kravetz return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, 59a01f4390SMike Kravetz 1 << order); 60a01f4390SMike Kravetz } 61a01f4390SMike Kravetz #else 62a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 63a01f4390SMike Kravetz { 64a01f4390SMike Kravetz return false; 65a01f4390SMike Kravetz } 66dbda8feaSBarry Song #endif 67dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 68cf11e85fSRoman Gushchin 6953ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7053ba51d2SJon Tollefson 71e5ff2159SAndi Kleen /* for command line parsing */ 72e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 73e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 749fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 75282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 76b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 77e5ff2159SAndi Kleen 783935baa9SDavid Gibson /* 7931caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8031caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 813935baa9SDavid Gibson */ 82c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 830bd0f9fbSEric Paris 848382d914SDavidlohr Bueso /* 858382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 868382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 878382d914SDavidlohr Bueso */ 888382d914SDavidlohr Bueso static int num_fault_mutexes; 89c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 908382d914SDavidlohr Bueso 917ca02d0aSMike Kravetz /* Forward declaration */ 927ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 937ca02d0aSMike Kravetz 941d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 951d88433bSMiaohe Lin { 961d88433bSMiaohe Lin if (spool->count) 971d88433bSMiaohe Lin return false; 981d88433bSMiaohe Lin if (spool->max_hpages != -1) 991d88433bSMiaohe Lin return spool->used_hpages == 0; 1001d88433bSMiaohe Lin if (spool->min_hpages != -1) 1011d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1021d88433bSMiaohe Lin 1031d88433bSMiaohe Lin return true; 1041d88433bSMiaohe Lin } 1051d88433bSMiaohe Lin 106db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 107db71ef79SMike Kravetz unsigned long irq_flags) 10890481622SDavid Gibson { 109db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11090481622SDavid Gibson 11190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1127c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1137ca02d0aSMike Kravetz * free the subpool */ 1141d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1157ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1167ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1177ca02d0aSMike Kravetz -spool->min_hpages); 11890481622SDavid Gibson kfree(spool); 11990481622SDavid Gibson } 1207ca02d0aSMike Kravetz } 12190481622SDavid Gibson 1227ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1237ca02d0aSMike Kravetz long min_hpages) 12490481622SDavid Gibson { 12590481622SDavid Gibson struct hugepage_subpool *spool; 12690481622SDavid Gibson 127c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 12890481622SDavid Gibson if (!spool) 12990481622SDavid Gibson return NULL; 13090481622SDavid Gibson 13190481622SDavid Gibson spin_lock_init(&spool->lock); 13290481622SDavid Gibson spool->count = 1; 1337ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1347ca02d0aSMike Kravetz spool->hstate = h; 1357ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1367ca02d0aSMike Kravetz 1377ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1387ca02d0aSMike Kravetz kfree(spool); 1397ca02d0aSMike Kravetz return NULL; 1407ca02d0aSMike Kravetz } 1417ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14290481622SDavid Gibson 14390481622SDavid Gibson return spool; 14490481622SDavid Gibson } 14590481622SDavid Gibson 14690481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 14790481622SDavid Gibson { 148db71ef79SMike Kravetz unsigned long flags; 149db71ef79SMike Kravetz 150db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15190481622SDavid Gibson BUG_ON(!spool->count); 15290481622SDavid Gibson spool->count--; 153db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 15490481622SDavid Gibson } 15590481622SDavid Gibson 1561c5ecae3SMike Kravetz /* 1571c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1581c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1599e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1601c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1611c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1627c8de358SEthon Paul * a subpool minimum size must be maintained. 1631c5ecae3SMike Kravetz */ 1641c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 16590481622SDavid Gibson long delta) 16690481622SDavid Gibson { 1671c5ecae3SMike Kravetz long ret = delta; 16890481622SDavid Gibson 16990481622SDavid Gibson if (!spool) 1701c5ecae3SMike Kravetz return ret; 17190481622SDavid Gibson 172db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 17390481622SDavid Gibson 1741c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1751c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1761c5ecae3SMike Kravetz spool->used_hpages += delta; 1771c5ecae3SMike Kravetz else { 1781c5ecae3SMike Kravetz ret = -ENOMEM; 1791c5ecae3SMike Kravetz goto unlock_ret; 1801c5ecae3SMike Kravetz } 1811c5ecae3SMike Kravetz } 1821c5ecae3SMike Kravetz 18309a95e29SMike Kravetz /* minimum size accounting */ 18409a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1851c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1861c5ecae3SMike Kravetz /* 1871c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1881c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1891c5ecae3SMike Kravetz */ 1901c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1911c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1921c5ecae3SMike Kravetz } else { 1931c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1941c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1951c5ecae3SMike Kravetz } 1961c5ecae3SMike Kravetz } 1971c5ecae3SMike Kravetz 1981c5ecae3SMike Kravetz unlock_ret: 199db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20090481622SDavid Gibson return ret; 20190481622SDavid Gibson } 20290481622SDavid Gibson 2031c5ecae3SMike Kravetz /* 2041c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2051c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2061c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2071c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2081c5ecae3SMike Kravetz */ 2091c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21090481622SDavid Gibson long delta) 21190481622SDavid Gibson { 2121c5ecae3SMike Kravetz long ret = delta; 213db71ef79SMike Kravetz unsigned long flags; 2141c5ecae3SMike Kravetz 21590481622SDavid Gibson if (!spool) 2161c5ecae3SMike Kravetz return delta; 21790481622SDavid Gibson 218db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2191c5ecae3SMike Kravetz 2201c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22190481622SDavid Gibson spool->used_hpages -= delta; 2221c5ecae3SMike Kravetz 22309a95e29SMike Kravetz /* minimum size accounting */ 22409a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2251c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2261c5ecae3SMike Kravetz ret = 0; 2271c5ecae3SMike Kravetz else 2281c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2291c5ecae3SMike Kravetz 2301c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2311c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2321c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2331c5ecae3SMike Kravetz } 2341c5ecae3SMike Kravetz 2351c5ecae3SMike Kravetz /* 2361c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2371c5ecae3SMike Kravetz * quota reference, free it now. 2381c5ecae3SMike Kravetz */ 239db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2401c5ecae3SMike Kravetz 2411c5ecae3SMike Kravetz return ret; 24290481622SDavid Gibson } 24390481622SDavid Gibson 24490481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 24590481622SDavid Gibson { 24690481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 24790481622SDavid Gibson } 24890481622SDavid Gibson 24990481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25090481622SDavid Gibson { 251496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25290481622SDavid Gibson } 25390481622SDavid Gibson 2540db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2550db9d74eSMina Almasry * it for use. 2560db9d74eSMina Almasry */ 2570db9d74eSMina Almasry static struct file_region * 2580db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2590db9d74eSMina Almasry { 2600db9d74eSMina Almasry struct file_region *nrg = NULL; 2610db9d74eSMina Almasry 2620db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2630db9d74eSMina Almasry 2640db9d74eSMina Almasry resv->region_cache_count--; 2650db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2660db9d74eSMina Almasry list_del(&nrg->link); 2670db9d74eSMina Almasry 2680db9d74eSMina Almasry nrg->from = from; 2690db9d74eSMina Almasry nrg->to = to; 2700db9d74eSMina Almasry 2710db9d74eSMina Almasry return nrg; 2720db9d74eSMina Almasry } 2730db9d74eSMina Almasry 274075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 275075a61d0SMina Almasry struct file_region *rg) 276075a61d0SMina Almasry { 277075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 278075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 279075a61d0SMina Almasry nrg->css = rg->css; 280075a61d0SMina Almasry if (rg->css) 281075a61d0SMina Almasry css_get(rg->css); 282075a61d0SMina Almasry #endif 283075a61d0SMina Almasry } 284075a61d0SMina Almasry 285075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 286075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 287075a61d0SMina Almasry struct hstate *h, 288075a61d0SMina Almasry struct resv_map *resv, 289075a61d0SMina Almasry struct file_region *nrg) 290075a61d0SMina Almasry { 291075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 292075a61d0SMina Almasry if (h_cg) { 293075a61d0SMina Almasry nrg->reservation_counter = 294075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 295075a61d0SMina Almasry nrg->css = &h_cg->css; 296d85aecf2SMiaohe Lin /* 297d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 298d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 299d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 300d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 301d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 302d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 303d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 304d85aecf2SMiaohe Lin * untouched. 305d85aecf2SMiaohe Lin */ 306d85aecf2SMiaohe Lin css_get(&h_cg->css); 307075a61d0SMina Almasry if (!resv->pages_per_hpage) 308075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 309075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 310075a61d0SMina Almasry * a resv_map. 311075a61d0SMina Almasry */ 312075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 313075a61d0SMina Almasry } else { 314075a61d0SMina Almasry nrg->reservation_counter = NULL; 315075a61d0SMina Almasry nrg->css = NULL; 316075a61d0SMina Almasry } 317075a61d0SMina Almasry #endif 318075a61d0SMina Almasry } 319075a61d0SMina Almasry 320d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 321d85aecf2SMiaohe Lin { 322d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 323d85aecf2SMiaohe Lin if (rg->css) 324d85aecf2SMiaohe Lin css_put(rg->css); 325d85aecf2SMiaohe Lin #endif 326d85aecf2SMiaohe Lin } 327d85aecf2SMiaohe Lin 328a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 329a9b3f867SMina Almasry struct file_region *org) 330a9b3f867SMina Almasry { 331a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 3320739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 333a9b3f867SMina Almasry rg->css == org->css; 334a9b3f867SMina Almasry 335a9b3f867SMina Almasry #else 336a9b3f867SMina Almasry return true; 337a9b3f867SMina Almasry #endif 338a9b3f867SMina Almasry } 339a9b3f867SMina Almasry 340a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 341a9b3f867SMina Almasry { 342a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 343a9b3f867SMina Almasry 344a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 345a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 346a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 347a9b3f867SMina Almasry prg->to = rg->to; 348a9b3f867SMina Almasry 349a9b3f867SMina Almasry list_del(&rg->link); 350d85aecf2SMiaohe Lin put_uncharge_info(rg); 351a9b3f867SMina Almasry kfree(rg); 352a9b3f867SMina Almasry 3537db5e7b6SWei Yang rg = prg; 354a9b3f867SMina Almasry } 355a9b3f867SMina Almasry 356a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 357a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 358a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 359a9b3f867SMina Almasry nrg->from = rg->from; 360a9b3f867SMina Almasry 361a9b3f867SMina Almasry list_del(&rg->link); 362d85aecf2SMiaohe Lin put_uncharge_info(rg); 363a9b3f867SMina Almasry kfree(rg); 364a9b3f867SMina Almasry } 365a9b3f867SMina Almasry } 366a9b3f867SMina Almasry 3672103cf9cSPeter Xu static inline long 36884448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 3692103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3702103cf9cSPeter Xu long *regions_needed) 3712103cf9cSPeter Xu { 3722103cf9cSPeter Xu struct file_region *nrg; 3732103cf9cSPeter Xu 3742103cf9cSPeter Xu if (!regions_needed) { 3752103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3762103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 37784448c8eSJakob Koschel list_add(&nrg->link, rg); 3782103cf9cSPeter Xu coalesce_file_region(map, nrg); 3792103cf9cSPeter Xu } else 3802103cf9cSPeter Xu *regions_needed += 1; 3812103cf9cSPeter Xu 3822103cf9cSPeter Xu return to - from; 3832103cf9cSPeter Xu } 3842103cf9cSPeter Xu 385972a3da3SWei Yang /* 386972a3da3SWei Yang * Must be called with resv->lock held. 387972a3da3SWei Yang * 388972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 389972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 390972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 391972a3da3SWei Yang * the regions for this range. 392d75c6af9SMina Almasry */ 393d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 394075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 395972a3da3SWei Yang struct hstate *h, long *regions_needed) 396d75c6af9SMina Almasry { 3970db9d74eSMina Almasry long add = 0; 398d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3990db9d74eSMina Almasry long last_accounted_offset = f; 40084448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 40184448c8eSJakob Koschel struct list_head *rg = NULL; 402d75c6af9SMina Almasry 4030db9d74eSMina Almasry if (regions_needed) 4040db9d74eSMina Almasry *regions_needed = 0; 405d75c6af9SMina Almasry 4060db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 40784448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 4080db9d74eSMina Almasry * bounds checking. 4090db9d74eSMina Almasry */ 41084448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 4110db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 41284448c8eSJakob Koschel if (iter->from < f) { 4130db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4140db9d74eSMina Almasry * then we need to update last_accounted_offset. 4150db9d74eSMina Almasry */ 41684448c8eSJakob Koschel if (iter->to > last_accounted_offset) 41784448c8eSJakob Koschel last_accounted_offset = iter->to; 4180db9d74eSMina Almasry continue; 4190db9d74eSMina Almasry } 420d75c6af9SMina Almasry 4210db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4220db9d74eSMina Almasry * finished. 4230db9d74eSMina Almasry */ 42484448c8eSJakob Koschel if (iter->from >= t) { 42584448c8eSJakob Koschel rg = iter->link.prev; 426d75c6af9SMina Almasry break; 42784448c8eSJakob Koschel } 428d75c6af9SMina Almasry 42984448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 4300db9d74eSMina Almasry * update last_accounted_offset. 431d75c6af9SMina Almasry */ 43284448c8eSJakob Koschel if (iter->from > last_accounted_offset) 43384448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 4342103cf9cSPeter Xu last_accounted_offset, 43584448c8eSJakob Koschel iter->from, h, h_cg, 4362103cf9cSPeter Xu regions_needed); 437d75c6af9SMina Almasry 43884448c8eSJakob Koschel last_accounted_offset = iter->to; 4390db9d74eSMina Almasry } 4400db9d74eSMina Almasry 4410db9d74eSMina Almasry /* Handle the case where our range extends beyond 4420db9d74eSMina Almasry * last_accounted_offset. 4430db9d74eSMina Almasry */ 44484448c8eSJakob Koschel if (!rg) 44584448c8eSJakob Koschel rg = head->prev; 4462103cf9cSPeter Xu if (last_accounted_offset < t) 4472103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4482103cf9cSPeter Xu t, h, h_cg, regions_needed); 4490db9d74eSMina Almasry 4500db9d74eSMina Almasry return add; 4510db9d74eSMina Almasry } 4520db9d74eSMina Almasry 4530db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4540db9d74eSMina Almasry */ 4550db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4560db9d74eSMina Almasry int regions_needed) 4570db9d74eSMina Almasry __must_hold(&resv->lock) 4580db9d74eSMina Almasry { 4590db9d74eSMina Almasry struct list_head allocated_regions; 4600db9d74eSMina Almasry int to_allocate = 0, i = 0; 4610db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4620db9d74eSMina Almasry 4630db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4640db9d74eSMina Almasry 4650db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4660db9d74eSMina Almasry 4670db9d74eSMina Almasry /* 4680db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4690db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4700db9d74eSMina Almasry * 4710db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4720db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4730db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4740db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4750db9d74eSMina Almasry */ 4760db9d74eSMina Almasry while (resv->region_cache_count < 4770db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4780db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4790db9d74eSMina Almasry resv->region_cache_count; 4800db9d74eSMina Almasry 4810db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 482f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4830db9d74eSMina Almasry * needing to allocate for regions_needed. 4840db9d74eSMina Almasry */ 4850db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4860db9d74eSMina Almasry 4870db9d74eSMina Almasry spin_unlock(&resv->lock); 4880db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4890db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4900db9d74eSMina Almasry if (!trg) 4910db9d74eSMina Almasry goto out_of_memory; 4920db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4930db9d74eSMina Almasry } 4940db9d74eSMina Almasry 4950db9d74eSMina Almasry spin_lock(&resv->lock); 4960db9d74eSMina Almasry 497d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 498d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4990db9d74eSMina Almasry } 5000db9d74eSMina Almasry 5010db9d74eSMina Almasry return 0; 5020db9d74eSMina Almasry 5030db9d74eSMina Almasry out_of_memory: 5040db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 505d75c6af9SMina Almasry list_del(&rg->link); 506d75c6af9SMina Almasry kfree(rg); 507d75c6af9SMina Almasry } 5080db9d74eSMina Almasry return -ENOMEM; 509d75c6af9SMina Almasry } 510d75c6af9SMina Almasry 5111dd308a7SMike Kravetz /* 5121dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5130db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5140db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5150db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5160db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5170db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 518cf3ad20bSMike Kravetz * 5190db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5200db9d74eSMina Almasry * 5210db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5220db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5237c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5240db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5250db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5260db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5271dd308a7SMike Kravetz */ 5280db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 529075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 530075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 53196822904SAndy Whitcroft { 5320db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53396822904SAndy Whitcroft 5347b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5350db9d74eSMina Almasry retry: 5360db9d74eSMina Almasry 5370db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 538972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 539972a3da3SWei Yang &actual_regions_needed); 54096822904SAndy Whitcroft 5415e911373SMike Kravetz /* 5420db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5430db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5440db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5450db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5460db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5470db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5480db9d74eSMina Almasry * operation. 5495e911373SMike Kravetz */ 5500db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5510db9d74eSMina Almasry resv->region_cache_count < 5520db9d74eSMina Almasry resv->adds_in_progress + 5530db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5540db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5550db9d74eSMina Almasry * allocate file_region entries. 5560db9d74eSMina Almasry */ 5570db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5585e911373SMike Kravetz 5590db9d74eSMina Almasry if (allocate_file_region_entries( 5600db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5610db9d74eSMina Almasry return -ENOMEM; 5625e911373SMike Kravetz } 5635e911373SMike Kravetz 5640db9d74eSMina Almasry goto retry; 5650db9d74eSMina Almasry } 566cf3ad20bSMike Kravetz 567972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5680db9d74eSMina Almasry 5690db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5700db9d74eSMina Almasry 5717b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 572cf3ad20bSMike Kravetz return add; 57396822904SAndy Whitcroft } 57496822904SAndy Whitcroft 5751dd308a7SMike Kravetz /* 5761dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5771dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5781dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5791dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5801dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5811dd308a7SMike Kravetz * not change the number of huge pages represented by the 5820db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5830db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5840db9d74eSMina Almasry * file_region structure is added. 5850db9d74eSMina Almasry * 5860db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5870db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5880db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5895e911373SMike Kravetz * 5905e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5915e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5925e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5935e911373SMike Kravetz * is needed and can not be allocated. 5941dd308a7SMike Kravetz */ 5950db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5960db9d74eSMina Almasry long *out_regions_needed) 59796822904SAndy Whitcroft { 59896822904SAndy Whitcroft long chg = 0; 59996822904SAndy Whitcroft 6007b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6015e911373SMike Kravetz 602972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 603075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 604972a3da3SWei Yang out_regions_needed); 6055e911373SMike Kravetz 6060db9d74eSMina Almasry if (*out_regions_needed == 0) 6070db9d74eSMina Almasry *out_regions_needed = 1; 6085e911373SMike Kravetz 6090db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6105e911373SMike Kravetz return -ENOMEM; 6115e911373SMike Kravetz 6120db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61396822904SAndy Whitcroft 6147b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61596822904SAndy Whitcroft return chg; 61696822904SAndy Whitcroft } 61796822904SAndy Whitcroft 6181dd308a7SMike Kravetz /* 6195e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6205e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6215e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6225e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6230db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6240db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6250db9d74eSMina Almasry * the adds_in_progress counter. 6265e911373SMike Kravetz * 6275e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6285e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6295e911373SMike Kravetz * arguments will match the associated region_chg call. 6305e911373SMike Kravetz */ 6310db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6320db9d74eSMina Almasry long regions_needed) 6335e911373SMike Kravetz { 6345e911373SMike Kravetz spin_lock(&resv->lock); 6355e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6360db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6375e911373SMike Kravetz spin_unlock(&resv->lock); 6385e911373SMike Kravetz } 6395e911373SMike Kravetz 6405e911373SMike Kravetz /* 641feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 642feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 643feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 644feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 645feba16e2SMike Kravetz * 646feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 647feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 648feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 649feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 650feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 651feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 652feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6531dd308a7SMike Kravetz */ 654feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65596822904SAndy Whitcroft { 6561406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65796822904SAndy Whitcroft struct file_region *rg, *trg; 658feba16e2SMike Kravetz struct file_region *nrg = NULL; 659feba16e2SMike Kravetz long del = 0; 66096822904SAndy Whitcroft 661feba16e2SMike Kravetz retry: 6627b24d861SDavidlohr Bueso spin_lock(&resv->lock); 663feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 664dbe409e4SMike Kravetz /* 665dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 666dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 667dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 668dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 669dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 670dbe409e4SMike Kravetz */ 671dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 672feba16e2SMike Kravetz continue; 673dbe409e4SMike Kravetz 674feba16e2SMike Kravetz if (rg->from >= t) 67596822904SAndy Whitcroft break; 67696822904SAndy Whitcroft 677feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 678feba16e2SMike Kravetz /* 679feba16e2SMike Kravetz * Check for an entry in the cache before dropping 680feba16e2SMike Kravetz * lock and attempting allocation. 681feba16e2SMike Kravetz */ 682feba16e2SMike Kravetz if (!nrg && 683feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 684feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 685feba16e2SMike Kravetz struct file_region, 686feba16e2SMike Kravetz link); 687feba16e2SMike Kravetz list_del(&nrg->link); 688feba16e2SMike Kravetz resv->region_cache_count--; 68996822904SAndy Whitcroft } 69096822904SAndy Whitcroft 691feba16e2SMike Kravetz if (!nrg) { 692feba16e2SMike Kravetz spin_unlock(&resv->lock); 693feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 694feba16e2SMike Kravetz if (!nrg) 695feba16e2SMike Kravetz return -ENOMEM; 696feba16e2SMike Kravetz goto retry; 697feba16e2SMike Kravetz } 698feba16e2SMike Kravetz 699feba16e2SMike Kravetz del += t - f; 70079aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 701d85aecf2SMiaohe Lin resv, rg, t - f, false); 702feba16e2SMike Kravetz 703feba16e2SMike Kravetz /* New entry for end of split region */ 704feba16e2SMike Kravetz nrg->from = t; 705feba16e2SMike Kravetz nrg->to = rg->to; 706075a61d0SMina Almasry 707075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 708075a61d0SMina Almasry 709feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 710feba16e2SMike Kravetz 711feba16e2SMike Kravetz /* Original entry is trimmed */ 712feba16e2SMike Kravetz rg->to = f; 713feba16e2SMike Kravetz 714feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 715feba16e2SMike Kravetz nrg = NULL; 71696822904SAndy Whitcroft break; 717feba16e2SMike Kravetz } 718feba16e2SMike Kravetz 719feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 720feba16e2SMike Kravetz del += rg->to - rg->from; 721075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 722d85aecf2SMiaohe Lin rg->to - rg->from, true); 72396822904SAndy Whitcroft list_del(&rg->link); 72496822904SAndy Whitcroft kfree(rg); 725feba16e2SMike Kravetz continue; 72696822904SAndy Whitcroft } 7277b24d861SDavidlohr Bueso 728feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 729075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 730d85aecf2SMiaohe Lin t - rg->from, false); 731075a61d0SMina Almasry 73279aa925bSMike Kravetz del += t - rg->from; 73379aa925bSMike Kravetz rg->from = t; 73479aa925bSMike Kravetz } else { /* Trim end of region */ 735075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 736d85aecf2SMiaohe Lin rg->to - f, false); 73779aa925bSMike Kravetz 73879aa925bSMike Kravetz del += rg->to - f; 73979aa925bSMike Kravetz rg->to = f; 740feba16e2SMike Kravetz } 741feba16e2SMike Kravetz } 742feba16e2SMike Kravetz 7437b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 744feba16e2SMike Kravetz kfree(nrg); 745feba16e2SMike Kravetz return del; 74696822904SAndy Whitcroft } 74796822904SAndy Whitcroft 7481dd308a7SMike Kravetz /* 749b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 750b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 751b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 752b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 753b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 754b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 755b5cec28dSMike Kravetz * counts. 756b5cec28dSMike Kravetz */ 75772e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 758b5cec28dSMike Kravetz { 759b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 760b5cec28dSMike Kravetz long rsv_adjust; 761da56388cSMiaohe Lin bool reserved = false; 762b5cec28dSMike Kravetz 763b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 764da56388cSMiaohe Lin if (rsv_adjust > 0) { 765b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 766b5cec28dSMike Kravetz 767da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 768da56388cSMiaohe Lin reserved = true; 769da56388cSMiaohe Lin } else if (!rsv_adjust) { 770da56388cSMiaohe Lin reserved = true; 771b5cec28dSMike Kravetz } 772da56388cSMiaohe Lin 773da56388cSMiaohe Lin if (!reserved) 774da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 775b5cec28dSMike Kravetz } 776b5cec28dSMike Kravetz 777b5cec28dSMike Kravetz /* 7781dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7791dd308a7SMike Kravetz * that intersect with the range [f, t). 7801dd308a7SMike Kravetz */ 7811406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78284afd99bSAndy Whitcroft { 7831406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78484afd99bSAndy Whitcroft struct file_region *rg; 78584afd99bSAndy Whitcroft long chg = 0; 78684afd99bSAndy Whitcroft 7877b24d861SDavidlohr Bueso spin_lock(&resv->lock); 78884afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 78984afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 790f2135a4aSWang Sheng-Hui long seg_from; 791f2135a4aSWang Sheng-Hui long seg_to; 79284afd99bSAndy Whitcroft 79384afd99bSAndy Whitcroft if (rg->to <= f) 79484afd99bSAndy Whitcroft continue; 79584afd99bSAndy Whitcroft if (rg->from >= t) 79684afd99bSAndy Whitcroft break; 79784afd99bSAndy Whitcroft 79884afd99bSAndy Whitcroft seg_from = max(rg->from, f); 79984afd99bSAndy Whitcroft seg_to = min(rg->to, t); 80084afd99bSAndy Whitcroft 80184afd99bSAndy Whitcroft chg += seg_to - seg_from; 80284afd99bSAndy Whitcroft } 8037b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80484afd99bSAndy Whitcroft 80584afd99bSAndy Whitcroft return chg; 80684afd99bSAndy Whitcroft } 80784afd99bSAndy Whitcroft 80896822904SAndy Whitcroft /* 809e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 810e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 811e7c4b0bfSAndy Whitcroft */ 812a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 813a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 814e7c4b0bfSAndy Whitcroft { 815a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 816a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 817e7c4b0bfSAndy Whitcroft } 818e7c4b0bfSAndy Whitcroft 8190fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8200fe6e20bSNaoya Horiguchi unsigned long address) 8210fe6e20bSNaoya Horiguchi { 8220fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8230fe6e20bSNaoya Horiguchi } 824dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8250fe6e20bSNaoya Horiguchi 82684afd99bSAndy Whitcroft /* 82708fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 82808fba699SMel Gorman * cases this will be same size as used by the page table entries. 82908fba699SMel Gorman */ 83008fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 83108fba699SMel Gorman { 83205ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83305ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83408fba699SMel Gorman return PAGE_SIZE; 83508fba699SMel Gorman } 836f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83708fba699SMel Gorman 83808fba699SMel Gorman /* 8393340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8403340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 84109135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84209135cc5SDan Williams * version of this symbol is required. 8433340289dSMel Gorman */ 84409135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8453340289dSMel Gorman { 8463340289dSMel Gorman return vma_kernel_pagesize(vma); 8473340289dSMel Gorman } 8483340289dSMel Gorman 8493340289dSMel Gorman /* 85084afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 85184afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85284afd99bSAndy Whitcroft * alignment. 85384afd99bSAndy Whitcroft */ 85484afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85584afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85604f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85784afd99bSAndy Whitcroft 858a1e78772SMel Gorman /* 859a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 860a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 861a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 862a1e78772SMel Gorman * 863a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 864a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 865a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 866a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86784afd99bSAndy Whitcroft * 86884afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 86984afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 87084afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 87184afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87284afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87384afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87484afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87584afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 876a1e78772SMel Gorman */ 877e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 878e7c4b0bfSAndy Whitcroft { 879e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 880e7c4b0bfSAndy Whitcroft } 881e7c4b0bfSAndy Whitcroft 882e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 883e7c4b0bfSAndy Whitcroft unsigned long value) 884e7c4b0bfSAndy Whitcroft { 885e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 886e7c4b0bfSAndy Whitcroft } 887e7c4b0bfSAndy Whitcroft 888e9fe92aeSMina Almasry static void 889e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 890e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 891e9fe92aeSMina Almasry struct hstate *h) 892e9fe92aeSMina Almasry { 893e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 894e9fe92aeSMina Almasry if (!h_cg || !h) { 895e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 896e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 897e9fe92aeSMina Almasry resv_map->css = NULL; 898e9fe92aeSMina Almasry } else { 899e9fe92aeSMina Almasry resv_map->reservation_counter = 900e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 901e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 902e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 903e9fe92aeSMina Almasry } 904e9fe92aeSMina Almasry #endif 905e9fe92aeSMina Almasry } 906e9fe92aeSMina Almasry 9079119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 90884afd99bSAndy Whitcroft { 90984afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9105e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9115e911373SMike Kravetz 9125e911373SMike Kravetz if (!resv_map || !rg) { 9135e911373SMike Kravetz kfree(resv_map); 9145e911373SMike Kravetz kfree(rg); 91584afd99bSAndy Whitcroft return NULL; 9165e911373SMike Kravetz } 91784afd99bSAndy Whitcroft 91884afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9197b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 92084afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 92184afd99bSAndy Whitcroft 9225e911373SMike Kravetz resv_map->adds_in_progress = 0; 923e9fe92aeSMina Almasry /* 924e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 925e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 926e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 927e9fe92aeSMina Almasry * reservations are to be un-charged from here. 928e9fe92aeSMina Almasry */ 929e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9305e911373SMike Kravetz 9315e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9325e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9335e911373SMike Kravetz resv_map->region_cache_count = 1; 9345e911373SMike Kravetz 93584afd99bSAndy Whitcroft return resv_map; 93684afd99bSAndy Whitcroft } 93784afd99bSAndy Whitcroft 9389119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 93984afd99bSAndy Whitcroft { 94084afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9415e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9425e911373SMike Kravetz struct file_region *rg, *trg; 94384afd99bSAndy Whitcroft 94484afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 945feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9465e911373SMike Kravetz 9475e911373SMike Kravetz /* ... and any entries left in the cache */ 9485e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9495e911373SMike Kravetz list_del(&rg->link); 9505e911373SMike Kravetz kfree(rg); 9515e911373SMike Kravetz } 9525e911373SMike Kravetz 9535e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9545e911373SMike Kravetz 95584afd99bSAndy Whitcroft kfree(resv_map); 95684afd99bSAndy Whitcroft } 95784afd99bSAndy Whitcroft 9584e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9594e35f483SJoonsoo Kim { 960f27a5136SMike Kravetz /* 961f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 962f27a5136SMike Kravetz * address space within the inode. This original address space 963f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 964f27a5136SMike Kravetz * address space embedded within the inode. 965f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 966f27a5136SMike Kravetz * this may not be true for device special inodes. 967f27a5136SMike Kravetz */ 968f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9694e35f483SJoonsoo Kim } 9704e35f483SJoonsoo Kim 97184afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 972a1e78772SMel Gorman { 97381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9744e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9754e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9764e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9774e35f483SJoonsoo Kim 9784e35f483SJoonsoo Kim return inode_resv_map(inode); 9794e35f483SJoonsoo Kim 9804e35f483SJoonsoo Kim } else { 98184afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98284afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9834e35f483SJoonsoo Kim } 984a1e78772SMel Gorman } 985a1e78772SMel Gorman 98684afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 987a1e78772SMel Gorman { 98881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 990a1e78772SMel Gorman 99184afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99284afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99304f2cbe3SMel Gorman } 99404f2cbe3SMel Gorman 99504f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99604f2cbe3SMel Gorman { 99781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 999e7c4b0bfSAndy Whitcroft 1000e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 100104f2cbe3SMel Gorman } 100204f2cbe3SMel Gorman 100304f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100404f2cbe3SMel Gorman { 100581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1006e7c4b0bfSAndy Whitcroft 1007e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1008a1e78772SMel Gorman } 1009a1e78772SMel Gorman 101004f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 1011a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1012a1e78772SMel Gorman { 101381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1014f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1015a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1016a1e78772SMel Gorman } 1017a1e78772SMel Gorman 1018550a7d60SMina Almasry /* 1019550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1020550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1021550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1022550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1023550a7d60SMina Almasry * reservation. 1024550a7d60SMina Almasry */ 1025550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1026550a7d60SMina Almasry { 1027550a7d60SMina Almasry /* 1028550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1029550a7d60SMina Almasry * It has already been transferred to new_vma. 1030550a7d60SMina Almasry * 1031550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1032550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1033550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1034550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1035550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1036550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1037550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1038550a7d60SMina Almasry */ 1039550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1040550a7d60SMina Almasry 1041afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1042afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1043550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1044afe041c2SBui Quang Minh } 1045550a7d60SMina Almasry 1046550a7d60SMina Almasry reset_vma_resv_huge_pages(vma); 1047550a7d60SMina Almasry } 1048550a7d60SMina Almasry 1049a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1050559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1051a1e78772SMel Gorman { 1052af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1053af0ed73eSJoonsoo Kim /* 1054af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1055af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1056af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1057af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1058af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1059af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1060af0ed73eSJoonsoo Kim * properly, so add work-around here. 1061af0ed73eSJoonsoo Kim */ 1062af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1063559ec2f8SNicholas Krause return true; 1064af0ed73eSJoonsoo Kim else 1065559ec2f8SNicholas Krause return false; 1066af0ed73eSJoonsoo Kim } 1067a63884e9SJoonsoo Kim 1068a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10691fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10701fb1b0e9SMike Kravetz /* 10711fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10721fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10731fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10747c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10751fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10761fb1b0e9SMike Kravetz */ 10771fb1b0e9SMike Kravetz if (chg) 10781fb1b0e9SMike Kravetz return false; 10791fb1b0e9SMike Kravetz else 1080559ec2f8SNicholas Krause return true; 10811fb1b0e9SMike Kravetz } 1082a63884e9SJoonsoo Kim 1083a63884e9SJoonsoo Kim /* 1084a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1085a63884e9SJoonsoo Kim * private mappings. 1086a63884e9SJoonsoo Kim */ 108767961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 108867961f9dSMike Kravetz /* 108967961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 109067961f9dSMike Kravetz * could have been performed on the private mapping. 109167961f9dSMike Kravetz * Examine the value of chg to determine if reserves 109267961f9dSMike Kravetz * actually exist or were previously consumed. 109367961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 109467961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 109567961f9dSMike Kravetz * private mappings has different (opposite) semantics 109667961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 109767961f9dSMike Kravetz * has already taken this difference in semantics into 109867961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 109967961f9dSMike Kravetz * as in the shared case above. Code could easily be 110067961f9dSMike Kravetz * combined, but keeping it separate draws attention to 110167961f9dSMike Kravetz * subtle differences. 110267961f9dSMike Kravetz */ 110367961f9dSMike Kravetz if (chg) 110467961f9dSMike Kravetz return false; 110567961f9dSMike Kravetz else 1106559ec2f8SNicholas Krause return true; 110767961f9dSMike Kravetz } 1108a63884e9SJoonsoo Kim 1109559ec2f8SNicholas Krause return false; 1110a1e78772SMel Gorman } 1111a1e78772SMel Gorman 1112a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11131da177e4SLinus Torvalds { 11141da177e4SLinus Torvalds int nid = page_to_nid(page); 11159487ca60SMike Kravetz 11169487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1117b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1118b65a4edaSMike Kravetz 11190edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1120a5516438SAndi Kleen h->free_huge_pages++; 1121a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11226c037149SMike Kravetz SetHPageFreed(page); 11231da177e4SLinus Torvalds } 11241da177e4SLinus Torvalds 112594310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1126bf50bab2SNaoya Horiguchi { 1127bf50bab2SNaoya Horiguchi struct page *page; 11281a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1129bf50bab2SNaoya Horiguchi 11309487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1131bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11326077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1133bbe88753SJoonsoo Kim continue; 1134bbe88753SJoonsoo Kim 11356664bfc8SWei Yang if (PageHWPoison(page)) 11366664bfc8SWei Yang continue; 1137bbe88753SJoonsoo Kim 11380edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1139a9869b83SNaoya Horiguchi set_page_refcounted(page); 11406c037149SMike Kravetz ClearHPageFreed(page); 1141bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1142bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1143bf50bab2SNaoya Horiguchi return page; 1144bf50bab2SNaoya Horiguchi } 1145bf50bab2SNaoya Horiguchi 11466664bfc8SWei Yang return NULL; 11476664bfc8SWei Yang } 11486664bfc8SWei Yang 11493e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11503e59fcb0SMichal Hocko nodemask_t *nmask) 115194310cbcSAnshuman Khandual { 11523e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11533e59fcb0SMichal Hocko struct zonelist *zonelist; 11543e59fcb0SMichal Hocko struct zone *zone; 11553e59fcb0SMichal Hocko struct zoneref *z; 115698fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11573e59fcb0SMichal Hocko 11583e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11593e59fcb0SMichal Hocko 11603e59fcb0SMichal Hocko retry_cpuset: 11613e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11623e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 116394310cbcSAnshuman Khandual struct page *page; 116494310cbcSAnshuman Khandual 11653e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11663e59fcb0SMichal Hocko continue; 11673e59fcb0SMichal Hocko /* 11683e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11693e59fcb0SMichal Hocko * zone aware 11703e59fcb0SMichal Hocko */ 11713e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11723e59fcb0SMichal Hocko continue; 11733e59fcb0SMichal Hocko node = zone_to_nid(zone); 117494310cbcSAnshuman Khandual 117594310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 117694310cbcSAnshuman Khandual if (page) 117794310cbcSAnshuman Khandual return page; 117894310cbcSAnshuman Khandual } 11793e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11803e59fcb0SMichal Hocko goto retry_cpuset; 11813e59fcb0SMichal Hocko 118294310cbcSAnshuman Khandual return NULL; 118394310cbcSAnshuman Khandual } 118494310cbcSAnshuman Khandual 1185a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1186a5516438SAndi Kleen struct vm_area_struct *vma, 1187af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1188af0ed73eSJoonsoo Kim long chg) 11891da177e4SLinus Torvalds { 1190cfcaa66fSBen Widawsky struct page *page = NULL; 1191480eccf9SLee Schermerhorn struct mempolicy *mpol; 119204ec6264SVlastimil Babka gfp_t gfp_mask; 11933e59fcb0SMichal Hocko nodemask_t *nodemask; 119404ec6264SVlastimil Babka int nid; 11951da177e4SLinus Torvalds 1196a1e78772SMel Gorman /* 1197a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1198a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1199a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1200a1e78772SMel Gorman */ 1201af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1202a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1203c0ff7453SMiao Xie goto err; 1204a1e78772SMel Gorman 120504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1206a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 12076eab04a8SJustin P. Mattock goto err; 120804f2cbe3SMel Gorman 120904ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 121004ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1211cfcaa66fSBen Widawsky 1212cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12133e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1214cfcaa66fSBen Widawsky 1215cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1216cfcaa66fSBen Widawsky nodemask = NULL; 1217cfcaa66fSBen Widawsky } 1218cfcaa66fSBen Widawsky 1219cfcaa66fSBen Widawsky if (!page) 1220cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1221cfcaa66fSBen Widawsky 12223e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1223d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1224a63884e9SJoonsoo Kim h->resv_huge_pages--; 1225bf50bab2SNaoya Horiguchi } 1226cc9a6c87SMel Gorman 1227cc9a6c87SMel Gorman mpol_cond_put(mpol); 1228cc9a6c87SMel Gorman return page; 1229cc9a6c87SMel Gorman 1230c0ff7453SMiao Xie err: 1231cc9a6c87SMel Gorman return NULL; 12321da177e4SLinus Torvalds } 12331da177e4SLinus Torvalds 12341cac6f2cSLuiz Capitulino /* 12351cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12361cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12371cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12381cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12391cac6f2cSLuiz Capitulino * node for alloc or free. 12401cac6f2cSLuiz Capitulino */ 12411cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12421cac6f2cSLuiz Capitulino { 12430edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12441cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12451cac6f2cSLuiz Capitulino 12461cac6f2cSLuiz Capitulino return nid; 12471cac6f2cSLuiz Capitulino } 12481cac6f2cSLuiz Capitulino 12491cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12501cac6f2cSLuiz Capitulino { 12511cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12521cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12531cac6f2cSLuiz Capitulino return nid; 12541cac6f2cSLuiz Capitulino } 12551cac6f2cSLuiz Capitulino 12561cac6f2cSLuiz Capitulino /* 12571cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12581cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12591cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12601cac6f2cSLuiz Capitulino * mask. 12611cac6f2cSLuiz Capitulino */ 12621cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12631cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12641cac6f2cSLuiz Capitulino { 12651cac6f2cSLuiz Capitulino int nid; 12661cac6f2cSLuiz Capitulino 12671cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12681cac6f2cSLuiz Capitulino 12691cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12701cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12711cac6f2cSLuiz Capitulino 12721cac6f2cSLuiz Capitulino return nid; 12731cac6f2cSLuiz Capitulino } 12741cac6f2cSLuiz Capitulino 12751cac6f2cSLuiz Capitulino /* 127610c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12771cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12781cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12791cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12801cac6f2cSLuiz Capitulino */ 12811cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12821cac6f2cSLuiz Capitulino { 12831cac6f2cSLuiz Capitulino int nid; 12841cac6f2cSLuiz Capitulino 12851cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12861cac6f2cSLuiz Capitulino 12871cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12881cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12891cac6f2cSLuiz Capitulino 12901cac6f2cSLuiz Capitulino return nid; 12911cac6f2cSLuiz Capitulino } 12921cac6f2cSLuiz Capitulino 12931cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12941cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12951cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12961cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12971cac6f2cSLuiz Capitulino nr_nodes--) 12981cac6f2cSLuiz Capitulino 12991cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 13001cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13011cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13021cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13031cac6f2cSLuiz Capitulino nr_nodes--) 13041cac6f2cSLuiz Capitulino 13058531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 130634d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 130734d9e35bSMike Kravetz unsigned int order, bool demote) 1308944d9fecSLuiz Capitulino { 1309944d9fecSLuiz Capitulino int i; 1310944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1311944d9fecSLuiz Capitulino struct page *p = page + 1; 1312944d9fecSLuiz Capitulino 1313c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 131447e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 131547e29d32SJohn Hubbard 1316944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1317a01f4390SMike Kravetz p->mapping = NULL; 13181d798ca3SKirill A. Shutemov clear_compound_head(p); 131934d9e35bSMike Kravetz if (!demote) 1320944d9fecSLuiz Capitulino set_page_refcounted(p); 1321944d9fecSLuiz Capitulino } 1322944d9fecSLuiz Capitulino 1323944d9fecSLuiz Capitulino set_compound_order(page, 0); 13245232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 1325ba9c1201SGerald Schaefer page[1].compound_nr = 0; 13265232c63fSMatthew Wilcox (Oracle) #endif 1327944d9fecSLuiz Capitulino __ClearPageHead(page); 1328944d9fecSLuiz Capitulino } 1329944d9fecSLuiz Capitulino 13308531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13318531fc6fSMike Kravetz unsigned int order) 13328531fc6fSMike Kravetz { 13338531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13348531fc6fSMike Kravetz } 13358531fc6fSMike Kravetz 13368531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 133734d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 133834d9e35bSMike Kravetz unsigned int order) 133934d9e35bSMike Kravetz { 134034d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 134134d9e35bSMike Kravetz } 134234d9e35bSMike Kravetz 1343d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1344944d9fecSLuiz Capitulino { 1345cf11e85fSRoman Gushchin /* 1346cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1347cf11e85fSRoman Gushchin * cma_release() returns false. 1348cf11e85fSRoman Gushchin */ 1349dbda8feaSBarry Song #ifdef CONFIG_CMA 1350dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1351cf11e85fSRoman Gushchin return; 1352dbda8feaSBarry Song #endif 1353cf11e85fSRoman Gushchin 1354944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1355944d9fecSLuiz Capitulino } 1356944d9fecSLuiz Capitulino 13574eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1358d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1359d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1360944d9fecSLuiz Capitulino { 136104adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1362953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1363953f064aSLi Xinhai nid = numa_mem_id(); 1364944d9fecSLuiz Capitulino 1365dbda8feaSBarry Song #ifdef CONFIG_CMA 1366dbda8feaSBarry Song { 1367cf11e85fSRoman Gushchin struct page *page; 1368cf11e85fSRoman Gushchin int node; 1369cf11e85fSRoman Gushchin 1370953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1371953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1372953f064aSLi Xinhai huge_page_order(h), true); 1373953f064aSLi Xinhai if (page) 1374953f064aSLi Xinhai return page; 1375953f064aSLi Xinhai } 1376953f064aSLi Xinhai 1377953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1378cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1379953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1380cf11e85fSRoman Gushchin continue; 1381cf11e85fSRoman Gushchin 1382cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1383cf11e85fSRoman Gushchin huge_page_order(h), true); 1384cf11e85fSRoman Gushchin if (page) 1385cf11e85fSRoman Gushchin return page; 1386cf11e85fSRoman Gushchin } 1387cf11e85fSRoman Gushchin } 1388953f064aSLi Xinhai } 1389dbda8feaSBarry Song #endif 1390cf11e85fSRoman Gushchin 13915e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1392944d9fecSLuiz Capitulino } 1393944d9fecSLuiz Capitulino 13944eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13954eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13964eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13974eb0716eSAlexandre Ghiti { 13984eb0716eSAlexandre Ghiti return NULL; 13994eb0716eSAlexandre Ghiti } 14004eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1401944d9fecSLuiz Capitulino 1402e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1403d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14044eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14054eb0716eSAlexandre Ghiti { 14064eb0716eSAlexandre Ghiti return NULL; 14074eb0716eSAlexandre Ghiti } 1408d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1409944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1410d00181b9SKirill A. Shutemov unsigned int order) { } 1411944d9fecSLuiz Capitulino #endif 1412944d9fecSLuiz Capitulino 14136eb4e88aSMike Kravetz /* 14146eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 141534d9e35bSMike Kravetz * as just a compound page. 141634d9e35bSMike Kravetz * 141734d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14186eb4e88aSMike Kravetz * 14196eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14206eb4e88aSMike Kravetz */ 142134d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 142234d9e35bSMike Kravetz bool adjust_surplus, 142334d9e35bSMike Kravetz bool demote) 14246eb4e88aSMike Kravetz { 14256eb4e88aSMike Kravetz int nid = page_to_nid(page); 14266eb4e88aSMike Kravetz 14276eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14286eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14296eb4e88aSMike Kravetz 14309487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14316eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14326eb4e88aSMike Kravetz return; 14336eb4e88aSMike Kravetz 14346eb4e88aSMike Kravetz list_del(&page->lru); 14356eb4e88aSMike Kravetz 14366eb4e88aSMike Kravetz if (HPageFreed(page)) { 14376eb4e88aSMike Kravetz h->free_huge_pages--; 14386eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14396eb4e88aSMike Kravetz } 14406eb4e88aSMike Kravetz if (adjust_surplus) { 14416eb4e88aSMike Kravetz h->surplus_huge_pages--; 14426eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14436eb4e88aSMike Kravetz } 14446eb4e88aSMike Kravetz 1445e32d20c0SMike Kravetz /* 1446e32d20c0SMike Kravetz * Very subtle 1447e32d20c0SMike Kravetz * 1448e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1449e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1450e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1451e32d20c0SMike Kravetz * their reference. 1452e32d20c0SMike Kravetz * 1453e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1454e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1455e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1456e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1457e32d20c0SMike Kravetz * apply. 1458e32d20c0SMike Kravetz * 1459e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1460e32d20c0SMike Kravetz * after update_and_free_page is called. 146134d9e35bSMike Kravetz * 146234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 146334d9e35bSMike Kravetz * be turned into a page of smaller size. 1464e32d20c0SMike Kravetz */ 146534d9e35bSMike Kravetz if (!demote) 14666eb4e88aSMike Kravetz set_page_refcounted(page); 1467e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14686eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1469e32d20c0SMike Kravetz else 1470e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14716eb4e88aSMike Kravetz 14726eb4e88aSMike Kravetz h->nr_huge_pages--; 14736eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14746eb4e88aSMike Kravetz } 14756eb4e88aSMike Kravetz 147634d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 147734d9e35bSMike Kravetz bool adjust_surplus) 147834d9e35bSMike Kravetz { 147934d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 148034d9e35bSMike Kravetz } 148134d9e35bSMike Kravetz 14828531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14838531fc6fSMike Kravetz bool adjust_surplus) 14848531fc6fSMike Kravetz { 14858531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 14868531fc6fSMike Kravetz } 14878531fc6fSMike Kravetz 1488ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1489ad2fa371SMuchun Song bool adjust_surplus) 1490ad2fa371SMuchun Song { 1491ad2fa371SMuchun Song int zeroed; 1492ad2fa371SMuchun Song int nid = page_to_nid(page); 1493ad2fa371SMuchun Song 1494ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1495ad2fa371SMuchun Song 1496ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1497ad2fa371SMuchun Song 1498ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1499ad2fa371SMuchun Song h->nr_huge_pages++; 1500ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1501ad2fa371SMuchun Song 1502ad2fa371SMuchun Song if (adjust_surplus) { 1503ad2fa371SMuchun Song h->surplus_huge_pages++; 1504ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1505ad2fa371SMuchun Song } 1506ad2fa371SMuchun Song 1507ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1508ad2fa371SMuchun Song set_page_private(page, 0); 1509ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1510ad2fa371SMuchun Song 1511ad2fa371SMuchun Song /* 1512b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1513b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1514b65a4edaSMike Kravetz * just in case. 1515ad2fa371SMuchun Song */ 1516ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1517b65a4edaSMike Kravetz if (!zeroed) 1518b65a4edaSMike Kravetz /* 1519b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1520b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1521b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1522b65a4edaSMike Kravetz * when this other ref is dropped. 1523b65a4edaSMike Kravetz */ 1524b65a4edaSMike Kravetz return; 1525b65a4edaSMike Kravetz 1526ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1527ad2fa371SMuchun Song enqueue_huge_page(h, page); 1528ad2fa371SMuchun Song } 1529ad2fa371SMuchun Song 1530b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15316af2acb6SAdam Litke { 15326af2acb6SAdam Litke int i; 1533dbfee5aeSMike Kravetz struct page *subpage = page; 1534a5516438SAndi Kleen 15354eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1536944d9fecSLuiz Capitulino return; 153718229df5SAndy Whitcroft 1538161df60eSNaoya Horiguchi /* 1539161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1540161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1541161df60eSNaoya Horiguchi */ 1542161df60eSNaoya Horiguchi if (HPageRawHwpUnreliable(page)) 1543161df60eSNaoya Horiguchi return; 1544161df60eSNaoya Horiguchi 15456213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1546ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1547ad2fa371SMuchun Song /* 1548ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1549ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1550ad2fa371SMuchun Song * as a surplus page. 1551ad2fa371SMuchun Song */ 1552ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1553ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1554ad2fa371SMuchun Song return; 1555ad2fa371SMuchun Song } 1556ad2fa371SMuchun Song 1557161df60eSNaoya Horiguchi /* 1558161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1559161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1560161df60eSNaoya Horiguchi */ 1561161df60eSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 1562161df60eSNaoya Horiguchi hugetlb_clear_page_hwpoison(page); 1563161df60eSNaoya Horiguchi 1564dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1565dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1566dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 156732f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1568a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1569a7407a27SLuiz Capitulino 1 << PG_writeback); 15706af2acb6SAdam Litke } 1571a01f4390SMike Kravetz 1572a01f4390SMike Kravetz /* 1573a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1574a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1575a01f4390SMike Kravetz */ 1576a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1577a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1578944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1579944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1580944d9fecSLuiz Capitulino } else { 1581a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 15826af2acb6SAdam Litke } 1583944d9fecSLuiz Capitulino } 15846af2acb6SAdam Litke 1585b65d4adbSMuchun Song /* 1586b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1587b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1588b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1589b65d4adbSMuchun Song * the vmemmap pages. 1590b65d4adbSMuchun Song * 1591b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1592b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1593b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1594b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1595b65d4adbSMuchun Song */ 1596b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1597b65d4adbSMuchun Song 1598b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1599b65d4adbSMuchun Song { 1600b65d4adbSMuchun Song struct llist_node *node; 1601b65d4adbSMuchun Song 1602b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1603b65d4adbSMuchun Song 1604b65d4adbSMuchun Song while (node) { 1605b65d4adbSMuchun Song struct page *page; 1606b65d4adbSMuchun Song struct hstate *h; 1607b65d4adbSMuchun Song 1608b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1609b65d4adbSMuchun Song struct page, mapping); 1610b65d4adbSMuchun Song node = node->next; 1611b65d4adbSMuchun Song page->mapping = NULL; 1612b65d4adbSMuchun Song /* 1613b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1614b65d4adbSMuchun Song * is going to trigger because a previous call to 1615b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1616b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1617b65d4adbSMuchun Song */ 1618b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1619b65d4adbSMuchun Song 1620b65d4adbSMuchun Song __update_and_free_page(h, page); 1621b65d4adbSMuchun Song 1622b65d4adbSMuchun Song cond_resched(); 1623b65d4adbSMuchun Song } 1624b65d4adbSMuchun Song } 1625b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1626b65d4adbSMuchun Song 1627b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1628b65d4adbSMuchun Song { 16296213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1630b65d4adbSMuchun Song flush_work(&free_hpage_work); 1631b65d4adbSMuchun Song } 1632b65d4adbSMuchun Song 1633b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1634b65d4adbSMuchun Song bool atomic) 1635b65d4adbSMuchun Song { 1636ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1637b65d4adbSMuchun Song __update_and_free_page(h, page); 1638b65d4adbSMuchun Song return; 1639b65d4adbSMuchun Song } 1640b65d4adbSMuchun Song 1641b65d4adbSMuchun Song /* 1642b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1643b65d4adbSMuchun Song * 1644b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1645b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1646b65d4adbSMuchun Song * hasn't retrieved the list yet. 1647b65d4adbSMuchun Song */ 1648b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1649b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1650b65d4adbSMuchun Song } 1651b65d4adbSMuchun Song 165210c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 165310c6ec49SMike Kravetz { 165410c6ec49SMike Kravetz struct page *page, *t_page; 165510c6ec49SMike Kravetz 165610c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1657b65d4adbSMuchun Song update_and_free_page(h, page, false); 165810c6ec49SMike Kravetz cond_resched(); 165910c6ec49SMike Kravetz } 166010c6ec49SMike Kravetz } 166110c6ec49SMike Kravetz 1662e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1663e5ff2159SAndi Kleen { 1664e5ff2159SAndi Kleen struct hstate *h; 1665e5ff2159SAndi Kleen 1666e5ff2159SAndi Kleen for_each_hstate(h) { 1667e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1668e5ff2159SAndi Kleen return h; 1669e5ff2159SAndi Kleen } 1670e5ff2159SAndi Kleen return NULL; 1671e5ff2159SAndi Kleen } 1672e5ff2159SAndi Kleen 1673db71ef79SMike Kravetz void free_huge_page(struct page *page) 167427a85ef1SDavid Gibson { 1675a5516438SAndi Kleen /* 1676a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1677a5516438SAndi Kleen * compound page destructor. 1678a5516438SAndi Kleen */ 1679e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16807893d1d5SAdam Litke int nid = page_to_nid(page); 1681d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 168207443a85SJoonsoo Kim bool restore_reserve; 1683db71ef79SMike Kravetz unsigned long flags; 168427a85ef1SDavid Gibson 1685b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1686b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 16878ace22bcSYongkai Wu 1688d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 168978fbe906SDavid Hildenbrand if (PageAnon(page)) 169078fbe906SDavid Hildenbrand __ClearPageAnonExclusive(page); 16918ace22bcSYongkai Wu page->mapping = NULL; 1692d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1693d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 169427a85ef1SDavid Gibson 16951c5ecae3SMike Kravetz /* 1696d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 16970919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 16980919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 16990919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 17006c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 17010919e1b6SMike Kravetz * remove the reserved page from the subpool. 17020919e1b6SMike Kravetz */ 17030919e1b6SMike Kravetz if (!restore_reserve) { 17040919e1b6SMike Kravetz /* 17050919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 17060919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 17070919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 17080919e1b6SMike Kravetz * operation. 17091c5ecae3SMike Kravetz */ 17101c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 17111c5ecae3SMike Kravetz restore_reserve = true; 17120919e1b6SMike Kravetz } 17131c5ecae3SMike Kravetz 1714db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 17158f251a3dSMike Kravetz ClearHPageMigratable(page); 17166d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 17176d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 171808cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 171908cf9fafSMina Almasry pages_per_huge_page(h), page); 172007443a85SJoonsoo Kim if (restore_reserve) 172107443a85SJoonsoo Kim h->resv_huge_pages++; 172207443a85SJoonsoo Kim 17239157c311SMike Kravetz if (HPageTemporary(page)) { 17246eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1725db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1726b65d4adbSMuchun Song update_and_free_page(h, page, true); 1727ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17280edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17296eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1730db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1731b65d4adbSMuchun Song update_and_free_page(h, page, true); 17327893d1d5SAdam Litke } else { 17335d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1734a5516438SAndi Kleen enqueue_huge_page(h, page); 1735db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 173627a85ef1SDavid Gibson } 17371121828aSMike Kravetz } 173827a85ef1SDavid Gibson 1739d3d99fccSOscar Salvador /* 1740d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1741d3d99fccSOscar Salvador */ 1742d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1743d3d99fccSOscar Salvador { 1744d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1745d3d99fccSOscar Salvador h->nr_huge_pages++; 1746d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1747d3d99fccSOscar Salvador } 1748d3d99fccSOscar Salvador 1749f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1750b7ba30c6SAndi Kleen { 17516213834cSMuchun Song hugetlb_vmemmap_optimize(h, page); 17520edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1753f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1754ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17559dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17561adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1757d3d99fccSOscar Salvador } 1758d3d99fccSOscar Salvador 1759d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1760d3d99fccSOscar Salvador { 1761f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1762db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1763d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1764db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1765b7ba30c6SAndi Kleen } 1766b7ba30c6SAndi Kleen 176734d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 176834d9e35bSMike Kravetz bool demote) 176920a0307cSWu Fengguang { 17707118fc29SMike Kravetz int i, j; 177120a0307cSWu Fengguang int nr_pages = 1 << order; 177220a0307cSWu Fengguang struct page *p = page + 1; 177320a0307cSWu Fengguang 177420a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 177520a0307cSWu Fengguang set_compound_order(page, order); 1776ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1777de09d31dSKirill A. Shutemov __SetPageHead(page); 177820a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1779ef5a22beSAndrea Arcangeli /* 1780ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1781ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1782ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 17837c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1784ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1785ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1786ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1787ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1788ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1789ef5a22beSAndrea Arcangeli * after get_user_pages(). 1790ef5a22beSAndrea Arcangeli */ 1791ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 17927118fc29SMike Kravetz /* 17937118fc29SMike Kravetz * Subtle and very unlikely 17947118fc29SMike Kravetz * 17957118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 17967118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 17977118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 17987118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 17997118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 18007118fc29SMike Kravetz * We need to respect any increased ref count, and only set 18017118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1802416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1803416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1804416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1805416d85edSMike Kravetz * pages using the appropriate free interface. 180634d9e35bSMike Kravetz * 180734d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 18087118fc29SMike Kravetz */ 180934d9e35bSMike Kravetz if (!demote) { 18107118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1811416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 18127118fc29SMike Kravetz goto out_error; 18137118fc29SMike Kravetz } 181434d9e35bSMike Kravetz } else { 181534d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 181634d9e35bSMike Kravetz } 18171d798ca3SKirill A. Shutemov set_compound_head(p, page); 181820a0307cSWu Fengguang } 1819b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 182047e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18217118fc29SMike Kravetz return true; 18227118fc29SMike Kravetz 18237118fc29SMike Kravetz out_error: 18247118fc29SMike Kravetz /* undo tail page modifications made above */ 18257118fc29SMike Kravetz p = page + 1; 18267118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 18277118fc29SMike Kravetz clear_compound_head(p); 18287118fc29SMike Kravetz set_page_refcounted(p); 18297118fc29SMike Kravetz } 18307118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 18317118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 18327118fc29SMike Kravetz __ClearPageReserved(p); 18337118fc29SMike Kravetz set_compound_order(page, 0); 18345232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 18357118fc29SMike Kravetz page[1].compound_nr = 0; 18365232c63fSMatthew Wilcox (Oracle) #endif 18377118fc29SMike Kravetz __ClearPageHead(page); 18387118fc29SMike Kravetz return false; 183920a0307cSWu Fengguang } 184020a0307cSWu Fengguang 184134d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 184234d9e35bSMike Kravetz { 184334d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 184434d9e35bSMike Kravetz } 184534d9e35bSMike Kravetz 18468531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18478531fc6fSMike Kravetz unsigned int order) 18488531fc6fSMike Kravetz { 18498531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18508531fc6fSMike Kravetz } 18518531fc6fSMike Kravetz 18527795912cSAndrew Morton /* 18537795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18547795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18557795912cSAndrew Morton * details. 18567795912cSAndrew Morton */ 185720a0307cSWu Fengguang int PageHuge(struct page *page) 185820a0307cSWu Fengguang { 185920a0307cSWu Fengguang if (!PageCompound(page)) 186020a0307cSWu Fengguang return 0; 186120a0307cSWu Fengguang 186220a0307cSWu Fengguang page = compound_head(page); 1863f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 186420a0307cSWu Fengguang } 186543131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 186643131e14SNaoya Horiguchi 186727c73ae7SAndrea Arcangeli /* 186827c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 186927c73ae7SAndrea Arcangeli * normal or transparent huge pages. 187027c73ae7SAndrea Arcangeli */ 187127c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 187227c73ae7SAndrea Arcangeli { 187327c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 187427c73ae7SAndrea Arcangeli return 0; 187527c73ae7SAndrea Arcangeli 1876d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 187727c73ae7SAndrea Arcangeli } 18784e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 187927c73ae7SAndrea Arcangeli 1880c0d0381aSMike Kravetz /* 1881c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1882c0d0381aSMike Kravetz * 1883336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1884336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1885336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1886c0d0381aSMike Kravetz */ 1887c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1888c0d0381aSMike Kravetz { 1889336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1890c0d0381aSMike Kravetz 1891c0d0381aSMike Kravetz if (!mapping) 1892c0d0381aSMike Kravetz return mapping; 1893c0d0381aSMike Kravetz 1894c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1895c0d0381aSMike Kravetz return mapping; 1896c0d0381aSMike Kravetz 1897c0d0381aSMike Kravetz return NULL; 1898c0d0381aSMike Kravetz } 1899c0d0381aSMike Kravetz 1900fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 190113d60f4bSZhang Yi { 190213d60f4bSZhang Yi struct page *page_head = compound_head(page); 190313d60f4bSZhang Yi pgoff_t index = page_index(page_head); 190413d60f4bSZhang Yi unsigned long compound_idx; 190513d60f4bSZhang Yi 190613d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 190713d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 190813d60f4bSZhang Yi else 190913d60f4bSZhang Yi compound_idx = page - page_head; 191013d60f4bSZhang Yi 191113d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 191213d60f4bSZhang Yi } 191313d60f4bSZhang Yi 19140c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1915f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1916f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19171da177e4SLinus Torvalds { 1918af0fb9dfSMichal Hocko int order = huge_page_order(h); 19191da177e4SLinus Torvalds struct page *page; 1920f60858f9SMike Kravetz bool alloc_try_hard = true; 1921f96efd58SJoe Jin 1922f60858f9SMike Kravetz /* 1923f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1924f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1925f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1926f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1927f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1928f60858f9SMike Kravetz */ 1929f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1930f60858f9SMike Kravetz alloc_try_hard = false; 1931f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1932f60858f9SMike Kravetz if (alloc_try_hard) 1933f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1934af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1935af0fb9dfSMichal Hocko nid = numa_mem_id(); 193684172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1937af0fb9dfSMichal Hocko if (page) 1938af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1939af0fb9dfSMichal Hocko else 1940af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 194163b4613cSNishanth Aravamudan 1942f60858f9SMike Kravetz /* 1943f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1944f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1945f60858f9SMike Kravetz * normal 'try hard' allocations. 1946f60858f9SMike Kravetz */ 1947f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1948f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1949f60858f9SMike Kravetz 1950f60858f9SMike Kravetz /* 1951f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1952f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1953f60858f9SMike Kravetz * overall state change. 1954f60858f9SMike Kravetz */ 1955f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1956f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1957f60858f9SMike Kravetz 195863b4613cSNishanth Aravamudan return page; 195963b4613cSNishanth Aravamudan } 196063b4613cSNishanth Aravamudan 1961af0fb9dfSMichal Hocko /* 19620c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 19630c397daeSMichal Hocko * should use this function to get new hugetlb pages 19640c397daeSMichal Hocko */ 19650c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1966f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1967f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19680c397daeSMichal Hocko { 19690c397daeSMichal Hocko struct page *page; 19707118fc29SMike Kravetz bool retry = false; 19710c397daeSMichal Hocko 19727118fc29SMike Kravetz retry: 19730c397daeSMichal Hocko if (hstate_is_gigantic(h)) 19740c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 19750c397daeSMichal Hocko else 19760c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1977f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 19780c397daeSMichal Hocko if (!page) 19790c397daeSMichal Hocko return NULL; 19800c397daeSMichal Hocko 19817118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 19827118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 19837118fc29SMike Kravetz /* 19847118fc29SMike Kravetz * Rare failure to convert pages to compound page. 19857118fc29SMike Kravetz * Free pages and try again - ONCE! 19867118fc29SMike Kravetz */ 19877118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 19887118fc29SMike Kravetz if (!retry) { 19897118fc29SMike Kravetz retry = true; 19907118fc29SMike Kravetz goto retry; 19917118fc29SMike Kravetz } 19927118fc29SMike Kravetz return NULL; 19937118fc29SMike Kravetz } 19947118fc29SMike Kravetz } 19950c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 19960c397daeSMichal Hocko 19970c397daeSMichal Hocko return page; 19980c397daeSMichal Hocko } 19990c397daeSMichal Hocko 20000c397daeSMichal Hocko /* 2001af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2002af0fb9dfSMichal Hocko * manner. 2003af0fb9dfSMichal Hocko */ 2004f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2005f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2006b2261026SJoonsoo Kim { 2007b2261026SJoonsoo Kim struct page *page; 2008b2261026SJoonsoo Kim int nr_nodes, node; 2009af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2010b2261026SJoonsoo Kim 2011b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2012f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 2013f60858f9SMike Kravetz node_alloc_noretry); 2014af0fb9dfSMichal Hocko if (page) 2015b2261026SJoonsoo Kim break; 2016b2261026SJoonsoo Kim } 2017b2261026SJoonsoo Kim 2018af0fb9dfSMichal Hocko if (!page) 2019af0fb9dfSMichal Hocko return 0; 2020b2261026SJoonsoo Kim 2021af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2022af0fb9dfSMichal Hocko 2023af0fb9dfSMichal Hocko return 1; 2024b2261026SJoonsoo Kim } 2025b2261026SJoonsoo Kim 2026e8c5c824SLee Schermerhorn /* 202710c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 202810c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 202910c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 203010c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2031e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2032e8c5c824SLee Schermerhorn */ 203310c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 203410c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20356ae11b27SLee Schermerhorn bool acct_surplus) 2036e8c5c824SLee Schermerhorn { 2037b2261026SJoonsoo Kim int nr_nodes, node; 203810c6ec49SMike Kravetz struct page *page = NULL; 2039e8c5c824SLee Schermerhorn 20409487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2041b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2042685f3457SLee Schermerhorn /* 2043685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2044685f3457SLee Schermerhorn * nodes with surplus pages. 2045685f3457SLee Schermerhorn */ 2046b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2047b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 204810c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2049e8c5c824SLee Schermerhorn struct page, lru); 20506eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20519a76db09SLee Schermerhorn break; 2052e8c5c824SLee Schermerhorn } 2053b2261026SJoonsoo Kim } 2054e8c5c824SLee Schermerhorn 205510c6ec49SMike Kravetz return page; 2056e8c5c824SLee Schermerhorn } 2057e8c5c824SLee Schermerhorn 2058c8721bbbSNaoya Horiguchi /* 2059c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2060faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2061faf53defSNaoya Horiguchi * This function returns values like below: 2062faf53defSNaoya Horiguchi * 2063ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2064ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2065ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2066ad2fa371SMuchun Song * is enabled. 2067faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2068faf53defSNaoya Horiguchi * (allocated or reserved.) 2069faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2070faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2071c8721bbbSNaoya Horiguchi */ 2072c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2073c8721bbbSNaoya Horiguchi { 20746bc9b564SNaoya Horiguchi int rc = -EBUSY; 2075082d5b6bSGerald Schaefer 20767ffddd49SMuchun Song retry: 2077faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2078faf53defSNaoya Horiguchi if (!PageHuge(page)) 2079faf53defSNaoya Horiguchi return 0; 2080faf53defSNaoya Horiguchi 2081db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2082faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2083faf53defSNaoya Horiguchi rc = 0; 2084faf53defSNaoya Horiguchi goto out; 2085faf53defSNaoya Horiguchi } 2086faf53defSNaoya Horiguchi 2087faf53defSNaoya Horiguchi if (!page_count(page)) { 20882247bb33SGerald Schaefer struct page *head = compound_head(page); 20892247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 20906bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 2091082d5b6bSGerald Schaefer goto out; 20927ffddd49SMuchun Song 20937ffddd49SMuchun Song /* 20947ffddd49SMuchun Song * We should make sure that the page is already on the free list 20957ffddd49SMuchun Song * when it is dissolved. 20967ffddd49SMuchun Song */ 20976c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2098db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20997ffddd49SMuchun Song cond_resched(); 21007ffddd49SMuchun Song 21017ffddd49SMuchun Song /* 21027ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 21037ffddd49SMuchun Song * encounter this race. In fact, we have a chance 21047ffddd49SMuchun Song * to successfully dissolve the page if we do a 21057ffddd49SMuchun Song * retry. Because the race window is quite small. 21067ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 21077ffddd49SMuchun Song * for increasing the success rate of dissolving page. 21087ffddd49SMuchun Song */ 21097ffddd49SMuchun Song goto retry; 21107ffddd49SMuchun Song } 21117ffddd49SMuchun Song 2112ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2113ad2fa371SMuchun Song h->max_huge_pages--; 2114ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2115ad2fa371SMuchun Song 2116c3114a84SAnshuman Khandual /* 2117ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2118ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2119ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2120ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2121ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2122ad2fa371SMuchun Song * appropriate action on failure. 2123ad2fa371SMuchun Song */ 21246213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, head); 2125ad2fa371SMuchun Song if (!rc) { 2126b65d4adbSMuchun Song update_and_free_page(h, head, false); 2127ad2fa371SMuchun Song } else { 2128ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2129ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2130ad2fa371SMuchun Song h->max_huge_pages++; 2131ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2132ad2fa371SMuchun Song } 2133ad2fa371SMuchun Song 2134ad2fa371SMuchun Song return rc; 2135c8721bbbSNaoya Horiguchi } 2136082d5b6bSGerald Schaefer out: 2137db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2138082d5b6bSGerald Schaefer return rc; 2139c8721bbbSNaoya Horiguchi } 2140c8721bbbSNaoya Horiguchi 2141c8721bbbSNaoya Horiguchi /* 2142c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2143c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21442247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21452247bb33SGerald Schaefer * part of it lies within the given range. 2146082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2147082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2148c8721bbbSNaoya Horiguchi */ 2149082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2150c8721bbbSNaoya Horiguchi { 2151c8721bbbSNaoya Horiguchi unsigned long pfn; 2152eb03aa00SGerald Schaefer struct page *page; 2153082d5b6bSGerald Schaefer int rc = 0; 2154dc2628f3SMuchun Song unsigned int order; 2155dc2628f3SMuchun Song struct hstate *h; 2156c8721bbbSNaoya Horiguchi 2157d0177639SLi Zhong if (!hugepages_supported()) 2158082d5b6bSGerald Schaefer return rc; 2159d0177639SLi Zhong 2160dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2161dc2628f3SMuchun Song for_each_hstate(h) 2162dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2163dc2628f3SMuchun Song 2164dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2165eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2166eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2167eb03aa00SGerald Schaefer if (rc) 2168082d5b6bSGerald Schaefer break; 2169eb03aa00SGerald Schaefer } 2170082d5b6bSGerald Schaefer 2171082d5b6bSGerald Schaefer return rc; 2172c8721bbbSNaoya Horiguchi } 2173c8721bbbSNaoya Horiguchi 2174ab5ac90aSMichal Hocko /* 2175ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2176ab5ac90aSMichal Hocko */ 21770c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2178b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 21797893d1d5SAdam Litke { 21809980d744SMichal Hocko struct page *page = NULL; 2181b65a4edaSMike Kravetz bool retry = false; 21827893d1d5SAdam Litke 2183bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2184aa888a74SAndi Kleen return NULL; 2185aa888a74SAndi Kleen 2186db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21879980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 21889980d744SMichal Hocko goto out_unlock; 2189db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2190d1c3fb1fSNishanth Aravamudan 2191b65a4edaSMike Kravetz retry: 2192f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 21939980d744SMichal Hocko if (!page) 21940c397daeSMichal Hocko return NULL; 2195d1c3fb1fSNishanth Aravamudan 2196db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21979980d744SMichal Hocko /* 21989980d744SMichal Hocko * We could have raced with the pool size change. 21999980d744SMichal Hocko * Double check that and simply deallocate the new page 22009980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 22019980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 22029980d744SMichal Hocko * codeflow 22039980d744SMichal Hocko */ 22049980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 22059157c311SMike Kravetz SetHPageTemporary(page); 2206db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22079980d744SMichal Hocko put_page(page); 22082bf753e6SKai Shen return NULL; 2209b65a4edaSMike Kravetz } 2210b65a4edaSMike Kravetz 2211b65a4edaSMike Kravetz if (zero_ref) { 2212b65a4edaSMike Kravetz /* 2213b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2214b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2215b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2216b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2217b65a4edaSMike Kravetz */ 2218b65a4edaSMike Kravetz SetHPageTemporary(page); 2219b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2220b65a4edaSMike Kravetz /* 2221b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2222b65a4edaSMike Kravetz * huge. Retry once. 2223b65a4edaSMike Kravetz */ 2224b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2225b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2226b65a4edaSMike Kravetz if (retry) 2227b65a4edaSMike Kravetz return NULL; 2228b65a4edaSMike Kravetz 2229b65a4edaSMike Kravetz retry = true; 2230b65a4edaSMike Kravetz goto retry; 2231b65a4edaSMike Kravetz } 2232b65a4edaSMike Kravetz ClearHPageTemporary(page); 2233b65a4edaSMike Kravetz } 2234b65a4edaSMike Kravetz 22359980d744SMichal Hocko h->surplus_huge_pages++; 22364704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22379980d744SMichal Hocko 22389980d744SMichal Hocko out_unlock: 2239db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22407893d1d5SAdam Litke 22417893d1d5SAdam Litke return page; 22427893d1d5SAdam Litke } 22437893d1d5SAdam Litke 2244bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2245ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2246ab5ac90aSMichal Hocko { 2247ab5ac90aSMichal Hocko struct page *page; 2248ab5ac90aSMichal Hocko 2249ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2250ab5ac90aSMichal Hocko return NULL; 2251ab5ac90aSMichal Hocko 2252f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2253ab5ac90aSMichal Hocko if (!page) 2254ab5ac90aSMichal Hocko return NULL; 2255ab5ac90aSMichal Hocko 2256ab5ac90aSMichal Hocko /* 2257ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2258ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2259ab5ac90aSMichal Hocko */ 22609157c311SMike Kravetz SetHPageTemporary(page); 2261ab5ac90aSMichal Hocko 2262ab5ac90aSMichal Hocko return page; 2263ab5ac90aSMichal Hocko } 2264ab5ac90aSMichal Hocko 2265e4e574b7SAdam Litke /* 2266099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2267099730d6SDave Hansen */ 2268e0ec90eeSDave Hansen static 22690c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2270099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2271099730d6SDave Hansen { 2272cfcaa66fSBen Widawsky struct page *page = NULL; 2273aaf14e40SMichal Hocko struct mempolicy *mpol; 2274aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2275aaf14e40SMichal Hocko int nid; 2276aaf14e40SMichal Hocko nodemask_t *nodemask; 2277aaf14e40SMichal Hocko 2278aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2279cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2280cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2281cfcaa66fSBen Widawsky 2282cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2283cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2284cfcaa66fSBen Widawsky 2285cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2286cfcaa66fSBen Widawsky nodemask = NULL; 2287cfcaa66fSBen Widawsky } 2288cfcaa66fSBen Widawsky 2289cfcaa66fSBen Widawsky if (!page) 2290b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2291aaf14e40SMichal Hocko mpol_cond_put(mpol); 2292aaf14e40SMichal Hocko return page; 2293099730d6SDave Hansen } 2294099730d6SDave Hansen 2295ab5ac90aSMichal Hocko /* page migration callback function */ 22963e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2297d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 22984db9b2efSMichal Hocko { 2299db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23004db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 23013e59fcb0SMichal Hocko struct page *page; 23023e59fcb0SMichal Hocko 23033e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 23043e59fcb0SMichal Hocko if (page) { 2305db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23063e59fcb0SMichal Hocko return page; 23074db9b2efSMichal Hocko } 23084db9b2efSMichal Hocko } 2309db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23104db9b2efSMichal Hocko 23110c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 23124db9b2efSMichal Hocko } 23134db9b2efSMichal Hocko 2314ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2315389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2316389c8178SMichal Hocko unsigned long address) 2317ebd63723SMichal Hocko { 2318ebd63723SMichal Hocko struct mempolicy *mpol; 2319ebd63723SMichal Hocko nodemask_t *nodemask; 2320ebd63723SMichal Hocko struct page *page; 2321ebd63723SMichal Hocko gfp_t gfp_mask; 2322ebd63723SMichal Hocko int node; 2323ebd63723SMichal Hocko 2324ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2325ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2326d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2327ebd63723SMichal Hocko mpol_cond_put(mpol); 2328ebd63723SMichal Hocko 2329ebd63723SMichal Hocko return page; 2330ebd63723SMichal Hocko } 2331ebd63723SMichal Hocko 2332bf50bab2SNaoya Horiguchi /* 233325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2334e4e574b7SAdam Litke * of size 'delta'. 2335e4e574b7SAdam Litke */ 23360a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23371b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2338e4e574b7SAdam Litke { 2339e4e574b7SAdam Litke struct list_head surplus_list; 2340e4e574b7SAdam Litke struct page *page, *tmp; 23410a4f3d1bSLiu Xiang int ret; 23420a4f3d1bSLiu Xiang long i; 23430a4f3d1bSLiu Xiang long needed, allocated; 234428073b02SHillf Danton bool alloc_ok = true; 2345e4e574b7SAdam Litke 23469487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2347a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2348ac09b3a1SAdam Litke if (needed <= 0) { 2349a5516438SAndi Kleen h->resv_huge_pages += delta; 2350e4e574b7SAdam Litke return 0; 2351ac09b3a1SAdam Litke } 2352e4e574b7SAdam Litke 2353e4e574b7SAdam Litke allocated = 0; 2354e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2355e4e574b7SAdam Litke 2356e4e574b7SAdam Litke ret = -ENOMEM; 2357e4e574b7SAdam Litke retry: 2358db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2359e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23600c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2361b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 236228073b02SHillf Danton if (!page) { 236328073b02SHillf Danton alloc_ok = false; 236428073b02SHillf Danton break; 236528073b02SHillf Danton } 2366e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 236769ed779aSDavid Rientjes cond_resched(); 2368e4e574b7SAdam Litke } 236928073b02SHillf Danton allocated += i; 2370e4e574b7SAdam Litke 2371e4e574b7SAdam Litke /* 2372e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2373e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2374e4e574b7SAdam Litke */ 2375db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2376a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2377a5516438SAndi Kleen (h->free_huge_pages + allocated); 237828073b02SHillf Danton if (needed > 0) { 237928073b02SHillf Danton if (alloc_ok) 2380e4e574b7SAdam Litke goto retry; 238128073b02SHillf Danton /* 238228073b02SHillf Danton * We were not able to allocate enough pages to 238328073b02SHillf Danton * satisfy the entire reservation so we free what 238428073b02SHillf Danton * we've allocated so far. 238528073b02SHillf Danton */ 238628073b02SHillf Danton goto free; 238728073b02SHillf Danton } 2388e4e574b7SAdam Litke /* 2389e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 239025985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2391e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2392ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2393ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2394ac09b3a1SAdam Litke * before they are reserved. 2395e4e574b7SAdam Litke */ 2396e4e574b7SAdam Litke needed += allocated; 2397a5516438SAndi Kleen h->resv_huge_pages += delta; 2398e4e574b7SAdam Litke ret = 0; 2399a9869b83SNaoya Horiguchi 240019fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 240119fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 240219fc3f0aSAdam Litke if ((--needed) < 0) 240319fc3f0aSAdam Litke break; 2404b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2405a5516438SAndi Kleen enqueue_huge_page(h, page); 240619fc3f0aSAdam Litke } 240728073b02SHillf Danton free: 2408db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 240919fc3f0aSAdam Litke 2410b65a4edaSMike Kravetz /* 2411b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2412b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2413b65a4edaSMike Kravetz */ 2414c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2415b65a4edaSMike Kravetz free_huge_page(page); 2416db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2417e4e574b7SAdam Litke 2418e4e574b7SAdam Litke return ret; 2419e4e574b7SAdam Litke } 2420e4e574b7SAdam Litke 2421e4e574b7SAdam Litke /* 2422e5bbc8a6SMike Kravetz * This routine has two main purposes: 2423e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2424e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2425e5bbc8a6SMike Kravetz * to the associated reservation map. 2426e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2427e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2428e4e574b7SAdam Litke */ 2429a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2430a5516438SAndi Kleen unsigned long unused_resv_pages) 2431e4e574b7SAdam Litke { 2432e4e574b7SAdam Litke unsigned long nr_pages; 243310c6ec49SMike Kravetz struct page *page; 243410c6ec49SMike Kravetz LIST_HEAD(page_list); 243510c6ec49SMike Kravetz 24369487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 243710c6ec49SMike Kravetz /* Uncommit the reservation */ 243810c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2439e4e574b7SAdam Litke 2440c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2441e5bbc8a6SMike Kravetz goto out; 2442aa888a74SAndi Kleen 2443e5bbc8a6SMike Kravetz /* 2444e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2445e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2446e5bbc8a6SMike Kravetz */ 2447a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2448e4e574b7SAdam Litke 2449685f3457SLee Schermerhorn /* 2450685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24519b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24529b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24539b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 245410c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24559b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2456685f3457SLee Schermerhorn */ 2457685f3457SLee Schermerhorn while (nr_pages--) { 245810c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 245910c6ec49SMike Kravetz if (!page) 2460e5bbc8a6SMike Kravetz goto out; 246110c6ec49SMike Kravetz 246210c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2463e4e574b7SAdam Litke } 2464e5bbc8a6SMike Kravetz 2465e5bbc8a6SMike Kravetz out: 2466db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 246710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2468db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2469e4e574b7SAdam Litke } 2470e4e574b7SAdam Litke 24715e911373SMike Kravetz 2472c37f9fb1SAndy Whitcroft /* 2473feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24745e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2475cf3ad20bSMike Kravetz * 2476cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2477cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2478cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2479cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2480cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2481feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2482feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2483feba16e2SMike Kravetz * is called in such cases. 2484cf3ad20bSMike Kravetz * 2485cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2486cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2487cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2488cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2489cf3ad20bSMike Kravetz * take appropriate action. 249096b96a96SMike Kravetz * 249196b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 249296b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 249396b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 249496b96a96SMike Kravetz * reservation exists. 2495846be085SMike Kravetz * 2496846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2497846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2498846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2499846be085SMike Kravetz * exists. 2500c37f9fb1SAndy Whitcroft */ 25015e911373SMike Kravetz enum vma_resv_mode { 25025e911373SMike Kravetz VMA_NEEDS_RESV, 25035e911373SMike Kravetz VMA_COMMIT_RESV, 2504feba16e2SMike Kravetz VMA_END_RESV, 250596b96a96SMike Kravetz VMA_ADD_RESV, 2506846be085SMike Kravetz VMA_DEL_RESV, 25075e911373SMike Kravetz }; 2508cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2509cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 25105e911373SMike Kravetz enum vma_resv_mode mode) 2511c37f9fb1SAndy Whitcroft { 25124e35f483SJoonsoo Kim struct resv_map *resv; 25134e35f483SJoonsoo Kim pgoff_t idx; 2514cf3ad20bSMike Kravetz long ret; 25150db9d74eSMina Almasry long dummy_out_regions_needed; 2516c37f9fb1SAndy Whitcroft 25174e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25184e35f483SJoonsoo Kim if (!resv) 2519c37f9fb1SAndy Whitcroft return 1; 2520c37f9fb1SAndy Whitcroft 25214e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25225e911373SMike Kravetz switch (mode) { 25235e911373SMike Kravetz case VMA_NEEDS_RESV: 25240db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25250db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25260db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25270db9d74eSMina Almasry * ever require 1 region. 25280db9d74eSMina Almasry */ 25290db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25305e911373SMike Kravetz break; 25315e911373SMike Kravetz case VMA_COMMIT_RESV: 2532075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25330db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25340db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25355e911373SMike Kravetz break; 2536feba16e2SMike Kravetz case VMA_END_RESV: 25370db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25385e911373SMike Kravetz ret = 0; 25395e911373SMike Kravetz break; 254096b96a96SMike Kravetz case VMA_ADD_RESV: 25410db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2542075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25430db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25440db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25450db9d74eSMina Almasry } else { 25460db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 254796b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 254896b96a96SMike Kravetz } 254996b96a96SMike Kravetz break; 2550846be085SMike Kravetz case VMA_DEL_RESV: 2551846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2552846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2553846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2554846be085SMike Kravetz } else { 2555846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2556846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2557846be085SMike Kravetz VM_BUG_ON(ret < 0); 2558846be085SMike Kravetz } 2559846be085SMike Kravetz break; 25605e911373SMike Kravetz default: 25615e911373SMike Kravetz BUG(); 25625e911373SMike Kravetz } 256384afd99bSAndy Whitcroft 2564846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2565cf3ad20bSMike Kravetz return ret; 256667961f9dSMike Kravetz /* 2567bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2568bf3d12b9SMiaohe Lin * 256967961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 257067961f9dSMike Kravetz * However, a file associated with mapping could have been 257167961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 257267961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 257367961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 257467961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 257567961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 257667961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 257767961f9dSMike Kravetz * reservation has already been consumed. As a result, the 257867961f9dSMike Kravetz * return value of this routine is the opposite of the 257967961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 258067961f9dSMike Kravetz */ 2581bf3d12b9SMiaohe Lin if (ret > 0) 258267961f9dSMike Kravetz return 0; 2583bf3d12b9SMiaohe Lin if (ret == 0) 258467961f9dSMike Kravetz return 1; 2585bf3d12b9SMiaohe Lin return ret; 258684afd99bSAndy Whitcroft } 2587cf3ad20bSMike Kravetz 2588cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2589a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2590c37f9fb1SAndy Whitcroft { 25915e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2592cf3ad20bSMike Kravetz } 2593c37f9fb1SAndy Whitcroft 2594cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2595cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2596cf3ad20bSMike Kravetz { 25975e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 25985e911373SMike Kravetz } 25995e911373SMike Kravetz 2600feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 26015e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 26025e911373SMike Kravetz { 2603feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2604c37f9fb1SAndy Whitcroft } 2605c37f9fb1SAndy Whitcroft 260696b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 260796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 260896b96a96SMike Kravetz { 260996b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 261096b96a96SMike Kravetz } 261196b96a96SMike Kravetz 2612846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2613846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 261496b96a96SMike Kravetz { 2615846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2616846be085SMike Kravetz } 2617846be085SMike Kravetz 2618846be085SMike Kravetz /* 2619846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2620846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2621846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2622846be085SMike Kravetz * 2623846be085SMike Kravetz * It handles two specific cases: 2624846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2625846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2626846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2627846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2628846be085SMike Kravetz * 2629846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2630846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2631846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2632846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2633846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2634846be085SMike Kravetz * reserve map indicates there is a reservation present. 2635846be085SMike Kravetz * 2636846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2637846be085SMike Kravetz */ 2638846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2639846be085SMike Kravetz unsigned long address, struct page *page) 2640846be085SMike Kravetz { 264196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 264296b96a96SMike Kravetz 2643846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2644846be085SMike Kravetz if (unlikely(rc < 0)) 264596b96a96SMike Kravetz /* 264696b96a96SMike Kravetz * Rare out of memory condition in reserve map 2647d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 264896b96a96SMike Kravetz * global reserve count will not be incremented 264996b96a96SMike Kravetz * by free_huge_page. This will make it appear 265096b96a96SMike Kravetz * as though the reservation for this page was 265196b96a96SMike Kravetz * consumed. This may prevent the task from 265296b96a96SMike Kravetz * faulting in the page at a later time. This 265396b96a96SMike Kravetz * is better than inconsistent global huge page 265496b96a96SMike Kravetz * accounting of reserve counts. 265596b96a96SMike Kravetz */ 2656d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2657846be085SMike Kravetz else if (rc) 2658846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2659846be085SMike Kravetz else 2660846be085SMike Kravetz vma_end_reservation(h, vma, address); 2661846be085SMike Kravetz } else { 2662846be085SMike Kravetz if (!rc) { 266396b96a96SMike Kravetz /* 2664846be085SMike Kravetz * This indicates there is an entry in the reserve map 2665c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2666846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2667846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2668846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2669846be085SMike Kravetz * does not consume a reservation. 267096b96a96SMike Kravetz */ 2671846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2672846be085SMike Kravetz if (rc < 0) 2673846be085SMike Kravetz /* 2674846be085SMike Kravetz * VERY rare out of memory condition. Since 2675846be085SMike Kravetz * we can not delete the entry, set 2676846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2677846be085SMike Kravetz * count will be incremented when the page 2678846be085SMike Kravetz * is freed. This reserve will be consumed 2679846be085SMike Kravetz * on a subsequent allocation. 2680846be085SMike Kravetz */ 2681846be085SMike Kravetz SetHPageRestoreReserve(page); 2682846be085SMike Kravetz } else if (rc < 0) { 2683846be085SMike Kravetz /* 2684846be085SMike Kravetz * Rare out of memory condition from 2685846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2686846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2687846be085SMike Kravetz * this implies there is not an entry in the 2688846be085SMike Kravetz * reserve map. 2689846be085SMike Kravetz * 2690846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2691846be085SMike Kravetz * no reservation. We are done. 2692846be085SMike Kravetz */ 2693846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2694846be085SMike Kravetz /* 2695846be085SMike Kravetz * For private mappings, no entry indicates 2696846be085SMike Kravetz * a reservation is present. Since we can 2697846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2698846be085SMike Kravetz * on the page so reserve count will be 2699846be085SMike Kravetz * incremented when freed. This reserve will 2700846be085SMike Kravetz * be consumed on a subsequent allocation. 2701846be085SMike Kravetz */ 2702846be085SMike Kravetz SetHPageRestoreReserve(page); 270396b96a96SMike Kravetz } else 2704846be085SMike Kravetz /* 2705846be085SMike Kravetz * No reservation present, do nothing 2706846be085SMike Kravetz */ 270796b96a96SMike Kravetz vma_end_reservation(h, vma, address); 270896b96a96SMike Kravetz } 270996b96a96SMike Kravetz } 271096b96a96SMike Kravetz 2711369fa227SOscar Salvador /* 2712369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2713369fa227SOscar Salvador * @h: struct hstate old page belongs to 2714369fa227SOscar Salvador * @old_page: Old page to dissolve 2715ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2716369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2717369fa227SOscar Salvador */ 2718ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2719ae37c7ffSOscar Salvador struct list_head *list) 2720369fa227SOscar Salvador { 2721369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2722369fa227SOscar Salvador int nid = page_to_nid(old_page); 2723b65a4edaSMike Kravetz bool alloc_retry = false; 2724369fa227SOscar Salvador struct page *new_page; 2725369fa227SOscar Salvador int ret = 0; 2726369fa227SOscar Salvador 2727369fa227SOscar Salvador /* 2728369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2729f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2730f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2731f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2732f41f2ed4SMuchun Song * under the lock. 2733369fa227SOscar Salvador */ 2734b65a4edaSMike Kravetz alloc_retry: 2735369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2736369fa227SOscar Salvador if (!new_page) 2737369fa227SOscar Salvador return -ENOMEM; 2738b65a4edaSMike Kravetz /* 2739b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2740b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2741b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2742b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2743b65a4edaSMike Kravetz * 2744b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2745b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2746b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2747b65a4edaSMike Kravetz */ 2748b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2749b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2750b65a4edaSMike Kravetz if (alloc_retry) 2751b65a4edaSMike Kravetz return -EBUSY; 2752b65a4edaSMike Kravetz 2753b65a4edaSMike Kravetz alloc_retry = true; 2754b65a4edaSMike Kravetz goto alloc_retry; 2755b65a4edaSMike Kravetz } 2756b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2757b65a4edaSMike Kravetz 2758f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2759369fa227SOscar Salvador 2760369fa227SOscar Salvador retry: 2761369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2762369fa227SOscar Salvador if (!PageHuge(old_page)) { 2763369fa227SOscar Salvador /* 2764369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2765369fa227SOscar Salvador */ 2766369fa227SOscar Salvador goto free_new; 2767369fa227SOscar Salvador } else if (page_count(old_page)) { 2768369fa227SOscar Salvador /* 2769ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2770ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2771369fa227SOscar Salvador */ 2772ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 27737ce82f4cSMiaohe Lin ret = isolate_hugetlb(old_page, list); 2774ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2775369fa227SOscar Salvador goto free_new; 2776369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2777369fa227SOscar Salvador /* 2778369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2779369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2780369fa227SOscar Salvador * we retry. 2781369fa227SOscar Salvador */ 2782369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2783369fa227SOscar Salvador cond_resched(); 2784369fa227SOscar Salvador goto retry; 2785369fa227SOscar Salvador } else { 2786369fa227SOscar Salvador /* 2787369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2788369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2789369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2790369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2791369fa227SOscar Salvador * stable since this happens under the lock. 2792369fa227SOscar Salvador */ 2793369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2794369fa227SOscar Salvador 2795369fa227SOscar Salvador /* 2796b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2797b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2798369fa227SOscar Salvador */ 2799369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2800369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2801369fa227SOscar Salvador 2802369fa227SOscar Salvador /* 2803369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2804369fa227SOscar Salvador */ 2805369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2806b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2807369fa227SOscar Salvador } 2808369fa227SOscar Salvador 2809369fa227SOscar Salvador return ret; 2810369fa227SOscar Salvador 2811369fa227SOscar Salvador free_new: 2812369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2813b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2814b65a4edaSMike Kravetz set_page_refcounted(new_page); 2815b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2816369fa227SOscar Salvador 2817369fa227SOscar Salvador return ret; 2818369fa227SOscar Salvador } 2819369fa227SOscar Salvador 2820ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2821369fa227SOscar Salvador { 2822369fa227SOscar Salvador struct hstate *h; 2823369fa227SOscar Salvador struct page *head; 2824ae37c7ffSOscar Salvador int ret = -EBUSY; 2825369fa227SOscar Salvador 2826369fa227SOscar Salvador /* 2827369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2828369fa227SOscar Salvador * to carefully check the state under the lock. 2829369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2830369fa227SOscar Salvador */ 2831369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2832369fa227SOscar Salvador if (PageHuge(page)) { 2833369fa227SOscar Salvador head = compound_head(page); 2834369fa227SOscar Salvador h = page_hstate(head); 2835369fa227SOscar Salvador } else { 2836369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2837369fa227SOscar Salvador return 0; 2838369fa227SOscar Salvador } 2839369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2840369fa227SOscar Salvador 2841369fa227SOscar Salvador /* 2842369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2843369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2844369fa227SOscar Salvador * of bailing out right away without further retrying. 2845369fa227SOscar Salvador */ 2846369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2847369fa227SOscar Salvador return -ENOMEM; 2848369fa227SOscar Salvador 28497ce82f4cSMiaohe Lin if (page_count(head) && !isolate_hugetlb(head, list)) 2850ae37c7ffSOscar Salvador ret = 0; 2851ae37c7ffSOscar Salvador else if (!page_count(head)) 2852ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2853ae37c7ffSOscar Salvador 2854ae37c7ffSOscar Salvador return ret; 2855369fa227SOscar Salvador } 2856369fa227SOscar Salvador 285770c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 285804f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2859348ea204SAdam Litke { 286090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2861a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2862348ea204SAdam Litke struct page *page; 2863d85f69b0SMike Kravetz long map_chg, map_commit; 2864d85f69b0SMike Kravetz long gbl_chg; 28656d76dcf4SAneesh Kumar K.V int ret, idx; 28666d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 286708cf9fafSMina Almasry bool deferred_reserve; 28682fc39cecSAdam Litke 28696d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2870a1e78772SMel Gorman /* 2871d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2872d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2873d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2874a1e78772SMel Gorman */ 2875d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2876d85f69b0SMike Kravetz if (map_chg < 0) 287776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2878d85f69b0SMike Kravetz 2879d85f69b0SMike Kravetz /* 2880d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2881d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2882d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2883d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2884d85f69b0SMike Kravetz * checked against any subpool limit. 2885d85f69b0SMike Kravetz */ 2886d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2887d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2888d85f69b0SMike Kravetz if (gbl_chg < 0) { 2889feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 289076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 28915e911373SMike Kravetz } 289290d8b7e6SAdam Litke 2893d85f69b0SMike Kravetz /* 2894d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2895d85f69b0SMike Kravetz * map, there could be reservations associated with the 2896d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2897d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2898d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2899d85f69b0SMike Kravetz * the subpool reservations. 2900d85f69b0SMike Kravetz */ 2901d85f69b0SMike Kravetz if (avoid_reserve) 2902d85f69b0SMike Kravetz gbl_chg = 1; 2903d85f69b0SMike Kravetz } 2904d85f69b0SMike Kravetz 290508cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 290608cf9fafSMina Almasry */ 29076501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 290808cf9fafSMina Almasry if (deferred_reserve) { 290908cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 291008cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 29118f34af6fSJianyu Zhan if (ret) 29128f34af6fSJianyu Zhan goto out_subpool_put; 291308cf9fafSMina Almasry } 291408cf9fafSMina Almasry 291508cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 291608cf9fafSMina Almasry if (ret) 291708cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29188f34af6fSJianyu Zhan 2919db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2920d85f69b0SMike Kravetz /* 2921d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2922d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2923d85f69b0SMike Kravetz * a reservation exists for the allocation. 2924d85f69b0SMike Kravetz */ 2925d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 292681a6fcaeSJoonsoo Kim if (!page) { 2927db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29280c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29298f34af6fSJianyu Zhan if (!page) 29308f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2931a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2932d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2933a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2934a88c7695SNaoya Horiguchi } 2935db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 293615a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 293781a6fcaeSJoonsoo Kim /* Fall through */ 2938a1e78772SMel Gorman } 293981a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 294008cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 294108cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 294208cf9fafSMina Almasry */ 294308cf9fafSMina Almasry if (deferred_reserve) { 294408cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 294508cf9fafSMina Almasry h_cg, page); 294608cf9fafSMina Almasry } 294708cf9fafSMina Almasry 2948db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2949a1e78772SMel Gorman 2950d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2951a1e78772SMel Gorman 2952d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2953d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 295433039678SMike Kravetz /* 295533039678SMike Kravetz * The page was added to the reservation map between 295633039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 295733039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 295833039678SMike Kravetz * Adjust for the subpool count incremented above AND 295933039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 296033039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 296133039678SMike Kravetz * no longer applies. 296233039678SMike Kravetz */ 296333039678SMike Kravetz long rsv_adjust; 296433039678SMike Kravetz 296533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 296633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 296779aa925bSMike Kravetz if (deferred_reserve) 296879aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 296979aa925bSMike Kravetz pages_per_huge_page(h), page); 297033039678SMike Kravetz } 29717893d1d5SAdam Litke return page; 29728f34af6fSJianyu Zhan 29738f34af6fSJianyu Zhan out_uncharge_cgroup: 29748f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 297508cf9fafSMina Almasry out_uncharge_cgroup_reservation: 297608cf9fafSMina Almasry if (deferred_reserve) 297708cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 297808cf9fafSMina Almasry h_cg); 29798f34af6fSJianyu Zhan out_subpool_put: 2980d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29818f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2982feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29838f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2984b45b5bd6SDavid Gibson } 2985b45b5bd6SDavid Gibson 2986b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 2987e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2988b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 2989aa888a74SAndi Kleen { 2990b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 2991b2261026SJoonsoo Kim int nr_nodes, node; 2992aa888a74SAndi Kleen 2993b5389086SZhenguo Yao /* do node specific alloc */ 2994b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 2995b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 2996b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 2997b5389086SZhenguo Yao if (!m) 2998b5389086SZhenguo Yao return 0; 2999b5389086SZhenguo Yao goto found; 3000b5389086SZhenguo Yao } 3001b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3002b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3003b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 30048b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 300597ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3006aa888a74SAndi Kleen /* 3007aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3008aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3009aa888a74SAndi Kleen * puts them into the mem_map). 3010aa888a74SAndi Kleen */ 3011b5389086SZhenguo Yao if (!m) 3012b5389086SZhenguo Yao return 0; 3013aa888a74SAndi Kleen goto found; 3014aa888a74SAndi Kleen } 3015aa888a74SAndi Kleen 3016aa888a74SAndi Kleen found: 3017aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3018330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3019aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3020aa888a74SAndi Kleen m->hstate = h; 3021aa888a74SAndi Kleen return 1; 3022aa888a74SAndi Kleen } 3023aa888a74SAndi Kleen 302448b8d744SMike Kravetz /* 302548b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 302648b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 302748b8d744SMike Kravetz */ 3028aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3029aa888a74SAndi Kleen { 3030aa888a74SAndi Kleen struct huge_bootmem_page *m; 3031aa888a74SAndi Kleen 3032aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 303340d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3034aa888a74SAndi Kleen struct hstate *h = m->hstate; 3035ee8f248dSBecky Bruce 303648b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3037aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30387118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3039ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3040aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30417118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 30427118fc29SMike Kravetz } else { 3043416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30447118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30457118fc29SMike Kravetz } 3046af0fb9dfSMichal Hocko 3047b0320c7bSRafael Aquini /* 304848b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 304948b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 305048b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3051b0320c7bSRafael Aquini */ 3052c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3053520495feSCannon Matthews cond_resched(); 3054aa888a74SAndi Kleen } 3055aa888a74SAndi Kleen } 3056b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3057b5389086SZhenguo Yao { 3058b5389086SZhenguo Yao unsigned long i; 3059b5389086SZhenguo Yao char buf[32]; 3060b5389086SZhenguo Yao 3061b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3062b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3063b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3064b5389086SZhenguo Yao break; 3065b5389086SZhenguo Yao } else { 3066b5389086SZhenguo Yao struct page *page; 3067b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3068b5389086SZhenguo Yao 3069b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3070b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3071b5389086SZhenguo Yao if (!page) 3072b5389086SZhenguo Yao break; 3073b5389086SZhenguo Yao put_page(page); /* free it into the hugepage allocator */ 3074b5389086SZhenguo Yao } 3075b5389086SZhenguo Yao cond_resched(); 3076b5389086SZhenguo Yao } 3077b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3078b5389086SZhenguo Yao return; 3079b5389086SZhenguo Yao 3080b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3081b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3082b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3083b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3084b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3085b5389086SZhenguo Yao } 3086aa888a74SAndi Kleen 30878faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 30881da177e4SLinus Torvalds { 30891da177e4SLinus Torvalds unsigned long i; 3090f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3091b5389086SZhenguo Yao bool node_specific_alloc = false; 3092f60858f9SMike Kravetz 3093b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3094b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3095b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3096b5389086SZhenguo Yao return; 3097b5389086SZhenguo Yao } 3098b5389086SZhenguo Yao 3099b5389086SZhenguo Yao /* do node specific alloc */ 31000a7a0f6fSPeng Liu for_each_online_node(i) { 3101b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3102b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3103b5389086SZhenguo Yao node_specific_alloc = true; 3104b5389086SZhenguo Yao } 3105b5389086SZhenguo Yao } 3106b5389086SZhenguo Yao 3107b5389086SZhenguo Yao if (node_specific_alloc) 3108b5389086SZhenguo Yao return; 3109b5389086SZhenguo Yao 3110b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3111f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3112f60858f9SMike Kravetz /* 3113f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3114f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3115f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3116f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3117f60858f9SMike Kravetz */ 3118f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3119f60858f9SMike Kravetz GFP_KERNEL); 3120f60858f9SMike Kravetz } else { 3121f60858f9SMike Kravetz /* allocations done at boot time */ 3122f60858f9SMike Kravetz node_alloc_noretry = NULL; 3123f60858f9SMike Kravetz } 3124f60858f9SMike Kravetz 3125f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3126f60858f9SMike Kravetz if (node_alloc_noretry) 3127f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31281da177e4SLinus Torvalds 3129e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3130bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3131b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3132aa888a74SAndi Kleen break; 31330c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3134f60858f9SMike Kravetz &node_states[N_MEMORY], 3135f60858f9SMike Kravetz node_alloc_noretry)) 31361da177e4SLinus Torvalds break; 313769ed779aSDavid Rientjes cond_resched(); 31381da177e4SLinus Torvalds } 3139d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3140d715cf80SLiam R. Howlett char buf[32]; 3141d715cf80SLiam R. Howlett 3142c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3143d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3144d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31458faa8b07SAndi Kleen h->max_huge_pages = i; 3146e5ff2159SAndi Kleen } 3147f60858f9SMike Kravetz kfree(node_alloc_noretry); 3148d715cf80SLiam R. Howlett } 3149e5ff2159SAndi Kleen 3150e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3151e5ff2159SAndi Kleen { 315279dfc695SMike Kravetz struct hstate *h, *h2; 3153e5ff2159SAndi Kleen 3154e5ff2159SAndi Kleen for_each_hstate(h) { 31558faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3156bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31578faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 315879dfc695SMike Kravetz 315979dfc695SMike Kravetz /* 316079dfc695SMike Kravetz * Set demote order for each hstate. Note that 316179dfc695SMike Kravetz * h->demote_order is initially 0. 316279dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 316379dfc695SMike Kravetz * is not supported, so skip this. 3164a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3165a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 316679dfc695SMike Kravetz */ 316779dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 316879dfc695SMike Kravetz continue; 3169a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3170a01f4390SMike Kravetz continue; 317179dfc695SMike Kravetz for_each_hstate(h2) { 317279dfc695SMike Kravetz if (h2 == h) 317379dfc695SMike Kravetz continue; 317479dfc695SMike Kravetz if (h2->order < h->order && 317579dfc695SMike Kravetz h2->order > h->demote_order) 317679dfc695SMike Kravetz h->demote_order = h2->order; 317779dfc695SMike Kravetz } 3178e5ff2159SAndi Kleen } 3179e5ff2159SAndi Kleen } 3180e5ff2159SAndi Kleen 3181e5ff2159SAndi Kleen static void __init report_hugepages(void) 3182e5ff2159SAndi Kleen { 3183e5ff2159SAndi Kleen struct hstate *h; 3184e5ff2159SAndi Kleen 3185e5ff2159SAndi Kleen for_each_hstate(h) { 31864abd32dbSAndi Kleen char buf[32]; 3187c6247f72SMatthew Wilcox 3188c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 31896213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3190c6247f72SMatthew Wilcox buf, h->free_huge_pages); 31916213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 31926213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3193e5ff2159SAndi Kleen } 3194e5ff2159SAndi Kleen } 3195e5ff2159SAndi Kleen 31961da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 31976ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 31986ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31991da177e4SLinus Torvalds { 32004415cc8dSChristoph Lameter int i; 32011121828aSMike Kravetz LIST_HEAD(page_list); 32024415cc8dSChristoph Lameter 32039487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3204bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3205aa888a74SAndi Kleen return; 3206aa888a74SAndi Kleen 32071121828aSMike Kravetz /* 32081121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 32091121828aSMike Kravetz */ 32106ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 321110c6ec49SMike Kravetz struct page *page, *next; 3212a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3213a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3214a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32151121828aSMike Kravetz goto out; 32161da177e4SLinus Torvalds if (PageHighMem(page)) 32171da177e4SLinus Torvalds continue; 32186eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32191121828aSMike Kravetz list_add(&page->lru, &page_list); 32201121828aSMike Kravetz } 32211121828aSMike Kravetz } 32221121828aSMike Kravetz 32231121828aSMike Kravetz out: 3224db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 322510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3226db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32271da177e4SLinus Torvalds } 32281da177e4SLinus Torvalds #else 32296ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32306ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32311da177e4SLinus Torvalds { 32321da177e4SLinus Torvalds } 32331da177e4SLinus Torvalds #endif 32341da177e4SLinus Torvalds 323520a0307cSWu Fengguang /* 323620a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 323720a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 323820a0307cSWu Fengguang * Returns 1 if an adjustment was made. 323920a0307cSWu Fengguang */ 32406ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32416ae11b27SLee Schermerhorn int delta) 324220a0307cSWu Fengguang { 3243b2261026SJoonsoo Kim int nr_nodes, node; 324420a0307cSWu Fengguang 32459487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 324620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 324720a0307cSWu Fengguang 3248e8c5c824SLee Schermerhorn if (delta < 0) { 3249b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3250b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3251b2261026SJoonsoo Kim goto found; 3252b2261026SJoonsoo Kim } 3253b2261026SJoonsoo Kim } else { 3254b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3255b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3256b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3257b2261026SJoonsoo Kim goto found; 3258e8c5c824SLee Schermerhorn } 32599a76db09SLee Schermerhorn } 3260b2261026SJoonsoo Kim return 0; 326120a0307cSWu Fengguang 3262b2261026SJoonsoo Kim found: 326320a0307cSWu Fengguang h->surplus_huge_pages += delta; 3264b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3265b2261026SJoonsoo Kim return 1; 326620a0307cSWu Fengguang } 326720a0307cSWu Fengguang 3268a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3269fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32706ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32711da177e4SLinus Torvalds { 32727893d1d5SAdam Litke unsigned long min_count, ret; 327310c6ec49SMike Kravetz struct page *page; 327410c6ec49SMike Kravetz LIST_HEAD(page_list); 3275f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3276f60858f9SMike Kravetz 3277f60858f9SMike Kravetz /* 3278f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3279f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3280f60858f9SMike Kravetz * the requested huge pages. 3281f60858f9SMike Kravetz */ 3282f60858f9SMike Kravetz if (node_alloc_noretry) 3283f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3284f60858f9SMike Kravetz else 3285f60858f9SMike Kravetz return -ENOMEM; 32861da177e4SLinus Torvalds 328729383967SMike Kravetz /* 328829383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 328929383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 329029383967SMike Kravetz */ 329129383967SMike Kravetz mutex_lock(&h->resize_lock); 3292b65d4adbSMuchun Song flush_free_hpage_work(h); 3293db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32944eb0716eSAlexandre Ghiti 32954eb0716eSAlexandre Ghiti /* 3296fd875dcaSMike Kravetz * Check for a node specific request. 3297fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3298fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3299fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3300fd875dcaSMike Kravetz */ 3301fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3302fd875dcaSMike Kravetz unsigned long old_count = count; 3303fd875dcaSMike Kravetz 3304fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3305fd875dcaSMike Kravetz /* 3306fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3307fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3308fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3309fd875dcaSMike Kravetz * largest possible value to align with their intention. 3310fd875dcaSMike Kravetz */ 3311fd875dcaSMike Kravetz if (count < old_count) 3312fd875dcaSMike Kravetz count = ULONG_MAX; 3313fd875dcaSMike Kravetz } 3314fd875dcaSMike Kravetz 3315fd875dcaSMike Kravetz /* 33164eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33174eb0716eSAlexandre Ghiti * page range allocation. 33184eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33194eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33204eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33214eb0716eSAlexandre Ghiti */ 33224eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33234eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3324db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 332529383967SMike Kravetz mutex_unlock(&h->resize_lock); 3326f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33274eb0716eSAlexandre Ghiti return -EINVAL; 33284eb0716eSAlexandre Ghiti } 33294eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33304eb0716eSAlexandre Ghiti } 3331aa888a74SAndi Kleen 33327893d1d5SAdam Litke /* 33337893d1d5SAdam Litke * Increase the pool size 33347893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33357893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3336d1c3fb1fSNishanth Aravamudan * 33370c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3338d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3339d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3340d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3341d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33427893d1d5SAdam Litke */ 3343a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33446ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33457893d1d5SAdam Litke break; 33467893d1d5SAdam Litke } 33477893d1d5SAdam Litke 3348a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33497893d1d5SAdam Litke /* 33507893d1d5SAdam Litke * If this allocation races such that we no longer need the 33517893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33527893d1d5SAdam Litke * and reducing the surplus. 33537893d1d5SAdam Litke */ 3354db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3355649920c6SJia He 3356649920c6SJia He /* yield cpu to avoid soft lockup */ 3357649920c6SJia He cond_resched(); 3358649920c6SJia He 3359f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3360f60858f9SMike Kravetz node_alloc_noretry); 3361db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33627893d1d5SAdam Litke if (!ret) 33637893d1d5SAdam Litke goto out; 33647893d1d5SAdam Litke 3365536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3366536240f2SMel Gorman if (signal_pending(current)) 3367536240f2SMel Gorman goto out; 33687893d1d5SAdam Litke } 33697893d1d5SAdam Litke 33707893d1d5SAdam Litke /* 33717893d1d5SAdam Litke * Decrease the pool size 33727893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33737893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33747893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33757893d1d5SAdam Litke * to the desired size as pages become free. 3376d1c3fb1fSNishanth Aravamudan * 3377d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3378d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3379d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33800c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3381d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3382d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3383d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33847893d1d5SAdam Litke */ 3385a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33866b0c880dSAdam Litke min_count = max(count, min_count); 33876ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 338810c6ec49SMike Kravetz 338910c6ec49SMike Kravetz /* 339010c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 339110c6ec49SMike Kravetz */ 3392a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 339310c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 339410c6ec49SMike Kravetz if (!page) 33951da177e4SLinus Torvalds break; 339610c6ec49SMike Kravetz 339710c6ec49SMike Kravetz list_add(&page->lru, &page_list); 33981da177e4SLinus Torvalds } 339910c6ec49SMike Kravetz /* free the pages after dropping lock */ 3400db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 340110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3402b65d4adbSMuchun Song flush_free_hpage_work(h); 3403db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 340410c6ec49SMike Kravetz 3405a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 34066ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 34077893d1d5SAdam Litke break; 34087893d1d5SAdam Litke } 34097893d1d5SAdam Litke out: 34104eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3411db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 341229383967SMike Kravetz mutex_unlock(&h->resize_lock); 34134eb0716eSAlexandre Ghiti 3414f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3415f60858f9SMike Kravetz 34164eb0716eSAlexandre Ghiti return 0; 34171da177e4SLinus Torvalds } 34181da177e4SLinus Torvalds 34198531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34208531fc6fSMike Kravetz { 34218531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34228531fc6fSMike Kravetz struct hstate *target_hstate; 34238531fc6fSMike Kravetz int rc = 0; 34248531fc6fSMike Kravetz 34258531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34268531fc6fSMike Kravetz 34278531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34288531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34298531fc6fSMike Kravetz 34306213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 34318531fc6fSMike Kravetz if (rc) { 34328531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34338531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34348531fc6fSMike Kravetz set_page_refcounted(page); 34358531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34368531fc6fSMike Kravetz return rc; 34378531fc6fSMike Kravetz } 34388531fc6fSMike Kravetz 34398531fc6fSMike Kravetz /* 34408531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34418531fc6fSMike Kravetz * sizes as it will not ref count pages. 34428531fc6fSMike Kravetz */ 34438531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34448531fc6fSMike Kravetz 34458531fc6fSMike Kravetz /* 34468531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34478531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34488531fc6fSMike Kravetz * as surplus. 34498531fc6fSMike Kravetz * 34508531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34518531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34528531fc6fSMike Kravetz */ 34538531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34548531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34558531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 34568531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 34578531fc6fSMike Kravetz prep_compound_gigantic_page_for_demote(page + i, 34588531fc6fSMike Kravetz target_hstate->order); 34598531fc6fSMike Kravetz else 34608531fc6fSMike Kravetz prep_compound_page(page + i, target_hstate->order); 34618531fc6fSMike Kravetz set_page_private(page + i, 0); 34628531fc6fSMike Kravetz set_page_refcounted(page + i); 34638531fc6fSMike Kravetz prep_new_huge_page(target_hstate, page + i, nid); 34648531fc6fSMike Kravetz put_page(page + i); 34658531fc6fSMike Kravetz } 34668531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34678531fc6fSMike Kravetz 34688531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34698531fc6fSMike Kravetz 34708531fc6fSMike Kravetz /* 34718531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34728531fc6fSMike Kravetz * based on pool changes for the demoted page. 34738531fc6fSMike Kravetz */ 34748531fc6fSMike Kravetz h->max_huge_pages--; 3475a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3476a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 34778531fc6fSMike Kravetz 34788531fc6fSMike Kravetz return rc; 34798531fc6fSMike Kravetz } 34808531fc6fSMike Kravetz 348179dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 348279dfc695SMike Kravetz __must_hold(&hugetlb_lock) 348379dfc695SMike Kravetz { 34848531fc6fSMike Kravetz int nr_nodes, node; 34858531fc6fSMike Kravetz struct page *page; 348679dfc695SMike Kravetz 348779dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 348879dfc695SMike Kravetz 348979dfc695SMike Kravetz /* We should never get here if no demote order */ 349079dfc695SMike Kravetz if (!h->demote_order) { 349179dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 349279dfc695SMike Kravetz return -EINVAL; /* internal error */ 349379dfc695SMike Kravetz } 349479dfc695SMike Kravetz 34958531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 34965a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 34975a317412SMike Kravetz if (PageHWPoison(page)) 34985a317412SMike Kravetz continue; 34995a317412SMike Kravetz 35005a317412SMike Kravetz return demote_free_huge_page(h, page); 35018531fc6fSMike Kravetz } 35028531fc6fSMike Kravetz } 35038531fc6fSMike Kravetz 35045a317412SMike Kravetz /* 35055a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 35065a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 35075a317412SMike Kravetz */ 35085a317412SMike Kravetz return -EBUSY; 350979dfc695SMike Kravetz } 351079dfc695SMike Kravetz 3511a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3512a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3513a3437870SNishanth Aravamudan 351479dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 351579dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 351679dfc695SMike Kravetz 3517a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 351898bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3519a3437870SNishanth Aravamudan 3520a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3521a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3522a3437870SNishanth Aravamudan 35239a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35249a305230SLee Schermerhorn 35259a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3526a3437870SNishanth Aravamudan { 3527a3437870SNishanth Aravamudan int i; 35289a305230SLee Schermerhorn 3529a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35309a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35319a305230SLee Schermerhorn if (nidp) 35329a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3533a3437870SNishanth Aravamudan return &hstates[i]; 35349a305230SLee Schermerhorn } 35359a305230SLee Schermerhorn 35369a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3537a3437870SNishanth Aravamudan } 3538a3437870SNishanth Aravamudan 353906808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3540a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3541a3437870SNishanth Aravamudan { 35429a305230SLee Schermerhorn struct hstate *h; 35439a305230SLee Schermerhorn unsigned long nr_huge_pages; 35449a305230SLee Schermerhorn int nid; 35459a305230SLee Schermerhorn 35469a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35479a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35489a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35499a305230SLee Schermerhorn else 35509a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35519a305230SLee Schermerhorn 3552ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3553a3437870SNishanth Aravamudan } 3554adbe8726SEric B Munson 3555238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3556238d3c13SDavid Rientjes struct hstate *h, int nid, 3557238d3c13SDavid Rientjes unsigned long count, size_t len) 3558a3437870SNishanth Aravamudan { 3559a3437870SNishanth Aravamudan int err; 35602d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3561a3437870SNishanth Aravamudan 35622d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35632d0adf7eSOscar Salvador return -EINVAL; 3564adbe8726SEric B Munson 35659a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35669a305230SLee Schermerhorn /* 35679a305230SLee Schermerhorn * global hstate attribute 35689a305230SLee Schermerhorn */ 35699a305230SLee Schermerhorn if (!(obey_mempolicy && 35702d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35712d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35722d0adf7eSOscar Salvador else 35732d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35742d0adf7eSOscar Salvador } else { 35759a305230SLee Schermerhorn /* 3576fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3577fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35789a305230SLee Schermerhorn */ 35792d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35802d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3581fd875dcaSMike Kravetz } 35829a305230SLee Schermerhorn 35832d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 358406808b08SLee Schermerhorn 35854eb0716eSAlexandre Ghiti return err ? err : len; 358606808b08SLee Schermerhorn } 358706808b08SLee Schermerhorn 3588238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3589238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3590238d3c13SDavid Rientjes size_t len) 3591238d3c13SDavid Rientjes { 3592238d3c13SDavid Rientjes struct hstate *h; 3593238d3c13SDavid Rientjes unsigned long count; 3594238d3c13SDavid Rientjes int nid; 3595238d3c13SDavid Rientjes int err; 3596238d3c13SDavid Rientjes 3597238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3598238d3c13SDavid Rientjes if (err) 3599238d3c13SDavid Rientjes return err; 3600238d3c13SDavid Rientjes 3601238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3602238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3603238d3c13SDavid Rientjes } 3604238d3c13SDavid Rientjes 360506808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 360606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 360706808b08SLee Schermerhorn { 360806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 360906808b08SLee Schermerhorn } 361006808b08SLee Schermerhorn 361106808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 361206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 361306808b08SLee Schermerhorn { 3614238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3615a3437870SNishanth Aravamudan } 3616a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3617a3437870SNishanth Aravamudan 361806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 361906808b08SLee Schermerhorn 362006808b08SLee Schermerhorn /* 362106808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 362206808b08SLee Schermerhorn * huge page alloc/free. 362306808b08SLee Schermerhorn */ 362406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3625ae7a927dSJoe Perches struct kobj_attribute *attr, 3626ae7a927dSJoe Perches char *buf) 362706808b08SLee Schermerhorn { 362806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 362906808b08SLee Schermerhorn } 363006808b08SLee Schermerhorn 363106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 363206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 363306808b08SLee Schermerhorn { 3634238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 363506808b08SLee Schermerhorn } 363606808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 363706808b08SLee Schermerhorn #endif 363806808b08SLee Schermerhorn 363906808b08SLee Schermerhorn 3640a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3641a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3642a3437870SNishanth Aravamudan { 36439a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3644ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3645a3437870SNishanth Aravamudan } 3646adbe8726SEric B Munson 3647a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3648a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3649a3437870SNishanth Aravamudan { 3650a3437870SNishanth Aravamudan int err; 3651a3437870SNishanth Aravamudan unsigned long input; 36529a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3653a3437870SNishanth Aravamudan 3654bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3655adbe8726SEric B Munson return -EINVAL; 3656adbe8726SEric B Munson 36573dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3658a3437870SNishanth Aravamudan if (err) 365973ae31e5SEric B Munson return err; 3660a3437870SNishanth Aravamudan 3661db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3662a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3663db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3664a3437870SNishanth Aravamudan 3665a3437870SNishanth Aravamudan return count; 3666a3437870SNishanth Aravamudan } 3667a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3668a3437870SNishanth Aravamudan 3669a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3670a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3671a3437870SNishanth Aravamudan { 36729a305230SLee Schermerhorn struct hstate *h; 36739a305230SLee Schermerhorn unsigned long free_huge_pages; 36749a305230SLee Schermerhorn int nid; 36759a305230SLee Schermerhorn 36769a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36779a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36789a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36799a305230SLee Schermerhorn else 36809a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36819a305230SLee Schermerhorn 3682ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3683a3437870SNishanth Aravamudan } 3684a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3685a3437870SNishanth Aravamudan 3686a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3687a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3688a3437870SNishanth Aravamudan { 36899a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3690ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3691a3437870SNishanth Aravamudan } 3692a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3693a3437870SNishanth Aravamudan 3694a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3695a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3696a3437870SNishanth Aravamudan { 36979a305230SLee Schermerhorn struct hstate *h; 36989a305230SLee Schermerhorn unsigned long surplus_huge_pages; 36999a305230SLee Schermerhorn int nid; 37009a305230SLee Schermerhorn 37019a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37029a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37039a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 37049a305230SLee Schermerhorn else 37059a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 37069a305230SLee Schermerhorn 3707ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3708a3437870SNishanth Aravamudan } 3709a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3710a3437870SNishanth Aravamudan 371179dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 371279dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 371379dfc695SMike Kravetz { 371479dfc695SMike Kravetz unsigned long nr_demote; 371579dfc695SMike Kravetz unsigned long nr_available; 371679dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 371779dfc695SMike Kravetz struct hstate *h; 371879dfc695SMike Kravetz int err = 0; 371979dfc695SMike Kravetz int nid; 372079dfc695SMike Kravetz 372179dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 372279dfc695SMike Kravetz if (err) 372379dfc695SMike Kravetz return err; 372479dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 372579dfc695SMike Kravetz 372679dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 372779dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 372879dfc695SMike Kravetz n_mask = &nodes_allowed; 372979dfc695SMike Kravetz } else { 373079dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 373179dfc695SMike Kravetz } 373279dfc695SMike Kravetz 373379dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 373479dfc695SMike Kravetz mutex_lock(&h->resize_lock); 373579dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 373679dfc695SMike Kravetz 373779dfc695SMike Kravetz while (nr_demote) { 373879dfc695SMike Kravetz /* 373979dfc695SMike Kravetz * Check for available pages to demote each time thorough the 374079dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 374179dfc695SMike Kravetz */ 374279dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 374379dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 374479dfc695SMike Kravetz else 374579dfc695SMike Kravetz nr_available = h->free_huge_pages; 374679dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 374779dfc695SMike Kravetz if (!nr_available) 374879dfc695SMike Kravetz break; 374979dfc695SMike Kravetz 375079dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 375179dfc695SMike Kravetz if (err) 375279dfc695SMike Kravetz break; 375379dfc695SMike Kravetz 375479dfc695SMike Kravetz nr_demote--; 375579dfc695SMike Kravetz } 375679dfc695SMike Kravetz 375779dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 375879dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 375979dfc695SMike Kravetz 376079dfc695SMike Kravetz if (err) 376179dfc695SMike Kravetz return err; 376279dfc695SMike Kravetz return len; 376379dfc695SMike Kravetz } 376479dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 376579dfc695SMike Kravetz 376679dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 376779dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 376879dfc695SMike Kravetz { 376979dfc695SMike Kravetz int nid; 377079dfc695SMike Kravetz struct hstate *h = kobj_to_hstate(kobj, &nid); 377179dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 377279dfc695SMike Kravetz 377379dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 377479dfc695SMike Kravetz } 377579dfc695SMike Kravetz 377679dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 377779dfc695SMike Kravetz struct kobj_attribute *attr, 377879dfc695SMike Kravetz const char *buf, size_t count) 377979dfc695SMike Kravetz { 378079dfc695SMike Kravetz struct hstate *h, *demote_hstate; 378179dfc695SMike Kravetz unsigned long demote_size; 378279dfc695SMike Kravetz unsigned int demote_order; 378379dfc695SMike Kravetz int nid; 378479dfc695SMike Kravetz 378579dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 378679dfc695SMike Kravetz 378779dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 378879dfc695SMike Kravetz if (!demote_hstate) 378979dfc695SMike Kravetz return -EINVAL; 379079dfc695SMike Kravetz demote_order = demote_hstate->order; 3791a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3792a01f4390SMike Kravetz return -EINVAL; 379379dfc695SMike Kravetz 379479dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 379579dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 379679dfc695SMike Kravetz if (demote_order >= h->order) 379779dfc695SMike Kravetz return -EINVAL; 379879dfc695SMike Kravetz 379979dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 380079dfc695SMike Kravetz mutex_lock(&h->resize_lock); 380179dfc695SMike Kravetz h->demote_order = demote_order; 380279dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 380379dfc695SMike Kravetz 380479dfc695SMike Kravetz return count; 380579dfc695SMike Kravetz } 380679dfc695SMike Kravetz HSTATE_ATTR(demote_size); 380779dfc695SMike Kravetz 3808a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3809a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3810a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3811a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3812a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3813a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 381406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 381506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 381606808b08SLee Schermerhorn #endif 3817a3437870SNishanth Aravamudan NULL, 3818a3437870SNishanth Aravamudan }; 3819a3437870SNishanth Aravamudan 382067e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3821a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3822a3437870SNishanth Aravamudan }; 3823a3437870SNishanth Aravamudan 382479dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 382579dfc695SMike Kravetz &demote_size_attr.attr, 382679dfc695SMike Kravetz &demote_attr.attr, 382779dfc695SMike Kravetz NULL, 382879dfc695SMike Kravetz }; 382979dfc695SMike Kravetz 383079dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 383179dfc695SMike Kravetz .attrs = hstate_demote_attrs, 383279dfc695SMike Kravetz }; 383379dfc695SMike Kravetz 3834094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38359a305230SLee Schermerhorn struct kobject **hstate_kobjs, 383667e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3837a3437870SNishanth Aravamudan { 3838a3437870SNishanth Aravamudan int retval; 3839972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3840a3437870SNishanth Aravamudan 38419a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38429a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3843a3437870SNishanth Aravamudan return -ENOMEM; 3844a3437870SNishanth Aravamudan 38459a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3846cc2205a6SMiaohe Lin if (retval) { 38479a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3848cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 38493a6bdda0SMiaohe Lin return retval; 3850cc2205a6SMiaohe Lin } 3851a3437870SNishanth Aravamudan 385279dfc695SMike Kravetz if (h->demote_order) { 385301088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 385401088a60SMiaohe Lin &hstate_demote_attr_group); 385501088a60SMiaohe Lin if (retval) { 385679dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 385701088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 385801088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 385901088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 386001088a60SMiaohe Lin return retval; 386101088a60SMiaohe Lin } 386279dfc695SMike Kravetz } 386379dfc695SMike Kravetz 386401088a60SMiaohe Lin return 0; 3865a3437870SNishanth Aravamudan } 3866a3437870SNishanth Aravamudan 3867a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3868a3437870SNishanth Aravamudan { 3869a3437870SNishanth Aravamudan struct hstate *h; 3870a3437870SNishanth Aravamudan int err; 3871a3437870SNishanth Aravamudan 3872a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3873a3437870SNishanth Aravamudan if (!hugepages_kobj) 3874a3437870SNishanth Aravamudan return; 3875a3437870SNishanth Aravamudan 3876a3437870SNishanth Aravamudan for_each_hstate(h) { 38779a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 38789a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3879a3437870SNishanth Aravamudan if (err) 3880282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3881a3437870SNishanth Aravamudan } 3882a3437870SNishanth Aravamudan } 3883a3437870SNishanth Aravamudan 38849a305230SLee Schermerhorn #ifdef CONFIG_NUMA 38859a305230SLee Schermerhorn 38869a305230SLee Schermerhorn /* 38879a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 388810fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 388910fbcf4cSKay Sievers * index of a node device or _hstate == node id. 389010fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38919a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38929a305230SLee Schermerhorn */ 38939a305230SLee Schermerhorn struct node_hstate { 38949a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38959a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38969a305230SLee Schermerhorn }; 3897b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38989a305230SLee Schermerhorn 38999a305230SLee Schermerhorn /* 390010fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 39019a305230SLee Schermerhorn */ 39029a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 39039a305230SLee Schermerhorn &nr_hugepages_attr.attr, 39049a305230SLee Schermerhorn &free_hugepages_attr.attr, 39059a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 39069a305230SLee Schermerhorn NULL, 39079a305230SLee Schermerhorn }; 39089a305230SLee Schermerhorn 390967e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 39109a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 39119a305230SLee Schermerhorn }; 39129a305230SLee Schermerhorn 39139a305230SLee Schermerhorn /* 391410fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 39159a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 39169a305230SLee Schermerhorn */ 39179a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39189a305230SLee Schermerhorn { 39199a305230SLee Schermerhorn int nid; 39209a305230SLee Schermerhorn 39219a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 39229a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39239a305230SLee Schermerhorn int i; 39249a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39259a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39269a305230SLee Schermerhorn if (nidp) 39279a305230SLee Schermerhorn *nidp = nid; 39289a305230SLee Schermerhorn return &hstates[i]; 39299a305230SLee Schermerhorn } 39309a305230SLee Schermerhorn } 39319a305230SLee Schermerhorn 39329a305230SLee Schermerhorn BUG(); 39339a305230SLee Schermerhorn return NULL; 39349a305230SLee Schermerhorn } 39359a305230SLee Schermerhorn 39369a305230SLee Schermerhorn /* 393710fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39389a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39399a305230SLee Schermerhorn */ 39403cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 39419a305230SLee Schermerhorn { 39429a305230SLee Schermerhorn struct hstate *h; 394310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39449a305230SLee Schermerhorn 39459a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39469b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39479a305230SLee Schermerhorn 3948972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3949972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 395001088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 395101088a60SMiaohe Lin 395201088a60SMiaohe Lin if (!hstate_kobj) 395301088a60SMiaohe Lin continue; 395401088a60SMiaohe Lin if (h->demote_order) 395501088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 395601088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 395701088a60SMiaohe Lin kobject_put(hstate_kobj); 3958972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3959972dc4deSAneesh Kumar K.V } 39609a305230SLee Schermerhorn 39619a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39629a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39639a305230SLee Schermerhorn } 39649a305230SLee Schermerhorn 39659a305230SLee Schermerhorn 39669a305230SLee Schermerhorn /* 396710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39689a305230SLee Schermerhorn * No-op if attributes already registered. 39699a305230SLee Schermerhorn */ 39703cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 39719a305230SLee Schermerhorn { 39729a305230SLee Schermerhorn struct hstate *h; 397310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39749a305230SLee Schermerhorn int err; 39759a305230SLee Schermerhorn 39769a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39779a305230SLee Schermerhorn return; /* already allocated */ 39789a305230SLee Schermerhorn 39799a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 398010fbcf4cSKay Sievers &node->dev.kobj); 39819a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39829a305230SLee Schermerhorn return; 39839a305230SLee Schermerhorn 39849a305230SLee Schermerhorn for_each_hstate(h) { 39859a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39869a305230SLee Schermerhorn nhs->hstate_kobjs, 39879a305230SLee Schermerhorn &per_node_hstate_attr_group); 39889a305230SLee Schermerhorn if (err) { 3989282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 399010fbcf4cSKay Sievers h->name, node->dev.id); 39919a305230SLee Schermerhorn hugetlb_unregister_node(node); 39929a305230SLee Schermerhorn break; 39939a305230SLee Schermerhorn } 39949a305230SLee Schermerhorn } 39959a305230SLee Schermerhorn } 39969a305230SLee Schermerhorn 39979a305230SLee Schermerhorn /* 39989b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 399910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 400010fbcf4cSKay Sievers * registered their associated device by this time. 40019a305230SLee Schermerhorn */ 40027d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 40039a305230SLee Schermerhorn { 40049a305230SLee Schermerhorn int nid; 40059a305230SLee Schermerhorn 40068cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 40078732794bSWen Congyang struct node *node = node_devices[nid]; 400810fbcf4cSKay Sievers if (node->dev.id == nid) 40099a305230SLee Schermerhorn hugetlb_register_node(node); 40109a305230SLee Schermerhorn } 40119a305230SLee Schermerhorn 40129a305230SLee Schermerhorn /* 401310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 40149a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 40159a305230SLee Schermerhorn */ 40169a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 40179a305230SLee Schermerhorn hugetlb_unregister_node); 40189a305230SLee Schermerhorn } 40199a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 40209a305230SLee Schermerhorn 40219a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40229a305230SLee Schermerhorn { 40239a305230SLee Schermerhorn BUG(); 40249a305230SLee Schermerhorn if (nidp) 40259a305230SLee Schermerhorn *nidp = -1; 40269a305230SLee Schermerhorn return NULL; 40279a305230SLee Schermerhorn } 40289a305230SLee Schermerhorn 40299a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40309a305230SLee Schermerhorn 40319a305230SLee Schermerhorn #endif 40329a305230SLee Schermerhorn 4033263b8998SMiaohe Lin #ifdef CONFIG_CMA 4034263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4035263b8998SMiaohe Lin #else 4036263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4037263b8998SMiaohe Lin { 4038263b8998SMiaohe Lin } 4039263b8998SMiaohe Lin #endif 4040263b8998SMiaohe Lin 4041a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4042a3437870SNishanth Aravamudan { 40438382d914SDavidlohr Bueso int i; 40448382d914SDavidlohr Bueso 4045d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4046d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4047d6995da3SMike Kravetz 4048c2833a5bSMike Kravetz if (!hugepages_supported()) { 4049c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4050c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40510ef89d25SBenjamin Herrenschmidt return 0; 4052d715cf80SLiam R. Howlett } 4053d715cf80SLiam R. Howlett 4054282f4214SMike Kravetz /* 4055282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4056282f4214SMike Kravetz * architectures depend on setup being done here. 4057282f4214SMike Kravetz */ 4058a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4059282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4060282f4214SMike Kravetz /* 4061282f4214SMike Kravetz * If we did not parse a default huge page size, set 4062282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4063282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4064282f4214SMike Kravetz * specified, set that here as well. 4065282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4066282f4214SMike Kravetz * setting. A warning will be printed in this case. 4067282f4214SMike Kravetz */ 4068282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4069f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4070282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4071282f4214SMike Kravetz char buf[32]; 4072282f4214SMike Kravetz 4073282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4074282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4075282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4076282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4077282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4078282f4214SMike Kravetz default_hstate_max_huge_pages); 4079282f4214SMike Kravetz } 4080282f4214SMike Kravetz default_hstate.max_huge_pages = 4081282f4214SMike Kravetz default_hstate_max_huge_pages; 4082b5389086SZhenguo Yao 40830a7a0f6fSPeng Liu for_each_online_node(i) 4084b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4085b5389086SZhenguo Yao default_hugepages_in_node[i]; 4086282f4214SMike Kravetz } 4087f8b74815SVaishali Thakkar } 4088a3437870SNishanth Aravamudan 4089cf11e85fSRoman Gushchin hugetlb_cma_check(); 4090a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4091aa888a74SAndi Kleen gather_bootmem_prealloc(); 4092a3437870SNishanth Aravamudan report_hugepages(); 4093a3437870SNishanth Aravamudan 4094a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 40959a305230SLee Schermerhorn hugetlb_register_all_nodes(); 40967179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 40979a305230SLee Schermerhorn 40988382d914SDavidlohr Bueso #ifdef CONFIG_SMP 40998382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 41008382d914SDavidlohr Bueso #else 41018382d914SDavidlohr Bueso num_fault_mutexes = 1; 41028382d914SDavidlohr Bueso #endif 4103c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 41046da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 41056da2ec56SKees Cook GFP_KERNEL); 4106c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 41078382d914SDavidlohr Bueso 41088382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4109c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4110a3437870SNishanth Aravamudan return 0; 4111a3437870SNishanth Aravamudan } 41123e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4113a3437870SNishanth Aravamudan 4114ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4115ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 41169fee021dSVaishali Thakkar { 4117ae94da89SMike Kravetz return size == HPAGE_SIZE; 41189fee021dSVaishali Thakkar } 41199fee021dSVaishali Thakkar 4120d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4121a3437870SNishanth Aravamudan { 4122a3437870SNishanth Aravamudan struct hstate *h; 41238faa8b07SAndi Kleen unsigned long i; 41248faa8b07SAndi Kleen 4125a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4126a3437870SNishanth Aravamudan return; 4127a3437870SNishanth Aravamudan } 412847d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4129a3437870SNishanth Aravamudan BUG_ON(order == 0); 413047d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 413129383967SMike Kravetz mutex_init(&h->resize_lock); 4132a3437870SNishanth Aravamudan h->order = order; 4133aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 41348faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 41358faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41360edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 413754f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 413854f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4139a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4140*c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 41418faa8b07SAndi Kleen 4142a3437870SNishanth Aravamudan parsed_hstate = h; 4143a3437870SNishanth Aravamudan } 4144a3437870SNishanth Aravamudan 4145b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4146b5389086SZhenguo Yao { 4147b5389086SZhenguo Yao return true; 4148b5389086SZhenguo Yao } 4149f87442f4SPeng Liu 4150f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4151f87442f4SPeng Liu { 4152f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4153f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4154f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 4155f87442f4SPeng Liu MAX_NUMNODES * sizeof(unsigned int)); 4156f87442f4SPeng Liu } else { 4157f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4158f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 4159f87442f4SPeng Liu MAX_NUMNODES * sizeof(unsigned int)); 4160f87442f4SPeng Liu } 4161f87442f4SPeng Liu } 4162f87442f4SPeng Liu 4163282f4214SMike Kravetz /* 4164282f4214SMike Kravetz * hugepages command line processing 4165282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4166282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4167282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4168282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4169282f4214SMike Kravetz */ 4170282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4171a3437870SNishanth Aravamudan { 4172a3437870SNishanth Aravamudan unsigned long *mhp; 41738faa8b07SAndi Kleen static unsigned long *last_mhp; 4174b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4175b5389086SZhenguo Yao int count; 4176b5389086SZhenguo Yao unsigned long tmp; 4177b5389086SZhenguo Yao char *p = s; 4178a3437870SNishanth Aravamudan 41799fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4180282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41819fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4182f81f6e4bSPeng Liu return 1; 41839fee021dSVaishali Thakkar } 4184282f4214SMike Kravetz 4185a3437870SNishanth Aravamudan /* 4186282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4187282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4188282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4189282f4214SMike Kravetz * default_hugepagesz. 4190a3437870SNishanth Aravamudan */ 41919fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4192a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4193a3437870SNishanth Aravamudan else 4194a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4195a3437870SNishanth Aravamudan 41968faa8b07SAndi Kleen if (mhp == last_mhp) { 4197282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4198f81f6e4bSPeng Liu return 1; 41998faa8b07SAndi Kleen } 42008faa8b07SAndi Kleen 4201b5389086SZhenguo Yao while (*p) { 4202b5389086SZhenguo Yao count = 0; 4203b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4204b5389086SZhenguo Yao goto invalid; 4205b5389086SZhenguo Yao /* Parameter is node format */ 4206b5389086SZhenguo Yao if (p[count] == ':') { 4207b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4208b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4209f81f6e4bSPeng Liu return 1; 4210b5389086SZhenguo Yao } 42110a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4212e79ce983SLiu Yuntao goto invalid; 42130a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4214b5389086SZhenguo Yao p += count + 1; 4215b5389086SZhenguo Yao /* Parse hugepages */ 4216b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4217b5389086SZhenguo Yao goto invalid; 4218b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4219b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4220b5389086SZhenguo Yao else 4221b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4222b5389086SZhenguo Yao *mhp += tmp; 4223b5389086SZhenguo Yao /* Go to parse next node*/ 4224b5389086SZhenguo Yao if (p[count] == ',') 4225b5389086SZhenguo Yao p += count + 1; 4226b5389086SZhenguo Yao else 4227b5389086SZhenguo Yao break; 4228b5389086SZhenguo Yao } else { 4229b5389086SZhenguo Yao if (p != s) 4230b5389086SZhenguo Yao goto invalid; 4231b5389086SZhenguo Yao *mhp = tmp; 4232b5389086SZhenguo Yao break; 4233b5389086SZhenguo Yao } 4234b5389086SZhenguo Yao } 4235a3437870SNishanth Aravamudan 42368faa8b07SAndi Kleen /* 42378faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 423804adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 42398faa8b07SAndi Kleen * use the bootmem allocator. 42408faa8b07SAndi Kleen */ 424104adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 42428faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 42438faa8b07SAndi Kleen 42448faa8b07SAndi Kleen last_mhp = mhp; 42458faa8b07SAndi Kleen 4246a3437870SNishanth Aravamudan return 1; 4247b5389086SZhenguo Yao 4248b5389086SZhenguo Yao invalid: 4249b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4250f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4251f81f6e4bSPeng Liu return 1; 4252a3437870SNishanth Aravamudan } 4253282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4254e11bfbfcSNick Piggin 4255282f4214SMike Kravetz /* 4256282f4214SMike Kravetz * hugepagesz command line processing 4257282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4258282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4259282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4260282f4214SMike Kravetz * hugepagesz argument was valid. 4261282f4214SMike Kravetz */ 4262359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4263e11bfbfcSNick Piggin { 4264359f2544SMike Kravetz unsigned long size; 4265282f4214SMike Kravetz struct hstate *h; 4266282f4214SMike Kravetz 4267282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4268359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4269359f2544SMike Kravetz 4270359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4271282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4272f81f6e4bSPeng Liu return 1; 4273359f2544SMike Kravetz } 4274359f2544SMike Kravetz 4275282f4214SMike Kravetz h = size_to_hstate(size); 4276282f4214SMike Kravetz if (h) { 4277282f4214SMike Kravetz /* 4278282f4214SMike Kravetz * hstate for this size already exists. This is normally 4279282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4280282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4281282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4282282f4214SMike Kravetz * previously specified. 4283282f4214SMike Kravetz */ 4284282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4285282f4214SMike Kravetz default_hstate.max_huge_pages) { 4286282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4287f81f6e4bSPeng Liu return 1; 428838237830SMike Kravetz } 428938237830SMike Kravetz 4290282f4214SMike Kravetz /* 4291282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4292282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4293282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4294282f4214SMike Kravetz */ 4295282f4214SMike Kravetz parsed_hstate = h; 4296282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4297e11bfbfcSNick Piggin return 1; 4298e11bfbfcSNick Piggin } 4299282f4214SMike Kravetz 4300359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4301282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4302359f2544SMike Kravetz return 1; 4303359f2544SMike Kravetz } 4304359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4305359f2544SMike Kravetz 4306282f4214SMike Kravetz /* 4307282f4214SMike Kravetz * default_hugepagesz command line input 4308282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4309282f4214SMike Kravetz */ 4310ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4311e11bfbfcSNick Piggin { 4312ae94da89SMike Kravetz unsigned long size; 4313b5389086SZhenguo Yao int i; 4314ae94da89SMike Kravetz 4315282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4316282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4317282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4318f81f6e4bSPeng Liu return 1; 4319282f4214SMike Kravetz } 4320282f4214SMike Kravetz 4321282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4322282f4214SMike Kravetz 4323282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4324282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4325f81f6e4bSPeng Liu return 1; 4326282f4214SMike Kravetz } 4327282f4214SMike Kravetz 4328282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4329282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4330282f4214SMike Kravetz parsed_default_hugepagesz = true; 4331282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4332282f4214SMike Kravetz 4333282f4214SMike Kravetz /* 4334282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4335282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4336282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4337282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4338282f4214SMike Kravetz * allocated here from bootmem allocator. 4339282f4214SMike Kravetz */ 4340282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4341282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 43420a7a0f6fSPeng Liu for_each_online_node(i) 4343b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4344b5389086SZhenguo Yao default_hugepages_in_node[i]; 4345282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4346282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4347282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4348282f4214SMike Kravetz } 4349282f4214SMike Kravetz 4350e11bfbfcSNick Piggin return 1; 4351e11bfbfcSNick Piggin } 4352ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4353a3437870SNishanth Aravamudan 4354d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4355d2226ebdSFeng Tang { 4356d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4357d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4358d2226ebdSFeng Tang 4359d2226ebdSFeng Tang /* 4360d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4361d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4362d2226ebdSFeng Tang */ 4363d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4364d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4365d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4366d2226ebdSFeng Tang return &mpol->nodes; 4367d2226ebdSFeng Tang #endif 4368d2226ebdSFeng Tang return NULL; 4369d2226ebdSFeng Tang } 4370d2226ebdSFeng Tang 43718ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43728a213460SNishanth Aravamudan { 43738a213460SNishanth Aravamudan int node; 43748a213460SNishanth Aravamudan unsigned int nr = 0; 4375d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 43768ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43778ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43788a213460SNishanth Aravamudan 4379d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 43808ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4381d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 43828a213460SNishanth Aravamudan nr += array[node]; 43838ca39e68SMuchun Song } 43848a213460SNishanth Aravamudan 43858a213460SNishanth Aravamudan return nr; 43868a213460SNishanth Aravamudan } 43878a213460SNishanth Aravamudan 43888a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 438917743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 439017743798SMuchun Song void *buffer, size_t *length, 439117743798SMuchun Song loff_t *ppos, unsigned long *out) 439217743798SMuchun Song { 439317743798SMuchun Song struct ctl_table dup_table; 439417743798SMuchun Song 439517743798SMuchun Song /* 439617743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 439717743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 439817743798SMuchun Song */ 439917743798SMuchun Song dup_table = *table; 440017743798SMuchun Song dup_table.data = out; 440117743798SMuchun Song 440217743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 440317743798SMuchun Song } 440417743798SMuchun Song 440506808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 440606808b08SLee Schermerhorn struct ctl_table *table, int write, 440732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 44081da177e4SLinus Torvalds { 4409e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4410238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 441108d4a246SMichal Hocko int ret; 4412e5ff2159SAndi Kleen 4413457c1b27SNishanth Aravamudan if (!hugepages_supported()) 441486613628SJan Stancek return -EOPNOTSUPP; 4415457c1b27SNishanth Aravamudan 441617743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 441717743798SMuchun Song &tmp); 441808d4a246SMichal Hocko if (ret) 441908d4a246SMichal Hocko goto out; 4420e5ff2159SAndi Kleen 4421238d3c13SDavid Rientjes if (write) 4422238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4423238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 442408d4a246SMichal Hocko out: 442508d4a246SMichal Hocko return ret; 44261da177e4SLinus Torvalds } 4427396faf03SMel Gorman 442806808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 442932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 443006808b08SLee Schermerhorn { 443106808b08SLee Schermerhorn 443206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 443306808b08SLee Schermerhorn buffer, length, ppos); 443406808b08SLee Schermerhorn } 443506808b08SLee Schermerhorn 443606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 443706808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 443832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 443906808b08SLee Schermerhorn { 444006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 444106808b08SLee Schermerhorn buffer, length, ppos); 444206808b08SLee Schermerhorn } 444306808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 444406808b08SLee Schermerhorn 4445a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 444632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4447a3d0c6aaSNishanth Aravamudan { 4448a5516438SAndi Kleen struct hstate *h = &default_hstate; 4449e5ff2159SAndi Kleen unsigned long tmp; 445008d4a246SMichal Hocko int ret; 4451e5ff2159SAndi Kleen 4452457c1b27SNishanth Aravamudan if (!hugepages_supported()) 445386613628SJan Stancek return -EOPNOTSUPP; 4454457c1b27SNishanth Aravamudan 4455e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4456e5ff2159SAndi Kleen 4457bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4458adbe8726SEric B Munson return -EINVAL; 4459adbe8726SEric B Munson 446017743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 446117743798SMuchun Song &tmp); 446208d4a246SMichal Hocko if (ret) 446308d4a246SMichal Hocko goto out; 4464e5ff2159SAndi Kleen 4465e5ff2159SAndi Kleen if (write) { 4466db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4467e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4468db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4469e5ff2159SAndi Kleen } 447008d4a246SMichal Hocko out: 447108d4a246SMichal Hocko return ret; 4472a3d0c6aaSNishanth Aravamudan } 4473a3d0c6aaSNishanth Aravamudan 44741da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44751da177e4SLinus Torvalds 4476e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44771da177e4SLinus Torvalds { 4478fcb2b0c5SRoman Gushchin struct hstate *h; 4479fcb2b0c5SRoman Gushchin unsigned long total = 0; 4480fcb2b0c5SRoman Gushchin 4481457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4482457c1b27SNishanth Aravamudan return; 4483fcb2b0c5SRoman Gushchin 4484fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4485fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4486fcb2b0c5SRoman Gushchin 4487aca78307SMiaohe Lin total += huge_page_size(h) * count; 4488fcb2b0c5SRoman Gushchin 4489fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4490e1759c21SAlexey Dobriyan seq_printf(m, 44911da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 44921da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4493b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 44947893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 44954f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4496fcb2b0c5SRoman Gushchin count, 4497a5516438SAndi Kleen h->free_huge_pages, 4498a5516438SAndi Kleen h->resv_huge_pages, 4499a5516438SAndi Kleen h->surplus_huge_pages, 4500aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4501fcb2b0c5SRoman Gushchin } 4502fcb2b0c5SRoman Gushchin 4503aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 45041da177e4SLinus Torvalds } 45051da177e4SLinus Torvalds 45067981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 45071da177e4SLinus Torvalds { 4508a5516438SAndi Kleen struct hstate *h = &default_hstate; 45097981593bSJoe Perches 4510457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4511457c1b27SNishanth Aravamudan return 0; 45127981593bSJoe Perches 45137981593bSJoe Perches return sysfs_emit_at(buf, len, 45141da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4515a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4516a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4517a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4518a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4519a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 45201da177e4SLinus Torvalds } 45211da177e4SLinus Torvalds 4522dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4523949f7ec5SDavid Rientjes { 4524949f7ec5SDavid Rientjes struct hstate *h; 4525949f7ec5SDavid Rientjes 4526457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4527457c1b27SNishanth Aravamudan return; 4528457c1b27SNishanth Aravamudan 4529949f7ec5SDavid Rientjes for_each_hstate(h) 4530dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4531949f7ec5SDavid Rientjes nid, 4532949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4533949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4534949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4535aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4536949f7ec5SDavid Rientjes } 4537949f7ec5SDavid Rientjes 45385d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 45395d317b2bSNaoya Horiguchi { 45405d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 45415d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 45425d317b2bSNaoya Horiguchi } 45435d317b2bSNaoya Horiguchi 45441da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 45451da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 45461da177e4SLinus Torvalds { 4547d0028588SWanpeng Li struct hstate *h; 4548d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4549d0028588SWanpeng Li 4550d0028588SWanpeng Li for_each_hstate(h) 4551d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4552d0028588SWanpeng Li return nr_total_pages; 45531da177e4SLinus Torvalds } 45541da177e4SLinus Torvalds 4555a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4556fc1b8a73SMel Gorman { 4557fc1b8a73SMel Gorman int ret = -ENOMEM; 4558fc1b8a73SMel Gorman 45590aa7f354SMiaohe Lin if (!delta) 45600aa7f354SMiaohe Lin return 0; 45610aa7f354SMiaohe Lin 4562db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4563fc1b8a73SMel Gorman /* 4564fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4565fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4566fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4567fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4568fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4569fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4570fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4571fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4572fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4573fc1b8a73SMel Gorman * 4574fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4575fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4576fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4577fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4578fc1b8a73SMel Gorman * semantics that cpuset has. 45798ca39e68SMuchun Song * 45808ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45818ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45828ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45838ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45848ca39e68SMuchun Song * above. 4585fc1b8a73SMel Gorman */ 4586fc1b8a73SMel Gorman if (delta > 0) { 4587a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4588fc1b8a73SMel Gorman goto out; 4589fc1b8a73SMel Gorman 45908ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4591a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4592fc1b8a73SMel Gorman goto out; 4593fc1b8a73SMel Gorman } 4594fc1b8a73SMel Gorman } 4595fc1b8a73SMel Gorman 4596fc1b8a73SMel Gorman ret = 0; 4597fc1b8a73SMel Gorman if (delta < 0) 4598a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4599fc1b8a73SMel Gorman 4600fc1b8a73SMel Gorman out: 4601db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4602fc1b8a73SMel Gorman return ret; 4603fc1b8a73SMel Gorman } 4604fc1b8a73SMel Gorman 460584afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 460684afd99bSAndy Whitcroft { 4607f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 460884afd99bSAndy Whitcroft 460984afd99bSAndy Whitcroft /* 461084afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 461184afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 461284afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 461325985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 461484afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 461584afd99bSAndy Whitcroft * new reference here without additional locking. 461684afd99bSAndy Whitcroft */ 461709a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 461809a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4619f522c3acSJoonsoo Kim kref_get(&resv->refs); 462084afd99bSAndy Whitcroft } 462109a26e83SMike Kravetz } 462284afd99bSAndy Whitcroft 4623a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4624a1e78772SMel Gorman { 4625a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4626f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 462790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 46284e35f483SJoonsoo Kim unsigned long reserve, start, end; 46291c5ecae3SMike Kravetz long gbl_reserve; 463084afd99bSAndy Whitcroft 46314e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 46324e35f483SJoonsoo Kim return; 46334e35f483SJoonsoo Kim 4634a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4635a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 463684afd99bSAndy Whitcroft 46374e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4638e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 46397251ff78SAdam Litke if (reserve) { 46401c5ecae3SMike Kravetz /* 46411c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 46421c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 46431c5ecae3SMike Kravetz */ 46441c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 46451c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 46467251ff78SAdam Litke } 4647e9fe92aeSMina Almasry 4648e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4649a1e78772SMel Gorman } 4650a1e78772SMel Gorman 465131383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 465231383c68SDan Williams { 465331383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 465431383c68SDan Williams return -EINVAL; 465531383c68SDan Williams return 0; 465631383c68SDan Williams } 465731383c68SDan Williams 465805ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 465905ea8860SDan Williams { 4660aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 466105ea8860SDan Williams } 466205ea8860SDan Williams 46631da177e4SLinus Torvalds /* 46641da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46651da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46666c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46671da177e4SLinus Torvalds * this far. 46681da177e4SLinus Torvalds */ 4669b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 46701da177e4SLinus Torvalds { 46711da177e4SLinus Torvalds BUG(); 4672d0217ac0SNick Piggin return 0; 46731da177e4SLinus Torvalds } 46741da177e4SLinus Torvalds 4675eec3636aSJane Chu /* 4676eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4677eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4678eec3636aSJane Chu * This is because under System V memory model, mappings created via 4679eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4680eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4681eec3636aSJane Chu */ 4682f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4683d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 468484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4685a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4686dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 468705ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 46881da177e4SLinus Torvalds }; 46891da177e4SLinus Torvalds 46901e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 46911e8f889bSDavid Gibson int writable) 469263551ae0SDavid Gibson { 469363551ae0SDavid Gibson pte_t entry; 469479c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 469563551ae0SDavid Gibson 46961e8f889bSDavid Gibson if (writable) { 4697106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4698106c992aSGerald Schaefer vma->vm_page_prot))); 469963551ae0SDavid Gibson } else { 4700106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4701106c992aSGerald Schaefer vma->vm_page_prot)); 470263551ae0SDavid Gibson } 470363551ae0SDavid Gibson entry = pte_mkyoung(entry); 470479c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 470563551ae0SDavid Gibson 470663551ae0SDavid Gibson return entry; 470763551ae0SDavid Gibson } 470863551ae0SDavid Gibson 47091e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 47101e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 47111e8f889bSDavid Gibson { 47121e8f889bSDavid Gibson pte_t entry; 47131e8f889bSDavid Gibson 4714106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 471532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 47164b3073e1SRussell King update_mmu_cache(vma, address, ptep); 47171e8f889bSDavid Gibson } 47181e8f889bSDavid Gibson 4719d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 47204a705fefSNaoya Horiguchi { 47214a705fefSNaoya Horiguchi swp_entry_t swp; 47224a705fefSNaoya Horiguchi 47234a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4724d5ed7444SAneesh Kumar K.V return false; 47254a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4726d79d176aSBaoquan He if (is_migration_entry(swp)) 4727d5ed7444SAneesh Kumar K.V return true; 47284a705fefSNaoya Horiguchi else 4729d5ed7444SAneesh Kumar K.V return false; 47304a705fefSNaoya Horiguchi } 47314a705fefSNaoya Horiguchi 47323e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 47334a705fefSNaoya Horiguchi { 47344a705fefSNaoya Horiguchi swp_entry_t swp; 47354a705fefSNaoya Horiguchi 47364a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 47373e5c3600SBaoquan He return false; 47384a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4739d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 47403e5c3600SBaoquan He return true; 47414a705fefSNaoya Horiguchi else 47423e5c3600SBaoquan He return false; 47434a705fefSNaoya Horiguchi } 47441e8f889bSDavid Gibson 47454eae4efaSPeter Xu static void 47464eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 47474eae4efaSPeter Xu struct page *new_page) 47484eae4efaSPeter Xu { 47494eae4efaSPeter Xu __SetPageUptodate(new_page); 47504eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 47511eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 47524eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 47534eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 47544eae4efaSPeter Xu SetHPageMigratable(new_page); 47554eae4efaSPeter Xu } 47564eae4efaSPeter Xu 475763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4758bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4759bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 476063551ae0SDavid Gibson { 47613aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 476263551ae0SDavid Gibson struct page *ptepage; 47631c59827dSHugh Dickins unsigned long addr; 4764bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4765bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4766a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47674eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4768bc70fbf2SPeter Xu struct address_space *mapping = src_vma->vm_file->f_mapping; 4769ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4770e95a9851SMike Kravetz unsigned long last_addr_mask; 4771e8569dd2SAndreas Sandberg int ret = 0; 47721e8f889bSDavid Gibson 4773ac46d4f3SJérôme Glisse if (cow) { 4774bc70fbf2SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, 4775bc70fbf2SPeter Xu src_vma->vm_start, 4776bc70fbf2SPeter Xu src_vma->vm_end); 4777ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4778623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4779623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 4780c0d0381aSMike Kravetz } else { 4781c0d0381aSMike Kravetz /* 4782c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4783c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4784c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4785c0d0381aSMike Kravetz * huge_pmd_unshare. 4786c0d0381aSMike Kravetz */ 4787c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4788ac46d4f3SJérôme Glisse } 4789e8569dd2SAndreas Sandberg 4790e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4791bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4792cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 47937868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4794e95a9851SMike Kravetz if (!src_pte) { 4795e95a9851SMike Kravetz addr |= last_addr_mask; 4796c74df32cSHugh Dickins continue; 4797e95a9851SMike Kravetz } 4798bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4799e8569dd2SAndreas Sandberg if (!dst_pte) { 4800e8569dd2SAndreas Sandberg ret = -ENOMEM; 4801e8569dd2SAndreas Sandberg break; 4802e8569dd2SAndreas Sandberg } 4803c5c99429SLarry Woodman 48045e41540cSMike Kravetz /* 48055e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 48065e41540cSMike Kravetz * 48073aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 48085e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 48093aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 48103aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 48115e41540cSMike Kravetz */ 48123aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 4813e95a9851SMike Kravetz addr |= last_addr_mask; 4814c5c99429SLarry Woodman continue; 4815e95a9851SMike Kravetz } 4816c5c99429SLarry Woodman 4817cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4818cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4819cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48204a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 48214eae4efaSPeter Xu again: 48223aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 48235e41540cSMike Kravetz /* 48243aa4ed80SMiaohe Lin * Skip if src entry none. 48255e41540cSMike Kravetz */ 48264a705fefSNaoya Horiguchi ; 4827c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 4828c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 4829c2cb0dccSNaoya Horiguchi 4830c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 4831c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 4832c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 4833c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 48344a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 4835bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 48364a705fefSNaoya Horiguchi 48376c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 48384a705fefSNaoya Horiguchi /* 48394a705fefSNaoya Horiguchi * COW mappings require pages in both 48404a705fefSNaoya Horiguchi * parent and child to be set to read. 48414a705fefSNaoya Horiguchi */ 48424dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 48434dd845b5SAlistair Popple swp_offset(swp_entry)); 48444a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4845bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 4846bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 484718f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 48484a705fefSNaoya Horiguchi } 4849bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 4850bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 485118f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 4852bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 4853bc70fbf2SPeter Xu /* 4854bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 4855bc70fbf2SPeter Xu * uffd-wp enabled. 4856bc70fbf2SPeter Xu */ 4857bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 4858bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 48594a705fefSNaoya Horiguchi } else { 48604eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48614eae4efaSPeter Xu ptepage = pte_page(entry); 48624eae4efaSPeter Xu get_page(ptepage); 48634eae4efaSPeter Xu 48644eae4efaSPeter Xu /* 4865fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 4866fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 4867fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 4868fb3d824dSDavid Hildenbrand * fork. 48694eae4efaSPeter Xu * 48704eae4efaSPeter Xu * When pre-allocating the page or copying data, we 48714eae4efaSPeter Xu * need to be without the pgtable locks since we could 48724eae4efaSPeter Xu * sleep during the process. 48734eae4efaSPeter Xu */ 4874fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 4875fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 4876bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 4877bc70fbf2SPeter Xu src_vma)) { 48784eae4efaSPeter Xu pte_t src_pte_old = entry; 48794eae4efaSPeter Xu struct page *new; 48804eae4efaSPeter Xu 48814eae4efaSPeter Xu spin_unlock(src_ptl); 48824eae4efaSPeter Xu spin_unlock(dst_ptl); 48834eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 4884bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 48854eae4efaSPeter Xu if (IS_ERR(new)) { 48864eae4efaSPeter Xu put_page(ptepage); 48874eae4efaSPeter Xu ret = PTR_ERR(new); 48884eae4efaSPeter Xu break; 48894eae4efaSPeter Xu } 4890bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 48914eae4efaSPeter Xu npages); 48924eae4efaSPeter Xu put_page(ptepage); 48934eae4efaSPeter Xu 48944eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 48954eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 48964eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 48974eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48984eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48994eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4900bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 4901846be085SMike Kravetz new); 49024eae4efaSPeter Xu put_page(new); 49033aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 49044eae4efaSPeter Xu goto again; 49054eae4efaSPeter Xu } 4906bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 49074eae4efaSPeter Xu spin_unlock(src_ptl); 49084eae4efaSPeter Xu spin_unlock(dst_ptl); 49094eae4efaSPeter Xu continue; 49104eae4efaSPeter Xu } 49114eae4efaSPeter Xu 491234ee645eSJoerg Roedel if (cow) { 49130f10851eSJérôme Glisse /* 49140f10851eSJérôme Glisse * No need to notify as we are downgrading page 49150f10851eSJérôme Glisse * table protection not changing it to point 49160f10851eSJérôme Glisse * to a new page. 49170f10851eSJérôme Glisse * 4918ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 49190f10851eSJérôme Glisse */ 49207f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 492184894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 492234ee645eSJoerg Roedel } 49234eae4efaSPeter Xu 492463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 49254eae4efaSPeter Xu hugetlb_count_add(npages, dst); 49261c59827dSHugh Dickins } 4927cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4928cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 492963551ae0SDavid Gibson } 493063551ae0SDavid Gibson 4931623a1ddfSDavid Hildenbrand if (cow) { 4932623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 4933ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4934623a1ddfSDavid Hildenbrand } else { 4935c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4936623a1ddfSDavid Hildenbrand } 4937e8569dd2SAndreas Sandberg 4938e8569dd2SAndreas Sandberg return ret; 493963551ae0SDavid Gibson } 494063551ae0SDavid Gibson 4941550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4942db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4943550a7d60SMina Almasry { 4944550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4945550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4946550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4947db110a99SAneesh Kumar K.V pte_t pte; 4948550a7d60SMina Almasry 4949550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4950550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4951550a7d60SMina Almasry 4952550a7d60SMina Almasry /* 4953550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4954550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4955550a7d60SMina Almasry */ 4956550a7d60SMina Almasry if (src_ptl != dst_ptl) 4957550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4958550a7d60SMina Almasry 4959550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4960550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4961550a7d60SMina Almasry 4962550a7d60SMina Almasry if (src_ptl != dst_ptl) 4963550a7d60SMina Almasry spin_unlock(src_ptl); 4964550a7d60SMina Almasry spin_unlock(dst_ptl); 4965550a7d60SMina Almasry } 4966550a7d60SMina Almasry 4967550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4968550a7d60SMina Almasry struct vm_area_struct *new_vma, 4969550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4970550a7d60SMina Almasry unsigned long len) 4971550a7d60SMina Almasry { 4972550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4973550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4974550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4975550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4976550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4977e95a9851SMike Kravetz unsigned long last_addr_mask; 4978550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4979550a7d60SMina Almasry struct mmu_notifier_range range; 49803d0b95cdSBaolin Wang bool shared_pmd = false; 4981550a7d60SMina Almasry 4982550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 4983550a7d60SMina Almasry old_end); 4984550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 49853d0b95cdSBaolin Wang /* 49863d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 49873d0b95cdSBaolin Wang * range. 49883d0b95cdSBaolin Wang */ 49893d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 49903d0b95cdSBaolin Wang 4991550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 4992e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4993550a7d60SMina Almasry /* Prevent race with file truncation */ 4994550a7d60SMina Almasry i_mmap_lock_write(mapping); 4995550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 4996550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 4997e95a9851SMike Kravetz if (!src_pte) { 4998e95a9851SMike Kravetz old_addr |= last_addr_mask; 4999e95a9851SMike Kravetz new_addr |= last_addr_mask; 5000550a7d60SMina Almasry continue; 5001e95a9851SMike Kravetz } 5002550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5003550a7d60SMina Almasry continue; 5004550a7d60SMina Almasry 50054ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 50063d0b95cdSBaolin Wang shared_pmd = true; 50074ddb4d91SMike Kravetz old_addr |= last_addr_mask; 50084ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5009550a7d60SMina Almasry continue; 50103d0b95cdSBaolin Wang } 5011550a7d60SMina Almasry 5012550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5013550a7d60SMina Almasry if (!dst_pte) 5014550a7d60SMina Almasry break; 5015550a7d60SMina Almasry 5016db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5017550a7d60SMina Almasry } 50183d0b95cdSBaolin Wang 50193d0b95cdSBaolin Wang if (shared_pmd) 50203d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 50213d0b95cdSBaolin Wang else 5022550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5023550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 502413e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 5025550a7d60SMina Almasry 5026550a7d60SMina Almasry return len + old_addr - old_end; 5027550a7d60SMina Almasry } 5028550a7d60SMina Almasry 502973c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 503024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 503105e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 503263551ae0SDavid Gibson { 503363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 503463551ae0SDavid Gibson unsigned long address; 5035c7546f8fSDavid Gibson pte_t *ptep; 503663551ae0SDavid Gibson pte_t pte; 5037cb900f41SKirill A. Shutemov spinlock_t *ptl; 503863551ae0SDavid Gibson struct page *page; 5039a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5040a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5041ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5042e95a9851SMike Kravetz unsigned long last_addr_mask; 5043a4a118f2SNadav Amit bool force_flush = false; 5044a5516438SAndi Kleen 504563551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5046a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5047a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 504863551ae0SDavid Gibson 504907e32661SAneesh Kumar K.V /* 505007e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 505107e32661SAneesh Kumar K.V * to huge page. 505207e32661SAneesh Kumar K.V */ 5053ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 505424669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5055dff11abeSMike Kravetz 5056dff11abeSMike Kravetz /* 5057dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 5058dff11abeSMike Kravetz */ 50596f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 50606f4f13e8SJérôme Glisse end); 5061ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5062ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5063e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5064569f48b8SHillf Danton address = start; 5065569f48b8SHillf Danton for (; address < end; address += sz) { 50667868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 5067e95a9851SMike Kravetz if (!ptep) { 5068e95a9851SMike Kravetz address |= last_addr_mask; 5069c7546f8fSDavid Gibson continue; 5070e95a9851SMike Kravetz } 5071c7546f8fSDavid Gibson 5072cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 50734ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 507431d49da5SAneesh Kumar K.V spin_unlock(ptl); 5075a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5076a4a118f2SNadav Amit force_flush = true; 50774ddb4d91SMike Kravetz address |= last_addr_mask; 507831d49da5SAneesh Kumar K.V continue; 507931d49da5SAneesh Kumar K.V } 508039dde65cSChen, Kenneth W 50816629326bSHillf Danton pte = huge_ptep_get(ptep); 508231d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 508331d49da5SAneesh Kumar K.V spin_unlock(ptl); 508431d49da5SAneesh Kumar K.V continue; 508531d49da5SAneesh Kumar K.V } 50866629326bSHillf Danton 50876629326bSHillf Danton /* 50889fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 50899fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 50906629326bSHillf Danton */ 50919fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 509205e90bd0SPeter Xu /* 509305e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 509405e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 509505e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 509605e90bd0SPeter Xu * pte with a marker. 509705e90bd0SPeter Xu */ 509805e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 509905e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 510005e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 510105e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 510205e90bd0SPeter Xu else 51039386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 510431d49da5SAneesh Kumar K.V spin_unlock(ptl); 510531d49da5SAneesh Kumar K.V continue; 51068c4894c6SNaoya Horiguchi } 51076629326bSHillf Danton 51086629326bSHillf Danton page = pte_page(pte); 510904f2cbe3SMel Gorman /* 511004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 511104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 511204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 511304f2cbe3SMel Gorman */ 511404f2cbe3SMel Gorman if (ref_page) { 511531d49da5SAneesh Kumar K.V if (page != ref_page) { 511631d49da5SAneesh Kumar K.V spin_unlock(ptl); 511731d49da5SAneesh Kumar K.V continue; 511831d49da5SAneesh Kumar K.V } 511904f2cbe3SMel Gorman /* 512004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 512104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 512204f2cbe3SMel Gorman * looking like data was lost 512304f2cbe3SMel Gorman */ 512404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 512504f2cbe3SMel Gorman } 512604f2cbe3SMel Gorman 5127c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5128b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5129106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 51306649a386SKen Chen set_page_dirty(page); 513105e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 513205e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 513305e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 513405e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 513505e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 51365d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5137cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 513831d49da5SAneesh Kumar K.V 5139cb900f41SKirill A. Shutemov spin_unlock(ptl); 5140e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 514124669e58SAneesh Kumar K.V /* 514231d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 514324669e58SAneesh Kumar K.V */ 514431d49da5SAneesh Kumar K.V if (ref_page) 514531d49da5SAneesh Kumar K.V break; 5146fe1668aeSChen, Kenneth W } 5147ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 514824669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5149a4a118f2SNadav Amit 5150a4a118f2SNadav Amit /* 5151a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5152a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5153a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5154a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5155a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5156a4a118f2SNadav Amit * dropped as well. 5157a4a118f2SNadav Amit * 5158a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5159a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5160a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5161a4a118f2SNadav Amit * Instead, do flush now. 5162a4a118f2SNadav Amit */ 5163a4a118f2SNadav Amit if (force_flush) 5164a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 51651da177e4SLinus Torvalds } 516663551ae0SDavid Gibson 5167d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5168d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 516905e90bd0SPeter Xu unsigned long end, struct page *ref_page, 517005e90bd0SPeter Xu zap_flags_t zap_flags) 5171d833352aSMel Gorman { 517205e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5173d833352aSMel Gorman 5174d833352aSMel Gorman /* 5175d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 5176d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 5177d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 5178d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 5179d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 5180c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 5181d833352aSMel Gorman * because in the context this is called, the VMA is about to be 5182c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 5183d833352aSMel Gorman */ 5184d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 5185d833352aSMel Gorman } 5186d833352aSMel Gorman 5187502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 518805e90bd0SPeter Xu unsigned long end, struct page *ref_page, 518905e90bd0SPeter Xu zap_flags_t zap_flags) 5190502717f4SChen, Kenneth W { 519124669e58SAneesh Kumar K.V struct mmu_gather tlb; 5192dff11abeSMike Kravetz 5193a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 519405e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5195ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5196502717f4SChen, Kenneth W } 5197502717f4SChen, Kenneth W 519804f2cbe3SMel Gorman /* 519904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5200578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 520104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 520204f2cbe3SMel Gorman * same region. 520304f2cbe3SMel Gorman */ 52042f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 52052a4b3dedSHarvey Harrison struct page *page, unsigned long address) 520604f2cbe3SMel Gorman { 52077526674dSAdam Litke struct hstate *h = hstate_vma(vma); 520804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 520904f2cbe3SMel Gorman struct address_space *mapping; 521004f2cbe3SMel Gorman pgoff_t pgoff; 521104f2cbe3SMel Gorman 521204f2cbe3SMel Gorman /* 521304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 521404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 521504f2cbe3SMel Gorman */ 52167526674dSAdam Litke address = address & huge_page_mask(h); 521736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 521836e4f20aSMichal Hocko vma->vm_pgoff; 521993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 522004f2cbe3SMel Gorman 52214eb2b1dcSMel Gorman /* 52224eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 52234eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 52244eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 52254eb2b1dcSMel Gorman */ 522683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 52276b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 522804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 522904f2cbe3SMel Gorman if (iter_vma == vma) 523004f2cbe3SMel Gorman continue; 523104f2cbe3SMel Gorman 523204f2cbe3SMel Gorman /* 52332f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 52342f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 52352f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 52362f84a899SMel Gorman */ 52372f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 52382f84a899SMel Gorman continue; 52392f84a899SMel Gorman 52402f84a899SMel Gorman /* 524104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 524204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 524304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 524404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 524504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 524604f2cbe3SMel Gorman */ 524704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 524824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 524905e90bd0SPeter Xu address + huge_page_size(h), page, 0); 525004f2cbe3SMel Gorman } 525183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 525204f2cbe3SMel Gorman } 525304f2cbe3SMel Gorman 52540fe6e20bSNaoya Horiguchi /* 5255c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5256aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5257ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5258ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 52590fe6e20bSNaoya Horiguchi */ 5260c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5261c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5262cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 52631e8f889bSDavid Gibson { 5264c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 52653999f52eSAneesh Kumar K.V pte_t pte; 5266a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52671e8f889bSDavid Gibson struct page *old_page, *new_page; 52682b740303SSouptick Joarder int outside_reserve = 0; 52692b740303SSouptick Joarder vm_fault_t ret = 0; 5270974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5271ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 52721e8f889bSDavid Gibson 5273c89357e2SDavid Hildenbrand VM_BUG_ON(unshare && (flags & FOLL_WRITE)); 5274c89357e2SDavid Hildenbrand VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); 5275c89357e2SDavid Hildenbrand 52761d8d1464SDavid Hildenbrand /* 52771d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 52781d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 52791d8d1464SDavid Hildenbrand */ 52801d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 52811d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 52821d8d1464SDavid Hildenbrand 52831d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 52841d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 52851d8d1464SDavid Hildenbrand if (unlikely(unshare)) 52861d8d1464SDavid Hildenbrand return 0; 52871d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 52881d8d1464SDavid Hildenbrand return 0; 52891d8d1464SDavid Hildenbrand } 52901d8d1464SDavid Hildenbrand 52913999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 52921e8f889bSDavid Gibson old_page = pte_page(pte); 52931e8f889bSDavid Gibson 5294662ce1dcSYang Yang delayacct_wpcopy_start(); 5295662ce1dcSYang Yang 529604f2cbe3SMel Gorman retry_avoidcopy: 5297c89357e2SDavid Hildenbrand /* 5298c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5299c89357e2SDavid Hildenbrand * owner and can reuse this page. 5300c89357e2SDavid Hildenbrand */ 530137a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5302c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 53035a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5304c89357e2SDavid Hildenbrand if (likely(!unshare)) 53055b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5306662ce1dcSYang Yang 5307662ce1dcSYang Yang delayacct_wpcopy_end(); 530883c54070SNick Piggin return 0; 53091e8f889bSDavid Gibson } 53106c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 53116c287605SDavid Hildenbrand old_page); 53121e8f889bSDavid Gibson 531304f2cbe3SMel Gorman /* 531404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 531504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 531604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 531704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 531804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 531904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 532004f2cbe3SMel Gorman * of the full address range. 532104f2cbe3SMel Gorman */ 53225944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 532304f2cbe3SMel Gorman old_page != pagecache_page) 532404f2cbe3SMel Gorman outside_reserve = 1; 532504f2cbe3SMel Gorman 532609cbfeafSKirill A. Shutemov get_page(old_page); 5327b76c8cfbSLarry Woodman 5328ad4404a2SDavidlohr Bueso /* 5329ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5330ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5331ad4404a2SDavidlohr Bueso */ 5332cb900f41SKirill A. Shutemov spin_unlock(ptl); 53335b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 53341e8f889bSDavid Gibson 53352fc39cecSAdam Litke if (IS_ERR(new_page)) { 533604f2cbe3SMel Gorman /* 533704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 533804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 533904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 534004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 534104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 534204f2cbe3SMel Gorman */ 534304f2cbe3SMel Gorman if (outside_reserve) { 5344e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 5345e7dd91c4SMike Kravetz pgoff_t idx; 5346e7dd91c4SMike Kravetz u32 hash; 5347e7dd91c4SMike Kravetz 534809cbfeafSKirill A. Shutemov put_page(old_page); 534904f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 5350e7dd91c4SMike Kravetz /* 5351e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 5352e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 5353e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 5354e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 5355e7dd91c4SMike Kravetz * PMD sharing. 5356e7dd91c4SMike Kravetz * 5357e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 5358e7dd91c4SMike Kravetz */ 5359e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5360e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 5361e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5362e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 5363e7dd91c4SMike Kravetz 53645b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 5365e7dd91c4SMike Kravetz 5366e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 5367e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 5368cb900f41SKirill A. Shutemov spin_lock(ptl); 53695b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5370a9af0c5dSNaoya Horiguchi if (likely(ptep && 5371a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 537204f2cbe3SMel Gorman goto retry_avoidcopy; 5373a734bcc8SHillf Danton /* 5374cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5375cb900f41SKirill A. Shutemov * lock, and our job is done. 5376a734bcc8SHillf Danton */ 5377662ce1dcSYang Yang delayacct_wpcopy_end(); 5378a734bcc8SHillf Danton return 0; 537904f2cbe3SMel Gorman } 538004f2cbe3SMel Gorman 53812b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5382ad4404a2SDavidlohr Bueso goto out_release_old; 53831e8f889bSDavid Gibson } 53841e8f889bSDavid Gibson 53850fe6e20bSNaoya Horiguchi /* 53860fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 53870fe6e20bSNaoya Horiguchi * anon_vma prepared. 53880fe6e20bSNaoya Horiguchi */ 538944e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5390ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5391ad4404a2SDavidlohr Bueso goto out_release_all; 539244e2aa93SDean Nelson } 53930fe6e20bSNaoya Horiguchi 5394974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 539547ad8475SAndrea Arcangeli pages_per_huge_page(h)); 53960ed361deSNick Piggin __SetPageUptodate(new_page); 53971e8f889bSDavid Gibson 53987269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 53996f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5400ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5401ad4404a2SDavidlohr Bueso 5402b76c8cfbSLarry Woodman /* 5403cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5404b76c8cfbSLarry Woodman * before the page tables are altered 5405b76c8cfbSLarry Woodman */ 5406cb900f41SKirill A. Shutemov spin_lock(ptl); 54075b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5408a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5409d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 541007443a85SJoonsoo Kim 5411c89357e2SDavid Hildenbrand /* Break COW or unshare */ 54125b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5413ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5414cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 54155b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 54161eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5417c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 54188f251a3dSMike Kravetz SetHPageMigratable(new_page); 54191e8f889bSDavid Gibson /* Make the old page be freed below */ 54201e8f889bSDavid Gibson new_page = old_page; 54211e8f889bSDavid Gibson } 5422cb900f41SKirill A. Shutemov spin_unlock(ptl); 5423ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5424ad4404a2SDavidlohr Bueso out_release_all: 5425c89357e2SDavid Hildenbrand /* 5426c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5427c89357e2SDavid Hildenbrand * unshare) 5428c89357e2SDavid Hildenbrand */ 5429c7b1850dSMike Kravetz if (new_page != old_page) 54305b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 543109cbfeafSKirill A. Shutemov put_page(new_page); 5432ad4404a2SDavidlohr Bueso out_release_old: 543309cbfeafSKirill A. Shutemov put_page(old_page); 54348312034fSJoonsoo Kim 5435ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5436662ce1dcSYang Yang 5437662ce1dcSYang Yang delayacct_wpcopy_end(); 5438ad4404a2SDavidlohr Bueso return ret; 54391e8f889bSDavid Gibson } 54401e8f889bSDavid Gibson 544104f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 5442a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 5443a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 544404f2cbe3SMel Gorman { 544504f2cbe3SMel Gorman struct address_space *mapping; 5446e7c4b0bfSAndy Whitcroft pgoff_t idx; 544704f2cbe3SMel Gorman 544804f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 5449a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 545004f2cbe3SMel Gorman 545104f2cbe3SMel Gorman return find_lock_page(mapping, idx); 545204f2cbe3SMel Gorman } 545304f2cbe3SMel Gorman 54543ae77f43SHugh Dickins /* 54553ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 54563ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 54573ae77f43SHugh Dickins */ 54583ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 54592a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 54602a15efc9SHugh Dickins { 54612a15efc9SHugh Dickins struct address_space *mapping; 54622a15efc9SHugh Dickins pgoff_t idx; 54632a15efc9SHugh Dickins struct page *page; 54642a15efc9SHugh Dickins 54652a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 54662a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 54672a15efc9SHugh Dickins 54682a15efc9SHugh Dickins page = find_get_page(mapping, idx); 54692a15efc9SHugh Dickins if (page) 54702a15efc9SHugh Dickins put_page(page); 54712a15efc9SHugh Dickins return page != NULL; 54722a15efc9SHugh Dickins } 54732a15efc9SHugh Dickins 5474ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 5475ab76ad54SMike Kravetz pgoff_t idx) 5476ab76ad54SMike Kravetz { 5477d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5478ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5479ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5480d9ef44deSMatthew Wilcox (Oracle) int err; 5481ab76ad54SMike Kravetz 5482d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5483d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5484d9ef44deSMatthew Wilcox (Oracle) 5485d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5486d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5487ab76ad54SMike Kravetz return err; 5488d9ef44deSMatthew Wilcox (Oracle) } 5489d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5490ab76ad54SMike Kravetz 549122146c3cSMike Kravetz /* 5492d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 549322146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 549422146c3cSMike Kravetz */ 5495d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 549622146c3cSMike Kravetz 5497ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5498ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5499ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5500ab76ad54SMike Kravetz return 0; 5501ab76ad54SMike Kravetz } 5502ab76ad54SMike Kravetz 55037677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 55047677f7fdSAxel Rasmussen struct address_space *mapping, 55057677f7fdSAxel Rasmussen pgoff_t idx, 55067677f7fdSAxel Rasmussen unsigned int flags, 55077677f7fdSAxel Rasmussen unsigned long haddr, 5508824ddc60SNadav Amit unsigned long addr, 55097677f7fdSAxel Rasmussen unsigned long reason) 55107677f7fdSAxel Rasmussen { 55117677f7fdSAxel Rasmussen vm_fault_t ret; 55127677f7fdSAxel Rasmussen u32 hash; 55137677f7fdSAxel Rasmussen struct vm_fault vmf = { 55147677f7fdSAxel Rasmussen .vma = vma, 55157677f7fdSAxel Rasmussen .address = haddr, 5516824ddc60SNadav Amit .real_address = addr, 55177677f7fdSAxel Rasmussen .flags = flags, 55187677f7fdSAxel Rasmussen 55197677f7fdSAxel Rasmussen /* 55207677f7fdSAxel Rasmussen * Hard to debug if it ends up being 55217677f7fdSAxel Rasmussen * used by a callee that assumes 55227677f7fdSAxel Rasmussen * something about the other 55237677f7fdSAxel Rasmussen * uninitialized fields... same as in 55247677f7fdSAxel Rasmussen * memory.c 55257677f7fdSAxel Rasmussen */ 55267677f7fdSAxel Rasmussen }; 55277677f7fdSAxel Rasmussen 55287677f7fdSAxel Rasmussen /* 55297677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 55307677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 55317677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 55327677f7fdSAxel Rasmussen */ 55337677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 55347677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 55357677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 55367677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 55377677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 55387677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 55397677f7fdSAxel Rasmussen 55407677f7fdSAxel Rasmussen return ret; 55417677f7fdSAxel Rasmussen } 55427677f7fdSAxel Rasmussen 55432b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 55442b740303SSouptick Joarder struct vm_area_struct *vma, 55458382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5546c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5547c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5548ac9b9c66SHugh Dickins { 5549a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55502b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5551409eb8c2SHillf Danton int anon_rmap = 0; 55524c887265SAdam Litke unsigned long size; 55534c887265SAdam Litke struct page *page; 55541e8f889bSDavid Gibson pte_t new_pte; 5555cb900f41SKirill A. Shutemov spinlock_t *ptl; 5556285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5557c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 55584c887265SAdam Litke 555904f2cbe3SMel Gorman /* 556004f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 556104f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5562c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5563c89357e2SDavid Hildenbrand * be obvious. 556404f2cbe3SMel Gorman */ 556504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5566910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 556704f2cbe3SMel Gorman current->pid); 556804f2cbe3SMel Gorman return ret; 556904f2cbe3SMel Gorman } 557004f2cbe3SMel Gorman 55714c887265SAdam Litke /* 557287bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 557387bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 557487bf91d3SMike Kravetz * once for faults beyond end of file. 55754c887265SAdam Litke */ 5576a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 5577ebed4bfcSHugh Dickins if (idx >= size) 5578ebed4bfcSHugh Dickins goto out; 55791a1aad8aSMike Kravetz 5580c7b1850dSMike Kravetz new_page = false; 558187bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 558287bf91d3SMike Kravetz if (!page) { 55837677f7fdSAxel Rasmussen /* Check for page in userfault range */ 55841a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 55857677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5586824ddc60SNadav Amit flags, haddr, address, 55877677f7fdSAxel Rasmussen VM_UFFD_MISSING); 55881a1aad8aSMike Kravetz goto out; 55891a1aad8aSMike Kravetz } 55901a1aad8aSMike Kravetz 5591285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 55922fc39cecSAdam Litke if (IS_ERR(page)) { 55934643d67eSMike Kravetz /* 55944643d67eSMike Kravetz * Returning error will result in faulting task being 55954643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 55964643d67eSMike Kravetz * tasks from racing to fault in the same page which 55974643d67eSMike Kravetz * could result in false unable to allocate errors. 55984643d67eSMike Kravetz * Page migration does not take the fault mutex, but 55994643d67eSMike Kravetz * does a clear then write of pte's under page table 56004643d67eSMike Kravetz * lock. Page fault code could race with migration, 56014643d67eSMike Kravetz * notice the clear pte and try to allocate a page 56024643d67eSMike Kravetz * here. Before returning error, get ptl and make 56034643d67eSMike Kravetz * sure there really is no pte entry. 56044643d67eSMike Kravetz */ 56054643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 56064643d67eSMike Kravetz ret = 0; 5607d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 56082b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5609d83e6c8aSMiaohe Lin spin_unlock(ptl); 56106bda666aSChristoph Lameter goto out; 56116bda666aSChristoph Lameter } 561247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 56130ed361deSNick Piggin __SetPageUptodate(page); 5614cb6acd01SMike Kravetz new_page = true; 5615ac9b9c66SHugh Dickins 5616f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 5617ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 56186bda666aSChristoph Lameter if (err) { 56193a5497a2SMiaohe Lin /* 56203a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 56213a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 56223a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 56233a5497a2SMiaohe Lin * to the page cache. So it's safe to call 56243a5497a2SMiaohe Lin * restore_reserve_on_error() here. 56253a5497a2SMiaohe Lin */ 56263a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 56276bda666aSChristoph Lameter put_page(page); 56286bda666aSChristoph Lameter goto out; 56296bda666aSChristoph Lameter } 5630c7b1850dSMike Kravetz new_pagecache_page = true; 563123be7468SMel Gorman } else { 56326bda666aSChristoph Lameter lock_page(page); 56330fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 56340fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 56350fe6e20bSNaoya Horiguchi goto backout_unlocked; 563623be7468SMel Gorman } 5637409eb8c2SHillf Danton anon_rmap = 1; 56380fe6e20bSNaoya Horiguchi } 56390fe6e20bSNaoya Horiguchi } else { 564057303d80SAndy Whitcroft /* 5641998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5642998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5643998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5644fd6a03edSNaoya Horiguchi */ 5645fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 56460eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5647972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5648fd6a03edSNaoya Horiguchi goto backout_unlocked; 56496bda666aSChristoph Lameter } 56507677f7fdSAxel Rasmussen 56517677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 56527677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 56537677f7fdSAxel Rasmussen unlock_page(page); 56547677f7fdSAxel Rasmussen put_page(page); 56557677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5656824ddc60SNadav Amit flags, haddr, address, 56577677f7fdSAxel Rasmussen VM_UFFD_MINOR); 56587677f7fdSAxel Rasmussen goto out; 56597677f7fdSAxel Rasmussen } 5660998b4382SNaoya Horiguchi } 56611e8f889bSDavid Gibson 566257303d80SAndy Whitcroft /* 566357303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 566457303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 566557303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 566657303d80SAndy Whitcroft * the spinlock. 566757303d80SAndy Whitcroft */ 56685e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5669285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 56702b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 56712b26736cSAndy Whitcroft goto backout_unlocked; 56722b26736cSAndy Whitcroft } 56735e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5674285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 56755e911373SMike Kravetz } 567657303d80SAndy Whitcroft 56778bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 567883c54070SNick Piggin ret = 0; 5679c64e912cSPeter Xu /* If pte changed from under us, retry */ 5680c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 56814c887265SAdam Litke goto backout; 56824c887265SAdam Litke 568307443a85SJoonsoo Kim if (anon_rmap) { 5684d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5685285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5686ac714904SChoi Gi-yong } else 5687fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 56881e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 56891e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5690c64e912cSPeter Xu /* 5691c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5692c64e912cSPeter Xu * if populated. 5693c64e912cSPeter Xu */ 5694c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5695c64e912cSPeter Xu new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); 5696285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 56971e8f889bSDavid Gibson 56985d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5699788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 57001e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5701c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 57021e8f889bSDavid Gibson } 57031e8f889bSDavid Gibson 5704cb900f41SKirill A. Shutemov spin_unlock(ptl); 5705cb6acd01SMike Kravetz 5706cb6acd01SMike Kravetz /* 57078f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 57088f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 57098f251a3dSMike Kravetz * been isolated for migration. 5710cb6acd01SMike Kravetz */ 5711cb6acd01SMike Kravetz if (new_page) 57128f251a3dSMike Kravetz SetHPageMigratable(page); 5713cb6acd01SMike Kravetz 57144c887265SAdam Litke unlock_page(page); 57154c887265SAdam Litke out: 5716ac9b9c66SHugh Dickins return ret; 57174c887265SAdam Litke 57184c887265SAdam Litke backout: 5719cb900f41SKirill A. Shutemov spin_unlock(ptl); 57202b26736cSAndy Whitcroft backout_unlocked: 57214c887265SAdam Litke unlock_page(page); 5722c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5723c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5724285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 57254c887265SAdam Litke put_page(page); 57264c887265SAdam Litke goto out; 5727ac9b9c66SHugh Dickins } 5728ac9b9c66SHugh Dickins 57298382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5730188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57318382d914SDavidlohr Bueso { 57328382d914SDavidlohr Bueso unsigned long key[2]; 57338382d914SDavidlohr Bueso u32 hash; 57348382d914SDavidlohr Bueso 57358382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 57368382d914SDavidlohr Bueso key[1] = idx; 57378382d914SDavidlohr Bueso 573855254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 57398382d914SDavidlohr Bueso 57408382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 57418382d914SDavidlohr Bueso } 57428382d914SDavidlohr Bueso #else 57438382d914SDavidlohr Bueso /* 57446c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 57458382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 57468382d914SDavidlohr Bueso */ 5747188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57488382d914SDavidlohr Bueso { 57498382d914SDavidlohr Bueso return 0; 57508382d914SDavidlohr Bueso } 57518382d914SDavidlohr Bueso #endif 57528382d914SDavidlohr Bueso 57532b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5754788c7df4SHugh Dickins unsigned long address, unsigned int flags) 575586e5216fSAdam Litke { 57568382d914SDavidlohr Bueso pte_t *ptep, entry; 5757cb900f41SKirill A. Shutemov spinlock_t *ptl; 57582b740303SSouptick Joarder vm_fault_t ret; 57598382d914SDavidlohr Bueso u32 hash; 57608382d914SDavidlohr Bueso pgoff_t idx; 57610fe6e20bSNaoya Horiguchi struct page *page = NULL; 576257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5763a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57648382d914SDavidlohr Bueso struct address_space *mapping; 57650f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5766285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 576786e5216fSAdam Litke 5768285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5769fd6a03edSNaoya Horiguchi if (ptep) { 5770c0d0381aSMike Kravetz /* 5771c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 5772c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 5773c0d0381aSMike Kravetz * not actually modifying content here. 5774c0d0381aSMike Kravetz */ 5775fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5776290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5777ad1ac596SMiaohe Lin migration_entry_wait_huge(vma, ptep); 5778290408d4SNaoya Horiguchi return 0; 5779290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5780aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5781972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5782b43a9990SMike Kravetz } 57838382d914SDavidlohr Bueso 5784c0d0381aSMike Kravetz /* 5785c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 578687bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 578787bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 578887bf91d3SMike Kravetz * and making the ptep no longer valid. 578987bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 5790c0d0381aSMike Kravetz * 5791c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5792c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5793c0d0381aSMike Kravetz * something has changed. 5794c0d0381aSMike Kravetz */ 5795ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5796c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5797aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5798c0d0381aSMike Kravetz if (!ptep) { 5799c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5800c0d0381aSMike Kravetz return VM_FAULT_OOM; 5801c0d0381aSMike Kravetz } 5802ddeaab32SMike Kravetz 58033935baa9SDavid Gibson /* 58043935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 58053935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 58063935baa9SDavid Gibson * the same page in the page cache. 58073935baa9SDavid Gibson */ 5808c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5809188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5810c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 58118382d914SDavidlohr Bueso 58127f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 5813c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 5814c64e912cSPeter Xu if (huge_pte_none_mostly(entry)) { 5815c64e912cSPeter Xu ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 5816c64e912cSPeter Xu entry, flags); 5817b4d1d99fSDavid Gibson goto out_mutex; 58183935baa9SDavid Gibson } 581986e5216fSAdam Litke 582083c54070SNick Piggin ret = 0; 58211e8f889bSDavid Gibson 582257303d80SAndy Whitcroft /* 58230f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 58240f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 58257c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 58267c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 58277c8de358SEthon Paul * properly handle it. 58280f792cf9SNaoya Horiguchi */ 58290f792cf9SNaoya Horiguchi if (!pte_present(entry)) 58300f792cf9SNaoya Horiguchi goto out_mutex; 58310f792cf9SNaoya Horiguchi 58320f792cf9SNaoya Horiguchi /* 5833c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 5834c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 583557303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 58361d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 58371d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 583857303d80SAndy Whitcroft */ 5839c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 58401d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 5841285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58422b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5843b4d1d99fSDavid Gibson goto out_mutex; 58442b26736cSAndy Whitcroft } 58455e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5846285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 584757303d80SAndy Whitcroft 58481d8d1464SDavid Hildenbrand pagecache_page = hugetlbfs_pagecache_page(h, vma, haddr); 584957303d80SAndy Whitcroft } 585057303d80SAndy Whitcroft 58510f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 58520fe6e20bSNaoya Horiguchi 5853c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 5854b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5855cb900f41SKirill A. Shutemov goto out_ptl; 5856b4d1d99fSDavid Gibson 5857166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 5858166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 5859166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5860166f3eccSPeter Xu struct vm_fault vmf = { 5861166f3eccSPeter Xu .vma = vma, 5862166f3eccSPeter Xu .address = haddr, 5863166f3eccSPeter Xu .real_address = address, 5864166f3eccSPeter Xu .flags = flags, 5865166f3eccSPeter Xu }; 5866166f3eccSPeter Xu 5867166f3eccSPeter Xu spin_unlock(ptl); 5868166f3eccSPeter Xu if (pagecache_page) { 5869166f3eccSPeter Xu unlock_page(pagecache_page); 5870166f3eccSPeter Xu put_page(pagecache_page); 5871166f3eccSPeter Xu } 5872166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5873166f3eccSPeter Xu i_mmap_unlock_read(mapping); 5874166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 5875166f3eccSPeter Xu } 5876166f3eccSPeter Xu 58770f792cf9SNaoya Horiguchi /* 5878c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 58790f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 58800f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 58810f792cf9SNaoya Horiguchi */ 58820f792cf9SNaoya Horiguchi page = pte_page(entry); 58830f792cf9SNaoya Horiguchi if (page != pagecache_page) 58840f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 58850f792cf9SNaoya Horiguchi need_wait_lock = 1; 58860f792cf9SNaoya Horiguchi goto out_ptl; 58870f792cf9SNaoya Horiguchi } 58880f792cf9SNaoya Horiguchi 58890f792cf9SNaoya Horiguchi get_page(page); 5890b4d1d99fSDavid Gibson 5891c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 5892106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5893c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 5894cb900f41SKirill A. Shutemov pagecache_page, ptl); 58950f792cf9SNaoya Horiguchi goto out_put_page; 5896c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 5897106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5898b4d1d99fSDavid Gibson } 5899c89357e2SDavid Hildenbrand } 5900b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5901285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5902788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5903285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 59040f792cf9SNaoya Horiguchi out_put_page: 59050f792cf9SNaoya Horiguchi if (page != pagecache_page) 59060f792cf9SNaoya Horiguchi unlock_page(page); 59070f792cf9SNaoya Horiguchi put_page(page); 5908cb900f41SKirill A. Shutemov out_ptl: 5909cb900f41SKirill A. Shutemov spin_unlock(ptl); 591057303d80SAndy Whitcroft 591157303d80SAndy Whitcroft if (pagecache_page) { 591257303d80SAndy Whitcroft unlock_page(pagecache_page); 591357303d80SAndy Whitcroft put_page(pagecache_page); 591457303d80SAndy Whitcroft } 5915b4d1d99fSDavid Gibson out_mutex: 5916c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5917c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 59180f792cf9SNaoya Horiguchi /* 59190f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 59200f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 59210f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 59220f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 59230f792cf9SNaoya Horiguchi * here without taking refcount. 59240f792cf9SNaoya Horiguchi */ 59250f792cf9SNaoya Horiguchi if (need_wait_lock) 59260f792cf9SNaoya Horiguchi wait_on_page_locked(page); 59271e8f889bSDavid Gibson return ret; 592886e5216fSAdam Litke } 592986e5216fSAdam Litke 5930714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 59318fb5debcSMike Kravetz /* 59328fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 59338fb5debcSMike Kravetz * modifications for huge pages. 59348fb5debcSMike Kravetz */ 59358fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 59368fb5debcSMike Kravetz pte_t *dst_pte, 59378fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 59388fb5debcSMike Kravetz unsigned long dst_addr, 59398fb5debcSMike Kravetz unsigned long src_addr, 5940f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 59416041c691SPeter Xu struct page **pagep, 59426041c691SPeter Xu bool wp_copy) 59438fb5debcSMike Kravetz { 5944f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 59458cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 59468cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 59478cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 59481e392147SAndrea Arcangeli unsigned long size; 59491c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 59508fb5debcSMike Kravetz pte_t _dst_pte; 59518fb5debcSMike Kravetz spinlock_t *ptl; 59528cc5fcbbSMina Almasry int ret = -ENOMEM; 59538fb5debcSMike Kravetz struct page *page; 5954f6191471SAxel Rasmussen int writable; 5955cc30042dSMina Almasry bool page_in_pagecache = false; 59568fb5debcSMike Kravetz 5957f6191471SAxel Rasmussen if (is_continue) { 5958f6191471SAxel Rasmussen ret = -EFAULT; 5959f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5960f6191471SAxel Rasmussen if (!page) 5961f6191471SAxel Rasmussen goto out; 5962cc30042dSMina Almasry page_in_pagecache = true; 5963f6191471SAxel Rasmussen } else if (!*pagep) { 5964d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5965d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5966d84cf06eSMina Almasry */ 5967d84cf06eSMina Almasry if (vm_shared && 5968d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5969d84cf06eSMina Almasry ret = -EEXIST; 59708fb5debcSMike Kravetz goto out; 5971d84cf06eSMina Almasry } 5972d84cf06eSMina Almasry 5973d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5974d84cf06eSMina Almasry if (IS_ERR(page)) { 5975d84cf06eSMina Almasry ret = -ENOMEM; 5976d84cf06eSMina Almasry goto out; 5977d84cf06eSMina Almasry } 59788fb5debcSMike Kravetz 59798fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 59808fb5debcSMike Kravetz (const void __user *) src_addr, 5981810a56b9SMike Kravetz pages_per_huge_page(h), false); 59828fb5debcSMike Kravetz 5983c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 59848fb5debcSMike Kravetz if (unlikely(ret)) { 59859e368259SAndrea Arcangeli ret = -ENOENT; 59868cc5fcbbSMina Almasry /* Free the allocated page which may have 59878cc5fcbbSMina Almasry * consumed a reservation. 59888cc5fcbbSMina Almasry */ 59898cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 59908cc5fcbbSMina Almasry put_page(page); 59918cc5fcbbSMina Almasry 59928cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 59938cc5fcbbSMina Almasry * contents. 59948cc5fcbbSMina Almasry */ 59958cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 59968cc5fcbbSMina Almasry if (!page) { 59978cc5fcbbSMina Almasry ret = -ENOMEM; 59988cc5fcbbSMina Almasry goto out; 59998cc5fcbbSMina Almasry } 60008fb5debcSMike Kravetz *pagep = page; 60018cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 60028cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 60038cc5fcbbSMina Almasry * page. 60048cc5fcbbSMina Almasry */ 60058fb5debcSMike Kravetz goto out; 60068fb5debcSMike Kravetz } 60078fb5debcSMike Kravetz } else { 60088cc5fcbbSMina Almasry if (vm_shared && 60098cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 60108cc5fcbbSMina Almasry put_page(*pagep); 60118cc5fcbbSMina Almasry ret = -EEXIST; 60128cc5fcbbSMina Almasry *pagep = NULL; 60138cc5fcbbSMina Almasry goto out; 60148cc5fcbbSMina Almasry } 60158cc5fcbbSMina Almasry 60168cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 60178cc5fcbbSMina Almasry if (IS_ERR(page)) { 6018da9a298fSMiaohe Lin put_page(*pagep); 60198cc5fcbbSMina Almasry ret = -ENOMEM; 60208cc5fcbbSMina Almasry *pagep = NULL; 60218cc5fcbbSMina Almasry goto out; 60228cc5fcbbSMina Almasry } 602334892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 602434892366SMuchun Song pages_per_huge_page(h)); 60258cc5fcbbSMina Almasry put_page(*pagep); 60268fb5debcSMike Kravetz *pagep = NULL; 60278fb5debcSMike Kravetz } 60288fb5debcSMike Kravetz 60298fb5debcSMike Kravetz /* 60308fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 60318fb5debcSMike Kravetz * preceding stores to the page contents become visible before 60328fb5debcSMike Kravetz * the set_pte_at() write. 60338fb5debcSMike Kravetz */ 60348fb5debcSMike Kravetz __SetPageUptodate(page); 60358fb5debcSMike Kravetz 6036f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6037f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 60381e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60391e392147SAndrea Arcangeli ret = -EFAULT; 60401e392147SAndrea Arcangeli if (idx >= size) 60411e392147SAndrea Arcangeli goto out_release_nounlock; 60421c9e8defSMike Kravetz 60431e392147SAndrea Arcangeli /* 60441e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 60451e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 60461e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 60471e392147SAndrea Arcangeli * the caller. 60481e392147SAndrea Arcangeli */ 60491c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 60501c9e8defSMike Kravetz if (ret) 60511c9e8defSMike Kravetz goto out_release_nounlock; 6052cc30042dSMina Almasry page_in_pagecache = true; 60531c9e8defSMike Kravetz } 60541c9e8defSMike Kravetz 60558fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 60568fb5debcSMike Kravetz spin_lock(ptl); 60578fb5debcSMike Kravetz 60581e392147SAndrea Arcangeli /* 60591e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 60601e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 60611e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 60621e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 60631e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 60641e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 60651e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 60661e392147SAndrea Arcangeli */ 60671e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60681e392147SAndrea Arcangeli ret = -EFAULT; 60691e392147SAndrea Arcangeli if (idx >= size) 60701e392147SAndrea Arcangeli goto out_release_unlock; 60711e392147SAndrea Arcangeli 60728fb5debcSMike Kravetz ret = -EEXIST; 60736041c691SPeter Xu /* 60746041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 60756041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 60766041c691SPeter Xu * page backing it, then access the page. 60776041c691SPeter Xu */ 60786041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 60798fb5debcSMike Kravetz goto out_release_unlock; 60808fb5debcSMike Kravetz 6081ab74ef70SMiaohe Lin if (page_in_pagecache) { 6082fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 60831c9e8defSMike Kravetz } else { 6084d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 60858fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 60861c9e8defSMike Kravetz } 60878fb5debcSMike Kravetz 60886041c691SPeter Xu /* 60896041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 60906041c691SPeter Xu * with wp flag set, don't set pte write bit. 60916041c691SPeter Xu */ 60926041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6093f6191471SAxel Rasmussen writable = 0; 6094f6191471SAxel Rasmussen else 6095f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6096f6191471SAxel Rasmussen 6097f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 60986041c691SPeter Xu /* 60996041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 61006041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 61016041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 61026041c691SPeter Xu * thrown away at will, even if write bit not set. 61036041c691SPeter Xu */ 61048fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 61058fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 61068fb5debcSMike Kravetz 61076041c691SPeter Xu if (wp_copy) 61086041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 61096041c691SPeter Xu 61108fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 61118fb5debcSMike Kravetz 61128fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 61138fb5debcSMike Kravetz 61148fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 61158fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 61168fb5debcSMike Kravetz 61178fb5debcSMike Kravetz spin_unlock(ptl); 6118f6191471SAxel Rasmussen if (!is_continue) 61198f251a3dSMike Kravetz SetHPageMigratable(page); 6120f6191471SAxel Rasmussen if (vm_shared || is_continue) 61211c9e8defSMike Kravetz unlock_page(page); 61228fb5debcSMike Kravetz ret = 0; 61238fb5debcSMike Kravetz out: 61248fb5debcSMike Kravetz return ret; 61258fb5debcSMike Kravetz out_release_unlock: 61268fb5debcSMike Kravetz spin_unlock(ptl); 6127f6191471SAxel Rasmussen if (vm_shared || is_continue) 61281c9e8defSMike Kravetz unlock_page(page); 61295af10dfdSAndrea Arcangeli out_release_nounlock: 6130cc30042dSMina Almasry if (!page_in_pagecache) 6131846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 61328fb5debcSMike Kravetz put_page(page); 61338fb5debcSMike Kravetz goto out; 61348fb5debcSMike Kravetz } 6135714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 61368fb5debcSMike Kravetz 613782e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 613882e5d378SJoao Martins int refs, struct page **pages, 613982e5d378SJoao Martins struct vm_area_struct **vmas) 614082e5d378SJoao Martins { 614182e5d378SJoao Martins int nr; 614282e5d378SJoao Martins 614382e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 614482e5d378SJoao Martins if (likely(pages)) 614582e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 614682e5d378SJoao Martins if (vmas) 614782e5d378SJoao Martins vmas[nr] = vma; 614882e5d378SJoao Martins } 614982e5d378SJoao Martins } 615082e5d378SJoao Martins 6151a7f22660SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, 6152a7f22660SDavid Hildenbrand bool *unshare) 6153a7f22660SDavid Hildenbrand { 6154a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6155a7f22660SDavid Hildenbrand 6156a7f22660SDavid Hildenbrand *unshare = false; 6157a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6158a7f22660SDavid Hildenbrand return true; 6159a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6160a7f22660SDavid Hildenbrand return false; 6161a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6162a7f22660SDavid Hildenbrand return true; 6163a7f22660SDavid Hildenbrand if (gup_must_unshare(flags, pte_page(pteval))) { 6164a7f22660SDavid Hildenbrand *unshare = true; 6165a7f22660SDavid Hildenbrand return true; 6166a7f22660SDavid Hildenbrand } 6167a7f22660SDavid Hildenbrand return false; 6168a7f22660SDavid Hildenbrand } 6169a7f22660SDavid Hildenbrand 617028a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 617163551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 617228a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 61734f6da934SPeter Xu long i, unsigned int flags, int *locked) 617463551ae0SDavid Gibson { 6175d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6176d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 617728a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6178a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 61790fa5bc40SJoao Martins int err = -EFAULT, refs; 618063551ae0SDavid Gibson 618163551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 618263551ae0SDavid Gibson pte_t *pte; 6183cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6184a7f22660SDavid Hildenbrand bool unshare = false; 61852a15efc9SHugh Dickins int absent; 618663551ae0SDavid Gibson struct page *page; 618763551ae0SDavid Gibson 61884c887265SAdam Litke /* 618902057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 619002057967SDavid Rientjes * potentially allocating memory. 619102057967SDavid Rientjes */ 6192fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 619302057967SDavid Rientjes remainder = 0; 619402057967SDavid Rientjes break; 619502057967SDavid Rientjes } 619602057967SDavid Rientjes 619702057967SDavid Rientjes /* 61984c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 61992a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 62004c887265SAdam Litke * first, for the page indexing below to work. 6201cb900f41SKirill A. Shutemov * 6202cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 62034c887265SAdam Litke */ 62047868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 62057868a208SPunit Agrawal huge_page_size(h)); 6206cb900f41SKirill A. Shutemov if (pte) 6207cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 62082a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 620963551ae0SDavid Gibson 62102a15efc9SHugh Dickins /* 62112a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 62123ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 62133ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 62143ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 62153ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 62162a15efc9SHugh Dickins */ 62173ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 62183ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6219cb900f41SKirill A. Shutemov if (pte) 6220cb900f41SKirill A. Shutemov spin_unlock(ptl); 62212a15efc9SHugh Dickins remainder = 0; 62222a15efc9SHugh Dickins break; 62232a15efc9SHugh Dickins } 62242a15efc9SHugh Dickins 62259cc3a5bdSNaoya Horiguchi /* 62269cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 62279cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 62289cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 62299cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 62309cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 62319cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 62329cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 62339cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 62349cc3a5bdSNaoya Horiguchi */ 6235a7f22660SDavid Hildenbrand if (absent || 6236a7f22660SDavid Hildenbrand __follow_hugetlb_must_fault(flags, pte, &unshare)) { 62372b740303SSouptick Joarder vm_fault_t ret; 623887ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 62394c887265SAdam Litke 6240cb900f41SKirill A. Shutemov if (pte) 6241cb900f41SKirill A. Shutemov spin_unlock(ptl); 624287ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 624387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6244a7f22660SDavid Hildenbrand else if (unshare) 6245a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 62464f6da934SPeter Xu if (locked) 624771335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 624871335f37SPeter Xu FAULT_FLAG_KILLABLE; 624987ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 625087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 625187ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 625287ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 62534426e945SPeter Xu /* 62544426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 62554426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 62564426e945SPeter Xu */ 625787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 625887ffc118SAndrea Arcangeli } 625987ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 626087ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 62612be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 62621c59827dSHugh Dickins remainder = 0; 62631c59827dSHugh Dickins break; 62641c59827dSHugh Dickins } 626587ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 62664f6da934SPeter Xu if (locked && 62671ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 62684f6da934SPeter Xu *locked = 0; 626987ffc118SAndrea Arcangeli *nr_pages = 0; 627087ffc118SAndrea Arcangeli /* 627187ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 627287ffc118SAndrea Arcangeli * error, it will return zero 627387ffc118SAndrea Arcangeli * instead. 627487ffc118SAndrea Arcangeli * 627587ffc118SAndrea Arcangeli * No need to update "position" as the 627687ffc118SAndrea Arcangeli * caller will not check it after 627787ffc118SAndrea Arcangeli * *nr_pages is set to 0. 627887ffc118SAndrea Arcangeli */ 627987ffc118SAndrea Arcangeli return i; 628087ffc118SAndrea Arcangeli } 628187ffc118SAndrea Arcangeli continue; 628287ffc118SAndrea Arcangeli } 628363551ae0SDavid Gibson 6284a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 62857f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 62868fde12caSLinus Torvalds 6287b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6288b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6289b6a2619cSDavid Hildenbrand 62908fde12caSLinus Torvalds /* 6291acbfb087SZhigang Lu * If subpage information not requested, update counters 6292acbfb087SZhigang Lu * and skip the same_page loop below. 6293acbfb087SZhigang Lu */ 6294acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6295acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6296acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6297acbfb087SZhigang Lu vaddr += huge_page_size(h); 6298acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6299acbfb087SZhigang Lu i += pages_per_huge_page(h); 6300acbfb087SZhigang Lu spin_unlock(ptl); 6301acbfb087SZhigang Lu continue; 6302acbfb087SZhigang Lu } 6303acbfb087SZhigang Lu 6304d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6305d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6306d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 63070fa5bc40SJoao Martins 630882e5d378SJoao Martins if (pages || vmas) 630982e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 631082e5d378SJoao Martins vma, refs, 631182e5d378SJoao Martins likely(pages) ? pages + i : NULL, 631282e5d378SJoao Martins vmas ? vmas + i : NULL); 631363551ae0SDavid Gibson 631482e5d378SJoao Martins if (pages) { 63150fa5bc40SJoao Martins /* 6316822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 63170fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 63180fa5bc40SJoao Martins * checked that the huge page is present in the page 63190fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 63200fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 63210fa5bc40SJoao Martins * head page and tail pages from being rearranged in 63220fa5bc40SJoao Martins * any way. So this page must be available at this 63230fa5bc40SJoao Martins * point, unless the page refcount overflowed: 63240fa5bc40SJoao Martins */ 6325822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 63260fa5bc40SJoao Martins flags))) { 63270fa5bc40SJoao Martins spin_unlock(ptl); 63280fa5bc40SJoao Martins remainder = 0; 63290fa5bc40SJoao Martins err = -ENOMEM; 63300fa5bc40SJoao Martins break; 63310fa5bc40SJoao Martins } 6332d5d4b0aaSChen, Kenneth W } 633382e5d378SJoao Martins 633482e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 633582e5d378SJoao Martins remainder -= refs; 633682e5d378SJoao Martins i += refs; 633782e5d378SJoao Martins 6338cb900f41SKirill A. Shutemov spin_unlock(ptl); 633963551ae0SDavid Gibson } 634028a35716SMichel Lespinasse *nr_pages = remainder; 634187ffc118SAndrea Arcangeli /* 634287ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 634387ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 634487ffc118SAndrea Arcangeli * branch. 634587ffc118SAndrea Arcangeli */ 634663551ae0SDavid Gibson *position = vaddr; 634763551ae0SDavid Gibson 63482be7cfedSDaniel Jordan return i ? i : err; 634963551ae0SDavid Gibson } 63508f860591SZhang, Yanmin 63517da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 63525a90d5a1SPeter Xu unsigned long address, unsigned long end, 63535a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 63548f860591SZhang, Yanmin { 63558f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 63568f860591SZhang, Yanmin unsigned long start = address; 63578f860591SZhang, Yanmin pte_t *ptep; 63588f860591SZhang, Yanmin pte_t pte; 6359a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 636060dfaad6SPeter Xu unsigned long pages = 0, psize = huge_page_size(h); 6361dff11abeSMike Kravetz bool shared_pmd = false; 6362ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6363e95a9851SMike Kravetz unsigned long last_addr_mask; 63645a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 63655a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6366dff11abeSMike Kravetz 6367dff11abeSMike Kravetz /* 6368dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6369ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6370dff11abeSMike Kravetz * range if PMD sharing is possible. 6371dff11abeSMike Kravetz */ 63727269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 63737269f999SJérôme Glisse 0, vma, mm, start, end); 6374ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 63758f860591SZhang, Yanmin 63768f860591SZhang, Yanmin BUG_ON(address >= end); 6377ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 63788f860591SZhang, Yanmin 6379ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 6380e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 638183cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 638260dfaad6SPeter Xu for (; address < end; address += psize) { 6383cb900f41SKirill A. Shutemov spinlock_t *ptl; 638460dfaad6SPeter Xu ptep = huge_pte_offset(mm, address, psize); 6385e95a9851SMike Kravetz if (!ptep) { 6386e95a9851SMike Kravetz address |= last_addr_mask; 63878f860591SZhang, Yanmin continue; 6388e95a9851SMike Kravetz } 6389cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 63904ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 639160dfaad6SPeter Xu /* 639260dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 639360dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 639460dfaad6SPeter Xu * happened due to some reason. 639560dfaad6SPeter Xu */ 639660dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 63977da4d641SPeter Zijlstra pages++; 6398cb900f41SKirill A. Shutemov spin_unlock(ptl); 6399dff11abeSMike Kravetz shared_pmd = true; 64004ddb4d91SMike Kravetz address |= last_addr_mask; 640139dde65cSChen, Kenneth W continue; 64027da4d641SPeter Zijlstra } 6403a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6404a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6405a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6406a8bda28dSNaoya Horiguchi continue; 6407a8bda28dSNaoya Horiguchi } 6408a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6409a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 64106c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 6411a8bda28dSNaoya Horiguchi 64126c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 6413a8bda28dSNaoya Horiguchi pte_t newpte; 6414a8bda28dSNaoya Horiguchi 64156c287605SDavid Hildenbrand if (PageAnon(page)) 64166c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 64176c287605SDavid Hildenbrand swp_offset(entry)); 64186c287605SDavid Hildenbrand else 64194dd845b5SAlistair Popple entry = make_readable_migration_entry( 64204dd845b5SAlistair Popple swp_offset(entry)); 6421a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 64225a90d5a1SPeter Xu if (uffd_wp) 64235a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 64245a90d5a1SPeter Xu else if (uffd_wp_resolve) 64255a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 642618f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 6427a8bda28dSNaoya Horiguchi pages++; 6428a8bda28dSNaoya Horiguchi } 6429a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6430a8bda28dSNaoya Horiguchi continue; 6431a8bda28dSNaoya Horiguchi } 643260dfaad6SPeter Xu if (unlikely(pte_marker_uffd_wp(pte))) { 643360dfaad6SPeter Xu /* 643460dfaad6SPeter Xu * This is changing a non-present pte into a none pte, 643560dfaad6SPeter Xu * no need for huge_ptep_modify_prot_start/commit(). 643660dfaad6SPeter Xu */ 643760dfaad6SPeter Xu if (uffd_wp_resolve) 643860dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 643960dfaad6SPeter Xu } 6440a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6441023bdd00SAneesh Kumar K.V pte_t old_pte; 644279c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6443023bdd00SAneesh Kumar K.V 6444023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 644516785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 644679c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 64475a90d5a1SPeter Xu if (uffd_wp) 64485a90d5a1SPeter Xu pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); 64495a90d5a1SPeter Xu else if (uffd_wp_resolve) 64505a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6451023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 64527da4d641SPeter Zijlstra pages++; 645360dfaad6SPeter Xu } else { 645460dfaad6SPeter Xu /* None pte */ 645560dfaad6SPeter Xu if (unlikely(uffd_wp)) 645660dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 645760dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 645860dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 64598f860591SZhang, Yanmin } 6460cb900f41SKirill A. Shutemov spin_unlock(ptl); 64618f860591SZhang, Yanmin } 6462d833352aSMel Gorman /* 6463c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6464d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6465c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6466dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6467dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6468d833352aSMel Gorman */ 6469dff11abeSMike Kravetz if (shared_pmd) 6470ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6471dff11abeSMike Kravetz else 64725491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 64730f10851eSJérôme Glisse /* 64740f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 64750f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 64760f10851eSJérôme Glisse * 6477ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 64780f10851eSJérôme Glisse */ 647983cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 6480ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 64817da4d641SPeter Zijlstra 64827da4d641SPeter Zijlstra return pages << h->order; 64838f860591SZhang, Yanmin } 64848f860591SZhang, Yanmin 648533b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 648633b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6487a1e78772SMel Gorman long from, long to, 64885a6fe125SMel Gorman struct vm_area_struct *vma, 6489ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6490e4e574b7SAdam Litke { 649133b8f84aSMike Kravetz long chg, add = -1; 6492a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 649390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 64949119a41eSJoonsoo Kim struct resv_map *resv_map; 6495075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 64960db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6497e4e574b7SAdam Litke 649863489f8eSMike Kravetz /* This should never happen */ 649963489f8eSMike Kravetz if (from > to) { 650063489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 650133b8f84aSMike Kravetz return false; 650263489f8eSMike Kravetz } 650363489f8eSMike Kravetz 6504a1e78772SMel Gorman /* 650517c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 650617c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 650790481622SDavid Gibson * without using reserves 650817c9d12eSMel Gorman */ 6509ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 651033b8f84aSMike Kravetz return true; 651117c9d12eSMel Gorman 651217c9d12eSMel Gorman /* 6513a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6514a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6515a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6516a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6517a1e78772SMel Gorman */ 65189119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6519f27a5136SMike Kravetz /* 6520f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6521f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6522f27a5136SMike Kravetz * hugetlbfs_get_inode). 6523f27a5136SMike Kravetz */ 65244e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 65259119a41eSJoonsoo Kim 65260db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 65279119a41eSJoonsoo Kim 65289119a41eSJoonsoo Kim } else { 6529e9fe92aeSMina Almasry /* Private mapping. */ 65309119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 65315a6fe125SMel Gorman if (!resv_map) 653233b8f84aSMike Kravetz return false; 65335a6fe125SMel Gorman 653417c9d12eSMel Gorman chg = to - from; 653517c9d12eSMel Gorman 65365a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 65375a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 65385a6fe125SMel Gorman } 65395a6fe125SMel Gorman 654033b8f84aSMike Kravetz if (chg < 0) 6541c50ac050SDave Hansen goto out_err; 654217c9d12eSMel Gorman 654333b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 654433b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6545075a61d0SMina Almasry goto out_err; 6546075a61d0SMina Almasry 6547075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6548075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6549075a61d0SMina Almasry * of the resv_map. 6550075a61d0SMina Almasry */ 6551075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6552075a61d0SMina Almasry } 6553075a61d0SMina Almasry 65541c5ecae3SMike Kravetz /* 65551c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 65561c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 65571c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 65581c5ecae3SMike Kravetz */ 65591c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 656033b8f84aSMike Kravetz if (gbl_reserve < 0) 6561075a61d0SMina Almasry goto out_uncharge_cgroup; 656217c9d12eSMel Gorman 656317c9d12eSMel Gorman /* 656417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 656590481622SDavid Gibson * Hand the pages back to the subpool if there are not 656617c9d12eSMel Gorman */ 656733b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6568075a61d0SMina Almasry goto out_put_pages; 656917c9d12eSMel Gorman 657017c9d12eSMel Gorman /* 657117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 657217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 657317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 657417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 657517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 657617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 657717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 657817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 657917c9d12eSMel Gorman * else has to be done for private mappings here 658017c9d12eSMel Gorman */ 658133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6582075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 658333039678SMike Kravetz 65840db9d74eSMina Almasry if (unlikely(add < 0)) { 65850db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6586075a61d0SMina Almasry goto out_put_pages; 65870db9d74eSMina Almasry } else if (unlikely(chg > add)) { 658833039678SMike Kravetz /* 658933039678SMike Kravetz * pages in this range were added to the reserve 659033039678SMike Kravetz * map between region_chg and region_add. This 659133039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 659233039678SMike Kravetz * the subpool and reserve counts modified above 659333039678SMike Kravetz * based on the difference. 659433039678SMike Kravetz */ 659533039678SMike Kravetz long rsv_adjust; 659633039678SMike Kravetz 6597d85aecf2SMiaohe Lin /* 6598d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6599d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6600d85aecf2SMiaohe Lin */ 6601075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6602075a61d0SMina Almasry hstate_index(h), 6603075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6604075a61d0SMina Almasry 660533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 660633039678SMike Kravetz chg - add); 660733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6608d85aecf2SMiaohe Lin } else if (h_cg) { 6609d85aecf2SMiaohe Lin /* 6610d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6611d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6612d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6613d85aecf2SMiaohe Lin * done. 6614d85aecf2SMiaohe Lin */ 6615d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 661633039678SMike Kravetz } 661733039678SMike Kravetz } 661833b8f84aSMike Kravetz return true; 661933b8f84aSMike Kravetz 6620075a61d0SMina Almasry out_put_pages: 6621075a61d0SMina Almasry /* put back original number of pages, chg */ 6622075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6623075a61d0SMina Almasry out_uncharge_cgroup: 6624075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6625075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6626c50ac050SDave Hansen out_err: 66275e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 66280db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 66290db9d74eSMina Almasry * region_add failed or didn't run. 66300db9d74eSMina Almasry */ 66310db9d74eSMina Almasry if (chg >= 0 && add < 0) 66320db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6633f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6634f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 663533b8f84aSMike Kravetz return false; 6636a43a8c39SChen, Kenneth W } 6637a43a8c39SChen, Kenneth W 6638b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6639b5cec28dSMike Kravetz long freed) 6640a43a8c39SChen, Kenneth W { 6641a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 66424e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 66439119a41eSJoonsoo Kim long chg = 0; 664490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66451c5ecae3SMike Kravetz long gbl_reserve; 664645c682a6SKen Chen 6647f27a5136SMike Kravetz /* 6648f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6649f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6650f27a5136SMike Kravetz */ 6651b5cec28dSMike Kravetz if (resv_map) { 6652b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6653b5cec28dSMike Kravetz /* 6654b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6655b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6656b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6657b5cec28dSMike Kravetz */ 6658b5cec28dSMike Kravetz if (chg < 0) 6659b5cec28dSMike Kravetz return chg; 6660b5cec28dSMike Kravetz } 6661b5cec28dSMike Kravetz 666245c682a6SKen Chen spin_lock(&inode->i_lock); 6663e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 666445c682a6SKen Chen spin_unlock(&inode->i_lock); 666545c682a6SKen Chen 66661c5ecae3SMike Kravetz /* 66671c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 66681c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6669dddf31a4SMiaohe Lin * 6670dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6671dddf31a4SMiaohe Lin * won't go negative. 66721c5ecae3SMike Kravetz */ 66731c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 66741c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6675b5cec28dSMike Kravetz 6676b5cec28dSMike Kravetz return 0; 6677a43a8c39SChen, Kenneth W } 667893f70f90SNaoya Horiguchi 66793212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 66803212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 66813212b535SSteve Capper struct vm_area_struct *vma, 66823212b535SSteve Capper unsigned long addr, pgoff_t idx) 66833212b535SSteve Capper { 66843212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 66853212b535SSteve Capper svma->vm_start; 66863212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 66873212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 66883212b535SSteve Capper 66893212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6690de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6691de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 66923212b535SSteve Capper 66933212b535SSteve Capper /* 66943212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 66953212b535SSteve Capper * page table page. 66963212b535SSteve Capper */ 66973212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 66983212b535SSteve Capper vm_flags != svm_flags || 669907e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 67003212b535SSteve Capper return 0; 67013212b535SSteve Capper 67023212b535SSteve Capper return saddr; 67033212b535SSteve Capper } 67043212b535SSteve Capper 670531aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 67063212b535SSteve Capper { 67073212b535SSteve Capper unsigned long base = addr & PUD_MASK; 67083212b535SSteve Capper unsigned long end = base + PUD_SIZE; 67093212b535SSteve Capper 67103212b535SSteve Capper /* 67113212b535SSteve Capper * check on proper vm_flags and page table alignment 67123212b535SSteve Capper */ 6713017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 671431aafb45SNicholas Krause return true; 671531aafb45SNicholas Krause return false; 67163212b535SSteve Capper } 67173212b535SSteve Capper 6718c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6719c1991e07SPeter Xu { 6720c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 6721c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 6722c1991e07SPeter Xu return false; 6723c1991e07SPeter Xu #endif 6724c1991e07SPeter Xu return vma_shareable(vma, addr); 6725c1991e07SPeter Xu } 6726c1991e07SPeter Xu 67273212b535SSteve Capper /* 6728017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6729017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6730017b1660SMike Kravetz * shared pmd mappings. 6731017b1660SMike Kravetz */ 6732017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6733017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6734017b1660SMike Kravetz { 6735a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6736a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6737017b1660SMike Kravetz 6738a1ba9da8SLi Xinhai /* 6739f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6740f0953a1bSIngo Molnar * must be at least partially within in. 6741a1ba9da8SLi Xinhai */ 6742a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6743a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6744017b1660SMike Kravetz return; 6745017b1660SMike Kravetz 674675802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6747a1ba9da8SLi Xinhai if (*start > v_start) 6748a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6749017b1660SMike Kravetz 6750a1ba9da8SLi Xinhai if (*end < v_end) 6751a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6752017b1660SMike Kravetz } 6753017b1660SMike Kravetz 6754017b1660SMike Kravetz /* 67553212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 67563212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 67573212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 6758c0d0381aSMike Kravetz * code much cleaner. 6759c0d0381aSMike Kravetz * 67600bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 67610bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 67620bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 67630bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 67643212b535SSteve Capper */ 6765aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6766aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 67673212b535SSteve Capper { 67683212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 67693212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 67703212b535SSteve Capper vma->vm_pgoff; 67713212b535SSteve Capper struct vm_area_struct *svma; 67723212b535SSteve Capper unsigned long saddr; 67733212b535SSteve Capper pte_t *spte = NULL; 67743212b535SSteve Capper pte_t *pte; 6775cb900f41SKirill A. Shutemov spinlock_t *ptl; 67763212b535SSteve Capper 67770bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 67783212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 67793212b535SSteve Capper if (svma == vma) 67803212b535SSteve Capper continue; 67813212b535SSteve Capper 67823212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 67833212b535SSteve Capper if (saddr) { 67847868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 67857868a208SPunit Agrawal vma_mmu_pagesize(svma)); 67863212b535SSteve Capper if (spte) { 67873212b535SSteve Capper get_page(virt_to_page(spte)); 67883212b535SSteve Capper break; 67893212b535SSteve Capper } 67903212b535SSteve Capper } 67913212b535SSteve Capper } 67923212b535SSteve Capper 67933212b535SSteve Capper if (!spte) 67943212b535SSteve Capper goto out; 67953212b535SSteve Capper 67968bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6797dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 67983212b535SSteve Capper pud_populate(mm, pud, 67993212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6800c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6801dc6c9a35SKirill A. Shutemov } else { 68023212b535SSteve Capper put_page(virt_to_page(spte)); 6803dc6c9a35SKirill A. Shutemov } 6804cb900f41SKirill A. Shutemov spin_unlock(ptl); 68053212b535SSteve Capper out: 68063212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 68073212b535SSteve Capper return pte; 68083212b535SSteve Capper } 68093212b535SSteve Capper 68103212b535SSteve Capper /* 68113212b535SSteve Capper * unmap huge page backed by shared pte. 68123212b535SSteve Capper * 68133212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 68143212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 68153212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 68163212b535SSteve Capper * 6817c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 68183212b535SSteve Capper * 68193212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 68203212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 68213212b535SSteve Capper */ 682234ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 68234ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 68243212b535SSteve Capper { 68254ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 68264ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 68274ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 68283212b535SSteve Capper 682934ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 68303212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 68313212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 68323212b535SSteve Capper return 0; 68333212b535SSteve Capper 68343212b535SSteve Capper pud_clear(pud); 68353212b535SSteve Capper put_page(virt_to_page(ptep)); 6836dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 68373212b535SSteve Capper return 1; 68383212b535SSteve Capper } 6839c1991e07SPeter Xu 68409e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 6841aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6842aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 68439e5fc74cSSteve Capper { 68449e5fc74cSSteve Capper return NULL; 68459e5fc74cSSteve Capper } 6846e81f2d22SZhang Zhen 684734ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 68484ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 6849e81f2d22SZhang Zhen { 6850e81f2d22SZhang Zhen return 0; 6851e81f2d22SZhang Zhen } 6852017b1660SMike Kravetz 6853017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6854017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6855017b1660SMike Kravetz { 6856017b1660SMike Kravetz } 6857c1991e07SPeter Xu 6858c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6859c1991e07SPeter Xu { 6860c1991e07SPeter Xu return false; 6861c1991e07SPeter Xu } 68623212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 68633212b535SSteve Capper 68649e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 6865aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 68669e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 68679e5fc74cSSteve Capper { 68689e5fc74cSSteve Capper pgd_t *pgd; 6869c2febafcSKirill A. Shutemov p4d_t *p4d; 68709e5fc74cSSteve Capper pud_t *pud; 68719e5fc74cSSteve Capper pte_t *pte = NULL; 68729e5fc74cSSteve Capper 68739e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6874f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 6875f4f0a3d8SKirill A. Shutemov if (!p4d) 6876f4f0a3d8SKirill A. Shutemov return NULL; 6877c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 68789e5fc74cSSteve Capper if (pud) { 68799e5fc74cSSteve Capper if (sz == PUD_SIZE) { 68809e5fc74cSSteve Capper pte = (pte_t *)pud; 68819e5fc74cSSteve Capper } else { 68829e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 6883c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 6884aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 68859e5fc74cSSteve Capper else 68869e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 68879e5fc74cSSteve Capper } 68889e5fc74cSSteve Capper } 68894e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 68909e5fc74cSSteve Capper 68919e5fc74cSSteve Capper return pte; 68929e5fc74cSSteve Capper } 68939e5fc74cSSteve Capper 68949b19df29SPunit Agrawal /* 68959b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 68969b19df29SPunit Agrawal * entry at address @addr 68979b19df29SPunit Agrawal * 68988ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 68998ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 69009b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 69019b19df29SPunit Agrawal * table. 69029b19df29SPunit Agrawal */ 69037868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 69047868a208SPunit Agrawal unsigned long addr, unsigned long sz) 69059e5fc74cSSteve Capper { 69069e5fc74cSSteve Capper pgd_t *pgd; 6907c2febafcSKirill A. Shutemov p4d_t *p4d; 69088ac0b81aSLi Xinhai pud_t *pud; 69098ac0b81aSLi Xinhai pmd_t *pmd; 69109e5fc74cSSteve Capper 69119e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6912c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6913c2febafcSKirill A. Shutemov return NULL; 6914c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6915c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6916c2febafcSKirill A. Shutemov return NULL; 69179b19df29SPunit Agrawal 6918c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 69198ac0b81aSLi Xinhai if (sz == PUD_SIZE) 69208ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 69219e5fc74cSSteve Capper return (pte_t *)pud; 69228ac0b81aSLi Xinhai if (!pud_present(*pud)) 69238ac0b81aSLi Xinhai return NULL; 69248ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 69259b19df29SPunit Agrawal 69269e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 69278ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 69289e5fc74cSSteve Capper return (pte_t *)pmd; 69299e5fc74cSSteve Capper } 69309e5fc74cSSteve Capper 6931e95a9851SMike Kravetz /* 6932e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 6933e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 6934e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 6935e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 6936e95a9851SMike Kravetz * version of this routine. 6937e95a9851SMike Kravetz */ 6938e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 6939e95a9851SMike Kravetz { 6940e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 6941e95a9851SMike Kravetz 6942e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 6943e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 6944e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 6945e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 6946e95a9851SMike Kravetz else 6947e95a9851SMike Kravetz return 0UL; 6948e95a9851SMike Kravetz } 6949e95a9851SMike Kravetz 6950e95a9851SMike Kravetz #else 6951e95a9851SMike Kravetz 6952e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 6953e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 6954e95a9851SMike Kravetz { 69554ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 69564ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 69574ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 69584ddb4d91SMike Kravetz #endif 6959e95a9851SMike Kravetz return 0UL; 6960e95a9851SMike Kravetz } 6961e95a9851SMike Kravetz 696261f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 696361f77edaSNaoya Horiguchi 696461f77edaSNaoya Horiguchi /* 696561f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 696661f77edaSNaoya Horiguchi * behavior. 696761f77edaSNaoya Horiguchi */ 696861f77edaSNaoya Horiguchi struct page * __weak 696961f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 697061f77edaSNaoya Horiguchi int write) 697161f77edaSNaoya Horiguchi { 697261f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 697361f77edaSNaoya Horiguchi } 697461f77edaSNaoya Horiguchi 697561f77edaSNaoya Horiguchi struct page * __weak 69764dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 69774dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 69784dc71451SAneesh Kumar K.V { 69794dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 69804dc71451SAneesh Kumar K.V return NULL; 69814dc71451SAneesh Kumar K.V } 69824dc71451SAneesh Kumar K.V 69834dc71451SAneesh Kumar K.V struct page * __weak 69849e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6985e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 69869e5fc74cSSteve Capper { 6987e66f17ffSNaoya Horiguchi struct page *page = NULL; 6988e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6989c9d398faSNaoya Horiguchi pte_t pte; 69903faa52c0SJohn Hubbard 69918909691bSDavid Hildenbrand /* 69928909691bSDavid Hildenbrand * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 69938909691bSDavid Hildenbrand * follow_hugetlb_page(). 69948909691bSDavid Hildenbrand */ 69958909691bSDavid Hildenbrand if (WARN_ON_ONCE(flags & FOLL_PIN)) 69963faa52c0SJohn Hubbard return NULL; 69973faa52c0SJohn Hubbard 6998e66f17ffSNaoya Horiguchi retry: 6999e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 7000e66f17ffSNaoya Horiguchi spin_lock(ptl); 7001e66f17ffSNaoya Horiguchi /* 7002e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 7003e66f17ffSNaoya Horiguchi * unmapped from other threads. 7004e66f17ffSNaoya Horiguchi */ 7005e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 7006e66f17ffSNaoya Horiguchi goto out; 7007c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 7008c9d398faSNaoya Horiguchi if (pte_present(pte)) { 700997534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 70103faa52c0SJohn Hubbard /* 70113faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 70123faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 70133faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 70143faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 70153faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 70163faa52c0SJohn Hubbard * unless the page refcount overflowed: 70173faa52c0SJohn Hubbard */ 70183faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 70193faa52c0SJohn Hubbard page = NULL; 70203faa52c0SJohn Hubbard goto out; 70213faa52c0SJohn Hubbard } 7022e66f17ffSNaoya Horiguchi } else { 7023c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 7024e66f17ffSNaoya Horiguchi spin_unlock(ptl); 7025ad1ac596SMiaohe Lin __migration_entry_wait_huge((pte_t *)pmd, ptl); 7026e66f17ffSNaoya Horiguchi goto retry; 7027e66f17ffSNaoya Horiguchi } 7028e66f17ffSNaoya Horiguchi /* 7029e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 7030e66f17ffSNaoya Horiguchi * follow_page_mask(). 7031e66f17ffSNaoya Horiguchi */ 7032e66f17ffSNaoya Horiguchi } 7033e66f17ffSNaoya Horiguchi out: 7034e66f17ffSNaoya Horiguchi spin_unlock(ptl); 70359e5fc74cSSteve Capper return page; 70369e5fc74cSSteve Capper } 70379e5fc74cSSteve Capper 703861f77edaSNaoya Horiguchi struct page * __weak 70399e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 7040e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 70419e5fc74cSSteve Capper { 70423a194f3fSNaoya Horiguchi struct page *page = NULL; 70433a194f3fSNaoya Horiguchi spinlock_t *ptl; 70443a194f3fSNaoya Horiguchi pte_t pte; 70453a194f3fSNaoya Horiguchi 70463a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(flags & FOLL_PIN)) 7047e66f17ffSNaoya Horiguchi return NULL; 70489e5fc74cSSteve Capper 70493a194f3fSNaoya Horiguchi retry: 70503a194f3fSNaoya Horiguchi ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud); 70513a194f3fSNaoya Horiguchi if (!pud_huge(*pud)) 70523a194f3fSNaoya Horiguchi goto out; 70533a194f3fSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pud); 70543a194f3fSNaoya Horiguchi if (pte_present(pte)) { 70553a194f3fSNaoya Horiguchi page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 70563a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 70573a194f3fSNaoya Horiguchi page = NULL; 70583a194f3fSNaoya Horiguchi goto out; 70593a194f3fSNaoya Horiguchi } 70603a194f3fSNaoya Horiguchi } else { 70613a194f3fSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 70623a194f3fSNaoya Horiguchi spin_unlock(ptl); 70633a194f3fSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pud, ptl); 70643a194f3fSNaoya Horiguchi goto retry; 70653a194f3fSNaoya Horiguchi } 70663a194f3fSNaoya Horiguchi /* 70673a194f3fSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 70683a194f3fSNaoya Horiguchi * follow_page_mask(). 70693a194f3fSNaoya Horiguchi */ 70703a194f3fSNaoya Horiguchi } 70713a194f3fSNaoya Horiguchi out: 70723a194f3fSNaoya Horiguchi spin_unlock(ptl); 70733a194f3fSNaoya Horiguchi return page; 70749e5fc74cSSteve Capper } 70759e5fc74cSSteve Capper 7076faaa5b62SAnshuman Khandual struct page * __weak 7077faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 7078faaa5b62SAnshuman Khandual { 70793faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 7080faaa5b62SAnshuman Khandual return NULL; 7081faaa5b62SAnshuman Khandual 7082faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 7083faaa5b62SAnshuman Khandual } 7084faaa5b62SAnshuman Khandual 70857ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 708631caf665SNaoya Horiguchi { 70877ce82f4cSMiaohe Lin int ret = 0; 7088bcc54222SNaoya Horiguchi 7089db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 70908f251a3dSMike Kravetz if (!PageHeadHuge(page) || 70918f251a3dSMike Kravetz !HPageMigratable(page) || 70920eb2df2bSMuchun Song !get_page_unless_zero(page)) { 70937ce82f4cSMiaohe Lin ret = -EBUSY; 7094bcc54222SNaoya Horiguchi goto unlock; 7095bcc54222SNaoya Horiguchi } 70968f251a3dSMike Kravetz ClearHPageMigratable(page); 709731caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7098bcc54222SNaoya Horiguchi unlock: 7099db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7100bcc54222SNaoya Horiguchi return ret; 710131caf665SNaoya Horiguchi } 710231caf665SNaoya Horiguchi 710325182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 710425182f05SNaoya Horiguchi { 710525182f05SNaoya Horiguchi int ret = 0; 710625182f05SNaoya Horiguchi 710725182f05SNaoya Horiguchi *hugetlb = false; 710825182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 710925182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 711025182f05SNaoya Horiguchi *hugetlb = true; 7111b283d983SNaoya Horiguchi if (HPageFreed(page)) 7112b283d983SNaoya Horiguchi ret = 0; 7113b283d983SNaoya Horiguchi else if (HPageMigratable(page)) 711425182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 71150ed950d1SNaoya Horiguchi else 71160ed950d1SNaoya Horiguchi ret = -EBUSY; 711725182f05SNaoya Horiguchi } 711825182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 711925182f05SNaoya Horiguchi return ret; 712025182f05SNaoya Horiguchi } 712125182f05SNaoya Horiguchi 7122405ce051SNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags) 7123405ce051SNaoya Horiguchi { 7124405ce051SNaoya Horiguchi int ret; 7125405ce051SNaoya Horiguchi 7126405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7127405ce051SNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags); 7128405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7129405ce051SNaoya Horiguchi return ret; 7130405ce051SNaoya Horiguchi } 7131405ce051SNaoya Horiguchi 713231caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 713331caf665SNaoya Horiguchi { 7134db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 71358f251a3dSMike Kravetz SetHPageMigratable(page); 713631caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7137db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 713831caf665SNaoya Horiguchi put_page(page); 713931caf665SNaoya Horiguchi } 7140ab5ac90aSMichal Hocko 7141ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 7142ab5ac90aSMichal Hocko { 7143ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 7144ab5ac90aSMichal Hocko 7145ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 7146ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 7147ab5ac90aSMichal Hocko 7148ab5ac90aSMichal Hocko /* 7149ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 7150ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7151ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7152ab5ac90aSMichal Hocko * the temporary status. 7153ab5ac90aSMichal Hocko * 7154ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7155ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7156ab5ac90aSMichal Hocko * the per-node's. 7157ab5ac90aSMichal Hocko */ 71589157c311SMike Kravetz if (HPageTemporary(newpage)) { 7159ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 7160ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 7161ab5ac90aSMichal Hocko 71629157c311SMike Kravetz SetHPageTemporary(oldpage); 71639157c311SMike Kravetz ClearHPageTemporary(newpage); 7164ab5ac90aSMichal Hocko 71655af1ab1dSMiaohe Lin /* 71665af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 71675af1ab1dSMiaohe Lin * when we do not cross the node. 71685af1ab1dSMiaohe Lin */ 71695af1ab1dSMiaohe Lin if (new_nid == old_nid) 71705af1ab1dSMiaohe Lin return; 7171db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7172ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7173ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7174ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7175ab5ac90aSMichal Hocko } 7176db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7177ab5ac90aSMichal Hocko } 7178ab5ac90aSMichal Hocko } 7179cf11e85fSRoman Gushchin 71806dfeaff9SPeter Xu /* 71816dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 71826dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 71836dfeaff9SPeter Xu */ 71846dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 71856dfeaff9SPeter Xu { 71866dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 71876dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 71886dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 71896dfeaff9SPeter Xu struct mmu_notifier_range range; 71906dfeaff9SPeter Xu unsigned long address, start, end; 71916dfeaff9SPeter Xu spinlock_t *ptl; 71926dfeaff9SPeter Xu pte_t *ptep; 71936dfeaff9SPeter Xu 71946dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 71956dfeaff9SPeter Xu return; 71966dfeaff9SPeter Xu 71976dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 71986dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 71996dfeaff9SPeter Xu 72006dfeaff9SPeter Xu if (start >= end) 72016dfeaff9SPeter Xu return; 72026dfeaff9SPeter Xu 72039c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 72046dfeaff9SPeter Xu /* 72056dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 72066dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 72076dfeaff9SPeter Xu */ 72086dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 72096dfeaff9SPeter Xu start, end); 72106dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 72116dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 72126dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 72136dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 72146dfeaff9SPeter Xu if (!ptep) 72156dfeaff9SPeter Xu continue; 72166dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 72174ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 72186dfeaff9SPeter Xu spin_unlock(ptl); 72196dfeaff9SPeter Xu } 72206dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 72216dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 72226dfeaff9SPeter Xu /* 72236dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7224ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 72256dfeaff9SPeter Xu */ 72266dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 72276dfeaff9SPeter Xu } 72286dfeaff9SPeter Xu 7229cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7230cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7231cf11e85fSRoman Gushchin 7232cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7233cf11e85fSRoman Gushchin { 723438e719abSBaolin Wang int nid, count = 0; 723538e719abSBaolin Wang unsigned long tmp; 723638e719abSBaolin Wang char *s = p; 723738e719abSBaolin Wang 723838e719abSBaolin Wang while (*s) { 723938e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 724038e719abSBaolin Wang break; 724138e719abSBaolin Wang 724238e719abSBaolin Wang if (s[count] == ':') { 7243f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 724438e719abSBaolin Wang break; 7245f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 724638e719abSBaolin Wang 724738e719abSBaolin Wang s += count + 1; 724838e719abSBaolin Wang tmp = memparse(s, &s); 724938e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 725038e719abSBaolin Wang hugetlb_cma_size += tmp; 725138e719abSBaolin Wang 725238e719abSBaolin Wang /* 725338e719abSBaolin Wang * Skip the separator if have one, otherwise 725438e719abSBaolin Wang * break the parsing. 725538e719abSBaolin Wang */ 725638e719abSBaolin Wang if (*s == ',') 725738e719abSBaolin Wang s++; 725838e719abSBaolin Wang else 725938e719abSBaolin Wang break; 726038e719abSBaolin Wang } else { 7261cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 726238e719abSBaolin Wang break; 726338e719abSBaolin Wang } 726438e719abSBaolin Wang } 726538e719abSBaolin Wang 7266cf11e85fSRoman Gushchin return 0; 7267cf11e85fSRoman Gushchin } 7268cf11e85fSRoman Gushchin 7269cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7270cf11e85fSRoman Gushchin 7271cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7272cf11e85fSRoman Gushchin { 7273cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 727438e719abSBaolin Wang bool node_specific_cma_alloc = false; 7275cf11e85fSRoman Gushchin int nid; 7276cf11e85fSRoman Gushchin 7277cf11e85fSRoman Gushchin cma_reserve_called = true; 7278cf11e85fSRoman Gushchin 7279cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7280cf11e85fSRoman Gushchin return; 7281cf11e85fSRoman Gushchin 728238e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 728338e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 728438e719abSBaolin Wang continue; 728538e719abSBaolin Wang 728630a51400SPeng Liu if (!node_online(nid)) { 728738e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 728838e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 728938e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 729038e719abSBaolin Wang continue; 729138e719abSBaolin Wang } 729238e719abSBaolin Wang 729338e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 729438e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 729538e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 729638e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 729738e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 729838e719abSBaolin Wang } else { 729938e719abSBaolin Wang node_specific_cma_alloc = true; 730038e719abSBaolin Wang } 730138e719abSBaolin Wang } 730238e719abSBaolin Wang 730338e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 730438e719abSBaolin Wang if (!hugetlb_cma_size) 730538e719abSBaolin Wang return; 730638e719abSBaolin Wang 7307cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7308cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7309cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7310a01f4390SMike Kravetz hugetlb_cma_size = 0; 7311cf11e85fSRoman Gushchin return; 7312cf11e85fSRoman Gushchin } 7313cf11e85fSRoman Gushchin 731438e719abSBaolin Wang if (!node_specific_cma_alloc) { 7315cf11e85fSRoman Gushchin /* 7316cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7317cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7318cf11e85fSRoman Gushchin */ 7319cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7320cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7321cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 732238e719abSBaolin Wang } 7323cf11e85fSRoman Gushchin 7324cf11e85fSRoman Gushchin reserved = 0; 732530a51400SPeng Liu for_each_online_node(nid) { 7326cf11e85fSRoman Gushchin int res; 73272281f797SBarry Song char name[CMA_MAX_NAME]; 7328cf11e85fSRoman Gushchin 732938e719abSBaolin Wang if (node_specific_cma_alloc) { 733038e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 733138e719abSBaolin Wang continue; 733238e719abSBaolin Wang 733338e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 733438e719abSBaolin Wang } else { 7335cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 733638e719abSBaolin Wang } 733738e719abSBaolin Wang 7338cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7339cf11e85fSRoman Gushchin 73402281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7341a01f4390SMike Kravetz /* 7342a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7343a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7344a01f4390SMike Kravetz * huge page demotion. 7345a01f4390SMike Kravetz */ 7346a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7347a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 734829d0f41dSBarry Song 0, false, name, 7349cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7350cf11e85fSRoman Gushchin if (res) { 7351cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7352cf11e85fSRoman Gushchin res, nid); 7353cf11e85fSRoman Gushchin continue; 7354cf11e85fSRoman Gushchin } 7355cf11e85fSRoman Gushchin 7356cf11e85fSRoman Gushchin reserved += size; 7357cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7358cf11e85fSRoman Gushchin size / SZ_1M, nid); 7359cf11e85fSRoman Gushchin 7360cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7361cf11e85fSRoman Gushchin break; 7362cf11e85fSRoman Gushchin } 7363a01f4390SMike Kravetz 7364a01f4390SMike Kravetz if (!reserved) 7365a01f4390SMike Kravetz /* 7366a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7367a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7368a01f4390SMike Kravetz */ 7369a01f4390SMike Kravetz hugetlb_cma_size = 0; 7370cf11e85fSRoman Gushchin } 7371cf11e85fSRoman Gushchin 7372263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7373cf11e85fSRoman Gushchin { 7374cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7375cf11e85fSRoman Gushchin return; 7376cf11e85fSRoman Gushchin 7377cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7378cf11e85fSRoman Gushchin } 7379cf11e85fSRoman Gushchin 7380cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7381