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 { 45934665341SMiaohe Lin 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 /* 4660db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4670db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4680db9d74eSMina Almasry * 4690db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4700db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4710db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4720db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4730db9d74eSMina Almasry */ 4740db9d74eSMina Almasry while (resv->region_cache_count < 4750db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4760db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4770db9d74eSMina Almasry resv->region_cache_count; 4780db9d74eSMina Almasry 4790db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 480f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4810db9d74eSMina Almasry * needing to allocate for regions_needed. 4820db9d74eSMina Almasry */ 4830db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4840db9d74eSMina Almasry 4850db9d74eSMina Almasry spin_unlock(&resv->lock); 4860db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4870db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4880db9d74eSMina Almasry if (!trg) 4890db9d74eSMina Almasry goto out_of_memory; 4900db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4910db9d74eSMina Almasry } 4920db9d74eSMina Almasry 4930db9d74eSMina Almasry spin_lock(&resv->lock); 4940db9d74eSMina Almasry 495d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 496d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4970db9d74eSMina Almasry } 4980db9d74eSMina Almasry 4990db9d74eSMina Almasry return 0; 5000db9d74eSMina Almasry 5010db9d74eSMina Almasry out_of_memory: 5020db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 503d75c6af9SMina Almasry list_del(&rg->link); 504d75c6af9SMina Almasry kfree(rg); 505d75c6af9SMina Almasry } 5060db9d74eSMina Almasry return -ENOMEM; 507d75c6af9SMina Almasry } 508d75c6af9SMina Almasry 5091dd308a7SMike Kravetz /* 5101dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5110db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5120db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5130db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5140db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5150db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 516cf3ad20bSMike Kravetz * 5170db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5180db9d74eSMina Almasry * 5190db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5200db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5217c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5220db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5230db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5240db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5251dd308a7SMike Kravetz */ 5260db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 527075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 528075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 52996822904SAndy Whitcroft { 5300db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53196822904SAndy Whitcroft 5327b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5330db9d74eSMina Almasry retry: 5340db9d74eSMina Almasry 5350db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 536972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 537972a3da3SWei Yang &actual_regions_needed); 53896822904SAndy Whitcroft 5395e911373SMike Kravetz /* 5400db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5410db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5420db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5430db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5440db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5450db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5460db9d74eSMina Almasry * operation. 5475e911373SMike Kravetz */ 5480db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5490db9d74eSMina Almasry resv->region_cache_count < 5500db9d74eSMina Almasry resv->adds_in_progress + 5510db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5520db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5530db9d74eSMina Almasry * allocate file_region entries. 5540db9d74eSMina Almasry */ 5550db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5565e911373SMike Kravetz 5570db9d74eSMina Almasry if (allocate_file_region_entries( 5580db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5590db9d74eSMina Almasry return -ENOMEM; 5605e911373SMike Kravetz } 5615e911373SMike Kravetz 5620db9d74eSMina Almasry goto retry; 5630db9d74eSMina Almasry } 564cf3ad20bSMike Kravetz 565972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5660db9d74eSMina Almasry 5670db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5680db9d74eSMina Almasry 5697b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 570cf3ad20bSMike Kravetz return add; 57196822904SAndy Whitcroft } 57296822904SAndy Whitcroft 5731dd308a7SMike Kravetz /* 5741dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5751dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5761dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5771dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5781dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5791dd308a7SMike Kravetz * not change the number of huge pages represented by the 5800db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5810db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5820db9d74eSMina Almasry * file_region structure is added. 5830db9d74eSMina Almasry * 5840db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5850db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5860db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5875e911373SMike Kravetz * 5885e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5895e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5905e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5915e911373SMike Kravetz * is needed and can not be allocated. 5921dd308a7SMike Kravetz */ 5930db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5940db9d74eSMina Almasry long *out_regions_needed) 59596822904SAndy Whitcroft { 59696822904SAndy Whitcroft long chg = 0; 59796822904SAndy Whitcroft 5987b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5995e911373SMike Kravetz 600972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 601075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 602972a3da3SWei Yang out_regions_needed); 6035e911373SMike Kravetz 6040db9d74eSMina Almasry if (*out_regions_needed == 0) 6050db9d74eSMina Almasry *out_regions_needed = 1; 6065e911373SMike Kravetz 6070db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6085e911373SMike Kravetz return -ENOMEM; 6095e911373SMike Kravetz 6100db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61196822904SAndy Whitcroft 6127b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61396822904SAndy Whitcroft return chg; 61496822904SAndy Whitcroft } 61596822904SAndy Whitcroft 6161dd308a7SMike Kravetz /* 6175e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6185e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6195e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6205e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6210db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6220db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6230db9d74eSMina Almasry * the adds_in_progress counter. 6245e911373SMike Kravetz * 6255e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6265e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6275e911373SMike Kravetz * arguments will match the associated region_chg call. 6285e911373SMike Kravetz */ 6290db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6300db9d74eSMina Almasry long regions_needed) 6315e911373SMike Kravetz { 6325e911373SMike Kravetz spin_lock(&resv->lock); 6335e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6340db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6355e911373SMike Kravetz spin_unlock(&resv->lock); 6365e911373SMike Kravetz } 6375e911373SMike Kravetz 6385e911373SMike Kravetz /* 639feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 640feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 641feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 642feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 643feba16e2SMike Kravetz * 644feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 645feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 646feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 647feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 648feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 649feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 650feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6511dd308a7SMike Kravetz */ 652feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65396822904SAndy Whitcroft { 6541406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65596822904SAndy Whitcroft struct file_region *rg, *trg; 656feba16e2SMike Kravetz struct file_region *nrg = NULL; 657feba16e2SMike Kravetz long del = 0; 65896822904SAndy Whitcroft 659feba16e2SMike Kravetz retry: 6607b24d861SDavidlohr Bueso spin_lock(&resv->lock); 661feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 662dbe409e4SMike Kravetz /* 663dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 664dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 665dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 666dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 667dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 668dbe409e4SMike Kravetz */ 669dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 670feba16e2SMike Kravetz continue; 671dbe409e4SMike Kravetz 672feba16e2SMike Kravetz if (rg->from >= t) 67396822904SAndy Whitcroft break; 67496822904SAndy Whitcroft 675feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 676feba16e2SMike Kravetz /* 677feba16e2SMike Kravetz * Check for an entry in the cache before dropping 678feba16e2SMike Kravetz * lock and attempting allocation. 679feba16e2SMike Kravetz */ 680feba16e2SMike Kravetz if (!nrg && 681feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 682feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 683feba16e2SMike Kravetz struct file_region, 684feba16e2SMike Kravetz link); 685feba16e2SMike Kravetz list_del(&nrg->link); 686feba16e2SMike Kravetz resv->region_cache_count--; 68796822904SAndy Whitcroft } 68896822904SAndy Whitcroft 689feba16e2SMike Kravetz if (!nrg) { 690feba16e2SMike Kravetz spin_unlock(&resv->lock); 691feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 692feba16e2SMike Kravetz if (!nrg) 693feba16e2SMike Kravetz return -ENOMEM; 694feba16e2SMike Kravetz goto retry; 695feba16e2SMike Kravetz } 696feba16e2SMike Kravetz 697feba16e2SMike Kravetz del += t - f; 69879aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 699d85aecf2SMiaohe Lin resv, rg, t - f, false); 700feba16e2SMike Kravetz 701feba16e2SMike Kravetz /* New entry for end of split region */ 702feba16e2SMike Kravetz nrg->from = t; 703feba16e2SMike Kravetz nrg->to = rg->to; 704075a61d0SMina Almasry 705075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 706075a61d0SMina Almasry 707feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 708feba16e2SMike Kravetz 709feba16e2SMike Kravetz /* Original entry is trimmed */ 710feba16e2SMike Kravetz rg->to = f; 711feba16e2SMike Kravetz 712feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 713feba16e2SMike Kravetz nrg = NULL; 71496822904SAndy Whitcroft break; 715feba16e2SMike Kravetz } 716feba16e2SMike Kravetz 717feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 718feba16e2SMike Kravetz del += rg->to - rg->from; 719075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 720d85aecf2SMiaohe Lin rg->to - rg->from, true); 72196822904SAndy Whitcroft list_del(&rg->link); 72296822904SAndy Whitcroft kfree(rg); 723feba16e2SMike Kravetz continue; 72496822904SAndy Whitcroft } 7257b24d861SDavidlohr Bueso 726feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 727075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 728d85aecf2SMiaohe Lin t - rg->from, false); 729075a61d0SMina Almasry 73079aa925bSMike Kravetz del += t - rg->from; 73179aa925bSMike Kravetz rg->from = t; 73279aa925bSMike Kravetz } else { /* Trim end of region */ 733075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 734d85aecf2SMiaohe Lin rg->to - f, false); 73579aa925bSMike Kravetz 73679aa925bSMike Kravetz del += rg->to - f; 73779aa925bSMike Kravetz rg->to = f; 738feba16e2SMike Kravetz } 739feba16e2SMike Kravetz } 740feba16e2SMike Kravetz 7417b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 742feba16e2SMike Kravetz kfree(nrg); 743feba16e2SMike Kravetz return del; 74496822904SAndy Whitcroft } 74596822904SAndy Whitcroft 7461dd308a7SMike Kravetz /* 747b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 748b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 749b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 750b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 751b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 752b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 753b5cec28dSMike Kravetz * counts. 754b5cec28dSMike Kravetz */ 75572e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 756b5cec28dSMike Kravetz { 757b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 758b5cec28dSMike Kravetz long rsv_adjust; 759da56388cSMiaohe Lin bool reserved = false; 760b5cec28dSMike Kravetz 761b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 762da56388cSMiaohe Lin if (rsv_adjust > 0) { 763b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 764b5cec28dSMike Kravetz 765da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 766da56388cSMiaohe Lin reserved = true; 767da56388cSMiaohe Lin } else if (!rsv_adjust) { 768da56388cSMiaohe Lin reserved = true; 769b5cec28dSMike Kravetz } 770da56388cSMiaohe Lin 771da56388cSMiaohe Lin if (!reserved) 772da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 773b5cec28dSMike Kravetz } 774b5cec28dSMike Kravetz 775b5cec28dSMike Kravetz /* 7761dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7771dd308a7SMike Kravetz * that intersect with the range [f, t). 7781dd308a7SMike Kravetz */ 7791406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78084afd99bSAndy Whitcroft { 7811406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78284afd99bSAndy Whitcroft struct file_region *rg; 78384afd99bSAndy Whitcroft long chg = 0; 78484afd99bSAndy Whitcroft 7857b24d861SDavidlohr Bueso spin_lock(&resv->lock); 78684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 78784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 788f2135a4aSWang Sheng-Hui long seg_from; 789f2135a4aSWang Sheng-Hui long seg_to; 79084afd99bSAndy Whitcroft 79184afd99bSAndy Whitcroft if (rg->to <= f) 79284afd99bSAndy Whitcroft continue; 79384afd99bSAndy Whitcroft if (rg->from >= t) 79484afd99bSAndy Whitcroft break; 79584afd99bSAndy Whitcroft 79684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 79784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 79884afd99bSAndy Whitcroft 79984afd99bSAndy Whitcroft chg += seg_to - seg_from; 80084afd99bSAndy Whitcroft } 8017b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80284afd99bSAndy Whitcroft 80384afd99bSAndy Whitcroft return chg; 80484afd99bSAndy Whitcroft } 80584afd99bSAndy Whitcroft 80696822904SAndy Whitcroft /* 807e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 808e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 809e7c4b0bfSAndy Whitcroft */ 810a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 811a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 812e7c4b0bfSAndy Whitcroft { 813a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 814a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 815e7c4b0bfSAndy Whitcroft } 816e7c4b0bfSAndy Whitcroft 8170fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8180fe6e20bSNaoya Horiguchi unsigned long address) 8190fe6e20bSNaoya Horiguchi { 8200fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8210fe6e20bSNaoya Horiguchi } 822dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8230fe6e20bSNaoya Horiguchi 82484afd99bSAndy Whitcroft /* 82508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 82608fba699SMel Gorman * cases this will be same size as used by the page table entries. 82708fba699SMel Gorman */ 82808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 82908fba699SMel Gorman { 83005ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83105ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83208fba699SMel Gorman return PAGE_SIZE; 83308fba699SMel Gorman } 834f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83508fba699SMel Gorman 83608fba699SMel Gorman /* 8373340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8383340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 83909135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84009135cc5SDan Williams * version of this symbol is required. 8413340289dSMel Gorman */ 84209135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8433340289dSMel Gorman { 8443340289dSMel Gorman return vma_kernel_pagesize(vma); 8453340289dSMel Gorman } 8463340289dSMel Gorman 8473340289dSMel Gorman /* 84884afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 84984afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85084afd99bSAndy Whitcroft * alignment. 85184afd99bSAndy Whitcroft */ 85284afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85384afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85404f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85584afd99bSAndy Whitcroft 856a1e78772SMel Gorman /* 857a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 858a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 859a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 860a1e78772SMel Gorman * 861a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 862a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 863a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 864a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86584afd99bSAndy Whitcroft * 86684afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 86784afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 86884afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 86984afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87084afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87184afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87284afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87384afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 874a1e78772SMel Gorman */ 875e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 876e7c4b0bfSAndy Whitcroft { 877e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 878e7c4b0bfSAndy Whitcroft } 879e7c4b0bfSAndy Whitcroft 880e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 881e7c4b0bfSAndy Whitcroft unsigned long value) 882e7c4b0bfSAndy Whitcroft { 883e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 884e7c4b0bfSAndy Whitcroft } 885e7c4b0bfSAndy Whitcroft 886e9fe92aeSMina Almasry static void 887e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 888e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 889e9fe92aeSMina Almasry struct hstate *h) 890e9fe92aeSMina Almasry { 891e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 892e9fe92aeSMina Almasry if (!h_cg || !h) { 893e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 894e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 895e9fe92aeSMina Almasry resv_map->css = NULL; 896e9fe92aeSMina Almasry } else { 897e9fe92aeSMina Almasry resv_map->reservation_counter = 898e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 899e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 900e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 901e9fe92aeSMina Almasry } 902e9fe92aeSMina Almasry #endif 903e9fe92aeSMina Almasry } 904e9fe92aeSMina Almasry 9059119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 90684afd99bSAndy Whitcroft { 90784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9085e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9095e911373SMike Kravetz 9105e911373SMike Kravetz if (!resv_map || !rg) { 9115e911373SMike Kravetz kfree(resv_map); 9125e911373SMike Kravetz kfree(rg); 91384afd99bSAndy Whitcroft return NULL; 9145e911373SMike Kravetz } 91584afd99bSAndy Whitcroft 91684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9177b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 91884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 91984afd99bSAndy Whitcroft 9205e911373SMike Kravetz resv_map->adds_in_progress = 0; 921e9fe92aeSMina Almasry /* 922e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 923e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 924e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 925e9fe92aeSMina Almasry * reservations are to be un-charged from here. 926e9fe92aeSMina Almasry */ 927e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9285e911373SMike Kravetz 9295e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9305e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9315e911373SMike Kravetz resv_map->region_cache_count = 1; 9325e911373SMike Kravetz 93384afd99bSAndy Whitcroft return resv_map; 93484afd99bSAndy Whitcroft } 93584afd99bSAndy Whitcroft 9369119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 93784afd99bSAndy Whitcroft { 93884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9395e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9405e911373SMike Kravetz struct file_region *rg, *trg; 94184afd99bSAndy Whitcroft 94284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 943feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9445e911373SMike Kravetz 9455e911373SMike Kravetz /* ... and any entries left in the cache */ 9465e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9475e911373SMike Kravetz list_del(&rg->link); 9485e911373SMike Kravetz kfree(rg); 9495e911373SMike Kravetz } 9505e911373SMike Kravetz 9515e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9525e911373SMike Kravetz 95384afd99bSAndy Whitcroft kfree(resv_map); 95484afd99bSAndy Whitcroft } 95584afd99bSAndy Whitcroft 9564e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9574e35f483SJoonsoo Kim { 958f27a5136SMike Kravetz /* 959f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 960f27a5136SMike Kravetz * address space within the inode. This original address space 961f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 962f27a5136SMike Kravetz * address space embedded within the inode. 963f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 964f27a5136SMike Kravetz * this may not be true for device special inodes. 965f27a5136SMike Kravetz */ 966f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9674e35f483SJoonsoo Kim } 9684e35f483SJoonsoo Kim 96984afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 970a1e78772SMel Gorman { 97181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9724e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9734e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9744e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9754e35f483SJoonsoo Kim 9764e35f483SJoonsoo Kim return inode_resv_map(inode); 9774e35f483SJoonsoo Kim 9784e35f483SJoonsoo Kim } else { 97984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9814e35f483SJoonsoo Kim } 982a1e78772SMel Gorman } 983a1e78772SMel Gorman 98484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 985a1e78772SMel Gorman { 98681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 988a1e78772SMel Gorman 98984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99104f2cbe3SMel Gorman } 99204f2cbe3SMel Gorman 99304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99404f2cbe3SMel Gorman { 99581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 997e7c4b0bfSAndy Whitcroft 998e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 99904f2cbe3SMel Gorman } 100004f2cbe3SMel Gorman 100104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100204f2cbe3SMel Gorman { 100381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1004e7c4b0bfSAndy Whitcroft 1005e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1006a1e78772SMel Gorman } 1007a1e78772SMel Gorman 100804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 1009a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1010a1e78772SMel Gorman { 101181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1012f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1013a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1014a1e78772SMel Gorman } 1015a1e78772SMel Gorman 1016550a7d60SMina Almasry /* 1017550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1018550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1019550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1020550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1021550a7d60SMina Almasry * reservation. 1022550a7d60SMina Almasry */ 1023550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1024550a7d60SMina Almasry { 1025550a7d60SMina Almasry /* 1026550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1027550a7d60SMina Almasry * It has already been transferred to new_vma. 1028550a7d60SMina Almasry * 1029550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1030550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1031550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1032550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1033550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1034550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1035550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1036550a7d60SMina Almasry */ 1037550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1038550a7d60SMina Almasry 1039afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1040afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1041550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1042afe041c2SBui Quang Minh } 1043550a7d60SMina Almasry 1044550a7d60SMina Almasry reset_vma_resv_huge_pages(vma); 1045550a7d60SMina Almasry } 1046550a7d60SMina Almasry 1047a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1048559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1049a1e78772SMel Gorman { 1050af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1051af0ed73eSJoonsoo Kim /* 1052af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1053af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1054af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1055af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1056af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1057af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1058af0ed73eSJoonsoo Kim * properly, so add work-around here. 1059af0ed73eSJoonsoo Kim */ 1060af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1061559ec2f8SNicholas Krause return true; 1062af0ed73eSJoonsoo Kim else 1063559ec2f8SNicholas Krause return false; 1064af0ed73eSJoonsoo Kim } 1065a63884e9SJoonsoo Kim 1066a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10671fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10681fb1b0e9SMike Kravetz /* 10691fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10701fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10711fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10727c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10731fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10741fb1b0e9SMike Kravetz */ 10751fb1b0e9SMike Kravetz if (chg) 10761fb1b0e9SMike Kravetz return false; 10771fb1b0e9SMike Kravetz else 1078559ec2f8SNicholas Krause return true; 10791fb1b0e9SMike Kravetz } 1080a63884e9SJoonsoo Kim 1081a63884e9SJoonsoo Kim /* 1082a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1083a63884e9SJoonsoo Kim * private mappings. 1084a63884e9SJoonsoo Kim */ 108567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 108667961f9dSMike Kravetz /* 108767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 108867961f9dSMike Kravetz * could have been performed on the private mapping. 108967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 109067961f9dSMike Kravetz * actually exist or were previously consumed. 109167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 109267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 109367961f9dSMike Kravetz * private mappings has different (opposite) semantics 109467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 109567961f9dSMike Kravetz * has already taken this difference in semantics into 109667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 109767961f9dSMike Kravetz * as in the shared case above. Code could easily be 109867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 109967961f9dSMike Kravetz * subtle differences. 110067961f9dSMike Kravetz */ 110167961f9dSMike Kravetz if (chg) 110267961f9dSMike Kravetz return false; 110367961f9dSMike Kravetz else 1104559ec2f8SNicholas Krause return true; 110567961f9dSMike Kravetz } 1106a63884e9SJoonsoo Kim 1107559ec2f8SNicholas Krause return false; 1108a1e78772SMel Gorman } 1109a1e78772SMel Gorman 1110a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11111da177e4SLinus Torvalds { 11121da177e4SLinus Torvalds int nid = page_to_nid(page); 11139487ca60SMike Kravetz 11149487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1115b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1116b65a4edaSMike Kravetz 11170edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1118a5516438SAndi Kleen h->free_huge_pages++; 1119a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11206c037149SMike Kravetz SetHPageFreed(page); 11211da177e4SLinus Torvalds } 11221da177e4SLinus Torvalds 112394310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1124bf50bab2SNaoya Horiguchi { 1125bf50bab2SNaoya Horiguchi struct page *page; 11261a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1127bf50bab2SNaoya Horiguchi 11289487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1129bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11306077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1131bbe88753SJoonsoo Kim continue; 1132bbe88753SJoonsoo Kim 11336664bfc8SWei Yang if (PageHWPoison(page)) 11346664bfc8SWei Yang continue; 1135bbe88753SJoonsoo Kim 11360edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1137a9869b83SNaoya Horiguchi set_page_refcounted(page); 11386c037149SMike Kravetz ClearHPageFreed(page); 1139bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1140bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1141bf50bab2SNaoya Horiguchi return page; 1142bf50bab2SNaoya Horiguchi } 1143bf50bab2SNaoya Horiguchi 11446664bfc8SWei Yang return NULL; 11456664bfc8SWei Yang } 11466664bfc8SWei Yang 11473e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11483e59fcb0SMichal Hocko nodemask_t *nmask) 114994310cbcSAnshuman Khandual { 11503e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11513e59fcb0SMichal Hocko struct zonelist *zonelist; 11523e59fcb0SMichal Hocko struct zone *zone; 11533e59fcb0SMichal Hocko struct zoneref *z; 115498fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11553e59fcb0SMichal Hocko 11563e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11573e59fcb0SMichal Hocko 11583e59fcb0SMichal Hocko retry_cpuset: 11593e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11603e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 116194310cbcSAnshuman Khandual struct page *page; 116294310cbcSAnshuman Khandual 11633e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11643e59fcb0SMichal Hocko continue; 11653e59fcb0SMichal Hocko /* 11663e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11673e59fcb0SMichal Hocko * zone aware 11683e59fcb0SMichal Hocko */ 11693e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11703e59fcb0SMichal Hocko continue; 11713e59fcb0SMichal Hocko node = zone_to_nid(zone); 117294310cbcSAnshuman Khandual 117394310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 117494310cbcSAnshuman Khandual if (page) 117594310cbcSAnshuman Khandual return page; 117694310cbcSAnshuman Khandual } 11773e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11783e59fcb0SMichal Hocko goto retry_cpuset; 11793e59fcb0SMichal Hocko 118094310cbcSAnshuman Khandual return NULL; 118194310cbcSAnshuman Khandual } 118294310cbcSAnshuman Khandual 1183a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1184a5516438SAndi Kleen struct vm_area_struct *vma, 1185af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1186af0ed73eSJoonsoo Kim long chg) 11871da177e4SLinus Torvalds { 1188cfcaa66fSBen Widawsky struct page *page = NULL; 1189480eccf9SLee Schermerhorn struct mempolicy *mpol; 119004ec6264SVlastimil Babka gfp_t gfp_mask; 11913e59fcb0SMichal Hocko nodemask_t *nodemask; 119204ec6264SVlastimil Babka int nid; 11931da177e4SLinus Torvalds 1194a1e78772SMel Gorman /* 1195a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1196a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1197a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1198a1e78772SMel Gorman */ 1199af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1200a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1201c0ff7453SMiao Xie goto err; 1202a1e78772SMel Gorman 120304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1204a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 12056eab04a8SJustin P. Mattock goto err; 120604f2cbe3SMel Gorman 120704ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 120804ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1209cfcaa66fSBen Widawsky 1210cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12113e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1212cfcaa66fSBen Widawsky 1213cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1214cfcaa66fSBen Widawsky nodemask = NULL; 1215cfcaa66fSBen Widawsky } 1216cfcaa66fSBen Widawsky 1217cfcaa66fSBen Widawsky if (!page) 1218cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1219cfcaa66fSBen Widawsky 12203e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1221d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1222a63884e9SJoonsoo Kim h->resv_huge_pages--; 1223bf50bab2SNaoya Horiguchi } 1224cc9a6c87SMel Gorman 1225cc9a6c87SMel Gorman mpol_cond_put(mpol); 1226cc9a6c87SMel Gorman return page; 1227cc9a6c87SMel Gorman 1228c0ff7453SMiao Xie err: 1229cc9a6c87SMel Gorman return NULL; 12301da177e4SLinus Torvalds } 12311da177e4SLinus Torvalds 12321cac6f2cSLuiz Capitulino /* 12331cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12341cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12351cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12361cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12371cac6f2cSLuiz Capitulino * node for alloc or free. 12381cac6f2cSLuiz Capitulino */ 12391cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12401cac6f2cSLuiz Capitulino { 12410edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12421cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12431cac6f2cSLuiz Capitulino 12441cac6f2cSLuiz Capitulino return nid; 12451cac6f2cSLuiz Capitulino } 12461cac6f2cSLuiz Capitulino 12471cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12481cac6f2cSLuiz Capitulino { 12491cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12501cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12511cac6f2cSLuiz Capitulino return nid; 12521cac6f2cSLuiz Capitulino } 12531cac6f2cSLuiz Capitulino 12541cac6f2cSLuiz Capitulino /* 12551cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12561cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12571cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12581cac6f2cSLuiz Capitulino * mask. 12591cac6f2cSLuiz Capitulino */ 12601cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12611cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12621cac6f2cSLuiz Capitulino { 12631cac6f2cSLuiz Capitulino int nid; 12641cac6f2cSLuiz Capitulino 12651cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12661cac6f2cSLuiz Capitulino 12671cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12681cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12691cac6f2cSLuiz Capitulino 12701cac6f2cSLuiz Capitulino return nid; 12711cac6f2cSLuiz Capitulino } 12721cac6f2cSLuiz Capitulino 12731cac6f2cSLuiz Capitulino /* 127410c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12751cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12761cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12771cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12781cac6f2cSLuiz Capitulino */ 12791cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12801cac6f2cSLuiz Capitulino { 12811cac6f2cSLuiz Capitulino int nid; 12821cac6f2cSLuiz Capitulino 12831cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12841cac6f2cSLuiz Capitulino 12851cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12861cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12871cac6f2cSLuiz Capitulino 12881cac6f2cSLuiz Capitulino return nid; 12891cac6f2cSLuiz Capitulino } 12901cac6f2cSLuiz Capitulino 12911cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12921cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12931cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12941cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12951cac6f2cSLuiz Capitulino nr_nodes--) 12961cac6f2cSLuiz Capitulino 12971cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12981cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12991cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13001cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13011cac6f2cSLuiz Capitulino nr_nodes--) 13021cac6f2cSLuiz Capitulino 13038531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 130434d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 130534d9e35bSMike Kravetz unsigned int order, bool demote) 1306944d9fecSLuiz Capitulino { 1307944d9fecSLuiz Capitulino int i; 1308944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1309944d9fecSLuiz Capitulino struct page *p = page + 1; 1310944d9fecSLuiz Capitulino 1311c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 131247e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 131347e29d32SJohn Hubbard 1314944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1315a01f4390SMike Kravetz p->mapping = NULL; 13161d798ca3SKirill A. Shutemov clear_compound_head(p); 131734d9e35bSMike Kravetz if (!demote) 1318944d9fecSLuiz Capitulino set_page_refcounted(p); 1319944d9fecSLuiz Capitulino } 1320944d9fecSLuiz Capitulino 1321944d9fecSLuiz Capitulino set_compound_order(page, 0); 13225232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 1323ba9c1201SGerald Schaefer page[1].compound_nr = 0; 13245232c63fSMatthew Wilcox (Oracle) #endif 1325944d9fecSLuiz Capitulino __ClearPageHead(page); 1326944d9fecSLuiz Capitulino } 1327944d9fecSLuiz Capitulino 13288531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13298531fc6fSMike Kravetz unsigned int order) 13308531fc6fSMike Kravetz { 13318531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13328531fc6fSMike Kravetz } 13338531fc6fSMike Kravetz 13348531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 133534d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 133634d9e35bSMike Kravetz unsigned int order) 133734d9e35bSMike Kravetz { 133834d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 133934d9e35bSMike Kravetz } 134034d9e35bSMike Kravetz 1341d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1342944d9fecSLuiz Capitulino { 1343cf11e85fSRoman Gushchin /* 1344cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1345cf11e85fSRoman Gushchin * cma_release() returns false. 1346cf11e85fSRoman Gushchin */ 1347dbda8feaSBarry Song #ifdef CONFIG_CMA 1348dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1349cf11e85fSRoman Gushchin return; 1350dbda8feaSBarry Song #endif 1351cf11e85fSRoman Gushchin 1352944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1353944d9fecSLuiz Capitulino } 1354944d9fecSLuiz Capitulino 13554eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1356d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1357d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1358944d9fecSLuiz Capitulino { 135904adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1360953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1361953f064aSLi Xinhai nid = numa_mem_id(); 1362944d9fecSLuiz Capitulino 1363dbda8feaSBarry Song #ifdef CONFIG_CMA 1364dbda8feaSBarry Song { 1365cf11e85fSRoman Gushchin struct page *page; 1366cf11e85fSRoman Gushchin int node; 1367cf11e85fSRoman Gushchin 1368953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1369953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1370953f064aSLi Xinhai huge_page_order(h), true); 1371953f064aSLi Xinhai if (page) 1372953f064aSLi Xinhai return page; 1373953f064aSLi Xinhai } 1374953f064aSLi Xinhai 1375953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1376cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1377953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1378cf11e85fSRoman Gushchin continue; 1379cf11e85fSRoman Gushchin 1380cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1381cf11e85fSRoman Gushchin huge_page_order(h), true); 1382cf11e85fSRoman Gushchin if (page) 1383cf11e85fSRoman Gushchin return page; 1384cf11e85fSRoman Gushchin } 1385cf11e85fSRoman Gushchin } 1386953f064aSLi Xinhai } 1387dbda8feaSBarry Song #endif 1388cf11e85fSRoman Gushchin 13895e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1390944d9fecSLuiz Capitulino } 1391944d9fecSLuiz Capitulino 13924eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13934eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13944eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13954eb0716eSAlexandre Ghiti { 13964eb0716eSAlexandre Ghiti return NULL; 13974eb0716eSAlexandre Ghiti } 13984eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1399944d9fecSLuiz Capitulino 1400e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1401d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14024eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14034eb0716eSAlexandre Ghiti { 14044eb0716eSAlexandre Ghiti return NULL; 14054eb0716eSAlexandre Ghiti } 1406d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1407944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1408d00181b9SKirill A. Shutemov unsigned int order) { } 1409944d9fecSLuiz Capitulino #endif 1410944d9fecSLuiz Capitulino 14116eb4e88aSMike Kravetz /* 14126eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 141334d9e35bSMike Kravetz * as just a compound page. 141434d9e35bSMike Kravetz * 141534d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14166eb4e88aSMike Kravetz * 14176eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14186eb4e88aSMike Kravetz */ 141934d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 142034d9e35bSMike Kravetz bool adjust_surplus, 142134d9e35bSMike Kravetz bool demote) 14226eb4e88aSMike Kravetz { 14236eb4e88aSMike Kravetz int nid = page_to_nid(page); 14246eb4e88aSMike Kravetz 14256eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14266eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14276eb4e88aSMike Kravetz 14289487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14296eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14306eb4e88aSMike Kravetz return; 14316eb4e88aSMike Kravetz 14326eb4e88aSMike Kravetz list_del(&page->lru); 14336eb4e88aSMike Kravetz 14346eb4e88aSMike Kravetz if (HPageFreed(page)) { 14356eb4e88aSMike Kravetz h->free_huge_pages--; 14366eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14376eb4e88aSMike Kravetz } 14386eb4e88aSMike Kravetz if (adjust_surplus) { 14396eb4e88aSMike Kravetz h->surplus_huge_pages--; 14406eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14416eb4e88aSMike Kravetz } 14426eb4e88aSMike Kravetz 1443e32d20c0SMike Kravetz /* 1444e32d20c0SMike Kravetz * Very subtle 1445e32d20c0SMike Kravetz * 1446e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1447e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1448e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1449e32d20c0SMike Kravetz * their reference. 1450e32d20c0SMike Kravetz * 1451e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1452e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1453e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1454e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1455e32d20c0SMike Kravetz * apply. 1456e32d20c0SMike Kravetz * 1457e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1458e32d20c0SMike Kravetz * after update_and_free_page is called. 145934d9e35bSMike Kravetz * 146034d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 146134d9e35bSMike Kravetz * be turned into a page of smaller size. 1462e32d20c0SMike Kravetz */ 146334d9e35bSMike Kravetz if (!demote) 14646eb4e88aSMike Kravetz set_page_refcounted(page); 1465e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14666eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1467e32d20c0SMike Kravetz else 1468e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14696eb4e88aSMike Kravetz 14706eb4e88aSMike Kravetz h->nr_huge_pages--; 14716eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14726eb4e88aSMike Kravetz } 14736eb4e88aSMike Kravetz 147434d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 147534d9e35bSMike Kravetz bool adjust_surplus) 147634d9e35bSMike Kravetz { 147734d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 147834d9e35bSMike Kravetz } 147934d9e35bSMike Kravetz 14808531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14818531fc6fSMike Kravetz bool adjust_surplus) 14828531fc6fSMike Kravetz { 14838531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 14848531fc6fSMike Kravetz } 14858531fc6fSMike Kravetz 1486ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1487ad2fa371SMuchun Song bool adjust_surplus) 1488ad2fa371SMuchun Song { 1489ad2fa371SMuchun Song int zeroed; 1490ad2fa371SMuchun Song int nid = page_to_nid(page); 1491ad2fa371SMuchun Song 1492ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1493ad2fa371SMuchun Song 1494ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1495ad2fa371SMuchun Song 1496ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1497ad2fa371SMuchun Song h->nr_huge_pages++; 1498ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1499ad2fa371SMuchun Song 1500ad2fa371SMuchun Song if (adjust_surplus) { 1501ad2fa371SMuchun Song h->surplus_huge_pages++; 1502ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1503ad2fa371SMuchun Song } 1504ad2fa371SMuchun Song 1505ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1506ad2fa371SMuchun Song set_page_private(page, 0); 1507ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1508ad2fa371SMuchun Song 1509ad2fa371SMuchun Song /* 1510b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1511b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1512b65a4edaSMike Kravetz * just in case. 1513ad2fa371SMuchun Song */ 1514ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1515b65a4edaSMike Kravetz if (!zeroed) 1516b65a4edaSMike Kravetz /* 1517b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1518b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1519b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1520b65a4edaSMike Kravetz * when this other ref is dropped. 1521b65a4edaSMike Kravetz */ 1522b65a4edaSMike Kravetz return; 1523b65a4edaSMike Kravetz 1524ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1525ad2fa371SMuchun Song enqueue_huge_page(h, page); 1526ad2fa371SMuchun Song } 1527ad2fa371SMuchun Song 1528b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15296af2acb6SAdam Litke { 15306af2acb6SAdam Litke int i; 1531dbfee5aeSMike Kravetz struct page *subpage = page; 1532a5516438SAndi Kleen 15334eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1534944d9fecSLuiz Capitulino return; 153518229df5SAndy Whitcroft 1536161df60eSNaoya Horiguchi /* 1537161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1538161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1539161df60eSNaoya Horiguchi */ 1540161df60eSNaoya Horiguchi if (HPageRawHwpUnreliable(page)) 1541161df60eSNaoya Horiguchi return; 1542161df60eSNaoya Horiguchi 15436213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1544ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1545ad2fa371SMuchun Song /* 1546ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1547ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1548ad2fa371SMuchun Song * as a surplus page. 1549ad2fa371SMuchun Song */ 1550ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1551ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1552ad2fa371SMuchun Song return; 1553ad2fa371SMuchun Song } 1554ad2fa371SMuchun Song 1555161df60eSNaoya Horiguchi /* 1556161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1557161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1558161df60eSNaoya Horiguchi */ 1559161df60eSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 1560161df60eSNaoya Horiguchi hugetlb_clear_page_hwpoison(page); 1561161df60eSNaoya Horiguchi 1562dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1563dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1564dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 156532f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1566a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1567a7407a27SLuiz Capitulino 1 << PG_writeback); 15686af2acb6SAdam Litke } 1569a01f4390SMike Kravetz 1570a01f4390SMike Kravetz /* 1571a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1572a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1573a01f4390SMike Kravetz */ 1574a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1575a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1576944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1577944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1578944d9fecSLuiz Capitulino } else { 1579a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 15806af2acb6SAdam Litke } 1581944d9fecSLuiz Capitulino } 15826af2acb6SAdam Litke 1583b65d4adbSMuchun Song /* 1584b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1585b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1586b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1587b65d4adbSMuchun Song * the vmemmap pages. 1588b65d4adbSMuchun Song * 1589b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1590b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1591b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1592b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1593b65d4adbSMuchun Song */ 1594b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1595b65d4adbSMuchun Song 1596b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1597b65d4adbSMuchun Song { 1598b65d4adbSMuchun Song struct llist_node *node; 1599b65d4adbSMuchun Song 1600b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1601b65d4adbSMuchun Song 1602b65d4adbSMuchun Song while (node) { 1603b65d4adbSMuchun Song struct page *page; 1604b65d4adbSMuchun Song struct hstate *h; 1605b65d4adbSMuchun Song 1606b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1607b65d4adbSMuchun Song struct page, mapping); 1608b65d4adbSMuchun Song node = node->next; 1609b65d4adbSMuchun Song page->mapping = NULL; 1610b65d4adbSMuchun Song /* 1611b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1612b65d4adbSMuchun Song * is going to trigger because a previous call to 1613b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1614b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1615b65d4adbSMuchun Song */ 1616b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1617b65d4adbSMuchun Song 1618b65d4adbSMuchun Song __update_and_free_page(h, page); 1619b65d4adbSMuchun Song 1620b65d4adbSMuchun Song cond_resched(); 1621b65d4adbSMuchun Song } 1622b65d4adbSMuchun Song } 1623b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1624b65d4adbSMuchun Song 1625b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1626b65d4adbSMuchun Song { 16276213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1628b65d4adbSMuchun Song flush_work(&free_hpage_work); 1629b65d4adbSMuchun Song } 1630b65d4adbSMuchun Song 1631b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1632b65d4adbSMuchun Song bool atomic) 1633b65d4adbSMuchun Song { 1634ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1635b65d4adbSMuchun Song __update_and_free_page(h, page); 1636b65d4adbSMuchun Song return; 1637b65d4adbSMuchun Song } 1638b65d4adbSMuchun Song 1639b65d4adbSMuchun Song /* 1640b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1641b65d4adbSMuchun Song * 1642b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1643b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1644b65d4adbSMuchun Song * hasn't retrieved the list yet. 1645b65d4adbSMuchun Song */ 1646b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1647b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1648b65d4adbSMuchun Song } 1649b65d4adbSMuchun Song 165010c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 165110c6ec49SMike Kravetz { 165210c6ec49SMike Kravetz struct page *page, *t_page; 165310c6ec49SMike Kravetz 165410c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1655b65d4adbSMuchun Song update_and_free_page(h, page, false); 165610c6ec49SMike Kravetz cond_resched(); 165710c6ec49SMike Kravetz } 165810c6ec49SMike Kravetz } 165910c6ec49SMike Kravetz 1660e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1661e5ff2159SAndi Kleen { 1662e5ff2159SAndi Kleen struct hstate *h; 1663e5ff2159SAndi Kleen 1664e5ff2159SAndi Kleen for_each_hstate(h) { 1665e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1666e5ff2159SAndi Kleen return h; 1667e5ff2159SAndi Kleen } 1668e5ff2159SAndi Kleen return NULL; 1669e5ff2159SAndi Kleen } 1670e5ff2159SAndi Kleen 1671db71ef79SMike Kravetz void free_huge_page(struct page *page) 167227a85ef1SDavid Gibson { 1673a5516438SAndi Kleen /* 1674a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1675a5516438SAndi Kleen * compound page destructor. 1676a5516438SAndi Kleen */ 1677e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16787893d1d5SAdam Litke int nid = page_to_nid(page); 1679d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 168007443a85SJoonsoo Kim bool restore_reserve; 1681db71ef79SMike Kravetz unsigned long flags; 168227a85ef1SDavid Gibson 1683b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1684b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 16858ace22bcSYongkai Wu 1686d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 168778fbe906SDavid Hildenbrand if (PageAnon(page)) 168878fbe906SDavid Hildenbrand __ClearPageAnonExclusive(page); 16898ace22bcSYongkai Wu page->mapping = NULL; 1690d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1691d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 169227a85ef1SDavid Gibson 16931c5ecae3SMike Kravetz /* 1694d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 16950919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 16960919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 16970919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 16986c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 16990919e1b6SMike Kravetz * remove the reserved page from the subpool. 17000919e1b6SMike Kravetz */ 17010919e1b6SMike Kravetz if (!restore_reserve) { 17020919e1b6SMike Kravetz /* 17030919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 17040919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 17050919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 17060919e1b6SMike Kravetz * operation. 17071c5ecae3SMike Kravetz */ 17081c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 17091c5ecae3SMike Kravetz restore_reserve = true; 17100919e1b6SMike Kravetz } 17111c5ecae3SMike Kravetz 1712db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 17138f251a3dSMike Kravetz ClearHPageMigratable(page); 17146d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 17156d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 171608cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 171708cf9fafSMina Almasry pages_per_huge_page(h), page); 171807443a85SJoonsoo Kim if (restore_reserve) 171907443a85SJoonsoo Kim h->resv_huge_pages++; 172007443a85SJoonsoo Kim 17219157c311SMike Kravetz if (HPageTemporary(page)) { 17226eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1723db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1724b65d4adbSMuchun Song update_and_free_page(h, page, true); 1725ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17260edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17276eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1728db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1729b65d4adbSMuchun Song update_and_free_page(h, page, true); 17307893d1d5SAdam Litke } else { 17315d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1732a5516438SAndi Kleen enqueue_huge_page(h, page); 1733db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 173427a85ef1SDavid Gibson } 17351121828aSMike Kravetz } 173627a85ef1SDavid Gibson 1737d3d99fccSOscar Salvador /* 1738d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1739d3d99fccSOscar Salvador */ 1740d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1741d3d99fccSOscar Salvador { 1742d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1743d3d99fccSOscar Salvador h->nr_huge_pages++; 1744d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1745d3d99fccSOscar Salvador } 1746d3d99fccSOscar Salvador 1747f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1748b7ba30c6SAndi Kleen { 17496213834cSMuchun Song hugetlb_vmemmap_optimize(h, page); 17500edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1751f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1752ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17539dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17541adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1755d3d99fccSOscar Salvador } 1756d3d99fccSOscar Salvador 1757d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1758d3d99fccSOscar Salvador { 1759f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1760db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1761d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1762db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1763b7ba30c6SAndi Kleen } 1764b7ba30c6SAndi Kleen 176534d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 176634d9e35bSMike Kravetz bool demote) 176720a0307cSWu Fengguang { 17687118fc29SMike Kravetz int i, j; 176920a0307cSWu Fengguang int nr_pages = 1 << order; 177020a0307cSWu Fengguang struct page *p = page + 1; 177120a0307cSWu Fengguang 177220a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 177320a0307cSWu Fengguang set_compound_order(page, order); 1774ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1775de09d31dSKirill A. Shutemov __SetPageHead(page); 177620a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1777ef5a22beSAndrea Arcangeli /* 1778ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1779ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1780ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 17817c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1782ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1783ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1784ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1785ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1786ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1787ef5a22beSAndrea Arcangeli * after get_user_pages(). 1788ef5a22beSAndrea Arcangeli */ 1789ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 17907118fc29SMike Kravetz /* 17917118fc29SMike Kravetz * Subtle and very unlikely 17927118fc29SMike Kravetz * 17937118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 17947118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 17957118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 17967118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 17977118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 17987118fc29SMike Kravetz * We need to respect any increased ref count, and only set 17997118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1800416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1801416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1802416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1803416d85edSMike Kravetz * pages using the appropriate free interface. 180434d9e35bSMike Kravetz * 180534d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 18067118fc29SMike Kravetz */ 180734d9e35bSMike Kravetz if (!demote) { 18087118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1809416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 18107118fc29SMike Kravetz goto out_error; 18117118fc29SMike Kravetz } 181234d9e35bSMike Kravetz } else { 181334d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 181434d9e35bSMike Kravetz } 18151d798ca3SKirill A. Shutemov set_compound_head(p, page); 181620a0307cSWu Fengguang } 1817b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 181847e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18197118fc29SMike Kravetz return true; 18207118fc29SMike Kravetz 18217118fc29SMike Kravetz out_error: 18227118fc29SMike Kravetz /* undo tail page modifications made above */ 18237118fc29SMike Kravetz p = page + 1; 18247118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 18257118fc29SMike Kravetz clear_compound_head(p); 18267118fc29SMike Kravetz set_page_refcounted(p); 18277118fc29SMike Kravetz } 18287118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 18297118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 18307118fc29SMike Kravetz __ClearPageReserved(p); 18317118fc29SMike Kravetz set_compound_order(page, 0); 18325232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 18337118fc29SMike Kravetz page[1].compound_nr = 0; 18345232c63fSMatthew Wilcox (Oracle) #endif 18357118fc29SMike Kravetz __ClearPageHead(page); 18367118fc29SMike Kravetz return false; 183720a0307cSWu Fengguang } 183820a0307cSWu Fengguang 183934d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 184034d9e35bSMike Kravetz { 184134d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 184234d9e35bSMike Kravetz } 184334d9e35bSMike Kravetz 18448531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18458531fc6fSMike Kravetz unsigned int order) 18468531fc6fSMike Kravetz { 18478531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18488531fc6fSMike Kravetz } 18498531fc6fSMike Kravetz 18507795912cSAndrew Morton /* 18517795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18527795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18537795912cSAndrew Morton * details. 18547795912cSAndrew Morton */ 185520a0307cSWu Fengguang int PageHuge(struct page *page) 185620a0307cSWu Fengguang { 185720a0307cSWu Fengguang if (!PageCompound(page)) 185820a0307cSWu Fengguang return 0; 185920a0307cSWu Fengguang 186020a0307cSWu Fengguang page = compound_head(page); 1861f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 186220a0307cSWu Fengguang } 186343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 186443131e14SNaoya Horiguchi 186527c73ae7SAndrea Arcangeli /* 186627c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 186727c73ae7SAndrea Arcangeli * normal or transparent huge pages. 186827c73ae7SAndrea Arcangeli */ 186927c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 187027c73ae7SAndrea Arcangeli { 187127c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 187227c73ae7SAndrea Arcangeli return 0; 187327c73ae7SAndrea Arcangeli 1874d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 187527c73ae7SAndrea Arcangeli } 18764e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 187727c73ae7SAndrea Arcangeli 1878c0d0381aSMike Kravetz /* 1879c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1880c0d0381aSMike Kravetz * 1881336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1882336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1883336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1884c0d0381aSMike Kravetz */ 1885c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1886c0d0381aSMike Kravetz { 1887336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1888c0d0381aSMike Kravetz 1889c0d0381aSMike Kravetz if (!mapping) 1890c0d0381aSMike Kravetz return mapping; 1891c0d0381aSMike Kravetz 1892c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1893c0d0381aSMike Kravetz return mapping; 1894c0d0381aSMike Kravetz 1895c0d0381aSMike Kravetz return NULL; 1896c0d0381aSMike Kravetz } 1897c0d0381aSMike Kravetz 1898fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 189913d60f4bSZhang Yi { 190013d60f4bSZhang Yi struct page *page_head = compound_head(page); 190113d60f4bSZhang Yi pgoff_t index = page_index(page_head); 190213d60f4bSZhang Yi unsigned long compound_idx; 190313d60f4bSZhang Yi 190413d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 190513d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 190613d60f4bSZhang Yi else 190713d60f4bSZhang Yi compound_idx = page - page_head; 190813d60f4bSZhang Yi 190913d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 191013d60f4bSZhang Yi } 191113d60f4bSZhang Yi 19120c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1913f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1914f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19151da177e4SLinus Torvalds { 1916af0fb9dfSMichal Hocko int order = huge_page_order(h); 19171da177e4SLinus Torvalds struct page *page; 1918f60858f9SMike Kravetz bool alloc_try_hard = true; 1919f96efd58SJoe Jin 1920f60858f9SMike Kravetz /* 1921f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1922f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1923f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1924f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1925f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1926f60858f9SMike Kravetz */ 1927f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1928f60858f9SMike Kravetz alloc_try_hard = false; 1929f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1930f60858f9SMike Kravetz if (alloc_try_hard) 1931f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1932af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1933af0fb9dfSMichal Hocko nid = numa_mem_id(); 193484172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1935af0fb9dfSMichal Hocko if (page) 1936af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1937af0fb9dfSMichal Hocko else 1938af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 193963b4613cSNishanth Aravamudan 1940f60858f9SMike Kravetz /* 1941f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1942f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1943f60858f9SMike Kravetz * normal 'try hard' allocations. 1944f60858f9SMike Kravetz */ 1945f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1946f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1947f60858f9SMike Kravetz 1948f60858f9SMike Kravetz /* 1949f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1950f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1951f60858f9SMike Kravetz * overall state change. 1952f60858f9SMike Kravetz */ 1953f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1954f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1955f60858f9SMike Kravetz 195663b4613cSNishanth Aravamudan return page; 195763b4613cSNishanth Aravamudan } 195863b4613cSNishanth Aravamudan 1959af0fb9dfSMichal Hocko /* 19600c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 19610c397daeSMichal Hocko * should use this function to get new hugetlb pages 19620c397daeSMichal Hocko */ 19630c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1964f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1965f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19660c397daeSMichal Hocko { 19670c397daeSMichal Hocko struct page *page; 19687118fc29SMike Kravetz bool retry = false; 19690c397daeSMichal Hocko 19707118fc29SMike Kravetz retry: 19710c397daeSMichal Hocko if (hstate_is_gigantic(h)) 19720c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 19730c397daeSMichal Hocko else 19740c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1975f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 19760c397daeSMichal Hocko if (!page) 19770c397daeSMichal Hocko return NULL; 19780c397daeSMichal Hocko 19797118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 19807118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 19817118fc29SMike Kravetz /* 19827118fc29SMike Kravetz * Rare failure to convert pages to compound page. 19837118fc29SMike Kravetz * Free pages and try again - ONCE! 19847118fc29SMike Kravetz */ 19857118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 19867118fc29SMike Kravetz if (!retry) { 19877118fc29SMike Kravetz retry = true; 19887118fc29SMike Kravetz goto retry; 19897118fc29SMike Kravetz } 19907118fc29SMike Kravetz return NULL; 19917118fc29SMike Kravetz } 19927118fc29SMike Kravetz } 19930c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 19940c397daeSMichal Hocko 19950c397daeSMichal Hocko return page; 19960c397daeSMichal Hocko } 19970c397daeSMichal Hocko 19980c397daeSMichal Hocko /* 1999af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2000af0fb9dfSMichal Hocko * manner. 2001af0fb9dfSMichal Hocko */ 2002f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2003f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2004b2261026SJoonsoo Kim { 2005b2261026SJoonsoo Kim struct page *page; 2006b2261026SJoonsoo Kim int nr_nodes, node; 2007af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2008b2261026SJoonsoo Kim 2009b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2010f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 2011f60858f9SMike Kravetz node_alloc_noretry); 2012af0fb9dfSMichal Hocko if (page) 2013b2261026SJoonsoo Kim break; 2014b2261026SJoonsoo Kim } 2015b2261026SJoonsoo Kim 2016af0fb9dfSMichal Hocko if (!page) 2017af0fb9dfSMichal Hocko return 0; 2018b2261026SJoonsoo Kim 2019af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2020af0fb9dfSMichal Hocko 2021af0fb9dfSMichal Hocko return 1; 2022b2261026SJoonsoo Kim } 2023b2261026SJoonsoo Kim 2024e8c5c824SLee Schermerhorn /* 202510c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 202610c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 202710c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 202810c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2029e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2030e8c5c824SLee Schermerhorn */ 203110c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 203210c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20336ae11b27SLee Schermerhorn bool acct_surplus) 2034e8c5c824SLee Schermerhorn { 2035b2261026SJoonsoo Kim int nr_nodes, node; 203610c6ec49SMike Kravetz struct page *page = NULL; 2037e8c5c824SLee Schermerhorn 20389487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2039b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2040685f3457SLee Schermerhorn /* 2041685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2042685f3457SLee Schermerhorn * nodes with surplus pages. 2043685f3457SLee Schermerhorn */ 2044b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2045b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 204610c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2047e8c5c824SLee Schermerhorn struct page, lru); 20486eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20499a76db09SLee Schermerhorn break; 2050e8c5c824SLee Schermerhorn } 2051b2261026SJoonsoo Kim } 2052e8c5c824SLee Schermerhorn 205310c6ec49SMike Kravetz return page; 2054e8c5c824SLee Schermerhorn } 2055e8c5c824SLee Schermerhorn 2056c8721bbbSNaoya Horiguchi /* 2057c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2058faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2059faf53defSNaoya Horiguchi * This function returns values like below: 2060faf53defSNaoya Horiguchi * 2061ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2062ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2063ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2064ad2fa371SMuchun Song * is enabled. 2065faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2066faf53defSNaoya Horiguchi * (allocated or reserved.) 2067faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2068faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2069c8721bbbSNaoya Horiguchi */ 2070c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2071c8721bbbSNaoya Horiguchi { 20726bc9b564SNaoya Horiguchi int rc = -EBUSY; 2073082d5b6bSGerald Schaefer 20747ffddd49SMuchun Song retry: 2075faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2076faf53defSNaoya Horiguchi if (!PageHuge(page)) 2077faf53defSNaoya Horiguchi return 0; 2078faf53defSNaoya Horiguchi 2079db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2080faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2081faf53defSNaoya Horiguchi rc = 0; 2082faf53defSNaoya Horiguchi goto out; 2083faf53defSNaoya Horiguchi } 2084faf53defSNaoya Horiguchi 2085faf53defSNaoya Horiguchi if (!page_count(page)) { 20862247bb33SGerald Schaefer struct page *head = compound_head(page); 20872247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 20886bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 2089082d5b6bSGerald Schaefer goto out; 20907ffddd49SMuchun Song 20917ffddd49SMuchun Song /* 20927ffddd49SMuchun Song * We should make sure that the page is already on the free list 20937ffddd49SMuchun Song * when it is dissolved. 20947ffddd49SMuchun Song */ 20956c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2096db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20977ffddd49SMuchun Song cond_resched(); 20987ffddd49SMuchun Song 20997ffddd49SMuchun Song /* 21007ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 21017ffddd49SMuchun Song * encounter this race. In fact, we have a chance 21027ffddd49SMuchun Song * to successfully dissolve the page if we do a 21037ffddd49SMuchun Song * retry. Because the race window is quite small. 21047ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 21057ffddd49SMuchun Song * for increasing the success rate of dissolving page. 21067ffddd49SMuchun Song */ 21077ffddd49SMuchun Song goto retry; 21087ffddd49SMuchun Song } 21097ffddd49SMuchun Song 2110ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2111ad2fa371SMuchun Song h->max_huge_pages--; 2112ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2113ad2fa371SMuchun Song 2114c3114a84SAnshuman Khandual /* 2115ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2116ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2117ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2118ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2119ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2120ad2fa371SMuchun Song * appropriate action on failure. 2121ad2fa371SMuchun Song */ 21226213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, head); 2123ad2fa371SMuchun Song if (!rc) { 2124b65d4adbSMuchun Song update_and_free_page(h, head, false); 2125ad2fa371SMuchun Song } else { 2126ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2127ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2128ad2fa371SMuchun Song h->max_huge_pages++; 2129ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2130ad2fa371SMuchun Song } 2131ad2fa371SMuchun Song 2132ad2fa371SMuchun Song return rc; 2133c8721bbbSNaoya Horiguchi } 2134082d5b6bSGerald Schaefer out: 2135db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2136082d5b6bSGerald Schaefer return rc; 2137c8721bbbSNaoya Horiguchi } 2138c8721bbbSNaoya Horiguchi 2139c8721bbbSNaoya Horiguchi /* 2140c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2141c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21422247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21432247bb33SGerald Schaefer * part of it lies within the given range. 2144082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2145082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2146c8721bbbSNaoya Horiguchi */ 2147082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2148c8721bbbSNaoya Horiguchi { 2149c8721bbbSNaoya Horiguchi unsigned long pfn; 2150eb03aa00SGerald Schaefer struct page *page; 2151082d5b6bSGerald Schaefer int rc = 0; 2152dc2628f3SMuchun Song unsigned int order; 2153dc2628f3SMuchun Song struct hstate *h; 2154c8721bbbSNaoya Horiguchi 2155d0177639SLi Zhong if (!hugepages_supported()) 2156082d5b6bSGerald Schaefer return rc; 2157d0177639SLi Zhong 2158dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2159dc2628f3SMuchun Song for_each_hstate(h) 2160dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2161dc2628f3SMuchun Song 2162dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2163eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2164eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2165eb03aa00SGerald Schaefer if (rc) 2166082d5b6bSGerald Schaefer break; 2167eb03aa00SGerald Schaefer } 2168082d5b6bSGerald Schaefer 2169082d5b6bSGerald Schaefer return rc; 2170c8721bbbSNaoya Horiguchi } 2171c8721bbbSNaoya Horiguchi 2172ab5ac90aSMichal Hocko /* 2173ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2174ab5ac90aSMichal Hocko */ 21750c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2176b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 21777893d1d5SAdam Litke { 21789980d744SMichal Hocko struct page *page = NULL; 2179b65a4edaSMike Kravetz bool retry = false; 21807893d1d5SAdam Litke 2181bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2182aa888a74SAndi Kleen return NULL; 2183aa888a74SAndi Kleen 2184db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21859980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 21869980d744SMichal Hocko goto out_unlock; 2187db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2188d1c3fb1fSNishanth Aravamudan 2189b65a4edaSMike Kravetz retry: 2190f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 21919980d744SMichal Hocko if (!page) 21920c397daeSMichal Hocko return NULL; 2193d1c3fb1fSNishanth Aravamudan 2194db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21959980d744SMichal Hocko /* 21969980d744SMichal Hocko * We could have raced with the pool size change. 21979980d744SMichal Hocko * Double check that and simply deallocate the new page 21989980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 21999980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 22009980d744SMichal Hocko * codeflow 22019980d744SMichal Hocko */ 22029980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 22039157c311SMike Kravetz SetHPageTemporary(page); 2204db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22059980d744SMichal Hocko put_page(page); 22062bf753e6SKai Shen return NULL; 2207b65a4edaSMike Kravetz } 2208b65a4edaSMike Kravetz 2209b65a4edaSMike Kravetz if (zero_ref) { 2210b65a4edaSMike Kravetz /* 2211b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2212b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2213b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2214b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2215b65a4edaSMike Kravetz */ 2216b65a4edaSMike Kravetz SetHPageTemporary(page); 2217b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2218b65a4edaSMike Kravetz /* 2219b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2220b65a4edaSMike Kravetz * huge. Retry once. 2221b65a4edaSMike Kravetz */ 2222b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2223b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2224b65a4edaSMike Kravetz if (retry) 2225b65a4edaSMike Kravetz return NULL; 2226b65a4edaSMike Kravetz 2227b65a4edaSMike Kravetz retry = true; 2228b65a4edaSMike Kravetz goto retry; 2229b65a4edaSMike Kravetz } 2230b65a4edaSMike Kravetz ClearHPageTemporary(page); 2231b65a4edaSMike Kravetz } 2232b65a4edaSMike Kravetz 22339980d744SMichal Hocko h->surplus_huge_pages++; 22344704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22359980d744SMichal Hocko 22369980d744SMichal Hocko out_unlock: 2237db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22387893d1d5SAdam Litke 22397893d1d5SAdam Litke return page; 22407893d1d5SAdam Litke } 22417893d1d5SAdam Litke 2242bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2243ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2244ab5ac90aSMichal Hocko { 2245ab5ac90aSMichal Hocko struct page *page; 2246ab5ac90aSMichal Hocko 2247ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2248ab5ac90aSMichal Hocko return NULL; 2249ab5ac90aSMichal Hocko 2250f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2251ab5ac90aSMichal Hocko if (!page) 2252ab5ac90aSMichal Hocko return NULL; 2253ab5ac90aSMichal Hocko 2254ab5ac90aSMichal Hocko /* 2255ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2256ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2257ab5ac90aSMichal Hocko */ 22589157c311SMike Kravetz SetHPageTemporary(page); 2259ab5ac90aSMichal Hocko 2260ab5ac90aSMichal Hocko return page; 2261ab5ac90aSMichal Hocko } 2262ab5ac90aSMichal Hocko 2263e4e574b7SAdam Litke /* 2264099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2265099730d6SDave Hansen */ 2266e0ec90eeSDave Hansen static 22670c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2268099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2269099730d6SDave Hansen { 2270cfcaa66fSBen Widawsky struct page *page = NULL; 2271aaf14e40SMichal Hocko struct mempolicy *mpol; 2272aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2273aaf14e40SMichal Hocko int nid; 2274aaf14e40SMichal Hocko nodemask_t *nodemask; 2275aaf14e40SMichal Hocko 2276aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2277cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2278cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2279cfcaa66fSBen Widawsky 2280cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2281cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2282cfcaa66fSBen Widawsky 2283cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2284cfcaa66fSBen Widawsky nodemask = NULL; 2285cfcaa66fSBen Widawsky } 2286cfcaa66fSBen Widawsky 2287cfcaa66fSBen Widawsky if (!page) 2288b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2289aaf14e40SMichal Hocko mpol_cond_put(mpol); 2290aaf14e40SMichal Hocko return page; 2291099730d6SDave Hansen } 2292099730d6SDave Hansen 2293ab5ac90aSMichal Hocko /* page migration callback function */ 22943e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2295d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 22964db9b2efSMichal Hocko { 2297db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22984db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 22993e59fcb0SMichal Hocko struct page *page; 23003e59fcb0SMichal Hocko 23013e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 23023e59fcb0SMichal Hocko if (page) { 2303db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23043e59fcb0SMichal Hocko return page; 23054db9b2efSMichal Hocko } 23064db9b2efSMichal Hocko } 2307db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23084db9b2efSMichal Hocko 23090c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 23104db9b2efSMichal Hocko } 23114db9b2efSMichal Hocko 2312ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2313389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2314389c8178SMichal Hocko unsigned long address) 2315ebd63723SMichal Hocko { 2316ebd63723SMichal Hocko struct mempolicy *mpol; 2317ebd63723SMichal Hocko nodemask_t *nodemask; 2318ebd63723SMichal Hocko struct page *page; 2319ebd63723SMichal Hocko gfp_t gfp_mask; 2320ebd63723SMichal Hocko int node; 2321ebd63723SMichal Hocko 2322ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2323ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2324d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2325ebd63723SMichal Hocko mpol_cond_put(mpol); 2326ebd63723SMichal Hocko 2327ebd63723SMichal Hocko return page; 2328ebd63723SMichal Hocko } 2329ebd63723SMichal Hocko 2330bf50bab2SNaoya Horiguchi /* 233125985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2332e4e574b7SAdam Litke * of size 'delta'. 2333e4e574b7SAdam Litke */ 23340a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23351b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2336e4e574b7SAdam Litke { 233734665341SMiaohe Lin LIST_HEAD(surplus_list); 2338e4e574b7SAdam Litke struct page *page, *tmp; 23390a4f3d1bSLiu Xiang int ret; 23400a4f3d1bSLiu Xiang long i; 23410a4f3d1bSLiu Xiang long needed, allocated; 234228073b02SHillf Danton bool alloc_ok = true; 2343e4e574b7SAdam Litke 23449487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2345a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2346ac09b3a1SAdam Litke if (needed <= 0) { 2347a5516438SAndi Kleen h->resv_huge_pages += delta; 2348e4e574b7SAdam Litke return 0; 2349ac09b3a1SAdam Litke } 2350e4e574b7SAdam Litke 2351e4e574b7SAdam Litke allocated = 0; 2352e4e574b7SAdam Litke 2353e4e574b7SAdam Litke ret = -ENOMEM; 2354e4e574b7SAdam Litke retry: 2355db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2356e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23570c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2358b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 235928073b02SHillf Danton if (!page) { 236028073b02SHillf Danton alloc_ok = false; 236128073b02SHillf Danton break; 236228073b02SHillf Danton } 2363e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 236469ed779aSDavid Rientjes cond_resched(); 2365e4e574b7SAdam Litke } 236628073b02SHillf Danton allocated += i; 2367e4e574b7SAdam Litke 2368e4e574b7SAdam Litke /* 2369e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2370e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2371e4e574b7SAdam Litke */ 2372db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2373a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2374a5516438SAndi Kleen (h->free_huge_pages + allocated); 237528073b02SHillf Danton if (needed > 0) { 237628073b02SHillf Danton if (alloc_ok) 2377e4e574b7SAdam Litke goto retry; 237828073b02SHillf Danton /* 237928073b02SHillf Danton * We were not able to allocate enough pages to 238028073b02SHillf Danton * satisfy the entire reservation so we free what 238128073b02SHillf Danton * we've allocated so far. 238228073b02SHillf Danton */ 238328073b02SHillf Danton goto free; 238428073b02SHillf Danton } 2385e4e574b7SAdam Litke /* 2386e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 238725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2388e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2389ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2390ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2391ac09b3a1SAdam Litke * before they are reserved. 2392e4e574b7SAdam Litke */ 2393e4e574b7SAdam Litke needed += allocated; 2394a5516438SAndi Kleen h->resv_huge_pages += delta; 2395e4e574b7SAdam Litke ret = 0; 2396a9869b83SNaoya Horiguchi 239719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 239819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 239919fc3f0aSAdam Litke if ((--needed) < 0) 240019fc3f0aSAdam Litke break; 2401b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2402a5516438SAndi Kleen enqueue_huge_page(h, page); 240319fc3f0aSAdam Litke } 240428073b02SHillf Danton free: 2405db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 240619fc3f0aSAdam Litke 2407b65a4edaSMike Kravetz /* 2408b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2409b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2410b65a4edaSMike Kravetz */ 2411c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2412b65a4edaSMike Kravetz free_huge_page(page); 2413db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2414e4e574b7SAdam Litke 2415e4e574b7SAdam Litke return ret; 2416e4e574b7SAdam Litke } 2417e4e574b7SAdam Litke 2418e4e574b7SAdam Litke /* 2419e5bbc8a6SMike Kravetz * This routine has two main purposes: 2420e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2421e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2422e5bbc8a6SMike Kravetz * to the associated reservation map. 2423e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2424e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2425e4e574b7SAdam Litke */ 2426a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2427a5516438SAndi Kleen unsigned long unused_resv_pages) 2428e4e574b7SAdam Litke { 2429e4e574b7SAdam Litke unsigned long nr_pages; 243010c6ec49SMike Kravetz struct page *page; 243110c6ec49SMike Kravetz LIST_HEAD(page_list); 243210c6ec49SMike Kravetz 24339487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 243410c6ec49SMike Kravetz /* Uncommit the reservation */ 243510c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2436e4e574b7SAdam Litke 2437c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2438e5bbc8a6SMike Kravetz goto out; 2439aa888a74SAndi Kleen 2440e5bbc8a6SMike Kravetz /* 2441e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2442e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2443e5bbc8a6SMike Kravetz */ 2444a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2445e4e574b7SAdam Litke 2446685f3457SLee Schermerhorn /* 2447685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24489b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24499b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24509b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 245110c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24529b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2453685f3457SLee Schermerhorn */ 2454685f3457SLee Schermerhorn while (nr_pages--) { 245510c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 245610c6ec49SMike Kravetz if (!page) 2457e5bbc8a6SMike Kravetz goto out; 245810c6ec49SMike Kravetz 245910c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2460e4e574b7SAdam Litke } 2461e5bbc8a6SMike Kravetz 2462e5bbc8a6SMike Kravetz out: 2463db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 246410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2465db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2466e4e574b7SAdam Litke } 2467e4e574b7SAdam Litke 24685e911373SMike Kravetz 2469c37f9fb1SAndy Whitcroft /* 2470feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24715e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2472cf3ad20bSMike Kravetz * 2473cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2474cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2475cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2476cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2477cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2478feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2479feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2480feba16e2SMike Kravetz * is called in such cases. 2481cf3ad20bSMike Kravetz * 2482cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2483cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2484cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2485cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2486cf3ad20bSMike Kravetz * take appropriate action. 248796b96a96SMike Kravetz * 248896b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 248996b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 249096b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 249196b96a96SMike Kravetz * reservation exists. 2492846be085SMike Kravetz * 2493846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2494846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2495846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2496846be085SMike Kravetz * exists. 2497c37f9fb1SAndy Whitcroft */ 24985e911373SMike Kravetz enum vma_resv_mode { 24995e911373SMike Kravetz VMA_NEEDS_RESV, 25005e911373SMike Kravetz VMA_COMMIT_RESV, 2501feba16e2SMike Kravetz VMA_END_RESV, 250296b96a96SMike Kravetz VMA_ADD_RESV, 2503846be085SMike Kravetz VMA_DEL_RESV, 25045e911373SMike Kravetz }; 2505cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2506cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 25075e911373SMike Kravetz enum vma_resv_mode mode) 2508c37f9fb1SAndy Whitcroft { 25094e35f483SJoonsoo Kim struct resv_map *resv; 25104e35f483SJoonsoo Kim pgoff_t idx; 2511cf3ad20bSMike Kravetz long ret; 25120db9d74eSMina Almasry long dummy_out_regions_needed; 2513c37f9fb1SAndy Whitcroft 25144e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25154e35f483SJoonsoo Kim if (!resv) 2516c37f9fb1SAndy Whitcroft return 1; 2517c37f9fb1SAndy Whitcroft 25184e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25195e911373SMike Kravetz switch (mode) { 25205e911373SMike Kravetz case VMA_NEEDS_RESV: 25210db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25220db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25230db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25240db9d74eSMina Almasry * ever require 1 region. 25250db9d74eSMina Almasry */ 25260db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25275e911373SMike Kravetz break; 25285e911373SMike Kravetz case VMA_COMMIT_RESV: 2529075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25300db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25310db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25325e911373SMike Kravetz break; 2533feba16e2SMike Kravetz case VMA_END_RESV: 25340db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25355e911373SMike Kravetz ret = 0; 25365e911373SMike Kravetz break; 253796b96a96SMike Kravetz case VMA_ADD_RESV: 25380db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2539075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25400db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25410db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25420db9d74eSMina Almasry } else { 25430db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 254496b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 254596b96a96SMike Kravetz } 254696b96a96SMike Kravetz break; 2547846be085SMike Kravetz case VMA_DEL_RESV: 2548846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2549846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2550846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2551846be085SMike Kravetz } else { 2552846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2553846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2554846be085SMike Kravetz VM_BUG_ON(ret < 0); 2555846be085SMike Kravetz } 2556846be085SMike Kravetz break; 25575e911373SMike Kravetz default: 25585e911373SMike Kravetz BUG(); 25595e911373SMike Kravetz } 256084afd99bSAndy Whitcroft 2561846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2562cf3ad20bSMike Kravetz return ret; 256367961f9dSMike Kravetz /* 2564bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2565bf3d12b9SMiaohe Lin * 256667961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 256767961f9dSMike Kravetz * However, a file associated with mapping could have been 256867961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 256967961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 257067961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 257167961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 257267961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 257367961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 257467961f9dSMike Kravetz * reservation has already been consumed. As a result, the 257567961f9dSMike Kravetz * return value of this routine is the opposite of the 257667961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 257767961f9dSMike Kravetz */ 2578bf3d12b9SMiaohe Lin if (ret > 0) 257967961f9dSMike Kravetz return 0; 2580bf3d12b9SMiaohe Lin if (ret == 0) 258167961f9dSMike Kravetz return 1; 2582bf3d12b9SMiaohe Lin return ret; 258384afd99bSAndy Whitcroft } 2584cf3ad20bSMike Kravetz 2585cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2586a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2587c37f9fb1SAndy Whitcroft { 25885e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2589cf3ad20bSMike Kravetz } 2590c37f9fb1SAndy Whitcroft 2591cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2592cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2593cf3ad20bSMike Kravetz { 25945e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 25955e911373SMike Kravetz } 25965e911373SMike Kravetz 2597feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 25985e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 25995e911373SMike Kravetz { 2600feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2601c37f9fb1SAndy Whitcroft } 2602c37f9fb1SAndy Whitcroft 260396b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 260496b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 260596b96a96SMike Kravetz { 260696b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 260796b96a96SMike Kravetz } 260896b96a96SMike Kravetz 2609846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2610846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 261196b96a96SMike Kravetz { 2612846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2613846be085SMike Kravetz } 2614846be085SMike Kravetz 2615846be085SMike Kravetz /* 2616846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2617846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2618846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2619846be085SMike Kravetz * 2620846be085SMike Kravetz * It handles two specific cases: 2621846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2622846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2623846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2624846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2625846be085SMike Kravetz * 2626846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2627846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2628846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2629846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2630846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2631846be085SMike Kravetz * reserve map indicates there is a reservation present. 2632846be085SMike Kravetz * 2633846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2634846be085SMike Kravetz */ 2635846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2636846be085SMike Kravetz unsigned long address, struct page *page) 2637846be085SMike Kravetz { 263896b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 263996b96a96SMike Kravetz 2640846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2641846be085SMike Kravetz if (unlikely(rc < 0)) 264296b96a96SMike Kravetz /* 264396b96a96SMike Kravetz * Rare out of memory condition in reserve map 2644d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 264596b96a96SMike Kravetz * global reserve count will not be incremented 264696b96a96SMike Kravetz * by free_huge_page. This will make it appear 264796b96a96SMike Kravetz * as though the reservation for this page was 264896b96a96SMike Kravetz * consumed. This may prevent the task from 264996b96a96SMike Kravetz * faulting in the page at a later time. This 265096b96a96SMike Kravetz * is better than inconsistent global huge page 265196b96a96SMike Kravetz * accounting of reserve counts. 265296b96a96SMike Kravetz */ 2653d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2654846be085SMike Kravetz else if (rc) 2655846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2656846be085SMike Kravetz else 2657846be085SMike Kravetz vma_end_reservation(h, vma, address); 2658846be085SMike Kravetz } else { 2659846be085SMike Kravetz if (!rc) { 266096b96a96SMike Kravetz /* 2661846be085SMike Kravetz * This indicates there is an entry in the reserve map 2662c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2663846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2664846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2665846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2666846be085SMike Kravetz * does not consume a reservation. 266796b96a96SMike Kravetz */ 2668846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2669846be085SMike Kravetz if (rc < 0) 2670846be085SMike Kravetz /* 2671846be085SMike Kravetz * VERY rare out of memory condition. Since 2672846be085SMike Kravetz * we can not delete the entry, set 2673846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2674846be085SMike Kravetz * count will be incremented when the page 2675846be085SMike Kravetz * is freed. This reserve will be consumed 2676846be085SMike Kravetz * on a subsequent allocation. 2677846be085SMike Kravetz */ 2678846be085SMike Kravetz SetHPageRestoreReserve(page); 2679846be085SMike Kravetz } else if (rc < 0) { 2680846be085SMike Kravetz /* 2681846be085SMike Kravetz * Rare out of memory condition from 2682846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2683846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2684846be085SMike Kravetz * this implies there is not an entry in the 2685846be085SMike Kravetz * reserve map. 2686846be085SMike Kravetz * 2687846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2688846be085SMike Kravetz * no reservation. We are done. 2689846be085SMike Kravetz */ 2690846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2691846be085SMike Kravetz /* 2692846be085SMike Kravetz * For private mappings, no entry indicates 2693846be085SMike Kravetz * a reservation is present. Since we can 2694846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2695846be085SMike Kravetz * on the page so reserve count will be 2696846be085SMike Kravetz * incremented when freed. This reserve will 2697846be085SMike Kravetz * be consumed on a subsequent allocation. 2698846be085SMike Kravetz */ 2699846be085SMike Kravetz SetHPageRestoreReserve(page); 270096b96a96SMike Kravetz } else 2701846be085SMike Kravetz /* 2702846be085SMike Kravetz * No reservation present, do nothing 2703846be085SMike Kravetz */ 270496b96a96SMike Kravetz vma_end_reservation(h, vma, address); 270596b96a96SMike Kravetz } 270696b96a96SMike Kravetz } 270796b96a96SMike Kravetz 2708369fa227SOscar Salvador /* 2709369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2710369fa227SOscar Salvador * @h: struct hstate old page belongs to 2711369fa227SOscar Salvador * @old_page: Old page to dissolve 2712ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2713369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2714369fa227SOscar Salvador */ 2715ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2716ae37c7ffSOscar Salvador struct list_head *list) 2717369fa227SOscar Salvador { 2718369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2719369fa227SOscar Salvador int nid = page_to_nid(old_page); 2720b65a4edaSMike Kravetz bool alloc_retry = false; 2721369fa227SOscar Salvador struct page *new_page; 2722369fa227SOscar Salvador int ret = 0; 2723369fa227SOscar Salvador 2724369fa227SOscar Salvador /* 2725369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2726f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2727f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2728f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2729f41f2ed4SMuchun Song * under the lock. 2730369fa227SOscar Salvador */ 2731b65a4edaSMike Kravetz alloc_retry: 2732369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2733369fa227SOscar Salvador if (!new_page) 2734369fa227SOscar Salvador return -ENOMEM; 2735b65a4edaSMike Kravetz /* 2736b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2737b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2738b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2739b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2740b65a4edaSMike Kravetz * 2741b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2742b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2743b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2744b65a4edaSMike Kravetz */ 2745b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2746b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2747b65a4edaSMike Kravetz if (alloc_retry) 2748b65a4edaSMike Kravetz return -EBUSY; 2749b65a4edaSMike Kravetz 2750b65a4edaSMike Kravetz alloc_retry = true; 2751b65a4edaSMike Kravetz goto alloc_retry; 2752b65a4edaSMike Kravetz } 2753b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2754b65a4edaSMike Kravetz 2755f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2756369fa227SOscar Salvador 2757369fa227SOscar Salvador retry: 2758369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2759369fa227SOscar Salvador if (!PageHuge(old_page)) { 2760369fa227SOscar Salvador /* 2761369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2762369fa227SOscar Salvador */ 2763369fa227SOscar Salvador goto free_new; 2764369fa227SOscar Salvador } else if (page_count(old_page)) { 2765369fa227SOscar Salvador /* 2766ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2767ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2768369fa227SOscar Salvador */ 2769ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 27707ce82f4cSMiaohe Lin ret = isolate_hugetlb(old_page, list); 2771ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2772369fa227SOscar Salvador goto free_new; 2773369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2774369fa227SOscar Salvador /* 2775369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2776369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2777369fa227SOscar Salvador * we retry. 2778369fa227SOscar Salvador */ 2779369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2780369fa227SOscar Salvador cond_resched(); 2781369fa227SOscar Salvador goto retry; 2782369fa227SOscar Salvador } else { 2783369fa227SOscar Salvador /* 2784369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2785369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2786369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2787369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2788369fa227SOscar Salvador * stable since this happens under the lock. 2789369fa227SOscar Salvador */ 2790369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2791369fa227SOscar Salvador 2792369fa227SOscar Salvador /* 2793b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2794b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2795369fa227SOscar Salvador */ 2796369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2797369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2798369fa227SOscar Salvador 2799369fa227SOscar Salvador /* 2800369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2801369fa227SOscar Salvador */ 2802369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2803b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2804369fa227SOscar Salvador } 2805369fa227SOscar Salvador 2806369fa227SOscar Salvador return ret; 2807369fa227SOscar Salvador 2808369fa227SOscar Salvador free_new: 2809369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2810b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2811b65a4edaSMike Kravetz set_page_refcounted(new_page); 2812b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2813369fa227SOscar Salvador 2814369fa227SOscar Salvador return ret; 2815369fa227SOscar Salvador } 2816369fa227SOscar Salvador 2817ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2818369fa227SOscar Salvador { 2819369fa227SOscar Salvador struct hstate *h; 2820369fa227SOscar Salvador struct page *head; 2821ae37c7ffSOscar Salvador int ret = -EBUSY; 2822369fa227SOscar Salvador 2823369fa227SOscar Salvador /* 2824369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2825369fa227SOscar Salvador * to carefully check the state under the lock. 2826369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2827369fa227SOscar Salvador */ 2828369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2829369fa227SOscar Salvador if (PageHuge(page)) { 2830369fa227SOscar Salvador head = compound_head(page); 2831369fa227SOscar Salvador h = page_hstate(head); 2832369fa227SOscar Salvador } else { 2833369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2834369fa227SOscar Salvador return 0; 2835369fa227SOscar Salvador } 2836369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2837369fa227SOscar Salvador 2838369fa227SOscar Salvador /* 2839369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2840369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2841369fa227SOscar Salvador * of bailing out right away without further retrying. 2842369fa227SOscar Salvador */ 2843369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2844369fa227SOscar Salvador return -ENOMEM; 2845369fa227SOscar Salvador 28467ce82f4cSMiaohe Lin if (page_count(head) && !isolate_hugetlb(head, list)) 2847ae37c7ffSOscar Salvador ret = 0; 2848ae37c7ffSOscar Salvador else if (!page_count(head)) 2849ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2850ae37c7ffSOscar Salvador 2851ae37c7ffSOscar Salvador return ret; 2852369fa227SOscar Salvador } 2853369fa227SOscar Salvador 285470c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 285504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2856348ea204SAdam Litke { 285790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2858a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2859348ea204SAdam Litke struct page *page; 2860d85f69b0SMike Kravetz long map_chg, map_commit; 2861d85f69b0SMike Kravetz long gbl_chg; 28626d76dcf4SAneesh Kumar K.V int ret, idx; 28636d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 286408cf9fafSMina Almasry bool deferred_reserve; 28652fc39cecSAdam Litke 28666d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2867a1e78772SMel Gorman /* 2868d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2869d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2870d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2871a1e78772SMel Gorman */ 2872d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2873d85f69b0SMike Kravetz if (map_chg < 0) 287476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2875d85f69b0SMike Kravetz 2876d85f69b0SMike Kravetz /* 2877d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2878d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2879d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2880d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2881d85f69b0SMike Kravetz * checked against any subpool limit. 2882d85f69b0SMike Kravetz */ 2883d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2884d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2885d85f69b0SMike Kravetz if (gbl_chg < 0) { 2886feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 288776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 28885e911373SMike Kravetz } 288990d8b7e6SAdam Litke 2890d85f69b0SMike Kravetz /* 2891d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2892d85f69b0SMike Kravetz * map, there could be reservations associated with the 2893d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2894d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2895d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2896d85f69b0SMike Kravetz * the subpool reservations. 2897d85f69b0SMike Kravetz */ 2898d85f69b0SMike Kravetz if (avoid_reserve) 2899d85f69b0SMike Kravetz gbl_chg = 1; 2900d85f69b0SMike Kravetz } 2901d85f69b0SMike Kravetz 290208cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 290308cf9fafSMina Almasry */ 29046501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 290508cf9fafSMina Almasry if (deferred_reserve) { 290608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 290708cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 29088f34af6fSJianyu Zhan if (ret) 29098f34af6fSJianyu Zhan goto out_subpool_put; 291008cf9fafSMina Almasry } 291108cf9fafSMina Almasry 291208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 291308cf9fafSMina Almasry if (ret) 291408cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29158f34af6fSJianyu Zhan 2916db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2917d85f69b0SMike Kravetz /* 2918d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2919d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2920d85f69b0SMike Kravetz * a reservation exists for the allocation. 2921d85f69b0SMike Kravetz */ 2922d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 292381a6fcaeSJoonsoo Kim if (!page) { 2924db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29250c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29268f34af6fSJianyu Zhan if (!page) 29278f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2928a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2929d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2930a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2931a88c7695SNaoya Horiguchi } 2932db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 293315a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 293481a6fcaeSJoonsoo Kim /* Fall through */ 2935a1e78772SMel Gorman } 293681a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 293708cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 293808cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 293908cf9fafSMina Almasry */ 294008cf9fafSMina Almasry if (deferred_reserve) { 294108cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 294208cf9fafSMina Almasry h_cg, page); 294308cf9fafSMina Almasry } 294408cf9fafSMina Almasry 2945db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2946a1e78772SMel Gorman 2947d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2948a1e78772SMel Gorman 2949d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2950d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 295133039678SMike Kravetz /* 295233039678SMike Kravetz * The page was added to the reservation map between 295333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 295433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 295533039678SMike Kravetz * Adjust for the subpool count incremented above AND 295633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 295733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 295833039678SMike Kravetz * no longer applies. 295933039678SMike Kravetz */ 296033039678SMike Kravetz long rsv_adjust; 296133039678SMike Kravetz 296233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 296333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 296479aa925bSMike Kravetz if (deferred_reserve) 296579aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 296679aa925bSMike Kravetz pages_per_huge_page(h), page); 296733039678SMike Kravetz } 29687893d1d5SAdam Litke return page; 29698f34af6fSJianyu Zhan 29708f34af6fSJianyu Zhan out_uncharge_cgroup: 29718f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 297208cf9fafSMina Almasry out_uncharge_cgroup_reservation: 297308cf9fafSMina Almasry if (deferred_reserve) 297408cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 297508cf9fafSMina Almasry h_cg); 29768f34af6fSJianyu Zhan out_subpool_put: 2977d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29788f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2979feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29808f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2981b45b5bd6SDavid Gibson } 2982b45b5bd6SDavid Gibson 2983b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 2984e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2985b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 2986aa888a74SAndi Kleen { 2987b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 2988b2261026SJoonsoo Kim int nr_nodes, node; 2989aa888a74SAndi Kleen 2990b5389086SZhenguo Yao /* do node specific alloc */ 2991b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 2992b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 2993b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 2994b5389086SZhenguo Yao if (!m) 2995b5389086SZhenguo Yao return 0; 2996b5389086SZhenguo Yao goto found; 2997b5389086SZhenguo Yao } 2998b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 2999b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3000b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 30018b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 300297ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3003aa888a74SAndi Kleen /* 3004aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3005aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3006aa888a74SAndi Kleen * puts them into the mem_map). 3007aa888a74SAndi Kleen */ 3008b5389086SZhenguo Yao if (!m) 3009b5389086SZhenguo Yao return 0; 3010aa888a74SAndi Kleen goto found; 3011aa888a74SAndi Kleen } 3012aa888a74SAndi Kleen 3013aa888a74SAndi Kleen found: 3014aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3015330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3016aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3017aa888a74SAndi Kleen m->hstate = h; 3018aa888a74SAndi Kleen return 1; 3019aa888a74SAndi Kleen } 3020aa888a74SAndi Kleen 302148b8d744SMike Kravetz /* 302248b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 302348b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 302448b8d744SMike Kravetz */ 3025aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3026aa888a74SAndi Kleen { 3027aa888a74SAndi Kleen struct huge_bootmem_page *m; 3028aa888a74SAndi Kleen 3029aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 303040d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3031aa888a74SAndi Kleen struct hstate *h = m->hstate; 3032ee8f248dSBecky Bruce 303348b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3034aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30357118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3036ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3037aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30387118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 30397118fc29SMike Kravetz } else { 3040416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30417118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30427118fc29SMike Kravetz } 3043af0fb9dfSMichal Hocko 3044b0320c7bSRafael Aquini /* 304548b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 304648b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 304748b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3048b0320c7bSRafael Aquini */ 3049c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3050520495feSCannon Matthews cond_resched(); 3051aa888a74SAndi Kleen } 3052aa888a74SAndi Kleen } 3053b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3054b5389086SZhenguo Yao { 3055b5389086SZhenguo Yao unsigned long i; 3056b5389086SZhenguo Yao char buf[32]; 3057b5389086SZhenguo Yao 3058b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3059b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3060b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3061b5389086SZhenguo Yao break; 3062b5389086SZhenguo Yao } else { 3063b5389086SZhenguo Yao struct page *page; 3064b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3065b5389086SZhenguo Yao 3066b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3067b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3068b5389086SZhenguo Yao if (!page) 3069b5389086SZhenguo Yao break; 3070b5389086SZhenguo Yao put_page(page); /* free it into the hugepage allocator */ 3071b5389086SZhenguo Yao } 3072b5389086SZhenguo Yao cond_resched(); 3073b5389086SZhenguo Yao } 3074b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3075b5389086SZhenguo Yao return; 3076b5389086SZhenguo Yao 3077b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3078b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3079b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3080b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3081b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3082b5389086SZhenguo Yao } 3083aa888a74SAndi Kleen 30848faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 30851da177e4SLinus Torvalds { 30861da177e4SLinus Torvalds unsigned long i; 3087f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3088b5389086SZhenguo Yao bool node_specific_alloc = false; 3089f60858f9SMike Kravetz 3090b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3091b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3092b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3093b5389086SZhenguo Yao return; 3094b5389086SZhenguo Yao } 3095b5389086SZhenguo Yao 3096b5389086SZhenguo Yao /* do node specific alloc */ 30970a7a0f6fSPeng Liu for_each_online_node(i) { 3098b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3099b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3100b5389086SZhenguo Yao node_specific_alloc = true; 3101b5389086SZhenguo Yao } 3102b5389086SZhenguo Yao } 3103b5389086SZhenguo Yao 3104b5389086SZhenguo Yao if (node_specific_alloc) 3105b5389086SZhenguo Yao return; 3106b5389086SZhenguo Yao 3107b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3108f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3109f60858f9SMike Kravetz /* 3110f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3111f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3112f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3113f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3114f60858f9SMike Kravetz */ 3115f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3116f60858f9SMike Kravetz GFP_KERNEL); 3117f60858f9SMike Kravetz } else { 3118f60858f9SMike Kravetz /* allocations done at boot time */ 3119f60858f9SMike Kravetz node_alloc_noretry = NULL; 3120f60858f9SMike Kravetz } 3121f60858f9SMike Kravetz 3122f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3123f60858f9SMike Kravetz if (node_alloc_noretry) 3124f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31251da177e4SLinus Torvalds 3126e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3127bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3128b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3129aa888a74SAndi Kleen break; 31300c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3131f60858f9SMike Kravetz &node_states[N_MEMORY], 3132f60858f9SMike Kravetz node_alloc_noretry)) 31331da177e4SLinus Torvalds break; 313469ed779aSDavid Rientjes cond_resched(); 31351da177e4SLinus Torvalds } 3136d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3137d715cf80SLiam R. Howlett char buf[32]; 3138d715cf80SLiam R. Howlett 3139c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3140d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3141d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31428faa8b07SAndi Kleen h->max_huge_pages = i; 3143e5ff2159SAndi Kleen } 3144f60858f9SMike Kravetz kfree(node_alloc_noretry); 3145d715cf80SLiam R. Howlett } 3146e5ff2159SAndi Kleen 3147e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3148e5ff2159SAndi Kleen { 314979dfc695SMike Kravetz struct hstate *h, *h2; 3150e5ff2159SAndi Kleen 3151e5ff2159SAndi Kleen for_each_hstate(h) { 31528faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3153bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31548faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 315579dfc695SMike Kravetz 315679dfc695SMike Kravetz /* 315779dfc695SMike Kravetz * Set demote order for each hstate. Note that 315879dfc695SMike Kravetz * h->demote_order is initially 0. 315979dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 316079dfc695SMike Kravetz * is not supported, so skip this. 3161a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3162a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 316379dfc695SMike Kravetz */ 316479dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 316579dfc695SMike Kravetz continue; 3166a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3167a01f4390SMike Kravetz continue; 316879dfc695SMike Kravetz for_each_hstate(h2) { 316979dfc695SMike Kravetz if (h2 == h) 317079dfc695SMike Kravetz continue; 317179dfc695SMike Kravetz if (h2->order < h->order && 317279dfc695SMike Kravetz h2->order > h->demote_order) 317379dfc695SMike Kravetz h->demote_order = h2->order; 317479dfc695SMike Kravetz } 3175e5ff2159SAndi Kleen } 3176e5ff2159SAndi Kleen } 3177e5ff2159SAndi Kleen 3178e5ff2159SAndi Kleen static void __init report_hugepages(void) 3179e5ff2159SAndi Kleen { 3180e5ff2159SAndi Kleen struct hstate *h; 3181e5ff2159SAndi Kleen 3182e5ff2159SAndi Kleen for_each_hstate(h) { 31834abd32dbSAndi Kleen char buf[32]; 3184c6247f72SMatthew Wilcox 3185c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 31866213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3187c6247f72SMatthew Wilcox buf, h->free_huge_pages); 31886213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 31896213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3190e5ff2159SAndi Kleen } 3191e5ff2159SAndi Kleen } 3192e5ff2159SAndi Kleen 31931da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 31946ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 31956ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31961da177e4SLinus Torvalds { 31974415cc8dSChristoph Lameter int i; 31981121828aSMike Kravetz LIST_HEAD(page_list); 31994415cc8dSChristoph Lameter 32009487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3201bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3202aa888a74SAndi Kleen return; 3203aa888a74SAndi Kleen 32041121828aSMike Kravetz /* 32051121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 32061121828aSMike Kravetz */ 32076ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 320810c6ec49SMike Kravetz struct page *page, *next; 3209a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3210a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3211a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32121121828aSMike Kravetz goto out; 32131da177e4SLinus Torvalds if (PageHighMem(page)) 32141da177e4SLinus Torvalds continue; 32156eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32161121828aSMike Kravetz list_add(&page->lru, &page_list); 32171121828aSMike Kravetz } 32181121828aSMike Kravetz } 32191121828aSMike Kravetz 32201121828aSMike Kravetz out: 3221db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 322210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3223db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32241da177e4SLinus Torvalds } 32251da177e4SLinus Torvalds #else 32266ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32276ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32281da177e4SLinus Torvalds { 32291da177e4SLinus Torvalds } 32301da177e4SLinus Torvalds #endif 32311da177e4SLinus Torvalds 323220a0307cSWu Fengguang /* 323320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 323420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 323520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 323620a0307cSWu Fengguang */ 32376ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32386ae11b27SLee Schermerhorn int delta) 323920a0307cSWu Fengguang { 3240b2261026SJoonsoo Kim int nr_nodes, node; 324120a0307cSWu Fengguang 32429487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 324320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 324420a0307cSWu Fengguang 3245e8c5c824SLee Schermerhorn if (delta < 0) { 3246b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3247b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3248b2261026SJoonsoo Kim goto found; 3249b2261026SJoonsoo Kim } 3250b2261026SJoonsoo Kim } else { 3251b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3252b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3253b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3254b2261026SJoonsoo Kim goto found; 3255e8c5c824SLee Schermerhorn } 32569a76db09SLee Schermerhorn } 3257b2261026SJoonsoo Kim return 0; 325820a0307cSWu Fengguang 3259b2261026SJoonsoo Kim found: 326020a0307cSWu Fengguang h->surplus_huge_pages += delta; 3261b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3262b2261026SJoonsoo Kim return 1; 326320a0307cSWu Fengguang } 326420a0307cSWu Fengguang 3265a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3266fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32676ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32681da177e4SLinus Torvalds { 32697893d1d5SAdam Litke unsigned long min_count, ret; 327010c6ec49SMike Kravetz struct page *page; 327110c6ec49SMike Kravetz LIST_HEAD(page_list); 3272f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3273f60858f9SMike Kravetz 3274f60858f9SMike Kravetz /* 3275f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3276f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3277f60858f9SMike Kravetz * the requested huge pages. 3278f60858f9SMike Kravetz */ 3279f60858f9SMike Kravetz if (node_alloc_noretry) 3280f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3281f60858f9SMike Kravetz else 3282f60858f9SMike Kravetz return -ENOMEM; 32831da177e4SLinus Torvalds 328429383967SMike Kravetz /* 328529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 328629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 328729383967SMike Kravetz */ 328829383967SMike Kravetz mutex_lock(&h->resize_lock); 3289b65d4adbSMuchun Song flush_free_hpage_work(h); 3290db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32914eb0716eSAlexandre Ghiti 32924eb0716eSAlexandre Ghiti /* 3293fd875dcaSMike Kravetz * Check for a node specific request. 3294fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3295fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3296fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3297fd875dcaSMike Kravetz */ 3298fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3299fd875dcaSMike Kravetz unsigned long old_count = count; 3300fd875dcaSMike Kravetz 3301fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3302fd875dcaSMike Kravetz /* 3303fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3304fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3305fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3306fd875dcaSMike Kravetz * largest possible value to align with their intention. 3307fd875dcaSMike Kravetz */ 3308fd875dcaSMike Kravetz if (count < old_count) 3309fd875dcaSMike Kravetz count = ULONG_MAX; 3310fd875dcaSMike Kravetz } 3311fd875dcaSMike Kravetz 3312fd875dcaSMike Kravetz /* 33134eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33144eb0716eSAlexandre Ghiti * page range allocation. 33154eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33164eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33174eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33184eb0716eSAlexandre Ghiti */ 33194eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33204eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3321db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 332229383967SMike Kravetz mutex_unlock(&h->resize_lock); 3323f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33244eb0716eSAlexandre Ghiti return -EINVAL; 33254eb0716eSAlexandre Ghiti } 33264eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33274eb0716eSAlexandre Ghiti } 3328aa888a74SAndi Kleen 33297893d1d5SAdam Litke /* 33307893d1d5SAdam Litke * Increase the pool size 33317893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33327893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3333d1c3fb1fSNishanth Aravamudan * 33340c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3335d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3336d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3337d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3338d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33397893d1d5SAdam Litke */ 3340a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33416ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33427893d1d5SAdam Litke break; 33437893d1d5SAdam Litke } 33447893d1d5SAdam Litke 3345a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33467893d1d5SAdam Litke /* 33477893d1d5SAdam Litke * If this allocation races such that we no longer need the 33487893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33497893d1d5SAdam Litke * and reducing the surplus. 33507893d1d5SAdam Litke */ 3351db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3352649920c6SJia He 3353649920c6SJia He /* yield cpu to avoid soft lockup */ 3354649920c6SJia He cond_resched(); 3355649920c6SJia He 3356f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3357f60858f9SMike Kravetz node_alloc_noretry); 3358db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33597893d1d5SAdam Litke if (!ret) 33607893d1d5SAdam Litke goto out; 33617893d1d5SAdam Litke 3362536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3363536240f2SMel Gorman if (signal_pending(current)) 3364536240f2SMel Gorman goto out; 33657893d1d5SAdam Litke } 33667893d1d5SAdam Litke 33677893d1d5SAdam Litke /* 33687893d1d5SAdam Litke * Decrease the pool size 33697893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33707893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33717893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33727893d1d5SAdam Litke * to the desired size as pages become free. 3373d1c3fb1fSNishanth Aravamudan * 3374d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3375d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3376d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33770c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3378d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3379d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3380d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33817893d1d5SAdam Litke */ 3382a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33836b0c880dSAdam Litke min_count = max(count, min_count); 33846ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 338510c6ec49SMike Kravetz 338610c6ec49SMike Kravetz /* 338710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 338810c6ec49SMike Kravetz */ 3389a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 339010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 339110c6ec49SMike Kravetz if (!page) 33921da177e4SLinus Torvalds break; 339310c6ec49SMike Kravetz 339410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 33951da177e4SLinus Torvalds } 339610c6ec49SMike Kravetz /* free the pages after dropping lock */ 3397db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 339810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3399b65d4adbSMuchun Song flush_free_hpage_work(h); 3400db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 340110c6ec49SMike Kravetz 3402a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 34036ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 34047893d1d5SAdam Litke break; 34057893d1d5SAdam Litke } 34067893d1d5SAdam Litke out: 34074eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3408db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 340929383967SMike Kravetz mutex_unlock(&h->resize_lock); 34104eb0716eSAlexandre Ghiti 3411f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3412f60858f9SMike Kravetz 34134eb0716eSAlexandre Ghiti return 0; 34141da177e4SLinus Torvalds } 34151da177e4SLinus Torvalds 34168531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34178531fc6fSMike Kravetz { 34188531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34198531fc6fSMike Kravetz struct hstate *target_hstate; 34208531fc6fSMike Kravetz int rc = 0; 34218531fc6fSMike Kravetz 34228531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34238531fc6fSMike Kravetz 34248531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34258531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34268531fc6fSMike Kravetz 34276213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 34288531fc6fSMike Kravetz if (rc) { 34298531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34308531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34318531fc6fSMike Kravetz set_page_refcounted(page); 34328531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34338531fc6fSMike Kravetz return rc; 34348531fc6fSMike Kravetz } 34358531fc6fSMike Kravetz 34368531fc6fSMike Kravetz /* 34378531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34388531fc6fSMike Kravetz * sizes as it will not ref count pages. 34398531fc6fSMike Kravetz */ 34408531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34418531fc6fSMike Kravetz 34428531fc6fSMike Kravetz /* 34438531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34448531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34458531fc6fSMike Kravetz * as surplus. 34468531fc6fSMike Kravetz * 34478531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34488531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34498531fc6fSMike Kravetz */ 34508531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34518531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34528531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 34538531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 34548531fc6fSMike Kravetz prep_compound_gigantic_page_for_demote(page + i, 34558531fc6fSMike Kravetz target_hstate->order); 34568531fc6fSMike Kravetz else 34578531fc6fSMike Kravetz prep_compound_page(page + i, target_hstate->order); 34588531fc6fSMike Kravetz set_page_private(page + i, 0); 34598531fc6fSMike Kravetz set_page_refcounted(page + i); 34608531fc6fSMike Kravetz prep_new_huge_page(target_hstate, page + i, nid); 34618531fc6fSMike Kravetz put_page(page + i); 34628531fc6fSMike Kravetz } 34638531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34648531fc6fSMike Kravetz 34658531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34668531fc6fSMike Kravetz 34678531fc6fSMike Kravetz /* 34688531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34698531fc6fSMike Kravetz * based on pool changes for the demoted page. 34708531fc6fSMike Kravetz */ 34718531fc6fSMike Kravetz h->max_huge_pages--; 3472a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3473a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 34748531fc6fSMike Kravetz 34758531fc6fSMike Kravetz return rc; 34768531fc6fSMike Kravetz } 34778531fc6fSMike Kravetz 347879dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 347979dfc695SMike Kravetz __must_hold(&hugetlb_lock) 348079dfc695SMike Kravetz { 34818531fc6fSMike Kravetz int nr_nodes, node; 34828531fc6fSMike Kravetz struct page *page; 348379dfc695SMike Kravetz 348479dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 348579dfc695SMike Kravetz 348679dfc695SMike Kravetz /* We should never get here if no demote order */ 348779dfc695SMike Kravetz if (!h->demote_order) { 348879dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 348979dfc695SMike Kravetz return -EINVAL; /* internal error */ 349079dfc695SMike Kravetz } 349179dfc695SMike Kravetz 34928531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 34935a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 34945a317412SMike Kravetz if (PageHWPoison(page)) 34955a317412SMike Kravetz continue; 34965a317412SMike Kravetz 34975a317412SMike Kravetz return demote_free_huge_page(h, page); 34988531fc6fSMike Kravetz } 34998531fc6fSMike Kravetz } 35008531fc6fSMike Kravetz 35015a317412SMike Kravetz /* 35025a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 35035a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 35045a317412SMike Kravetz */ 35055a317412SMike Kravetz return -EBUSY; 350679dfc695SMike Kravetz } 350779dfc695SMike Kravetz 3508a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3509a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3510a3437870SNishanth Aravamudan 351179dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 351279dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 351379dfc695SMike Kravetz 3514a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 351598bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3516a3437870SNishanth Aravamudan 3517a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3518a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3519a3437870SNishanth Aravamudan 35209a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35219a305230SLee Schermerhorn 35229a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3523a3437870SNishanth Aravamudan { 3524a3437870SNishanth Aravamudan int i; 35259a305230SLee Schermerhorn 3526a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35279a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35289a305230SLee Schermerhorn if (nidp) 35299a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3530a3437870SNishanth Aravamudan return &hstates[i]; 35319a305230SLee Schermerhorn } 35329a305230SLee Schermerhorn 35339a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3534a3437870SNishanth Aravamudan } 3535a3437870SNishanth Aravamudan 353606808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3537a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3538a3437870SNishanth Aravamudan { 35399a305230SLee Schermerhorn struct hstate *h; 35409a305230SLee Schermerhorn unsigned long nr_huge_pages; 35419a305230SLee Schermerhorn int nid; 35429a305230SLee Schermerhorn 35439a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35449a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35459a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35469a305230SLee Schermerhorn else 35479a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35489a305230SLee Schermerhorn 3549ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3550a3437870SNishanth Aravamudan } 3551adbe8726SEric B Munson 3552238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3553238d3c13SDavid Rientjes struct hstate *h, int nid, 3554238d3c13SDavid Rientjes unsigned long count, size_t len) 3555a3437870SNishanth Aravamudan { 3556a3437870SNishanth Aravamudan int err; 35572d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3558a3437870SNishanth Aravamudan 35592d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35602d0adf7eSOscar Salvador return -EINVAL; 3561adbe8726SEric B Munson 35629a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35639a305230SLee Schermerhorn /* 35649a305230SLee Schermerhorn * global hstate attribute 35659a305230SLee Schermerhorn */ 35669a305230SLee Schermerhorn if (!(obey_mempolicy && 35672d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35682d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35692d0adf7eSOscar Salvador else 35702d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35712d0adf7eSOscar Salvador } else { 35729a305230SLee Schermerhorn /* 3573fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3574fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35759a305230SLee Schermerhorn */ 35762d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35772d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3578fd875dcaSMike Kravetz } 35799a305230SLee Schermerhorn 35802d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 358106808b08SLee Schermerhorn 35824eb0716eSAlexandre Ghiti return err ? err : len; 358306808b08SLee Schermerhorn } 358406808b08SLee Schermerhorn 3585238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3586238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3587238d3c13SDavid Rientjes size_t len) 3588238d3c13SDavid Rientjes { 3589238d3c13SDavid Rientjes struct hstate *h; 3590238d3c13SDavid Rientjes unsigned long count; 3591238d3c13SDavid Rientjes int nid; 3592238d3c13SDavid Rientjes int err; 3593238d3c13SDavid Rientjes 3594238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3595238d3c13SDavid Rientjes if (err) 3596238d3c13SDavid Rientjes return err; 3597238d3c13SDavid Rientjes 3598238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3599238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3600238d3c13SDavid Rientjes } 3601238d3c13SDavid Rientjes 360206808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 360306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 360406808b08SLee Schermerhorn { 360506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 360606808b08SLee Schermerhorn } 360706808b08SLee Schermerhorn 360806808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 360906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 361006808b08SLee Schermerhorn { 3611238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3612a3437870SNishanth Aravamudan } 3613a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3614a3437870SNishanth Aravamudan 361506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 361606808b08SLee Schermerhorn 361706808b08SLee Schermerhorn /* 361806808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 361906808b08SLee Schermerhorn * huge page alloc/free. 362006808b08SLee Schermerhorn */ 362106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3622ae7a927dSJoe Perches struct kobj_attribute *attr, 3623ae7a927dSJoe Perches char *buf) 362406808b08SLee Schermerhorn { 362506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 362606808b08SLee Schermerhorn } 362706808b08SLee Schermerhorn 362806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 362906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 363006808b08SLee Schermerhorn { 3631238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 363206808b08SLee Schermerhorn } 363306808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 363406808b08SLee Schermerhorn #endif 363506808b08SLee Schermerhorn 363606808b08SLee Schermerhorn 3637a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3638a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3639a3437870SNishanth Aravamudan { 36409a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3641ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3642a3437870SNishanth Aravamudan } 3643adbe8726SEric B Munson 3644a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3645a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3646a3437870SNishanth Aravamudan { 3647a3437870SNishanth Aravamudan int err; 3648a3437870SNishanth Aravamudan unsigned long input; 36499a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3650a3437870SNishanth Aravamudan 3651bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3652adbe8726SEric B Munson return -EINVAL; 3653adbe8726SEric B Munson 36543dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3655a3437870SNishanth Aravamudan if (err) 365673ae31e5SEric B Munson return err; 3657a3437870SNishanth Aravamudan 3658db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3659a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3660db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3661a3437870SNishanth Aravamudan 3662a3437870SNishanth Aravamudan return count; 3663a3437870SNishanth Aravamudan } 3664a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3665a3437870SNishanth Aravamudan 3666a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3667a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3668a3437870SNishanth Aravamudan { 36699a305230SLee Schermerhorn struct hstate *h; 36709a305230SLee Schermerhorn unsigned long free_huge_pages; 36719a305230SLee Schermerhorn int nid; 36729a305230SLee Schermerhorn 36739a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36749a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36759a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36769a305230SLee Schermerhorn else 36779a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36789a305230SLee Schermerhorn 3679ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3680a3437870SNishanth Aravamudan } 3681a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3682a3437870SNishanth Aravamudan 3683a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3684a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3685a3437870SNishanth Aravamudan { 36869a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3687ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3688a3437870SNishanth Aravamudan } 3689a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3690a3437870SNishanth Aravamudan 3691a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3692a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3693a3437870SNishanth Aravamudan { 36949a305230SLee Schermerhorn struct hstate *h; 36959a305230SLee Schermerhorn unsigned long surplus_huge_pages; 36969a305230SLee Schermerhorn int nid; 36979a305230SLee Schermerhorn 36989a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36999a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37009a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 37019a305230SLee Schermerhorn else 37029a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 37039a305230SLee Schermerhorn 3704ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3705a3437870SNishanth Aravamudan } 3706a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3707a3437870SNishanth Aravamudan 370879dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 370979dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 371079dfc695SMike Kravetz { 371179dfc695SMike Kravetz unsigned long nr_demote; 371279dfc695SMike Kravetz unsigned long nr_available; 371379dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 371479dfc695SMike Kravetz struct hstate *h; 371579dfc695SMike Kravetz int err = 0; 371679dfc695SMike Kravetz int nid; 371779dfc695SMike Kravetz 371879dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 371979dfc695SMike Kravetz if (err) 372079dfc695SMike Kravetz return err; 372179dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 372279dfc695SMike Kravetz 372379dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 372479dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 372579dfc695SMike Kravetz n_mask = &nodes_allowed; 372679dfc695SMike Kravetz } else { 372779dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 372879dfc695SMike Kravetz } 372979dfc695SMike Kravetz 373079dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 373179dfc695SMike Kravetz mutex_lock(&h->resize_lock); 373279dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 373379dfc695SMike Kravetz 373479dfc695SMike Kravetz while (nr_demote) { 373579dfc695SMike Kravetz /* 373679dfc695SMike Kravetz * Check for available pages to demote each time thorough the 373779dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 373879dfc695SMike Kravetz */ 373979dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 374079dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 374179dfc695SMike Kravetz else 374279dfc695SMike Kravetz nr_available = h->free_huge_pages; 374379dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 374479dfc695SMike Kravetz if (!nr_available) 374579dfc695SMike Kravetz break; 374679dfc695SMike Kravetz 374779dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 374879dfc695SMike Kravetz if (err) 374979dfc695SMike Kravetz break; 375079dfc695SMike Kravetz 375179dfc695SMike Kravetz nr_demote--; 375279dfc695SMike Kravetz } 375379dfc695SMike Kravetz 375479dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 375579dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 375679dfc695SMike Kravetz 375779dfc695SMike Kravetz if (err) 375879dfc695SMike Kravetz return err; 375979dfc695SMike Kravetz return len; 376079dfc695SMike Kravetz } 376179dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 376279dfc695SMike Kravetz 376379dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 376479dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 376579dfc695SMike Kravetz { 3766*12658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 376779dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 376879dfc695SMike Kravetz 376979dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 377079dfc695SMike Kravetz } 377179dfc695SMike Kravetz 377279dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 377379dfc695SMike Kravetz struct kobj_attribute *attr, 377479dfc695SMike Kravetz const char *buf, size_t count) 377579dfc695SMike Kravetz { 377679dfc695SMike Kravetz struct hstate *h, *demote_hstate; 377779dfc695SMike Kravetz unsigned long demote_size; 377879dfc695SMike Kravetz unsigned int demote_order; 377979dfc695SMike Kravetz 378079dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 378179dfc695SMike Kravetz 378279dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 378379dfc695SMike Kravetz if (!demote_hstate) 378479dfc695SMike Kravetz return -EINVAL; 378579dfc695SMike Kravetz demote_order = demote_hstate->order; 3786a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3787a01f4390SMike Kravetz return -EINVAL; 378879dfc695SMike Kravetz 378979dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 3790*12658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 379179dfc695SMike Kravetz if (demote_order >= h->order) 379279dfc695SMike Kravetz return -EINVAL; 379379dfc695SMike Kravetz 379479dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 379579dfc695SMike Kravetz mutex_lock(&h->resize_lock); 379679dfc695SMike Kravetz h->demote_order = demote_order; 379779dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 379879dfc695SMike Kravetz 379979dfc695SMike Kravetz return count; 380079dfc695SMike Kravetz } 380179dfc695SMike Kravetz HSTATE_ATTR(demote_size); 380279dfc695SMike Kravetz 3803a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3804a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3805a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3806a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3807a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3808a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 380906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 381006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 381106808b08SLee Schermerhorn #endif 3812a3437870SNishanth Aravamudan NULL, 3813a3437870SNishanth Aravamudan }; 3814a3437870SNishanth Aravamudan 381567e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3816a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3817a3437870SNishanth Aravamudan }; 3818a3437870SNishanth Aravamudan 381979dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 382079dfc695SMike Kravetz &demote_size_attr.attr, 382179dfc695SMike Kravetz &demote_attr.attr, 382279dfc695SMike Kravetz NULL, 382379dfc695SMike Kravetz }; 382479dfc695SMike Kravetz 382579dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 382679dfc695SMike Kravetz .attrs = hstate_demote_attrs, 382779dfc695SMike Kravetz }; 382879dfc695SMike Kravetz 3829094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38309a305230SLee Schermerhorn struct kobject **hstate_kobjs, 383167e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3832a3437870SNishanth Aravamudan { 3833a3437870SNishanth Aravamudan int retval; 3834972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3835a3437870SNishanth Aravamudan 38369a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38379a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3838a3437870SNishanth Aravamudan return -ENOMEM; 3839a3437870SNishanth Aravamudan 38409a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3841cc2205a6SMiaohe Lin if (retval) { 38429a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3843cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 38443a6bdda0SMiaohe Lin return retval; 3845cc2205a6SMiaohe Lin } 3846a3437870SNishanth Aravamudan 384779dfc695SMike Kravetz if (h->demote_order) { 384801088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 384901088a60SMiaohe Lin &hstate_demote_attr_group); 385001088a60SMiaohe Lin if (retval) { 385179dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 385201088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 385301088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 385401088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 385501088a60SMiaohe Lin return retval; 385601088a60SMiaohe Lin } 385779dfc695SMike Kravetz } 385879dfc695SMike Kravetz 385901088a60SMiaohe Lin return 0; 3860a3437870SNishanth Aravamudan } 3861a3437870SNishanth Aravamudan 3862a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3863a3437870SNishanth Aravamudan { 3864a3437870SNishanth Aravamudan struct hstate *h; 3865a3437870SNishanth Aravamudan int err; 3866a3437870SNishanth Aravamudan 3867a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3868a3437870SNishanth Aravamudan if (!hugepages_kobj) 3869a3437870SNishanth Aravamudan return; 3870a3437870SNishanth Aravamudan 3871a3437870SNishanth Aravamudan for_each_hstate(h) { 38729a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 38739a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3874a3437870SNishanth Aravamudan if (err) 3875282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3876a3437870SNishanth Aravamudan } 3877a3437870SNishanth Aravamudan } 3878a3437870SNishanth Aravamudan 38799a305230SLee Schermerhorn #ifdef CONFIG_NUMA 38809a305230SLee Schermerhorn 38819a305230SLee Schermerhorn /* 38829a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 388310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 388410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 388510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38869a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38879a305230SLee Schermerhorn */ 38889a305230SLee Schermerhorn struct node_hstate { 38899a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38909a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38919a305230SLee Schermerhorn }; 3892b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38939a305230SLee Schermerhorn 38949a305230SLee Schermerhorn /* 389510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 38969a305230SLee Schermerhorn */ 38979a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 38989a305230SLee Schermerhorn &nr_hugepages_attr.attr, 38999a305230SLee Schermerhorn &free_hugepages_attr.attr, 39009a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 39019a305230SLee Schermerhorn NULL, 39029a305230SLee Schermerhorn }; 39039a305230SLee Schermerhorn 390467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 39059a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 39069a305230SLee Schermerhorn }; 39079a305230SLee Schermerhorn 39089a305230SLee Schermerhorn /* 390910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 39109a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 39119a305230SLee Schermerhorn */ 39129a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39139a305230SLee Schermerhorn { 39149a305230SLee Schermerhorn int nid; 39159a305230SLee Schermerhorn 39169a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 39179a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39189a305230SLee Schermerhorn int i; 39199a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39209a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39219a305230SLee Schermerhorn if (nidp) 39229a305230SLee Schermerhorn *nidp = nid; 39239a305230SLee Schermerhorn return &hstates[i]; 39249a305230SLee Schermerhorn } 39259a305230SLee Schermerhorn } 39269a305230SLee Schermerhorn 39279a305230SLee Schermerhorn BUG(); 39289a305230SLee Schermerhorn return NULL; 39299a305230SLee Schermerhorn } 39309a305230SLee Schermerhorn 39319a305230SLee Schermerhorn /* 393210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39339a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39349a305230SLee Schermerhorn */ 39353cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 39369a305230SLee Schermerhorn { 39379a305230SLee Schermerhorn struct hstate *h; 393810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39399a305230SLee Schermerhorn 39409a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39419b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39429a305230SLee Schermerhorn 3943972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3944972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 394501088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 394601088a60SMiaohe Lin 394701088a60SMiaohe Lin if (!hstate_kobj) 394801088a60SMiaohe Lin continue; 394901088a60SMiaohe Lin if (h->demote_order) 395001088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 395101088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 395201088a60SMiaohe Lin kobject_put(hstate_kobj); 3953972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3954972dc4deSAneesh Kumar K.V } 39559a305230SLee Schermerhorn 39569a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39579a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39589a305230SLee Schermerhorn } 39599a305230SLee Schermerhorn 39609a305230SLee Schermerhorn 39619a305230SLee Schermerhorn /* 396210fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39639a305230SLee Schermerhorn * No-op if attributes already registered. 39649a305230SLee Schermerhorn */ 39653cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 39669a305230SLee Schermerhorn { 39679a305230SLee Schermerhorn struct hstate *h; 396810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39699a305230SLee Schermerhorn int err; 39709a305230SLee Schermerhorn 39719a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39729a305230SLee Schermerhorn return; /* already allocated */ 39739a305230SLee Schermerhorn 39749a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 397510fbcf4cSKay Sievers &node->dev.kobj); 39769a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39779a305230SLee Schermerhorn return; 39789a305230SLee Schermerhorn 39799a305230SLee Schermerhorn for_each_hstate(h) { 39809a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39819a305230SLee Schermerhorn nhs->hstate_kobjs, 39829a305230SLee Schermerhorn &per_node_hstate_attr_group); 39839a305230SLee Schermerhorn if (err) { 3984282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 398510fbcf4cSKay Sievers h->name, node->dev.id); 39869a305230SLee Schermerhorn hugetlb_unregister_node(node); 39879a305230SLee Schermerhorn break; 39889a305230SLee Schermerhorn } 39899a305230SLee Schermerhorn } 39909a305230SLee Schermerhorn } 39919a305230SLee Schermerhorn 39929a305230SLee Schermerhorn /* 39939b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 399410fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 399510fbcf4cSKay Sievers * registered their associated device by this time. 39969a305230SLee Schermerhorn */ 39977d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 39989a305230SLee Schermerhorn { 39999a305230SLee Schermerhorn int nid; 40009a305230SLee Schermerhorn 40018cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 40028732794bSWen Congyang struct node *node = node_devices[nid]; 400310fbcf4cSKay Sievers if (node->dev.id == nid) 40049a305230SLee Schermerhorn hugetlb_register_node(node); 40059a305230SLee Schermerhorn } 40069a305230SLee Schermerhorn 40079a305230SLee Schermerhorn /* 400810fbcf4cSKay Sievers * Let the node device driver know we're here so it can 40099a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 40109a305230SLee Schermerhorn */ 40119a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 40129a305230SLee Schermerhorn hugetlb_unregister_node); 40139a305230SLee Schermerhorn } 40149a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 40159a305230SLee Schermerhorn 40169a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40179a305230SLee Schermerhorn { 40189a305230SLee Schermerhorn BUG(); 40199a305230SLee Schermerhorn if (nidp) 40209a305230SLee Schermerhorn *nidp = -1; 40219a305230SLee Schermerhorn return NULL; 40229a305230SLee Schermerhorn } 40239a305230SLee Schermerhorn 40249a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40259a305230SLee Schermerhorn 40269a305230SLee Schermerhorn #endif 40279a305230SLee Schermerhorn 4028263b8998SMiaohe Lin #ifdef CONFIG_CMA 4029263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4030263b8998SMiaohe Lin #else 4031263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4032263b8998SMiaohe Lin { 4033263b8998SMiaohe Lin } 4034263b8998SMiaohe Lin #endif 4035263b8998SMiaohe Lin 4036a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4037a3437870SNishanth Aravamudan { 40388382d914SDavidlohr Bueso int i; 40398382d914SDavidlohr Bueso 4040d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4041d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4042d6995da3SMike Kravetz 4043c2833a5bSMike Kravetz if (!hugepages_supported()) { 4044c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4045c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40460ef89d25SBenjamin Herrenschmidt return 0; 4047d715cf80SLiam R. Howlett } 4048d715cf80SLiam R. Howlett 4049282f4214SMike Kravetz /* 4050282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4051282f4214SMike Kravetz * architectures depend on setup being done here. 4052282f4214SMike Kravetz */ 4053a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4054282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4055282f4214SMike Kravetz /* 4056282f4214SMike Kravetz * If we did not parse a default huge page size, set 4057282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4058282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4059282f4214SMike Kravetz * specified, set that here as well. 4060282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4061282f4214SMike Kravetz * setting. A warning will be printed in this case. 4062282f4214SMike Kravetz */ 4063282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4064f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4065282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4066282f4214SMike Kravetz char buf[32]; 4067282f4214SMike Kravetz 4068282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4069282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4070282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4071282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4072282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4073282f4214SMike Kravetz default_hstate_max_huge_pages); 4074282f4214SMike Kravetz } 4075282f4214SMike Kravetz default_hstate.max_huge_pages = 4076282f4214SMike Kravetz default_hstate_max_huge_pages; 4077b5389086SZhenguo Yao 40780a7a0f6fSPeng Liu for_each_online_node(i) 4079b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4080b5389086SZhenguo Yao default_hugepages_in_node[i]; 4081282f4214SMike Kravetz } 4082f8b74815SVaishali Thakkar } 4083a3437870SNishanth Aravamudan 4084cf11e85fSRoman Gushchin hugetlb_cma_check(); 4085a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4086aa888a74SAndi Kleen gather_bootmem_prealloc(); 4087a3437870SNishanth Aravamudan report_hugepages(); 4088a3437870SNishanth Aravamudan 4089a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 40909a305230SLee Schermerhorn hugetlb_register_all_nodes(); 40917179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 40929a305230SLee Schermerhorn 40938382d914SDavidlohr Bueso #ifdef CONFIG_SMP 40948382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 40958382d914SDavidlohr Bueso #else 40968382d914SDavidlohr Bueso num_fault_mutexes = 1; 40978382d914SDavidlohr Bueso #endif 4098c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 40996da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 41006da2ec56SKees Cook GFP_KERNEL); 4101c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 41028382d914SDavidlohr Bueso 41038382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4104c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4105a3437870SNishanth Aravamudan return 0; 4106a3437870SNishanth Aravamudan } 41073e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4108a3437870SNishanth Aravamudan 4109ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4110ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 41119fee021dSVaishali Thakkar { 4112ae94da89SMike Kravetz return size == HPAGE_SIZE; 41139fee021dSVaishali Thakkar } 41149fee021dSVaishali Thakkar 4115d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4116a3437870SNishanth Aravamudan { 4117a3437870SNishanth Aravamudan struct hstate *h; 41188faa8b07SAndi Kleen unsigned long i; 41198faa8b07SAndi Kleen 4120a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4121a3437870SNishanth Aravamudan return; 4122a3437870SNishanth Aravamudan } 412347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4124a3437870SNishanth Aravamudan BUG_ON(order == 0); 412547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 412629383967SMike Kravetz mutex_init(&h->resize_lock); 4127a3437870SNishanth Aravamudan h->order = order; 4128aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 41298faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 41308faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41310edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 413254f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 413354f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4134a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4135c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 41368faa8b07SAndi Kleen 4137a3437870SNishanth Aravamudan parsed_hstate = h; 4138a3437870SNishanth Aravamudan } 4139a3437870SNishanth Aravamudan 4140b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4141b5389086SZhenguo Yao { 4142b5389086SZhenguo Yao return true; 4143b5389086SZhenguo Yao } 4144f87442f4SPeng Liu 4145f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4146f87442f4SPeng Liu { 4147f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4148f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4149f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 415010395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4151f87442f4SPeng Liu } else { 4152f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4153f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 415410395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4155f87442f4SPeng Liu } 4156f87442f4SPeng Liu } 4157f87442f4SPeng Liu 4158282f4214SMike Kravetz /* 4159282f4214SMike Kravetz * hugepages command line processing 4160282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4161282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4162282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4163282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4164282f4214SMike Kravetz */ 4165282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4166a3437870SNishanth Aravamudan { 4167a3437870SNishanth Aravamudan unsigned long *mhp; 41688faa8b07SAndi Kleen static unsigned long *last_mhp; 4169b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4170b5389086SZhenguo Yao int count; 4171b5389086SZhenguo Yao unsigned long tmp; 4172b5389086SZhenguo Yao char *p = s; 4173a3437870SNishanth Aravamudan 41749fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4175282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41769fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4177f81f6e4bSPeng Liu return 1; 41789fee021dSVaishali Thakkar } 4179282f4214SMike Kravetz 4180a3437870SNishanth Aravamudan /* 4181282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4182282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4183282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4184282f4214SMike Kravetz * default_hugepagesz. 4185a3437870SNishanth Aravamudan */ 41869fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4187a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4188a3437870SNishanth Aravamudan else 4189a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4190a3437870SNishanth Aravamudan 41918faa8b07SAndi Kleen if (mhp == last_mhp) { 4192282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4193f81f6e4bSPeng Liu return 1; 41948faa8b07SAndi Kleen } 41958faa8b07SAndi Kleen 4196b5389086SZhenguo Yao while (*p) { 4197b5389086SZhenguo Yao count = 0; 4198b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4199b5389086SZhenguo Yao goto invalid; 4200b5389086SZhenguo Yao /* Parameter is node format */ 4201b5389086SZhenguo Yao if (p[count] == ':') { 4202b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4203b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4204f81f6e4bSPeng Liu return 1; 4205b5389086SZhenguo Yao } 42060a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4207e79ce983SLiu Yuntao goto invalid; 42080a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4209b5389086SZhenguo Yao p += count + 1; 4210b5389086SZhenguo Yao /* Parse hugepages */ 4211b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4212b5389086SZhenguo Yao goto invalid; 4213b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4214b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4215b5389086SZhenguo Yao else 4216b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4217b5389086SZhenguo Yao *mhp += tmp; 4218b5389086SZhenguo Yao /* Go to parse next node*/ 4219b5389086SZhenguo Yao if (p[count] == ',') 4220b5389086SZhenguo Yao p += count + 1; 4221b5389086SZhenguo Yao else 4222b5389086SZhenguo Yao break; 4223b5389086SZhenguo Yao } else { 4224b5389086SZhenguo Yao if (p != s) 4225b5389086SZhenguo Yao goto invalid; 4226b5389086SZhenguo Yao *mhp = tmp; 4227b5389086SZhenguo Yao break; 4228b5389086SZhenguo Yao } 4229b5389086SZhenguo Yao } 4230a3437870SNishanth Aravamudan 42318faa8b07SAndi Kleen /* 42328faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 423304adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 42348faa8b07SAndi Kleen * use the bootmem allocator. 42358faa8b07SAndi Kleen */ 423604adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 42378faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 42388faa8b07SAndi Kleen 42398faa8b07SAndi Kleen last_mhp = mhp; 42408faa8b07SAndi Kleen 4241a3437870SNishanth Aravamudan return 1; 4242b5389086SZhenguo Yao 4243b5389086SZhenguo Yao invalid: 4244b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4245f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4246f81f6e4bSPeng Liu return 1; 4247a3437870SNishanth Aravamudan } 4248282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4249e11bfbfcSNick Piggin 4250282f4214SMike Kravetz /* 4251282f4214SMike Kravetz * hugepagesz command line processing 4252282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4253282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4254282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4255282f4214SMike Kravetz * hugepagesz argument was valid. 4256282f4214SMike Kravetz */ 4257359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4258e11bfbfcSNick Piggin { 4259359f2544SMike Kravetz unsigned long size; 4260282f4214SMike Kravetz struct hstate *h; 4261282f4214SMike Kravetz 4262282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4263359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4264359f2544SMike Kravetz 4265359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4266282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4267f81f6e4bSPeng Liu return 1; 4268359f2544SMike Kravetz } 4269359f2544SMike Kravetz 4270282f4214SMike Kravetz h = size_to_hstate(size); 4271282f4214SMike Kravetz if (h) { 4272282f4214SMike Kravetz /* 4273282f4214SMike Kravetz * hstate for this size already exists. This is normally 4274282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4275282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4276282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4277282f4214SMike Kravetz * previously specified. 4278282f4214SMike Kravetz */ 4279282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4280282f4214SMike Kravetz default_hstate.max_huge_pages) { 4281282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4282f81f6e4bSPeng Liu return 1; 428338237830SMike Kravetz } 428438237830SMike Kravetz 4285282f4214SMike Kravetz /* 4286282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4287282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4288282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4289282f4214SMike Kravetz */ 4290282f4214SMike Kravetz parsed_hstate = h; 4291282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4292e11bfbfcSNick Piggin return 1; 4293e11bfbfcSNick Piggin } 4294282f4214SMike Kravetz 4295359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4296282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4297359f2544SMike Kravetz return 1; 4298359f2544SMike Kravetz } 4299359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4300359f2544SMike Kravetz 4301282f4214SMike Kravetz /* 4302282f4214SMike Kravetz * default_hugepagesz command line input 4303282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4304282f4214SMike Kravetz */ 4305ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4306e11bfbfcSNick Piggin { 4307ae94da89SMike Kravetz unsigned long size; 4308b5389086SZhenguo Yao int i; 4309ae94da89SMike Kravetz 4310282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4311282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4312282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4313f81f6e4bSPeng Liu return 1; 4314282f4214SMike Kravetz } 4315282f4214SMike Kravetz 4316282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4317282f4214SMike Kravetz 4318282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4319282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4320f81f6e4bSPeng Liu return 1; 4321282f4214SMike Kravetz } 4322282f4214SMike Kravetz 4323282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4324282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4325282f4214SMike Kravetz parsed_default_hugepagesz = true; 4326282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4327282f4214SMike Kravetz 4328282f4214SMike Kravetz /* 4329282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4330282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4331282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4332282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4333282f4214SMike Kravetz * allocated here from bootmem allocator. 4334282f4214SMike Kravetz */ 4335282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4336282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 43370a7a0f6fSPeng Liu for_each_online_node(i) 4338b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4339b5389086SZhenguo Yao default_hugepages_in_node[i]; 4340282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4341282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4342282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4343282f4214SMike Kravetz } 4344282f4214SMike Kravetz 4345e11bfbfcSNick Piggin return 1; 4346e11bfbfcSNick Piggin } 4347ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4348a3437870SNishanth Aravamudan 4349d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4350d2226ebdSFeng Tang { 4351d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4352d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4353d2226ebdSFeng Tang 4354d2226ebdSFeng Tang /* 4355d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4356d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4357d2226ebdSFeng Tang */ 4358d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4359d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4360d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4361d2226ebdSFeng Tang return &mpol->nodes; 4362d2226ebdSFeng Tang #endif 4363d2226ebdSFeng Tang return NULL; 4364d2226ebdSFeng Tang } 4365d2226ebdSFeng Tang 43668ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43678a213460SNishanth Aravamudan { 43688a213460SNishanth Aravamudan int node; 43698a213460SNishanth Aravamudan unsigned int nr = 0; 4370d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 43718ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43728ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43738a213460SNishanth Aravamudan 4374d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 43758ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4376d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 43778a213460SNishanth Aravamudan nr += array[node]; 43788ca39e68SMuchun Song } 43798a213460SNishanth Aravamudan 43808a213460SNishanth Aravamudan return nr; 43818a213460SNishanth Aravamudan } 43828a213460SNishanth Aravamudan 43838a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 438417743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 438517743798SMuchun Song void *buffer, size_t *length, 438617743798SMuchun Song loff_t *ppos, unsigned long *out) 438717743798SMuchun Song { 438817743798SMuchun Song struct ctl_table dup_table; 438917743798SMuchun Song 439017743798SMuchun Song /* 439117743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 439217743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 439317743798SMuchun Song */ 439417743798SMuchun Song dup_table = *table; 439517743798SMuchun Song dup_table.data = out; 439617743798SMuchun Song 439717743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 439817743798SMuchun Song } 439917743798SMuchun Song 440006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 440106808b08SLee Schermerhorn struct ctl_table *table, int write, 440232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 44031da177e4SLinus Torvalds { 4404e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4405238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 440608d4a246SMichal Hocko int ret; 4407e5ff2159SAndi Kleen 4408457c1b27SNishanth Aravamudan if (!hugepages_supported()) 440986613628SJan Stancek return -EOPNOTSUPP; 4410457c1b27SNishanth Aravamudan 441117743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 441217743798SMuchun Song &tmp); 441308d4a246SMichal Hocko if (ret) 441408d4a246SMichal Hocko goto out; 4415e5ff2159SAndi Kleen 4416238d3c13SDavid Rientjes if (write) 4417238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4418238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 441908d4a246SMichal Hocko out: 442008d4a246SMichal Hocko return ret; 44211da177e4SLinus Torvalds } 4422396faf03SMel Gorman 442306808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 442432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 442506808b08SLee Schermerhorn { 442606808b08SLee Schermerhorn 442706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 442806808b08SLee Schermerhorn buffer, length, ppos); 442906808b08SLee Schermerhorn } 443006808b08SLee Schermerhorn 443106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 443206808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 443332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 443406808b08SLee Schermerhorn { 443506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 443606808b08SLee Schermerhorn buffer, length, ppos); 443706808b08SLee Schermerhorn } 443806808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 443906808b08SLee Schermerhorn 4440a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 444132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4442a3d0c6aaSNishanth Aravamudan { 4443a5516438SAndi Kleen struct hstate *h = &default_hstate; 4444e5ff2159SAndi Kleen unsigned long tmp; 444508d4a246SMichal Hocko int ret; 4446e5ff2159SAndi Kleen 4447457c1b27SNishanth Aravamudan if (!hugepages_supported()) 444886613628SJan Stancek return -EOPNOTSUPP; 4449457c1b27SNishanth Aravamudan 4450e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4451e5ff2159SAndi Kleen 4452bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4453adbe8726SEric B Munson return -EINVAL; 4454adbe8726SEric B Munson 445517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 445617743798SMuchun Song &tmp); 445708d4a246SMichal Hocko if (ret) 445808d4a246SMichal Hocko goto out; 4459e5ff2159SAndi Kleen 4460e5ff2159SAndi Kleen if (write) { 4461db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4462e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4463db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4464e5ff2159SAndi Kleen } 446508d4a246SMichal Hocko out: 446608d4a246SMichal Hocko return ret; 4467a3d0c6aaSNishanth Aravamudan } 4468a3d0c6aaSNishanth Aravamudan 44691da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44701da177e4SLinus Torvalds 4471e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44721da177e4SLinus Torvalds { 4473fcb2b0c5SRoman Gushchin struct hstate *h; 4474fcb2b0c5SRoman Gushchin unsigned long total = 0; 4475fcb2b0c5SRoman Gushchin 4476457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4477457c1b27SNishanth Aravamudan return; 4478fcb2b0c5SRoman Gushchin 4479fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4480fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4481fcb2b0c5SRoman Gushchin 4482aca78307SMiaohe Lin total += huge_page_size(h) * count; 4483fcb2b0c5SRoman Gushchin 4484fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4485e1759c21SAlexey Dobriyan seq_printf(m, 44861da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 44871da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4488b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 44897893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 44904f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4491fcb2b0c5SRoman Gushchin count, 4492a5516438SAndi Kleen h->free_huge_pages, 4493a5516438SAndi Kleen h->resv_huge_pages, 4494a5516438SAndi Kleen h->surplus_huge_pages, 4495aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4496fcb2b0c5SRoman Gushchin } 4497fcb2b0c5SRoman Gushchin 4498aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 44991da177e4SLinus Torvalds } 45001da177e4SLinus Torvalds 45017981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 45021da177e4SLinus Torvalds { 4503a5516438SAndi Kleen struct hstate *h = &default_hstate; 45047981593bSJoe Perches 4505457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4506457c1b27SNishanth Aravamudan return 0; 45077981593bSJoe Perches 45087981593bSJoe Perches return sysfs_emit_at(buf, len, 45091da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4510a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4511a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4512a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4513a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4514a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 45151da177e4SLinus Torvalds } 45161da177e4SLinus Torvalds 4517dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4518949f7ec5SDavid Rientjes { 4519949f7ec5SDavid Rientjes struct hstate *h; 4520949f7ec5SDavid Rientjes 4521457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4522457c1b27SNishanth Aravamudan return; 4523457c1b27SNishanth Aravamudan 4524949f7ec5SDavid Rientjes for_each_hstate(h) 4525dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4526949f7ec5SDavid Rientjes nid, 4527949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4528949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4529949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4530aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4531949f7ec5SDavid Rientjes } 4532949f7ec5SDavid Rientjes 45335d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 45345d317b2bSNaoya Horiguchi { 45355d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 45365d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 45375d317b2bSNaoya Horiguchi } 45385d317b2bSNaoya Horiguchi 45391da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 45401da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 45411da177e4SLinus Torvalds { 4542d0028588SWanpeng Li struct hstate *h; 4543d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4544d0028588SWanpeng Li 4545d0028588SWanpeng Li for_each_hstate(h) 4546d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4547d0028588SWanpeng Li return nr_total_pages; 45481da177e4SLinus Torvalds } 45491da177e4SLinus Torvalds 4550a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4551fc1b8a73SMel Gorman { 4552fc1b8a73SMel Gorman int ret = -ENOMEM; 4553fc1b8a73SMel Gorman 45540aa7f354SMiaohe Lin if (!delta) 45550aa7f354SMiaohe Lin return 0; 45560aa7f354SMiaohe Lin 4557db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4558fc1b8a73SMel Gorman /* 4559fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4560fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4561fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4562fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4563fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4564fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4565fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4566fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4567fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4568fc1b8a73SMel Gorman * 4569fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4570fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4571fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4572fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4573fc1b8a73SMel Gorman * semantics that cpuset has. 45748ca39e68SMuchun Song * 45758ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45768ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45778ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45788ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45798ca39e68SMuchun Song * above. 4580fc1b8a73SMel Gorman */ 4581fc1b8a73SMel Gorman if (delta > 0) { 4582a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4583fc1b8a73SMel Gorman goto out; 4584fc1b8a73SMel Gorman 45858ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4586a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4587fc1b8a73SMel Gorman goto out; 4588fc1b8a73SMel Gorman } 4589fc1b8a73SMel Gorman } 4590fc1b8a73SMel Gorman 4591fc1b8a73SMel Gorman ret = 0; 4592fc1b8a73SMel Gorman if (delta < 0) 4593a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4594fc1b8a73SMel Gorman 4595fc1b8a73SMel Gorman out: 4596db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4597fc1b8a73SMel Gorman return ret; 4598fc1b8a73SMel Gorman } 4599fc1b8a73SMel Gorman 460084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 460184afd99bSAndy Whitcroft { 4602f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 460384afd99bSAndy Whitcroft 460484afd99bSAndy Whitcroft /* 460584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 460684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 460784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 460825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 460984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 461084afd99bSAndy Whitcroft * new reference here without additional locking. 461184afd99bSAndy Whitcroft */ 461209a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 461309a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4614f522c3acSJoonsoo Kim kref_get(&resv->refs); 461584afd99bSAndy Whitcroft } 461609a26e83SMike Kravetz } 461784afd99bSAndy Whitcroft 4618a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4619a1e78772SMel Gorman { 4620a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4621f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 462290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 46234e35f483SJoonsoo Kim unsigned long reserve, start, end; 46241c5ecae3SMike Kravetz long gbl_reserve; 462584afd99bSAndy Whitcroft 46264e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 46274e35f483SJoonsoo Kim return; 46284e35f483SJoonsoo Kim 4629a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4630a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 463184afd99bSAndy Whitcroft 46324e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4633e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 46347251ff78SAdam Litke if (reserve) { 46351c5ecae3SMike Kravetz /* 46361c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 46371c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 46381c5ecae3SMike Kravetz */ 46391c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 46401c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 46417251ff78SAdam Litke } 4642e9fe92aeSMina Almasry 4643e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4644a1e78772SMel Gorman } 4645a1e78772SMel Gorman 464631383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 464731383c68SDan Williams { 464831383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 464931383c68SDan Williams return -EINVAL; 465031383c68SDan Williams return 0; 465131383c68SDan Williams } 465231383c68SDan Williams 465305ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 465405ea8860SDan Williams { 4655aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 465605ea8860SDan Williams } 465705ea8860SDan Williams 46581da177e4SLinus Torvalds /* 46591da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46601da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46616c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46621da177e4SLinus Torvalds * this far. 46631da177e4SLinus Torvalds */ 4664b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 46651da177e4SLinus Torvalds { 46661da177e4SLinus Torvalds BUG(); 4667d0217ac0SNick Piggin return 0; 46681da177e4SLinus Torvalds } 46691da177e4SLinus Torvalds 4670eec3636aSJane Chu /* 4671eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4672eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4673eec3636aSJane Chu * This is because under System V memory model, mappings created via 4674eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4675eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4676eec3636aSJane Chu */ 4677f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4678d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 467984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4680a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4681dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 468205ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 46831da177e4SLinus Torvalds }; 46841da177e4SLinus Torvalds 46851e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 46861e8f889bSDavid Gibson int writable) 468763551ae0SDavid Gibson { 468863551ae0SDavid Gibson pte_t entry; 468979c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 469063551ae0SDavid Gibson 46911e8f889bSDavid Gibson if (writable) { 4692106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4693106c992aSGerald Schaefer vma->vm_page_prot))); 469463551ae0SDavid Gibson } else { 4695106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4696106c992aSGerald Schaefer vma->vm_page_prot)); 469763551ae0SDavid Gibson } 469863551ae0SDavid Gibson entry = pte_mkyoung(entry); 469979c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 470063551ae0SDavid Gibson 470163551ae0SDavid Gibson return entry; 470263551ae0SDavid Gibson } 470363551ae0SDavid Gibson 47041e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 47051e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 47061e8f889bSDavid Gibson { 47071e8f889bSDavid Gibson pte_t entry; 47081e8f889bSDavid Gibson 4709106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 471032f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 47114b3073e1SRussell King update_mmu_cache(vma, address, ptep); 47121e8f889bSDavid Gibson } 47131e8f889bSDavid Gibson 4714d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 47154a705fefSNaoya Horiguchi { 47164a705fefSNaoya Horiguchi swp_entry_t swp; 47174a705fefSNaoya Horiguchi 47184a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4719d5ed7444SAneesh Kumar K.V return false; 47204a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4721d79d176aSBaoquan He if (is_migration_entry(swp)) 4722d5ed7444SAneesh Kumar K.V return true; 47234a705fefSNaoya Horiguchi else 4724d5ed7444SAneesh Kumar K.V return false; 47254a705fefSNaoya Horiguchi } 47264a705fefSNaoya Horiguchi 47273e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 47284a705fefSNaoya Horiguchi { 47294a705fefSNaoya Horiguchi swp_entry_t swp; 47304a705fefSNaoya Horiguchi 47314a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 47323e5c3600SBaoquan He return false; 47334a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4734d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 47353e5c3600SBaoquan He return true; 47364a705fefSNaoya Horiguchi else 47373e5c3600SBaoquan He return false; 47384a705fefSNaoya Horiguchi } 47391e8f889bSDavid Gibson 47404eae4efaSPeter Xu static void 47414eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 47424eae4efaSPeter Xu struct page *new_page) 47434eae4efaSPeter Xu { 47444eae4efaSPeter Xu __SetPageUptodate(new_page); 47454eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 47461eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 47474eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 47484eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 47494eae4efaSPeter Xu SetHPageMigratable(new_page); 47504eae4efaSPeter Xu } 47514eae4efaSPeter Xu 475263551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4753bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4754bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 475563551ae0SDavid Gibson { 47563aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 475763551ae0SDavid Gibson struct page *ptepage; 47581c59827dSHugh Dickins unsigned long addr; 4759bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4760bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4761a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47624eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4763bc70fbf2SPeter Xu struct address_space *mapping = src_vma->vm_file->f_mapping; 4764ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4765e95a9851SMike Kravetz unsigned long last_addr_mask; 4766e8569dd2SAndreas Sandberg int ret = 0; 47671e8f889bSDavid Gibson 4768ac46d4f3SJérôme Glisse if (cow) { 4769bc70fbf2SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, 4770bc70fbf2SPeter Xu src_vma->vm_start, 4771bc70fbf2SPeter Xu src_vma->vm_end); 4772ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4773623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4774623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 4775c0d0381aSMike Kravetz } else { 4776c0d0381aSMike Kravetz /* 4777c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4778c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4779c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4780c0d0381aSMike Kravetz * huge_pmd_unshare. 4781c0d0381aSMike Kravetz */ 4782c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4783ac46d4f3SJérôme Glisse } 4784e8569dd2SAndreas Sandberg 4785e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4786bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4787cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 47887868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4789e95a9851SMike Kravetz if (!src_pte) { 4790e95a9851SMike Kravetz addr |= last_addr_mask; 4791c74df32cSHugh Dickins continue; 4792e95a9851SMike Kravetz } 4793bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4794e8569dd2SAndreas Sandberg if (!dst_pte) { 4795e8569dd2SAndreas Sandberg ret = -ENOMEM; 4796e8569dd2SAndreas Sandberg break; 4797e8569dd2SAndreas Sandberg } 4798c5c99429SLarry Woodman 47995e41540cSMike Kravetz /* 48005e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 48015e41540cSMike Kravetz * 48023aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 48035e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 48043aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 48053aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 48065e41540cSMike Kravetz */ 48073aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 4808e95a9851SMike Kravetz addr |= last_addr_mask; 4809c5c99429SLarry Woodman continue; 4810e95a9851SMike Kravetz } 4811c5c99429SLarry Woodman 4812cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4813cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4814cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48154a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 48164eae4efaSPeter Xu again: 48173aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 48185e41540cSMike Kravetz /* 48193aa4ed80SMiaohe Lin * Skip if src entry none. 48205e41540cSMike Kravetz */ 48214a705fefSNaoya Horiguchi ; 4822c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 4823c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 4824c2cb0dccSNaoya Horiguchi 4825c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 4826c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 4827c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 4828c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 48294a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 4830bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 48314a705fefSNaoya Horiguchi 48326c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 48334a705fefSNaoya Horiguchi /* 48344a705fefSNaoya Horiguchi * COW mappings require pages in both 48354a705fefSNaoya Horiguchi * parent and child to be set to read. 48364a705fefSNaoya Horiguchi */ 48374dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 48384dd845b5SAlistair Popple swp_offset(swp_entry)); 48394a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4840bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 4841bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 484218f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 48434a705fefSNaoya Horiguchi } 4844bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 4845bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 484618f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 4847bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 4848bc70fbf2SPeter Xu /* 4849bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 4850bc70fbf2SPeter Xu * uffd-wp enabled. 4851bc70fbf2SPeter Xu */ 4852bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 4853bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 48544a705fefSNaoya Horiguchi } else { 48554eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48564eae4efaSPeter Xu ptepage = pte_page(entry); 48574eae4efaSPeter Xu get_page(ptepage); 48584eae4efaSPeter Xu 48594eae4efaSPeter Xu /* 4860fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 4861fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 4862fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 4863fb3d824dSDavid Hildenbrand * fork. 48644eae4efaSPeter Xu * 48654eae4efaSPeter Xu * When pre-allocating the page or copying data, we 48664eae4efaSPeter Xu * need to be without the pgtable locks since we could 48674eae4efaSPeter Xu * sleep during the process. 48684eae4efaSPeter Xu */ 4869fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 4870fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 4871bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 4872bc70fbf2SPeter Xu src_vma)) { 48734eae4efaSPeter Xu pte_t src_pte_old = entry; 48744eae4efaSPeter Xu struct page *new; 48754eae4efaSPeter Xu 48764eae4efaSPeter Xu spin_unlock(src_ptl); 48774eae4efaSPeter Xu spin_unlock(dst_ptl); 48784eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 4879bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 48804eae4efaSPeter Xu if (IS_ERR(new)) { 48814eae4efaSPeter Xu put_page(ptepage); 48824eae4efaSPeter Xu ret = PTR_ERR(new); 48834eae4efaSPeter Xu break; 48844eae4efaSPeter Xu } 4885bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 48864eae4efaSPeter Xu npages); 48874eae4efaSPeter Xu put_page(ptepage); 48884eae4efaSPeter Xu 48894eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 48904eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 48914eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 48924eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48934eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48944eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4895bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 4896846be085SMike Kravetz new); 48974eae4efaSPeter Xu put_page(new); 48983aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 48994eae4efaSPeter Xu goto again; 49004eae4efaSPeter Xu } 4901bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 49024eae4efaSPeter Xu spin_unlock(src_ptl); 49034eae4efaSPeter Xu spin_unlock(dst_ptl); 49044eae4efaSPeter Xu continue; 49054eae4efaSPeter Xu } 49064eae4efaSPeter Xu 490734ee645eSJoerg Roedel if (cow) { 49080f10851eSJérôme Glisse /* 49090f10851eSJérôme Glisse * No need to notify as we are downgrading page 49100f10851eSJérôme Glisse * table protection not changing it to point 49110f10851eSJérôme Glisse * to a new page. 49120f10851eSJérôme Glisse * 4913ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 49140f10851eSJérôme Glisse */ 49157f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 491684894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 491734ee645eSJoerg Roedel } 49184eae4efaSPeter Xu 491963551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 49204eae4efaSPeter Xu hugetlb_count_add(npages, dst); 49211c59827dSHugh Dickins } 4922cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4923cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 492463551ae0SDavid Gibson } 492563551ae0SDavid Gibson 4926623a1ddfSDavid Hildenbrand if (cow) { 4927623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 4928ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4929623a1ddfSDavid Hildenbrand } else { 4930c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4931623a1ddfSDavid Hildenbrand } 4932e8569dd2SAndreas Sandberg 4933e8569dd2SAndreas Sandberg return ret; 493463551ae0SDavid Gibson } 493563551ae0SDavid Gibson 4936550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4937db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4938550a7d60SMina Almasry { 4939550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4940550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4941550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4942db110a99SAneesh Kumar K.V pte_t pte; 4943550a7d60SMina Almasry 4944550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4945550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4946550a7d60SMina Almasry 4947550a7d60SMina Almasry /* 4948550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4949550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4950550a7d60SMina Almasry */ 4951550a7d60SMina Almasry if (src_ptl != dst_ptl) 4952550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4953550a7d60SMina Almasry 4954550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4955550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4956550a7d60SMina Almasry 4957550a7d60SMina Almasry if (src_ptl != dst_ptl) 4958550a7d60SMina Almasry spin_unlock(src_ptl); 4959550a7d60SMina Almasry spin_unlock(dst_ptl); 4960550a7d60SMina Almasry } 4961550a7d60SMina Almasry 4962550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4963550a7d60SMina Almasry struct vm_area_struct *new_vma, 4964550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4965550a7d60SMina Almasry unsigned long len) 4966550a7d60SMina Almasry { 4967550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4968550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4969550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4970550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4971550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4972e95a9851SMike Kravetz unsigned long last_addr_mask; 4973550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4974550a7d60SMina Almasry struct mmu_notifier_range range; 49753d0b95cdSBaolin Wang bool shared_pmd = false; 4976550a7d60SMina Almasry 4977550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 4978550a7d60SMina Almasry old_end); 4979550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 49803d0b95cdSBaolin Wang /* 49813d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 49823d0b95cdSBaolin Wang * range. 49833d0b95cdSBaolin Wang */ 49843d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 49853d0b95cdSBaolin Wang 4986550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 4987e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4988550a7d60SMina Almasry /* Prevent race with file truncation */ 4989550a7d60SMina Almasry i_mmap_lock_write(mapping); 4990550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 4991550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 4992e95a9851SMike Kravetz if (!src_pte) { 4993e95a9851SMike Kravetz old_addr |= last_addr_mask; 4994e95a9851SMike Kravetz new_addr |= last_addr_mask; 4995550a7d60SMina Almasry continue; 4996e95a9851SMike Kravetz } 4997550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 4998550a7d60SMina Almasry continue; 4999550a7d60SMina Almasry 50004ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 50013d0b95cdSBaolin Wang shared_pmd = true; 50024ddb4d91SMike Kravetz old_addr |= last_addr_mask; 50034ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5004550a7d60SMina Almasry continue; 50053d0b95cdSBaolin Wang } 5006550a7d60SMina Almasry 5007550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5008550a7d60SMina Almasry if (!dst_pte) 5009550a7d60SMina Almasry break; 5010550a7d60SMina Almasry 5011db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5012550a7d60SMina Almasry } 50133d0b95cdSBaolin Wang 50143d0b95cdSBaolin Wang if (shared_pmd) 50153d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 50163d0b95cdSBaolin Wang else 5017550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5018550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 501913e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 5020550a7d60SMina Almasry 5021550a7d60SMina Almasry return len + old_addr - old_end; 5022550a7d60SMina Almasry } 5023550a7d60SMina Almasry 502473c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 502524669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 502605e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 502763551ae0SDavid Gibson { 502863551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 502963551ae0SDavid Gibson unsigned long address; 5030c7546f8fSDavid Gibson pte_t *ptep; 503163551ae0SDavid Gibson pte_t pte; 5032cb900f41SKirill A. Shutemov spinlock_t *ptl; 503363551ae0SDavid Gibson struct page *page; 5034a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5035a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5036ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5037e95a9851SMike Kravetz unsigned long last_addr_mask; 5038a4a118f2SNadav Amit bool force_flush = false; 5039a5516438SAndi Kleen 504063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5041a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5042a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 504363551ae0SDavid Gibson 504407e32661SAneesh Kumar K.V /* 504507e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 504607e32661SAneesh Kumar K.V * to huge page. 504707e32661SAneesh Kumar K.V */ 5048ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 504924669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5050dff11abeSMike Kravetz 5051dff11abeSMike Kravetz /* 5052dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 5053dff11abeSMike Kravetz */ 50546f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 50556f4f13e8SJérôme Glisse end); 5056ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5057ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5058e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5059569f48b8SHillf Danton address = start; 5060569f48b8SHillf Danton for (; address < end; address += sz) { 50617868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 5062e95a9851SMike Kravetz if (!ptep) { 5063e95a9851SMike Kravetz address |= last_addr_mask; 5064c7546f8fSDavid Gibson continue; 5065e95a9851SMike Kravetz } 5066c7546f8fSDavid Gibson 5067cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 50684ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 506931d49da5SAneesh Kumar K.V spin_unlock(ptl); 5070a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5071a4a118f2SNadav Amit force_flush = true; 50724ddb4d91SMike Kravetz address |= last_addr_mask; 507331d49da5SAneesh Kumar K.V continue; 507431d49da5SAneesh Kumar K.V } 507539dde65cSChen, Kenneth W 50766629326bSHillf Danton pte = huge_ptep_get(ptep); 507731d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 507831d49da5SAneesh Kumar K.V spin_unlock(ptl); 507931d49da5SAneesh Kumar K.V continue; 508031d49da5SAneesh Kumar K.V } 50816629326bSHillf Danton 50826629326bSHillf Danton /* 50839fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 50849fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 50856629326bSHillf Danton */ 50869fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 508705e90bd0SPeter Xu /* 508805e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 508905e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 509005e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 509105e90bd0SPeter Xu * pte with a marker. 509205e90bd0SPeter Xu */ 509305e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 509405e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 509505e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 509605e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 509705e90bd0SPeter Xu else 50989386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 509931d49da5SAneesh Kumar K.V spin_unlock(ptl); 510031d49da5SAneesh Kumar K.V continue; 51018c4894c6SNaoya Horiguchi } 51026629326bSHillf Danton 51036629326bSHillf Danton page = pte_page(pte); 510404f2cbe3SMel Gorman /* 510504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 510604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 510704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 510804f2cbe3SMel Gorman */ 510904f2cbe3SMel Gorman if (ref_page) { 511031d49da5SAneesh Kumar K.V if (page != ref_page) { 511131d49da5SAneesh Kumar K.V spin_unlock(ptl); 511231d49da5SAneesh Kumar K.V continue; 511331d49da5SAneesh Kumar K.V } 511404f2cbe3SMel Gorman /* 511504f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 511604f2cbe3SMel Gorman * future faults in this VMA will fail rather than 511704f2cbe3SMel Gorman * looking like data was lost 511804f2cbe3SMel Gorman */ 511904f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 512004f2cbe3SMel Gorman } 512104f2cbe3SMel Gorman 5122c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5123b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5124106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 51256649a386SKen Chen set_page_dirty(page); 512605e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 512705e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 512805e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 512905e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 513005e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 51315d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5132cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 513331d49da5SAneesh Kumar K.V 5134cb900f41SKirill A. Shutemov spin_unlock(ptl); 5135e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 513624669e58SAneesh Kumar K.V /* 513731d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 513824669e58SAneesh Kumar K.V */ 513931d49da5SAneesh Kumar K.V if (ref_page) 514031d49da5SAneesh Kumar K.V break; 5141fe1668aeSChen, Kenneth W } 5142ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 514324669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5144a4a118f2SNadav Amit 5145a4a118f2SNadav Amit /* 5146a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5147a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5148a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5149a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5150a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5151a4a118f2SNadav Amit * dropped as well. 5152a4a118f2SNadav Amit * 5153a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5154a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5155a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5156a4a118f2SNadav Amit * Instead, do flush now. 5157a4a118f2SNadav Amit */ 5158a4a118f2SNadav Amit if (force_flush) 5159a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 51601da177e4SLinus Torvalds } 516163551ae0SDavid Gibson 5162d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5163d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 516405e90bd0SPeter Xu unsigned long end, struct page *ref_page, 516505e90bd0SPeter Xu zap_flags_t zap_flags) 5166d833352aSMel Gorman { 516705e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5168d833352aSMel Gorman 5169d833352aSMel Gorman /* 5170d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 5171d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 5172d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 5173d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 5174d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 5175c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 5176d833352aSMel Gorman * because in the context this is called, the VMA is about to be 5177c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 5178d833352aSMel Gorman */ 5179d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 5180d833352aSMel Gorman } 5181d833352aSMel Gorman 5182502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 518305e90bd0SPeter Xu unsigned long end, struct page *ref_page, 518405e90bd0SPeter Xu zap_flags_t zap_flags) 5185502717f4SChen, Kenneth W { 518624669e58SAneesh Kumar K.V struct mmu_gather tlb; 5187dff11abeSMike Kravetz 5188a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 518905e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5190ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5191502717f4SChen, Kenneth W } 5192502717f4SChen, Kenneth W 519304f2cbe3SMel Gorman /* 519404f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5195578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 519604f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 519704f2cbe3SMel Gorman * same region. 519804f2cbe3SMel Gorman */ 51992f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 52002a4b3dedSHarvey Harrison struct page *page, unsigned long address) 520104f2cbe3SMel Gorman { 52027526674dSAdam Litke struct hstate *h = hstate_vma(vma); 520304f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 520404f2cbe3SMel Gorman struct address_space *mapping; 520504f2cbe3SMel Gorman pgoff_t pgoff; 520604f2cbe3SMel Gorman 520704f2cbe3SMel Gorman /* 520804f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 520904f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 521004f2cbe3SMel Gorman */ 52117526674dSAdam Litke address = address & huge_page_mask(h); 521236e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 521336e4f20aSMichal Hocko vma->vm_pgoff; 521493c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 521504f2cbe3SMel Gorman 52164eb2b1dcSMel Gorman /* 52174eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 52184eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 52194eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 52204eb2b1dcSMel Gorman */ 522183cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 52226b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 522304f2cbe3SMel Gorman /* Do not unmap the current VMA */ 522404f2cbe3SMel Gorman if (iter_vma == vma) 522504f2cbe3SMel Gorman continue; 522604f2cbe3SMel Gorman 522704f2cbe3SMel Gorman /* 52282f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 52292f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 52302f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 52312f84a899SMel Gorman */ 52322f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 52332f84a899SMel Gorman continue; 52342f84a899SMel Gorman 52352f84a899SMel Gorman /* 523604f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 523704f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 523804f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 523904f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 524004f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 524104f2cbe3SMel Gorman */ 524204f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 524324669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 524405e90bd0SPeter Xu address + huge_page_size(h), page, 0); 524504f2cbe3SMel Gorman } 524683cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 524704f2cbe3SMel Gorman } 524804f2cbe3SMel Gorman 52490fe6e20bSNaoya Horiguchi /* 5250c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5251aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5252ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5253ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 52540fe6e20bSNaoya Horiguchi */ 5255c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5256c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5257cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 52581e8f889bSDavid Gibson { 5259c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 52603999f52eSAneesh Kumar K.V pte_t pte; 5261a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52621e8f889bSDavid Gibson struct page *old_page, *new_page; 52632b740303SSouptick Joarder int outside_reserve = 0; 52642b740303SSouptick Joarder vm_fault_t ret = 0; 5265974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5266ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 52671e8f889bSDavid Gibson 5268c89357e2SDavid Hildenbrand VM_BUG_ON(unshare && (flags & FOLL_WRITE)); 5269c89357e2SDavid Hildenbrand VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); 5270c89357e2SDavid Hildenbrand 52711d8d1464SDavid Hildenbrand /* 52721d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 52731d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 52741d8d1464SDavid Hildenbrand */ 52751d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 52761d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 52771d8d1464SDavid Hildenbrand 52781d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 52791d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 52801d8d1464SDavid Hildenbrand if (unlikely(unshare)) 52811d8d1464SDavid Hildenbrand return 0; 52821d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 52831d8d1464SDavid Hildenbrand return 0; 52841d8d1464SDavid Hildenbrand } 52851d8d1464SDavid Hildenbrand 52863999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 52871e8f889bSDavid Gibson old_page = pte_page(pte); 52881e8f889bSDavid Gibson 5289662ce1dcSYang Yang delayacct_wpcopy_start(); 5290662ce1dcSYang Yang 529104f2cbe3SMel Gorman retry_avoidcopy: 5292c89357e2SDavid Hildenbrand /* 5293c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5294c89357e2SDavid Hildenbrand * owner and can reuse this page. 5295c89357e2SDavid Hildenbrand */ 529637a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5297c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 52985a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5299c89357e2SDavid Hildenbrand if (likely(!unshare)) 53005b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5301662ce1dcSYang Yang 5302662ce1dcSYang Yang delayacct_wpcopy_end(); 530383c54070SNick Piggin return 0; 53041e8f889bSDavid Gibson } 53056c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 53066c287605SDavid Hildenbrand old_page); 53071e8f889bSDavid Gibson 530804f2cbe3SMel Gorman /* 530904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 531004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 531104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 531204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 531304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 531404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 531504f2cbe3SMel Gorman * of the full address range. 531604f2cbe3SMel Gorman */ 53175944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 531804f2cbe3SMel Gorman old_page != pagecache_page) 531904f2cbe3SMel Gorman outside_reserve = 1; 532004f2cbe3SMel Gorman 532109cbfeafSKirill A. Shutemov get_page(old_page); 5322b76c8cfbSLarry Woodman 5323ad4404a2SDavidlohr Bueso /* 5324ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5325ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5326ad4404a2SDavidlohr Bueso */ 5327cb900f41SKirill A. Shutemov spin_unlock(ptl); 53285b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 53291e8f889bSDavid Gibson 53302fc39cecSAdam Litke if (IS_ERR(new_page)) { 533104f2cbe3SMel Gorman /* 533204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 533304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 533404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 533504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 533604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 533704f2cbe3SMel Gorman */ 533804f2cbe3SMel Gorman if (outside_reserve) { 5339e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 5340e7dd91c4SMike Kravetz pgoff_t idx; 5341e7dd91c4SMike Kravetz u32 hash; 5342e7dd91c4SMike Kravetz 534309cbfeafSKirill A. Shutemov put_page(old_page); 534404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 5345e7dd91c4SMike Kravetz /* 5346e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 5347e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 5348e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 5349e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 5350e7dd91c4SMike Kravetz * PMD sharing. 5351e7dd91c4SMike Kravetz * 5352e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 5353e7dd91c4SMike Kravetz */ 5354e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5355e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 5356e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5357e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 5358e7dd91c4SMike Kravetz 53595b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 5360e7dd91c4SMike Kravetz 5361e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 5362e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 5363cb900f41SKirill A. Shutemov spin_lock(ptl); 53645b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5365a9af0c5dSNaoya Horiguchi if (likely(ptep && 5366a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 536704f2cbe3SMel Gorman goto retry_avoidcopy; 5368a734bcc8SHillf Danton /* 5369cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5370cb900f41SKirill A. Shutemov * lock, and our job is done. 5371a734bcc8SHillf Danton */ 5372662ce1dcSYang Yang delayacct_wpcopy_end(); 5373a734bcc8SHillf Danton return 0; 537404f2cbe3SMel Gorman } 537504f2cbe3SMel Gorman 53762b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5377ad4404a2SDavidlohr Bueso goto out_release_old; 53781e8f889bSDavid Gibson } 53791e8f889bSDavid Gibson 53800fe6e20bSNaoya Horiguchi /* 53810fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 53820fe6e20bSNaoya Horiguchi * anon_vma prepared. 53830fe6e20bSNaoya Horiguchi */ 538444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5385ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5386ad4404a2SDavidlohr Bueso goto out_release_all; 538744e2aa93SDean Nelson } 53880fe6e20bSNaoya Horiguchi 5389974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 539047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 53910ed361deSNick Piggin __SetPageUptodate(new_page); 53921e8f889bSDavid Gibson 53937269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 53946f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5395ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5396ad4404a2SDavidlohr Bueso 5397b76c8cfbSLarry Woodman /* 5398cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5399b76c8cfbSLarry Woodman * before the page tables are altered 5400b76c8cfbSLarry Woodman */ 5401cb900f41SKirill A. Shutemov spin_lock(ptl); 54025b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5403a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5404d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 540507443a85SJoonsoo Kim 5406c89357e2SDavid Hildenbrand /* Break COW or unshare */ 54075b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5408ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5409cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 54105b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 54111eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5412c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 54138f251a3dSMike Kravetz SetHPageMigratable(new_page); 54141e8f889bSDavid Gibson /* Make the old page be freed below */ 54151e8f889bSDavid Gibson new_page = old_page; 54161e8f889bSDavid Gibson } 5417cb900f41SKirill A. Shutemov spin_unlock(ptl); 5418ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5419ad4404a2SDavidlohr Bueso out_release_all: 5420c89357e2SDavid Hildenbrand /* 5421c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5422c89357e2SDavid Hildenbrand * unshare) 5423c89357e2SDavid Hildenbrand */ 5424c7b1850dSMike Kravetz if (new_page != old_page) 54255b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 542609cbfeafSKirill A. Shutemov put_page(new_page); 5427ad4404a2SDavidlohr Bueso out_release_old: 542809cbfeafSKirill A. Shutemov put_page(old_page); 54298312034fSJoonsoo Kim 5430ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5431662ce1dcSYang Yang 5432662ce1dcSYang Yang delayacct_wpcopy_end(); 5433ad4404a2SDavidlohr Bueso return ret; 54341e8f889bSDavid Gibson } 54351e8f889bSDavid Gibson 543604f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 5437a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 5438a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 543904f2cbe3SMel Gorman { 544004f2cbe3SMel Gorman struct address_space *mapping; 5441e7c4b0bfSAndy Whitcroft pgoff_t idx; 544204f2cbe3SMel Gorman 544304f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 5444a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 544504f2cbe3SMel Gorman 544604f2cbe3SMel Gorman return find_lock_page(mapping, idx); 544704f2cbe3SMel Gorman } 544804f2cbe3SMel Gorman 54493ae77f43SHugh Dickins /* 54503ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 54513ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 54523ae77f43SHugh Dickins */ 54533ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 54542a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 54552a15efc9SHugh Dickins { 54562a15efc9SHugh Dickins struct address_space *mapping; 54572a15efc9SHugh Dickins pgoff_t idx; 54582a15efc9SHugh Dickins struct page *page; 54592a15efc9SHugh Dickins 54602a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 54612a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 54622a15efc9SHugh Dickins 54632a15efc9SHugh Dickins page = find_get_page(mapping, idx); 54642a15efc9SHugh Dickins if (page) 54652a15efc9SHugh Dickins put_page(page); 54662a15efc9SHugh Dickins return page != NULL; 54672a15efc9SHugh Dickins } 54682a15efc9SHugh Dickins 5469ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 5470ab76ad54SMike Kravetz pgoff_t idx) 5471ab76ad54SMike Kravetz { 5472d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5473ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5474ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5475d9ef44deSMatthew Wilcox (Oracle) int err; 5476ab76ad54SMike Kravetz 5477d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5478d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5479d9ef44deSMatthew Wilcox (Oracle) 5480d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5481d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5482ab76ad54SMike Kravetz return err; 5483d9ef44deSMatthew Wilcox (Oracle) } 5484d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5485ab76ad54SMike Kravetz 548622146c3cSMike Kravetz /* 5487d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 548822146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 548922146c3cSMike Kravetz */ 5490d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 549122146c3cSMike Kravetz 5492ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5493ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5494ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5495ab76ad54SMike Kravetz return 0; 5496ab76ad54SMike Kravetz } 5497ab76ad54SMike Kravetz 54987677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 54997677f7fdSAxel Rasmussen struct address_space *mapping, 55007677f7fdSAxel Rasmussen pgoff_t idx, 55017677f7fdSAxel Rasmussen unsigned int flags, 55027677f7fdSAxel Rasmussen unsigned long haddr, 5503824ddc60SNadav Amit unsigned long addr, 55047677f7fdSAxel Rasmussen unsigned long reason) 55057677f7fdSAxel Rasmussen { 55067677f7fdSAxel Rasmussen vm_fault_t ret; 55077677f7fdSAxel Rasmussen u32 hash; 55087677f7fdSAxel Rasmussen struct vm_fault vmf = { 55097677f7fdSAxel Rasmussen .vma = vma, 55107677f7fdSAxel Rasmussen .address = haddr, 5511824ddc60SNadav Amit .real_address = addr, 55127677f7fdSAxel Rasmussen .flags = flags, 55137677f7fdSAxel Rasmussen 55147677f7fdSAxel Rasmussen /* 55157677f7fdSAxel Rasmussen * Hard to debug if it ends up being 55167677f7fdSAxel Rasmussen * used by a callee that assumes 55177677f7fdSAxel Rasmussen * something about the other 55187677f7fdSAxel Rasmussen * uninitialized fields... same as in 55197677f7fdSAxel Rasmussen * memory.c 55207677f7fdSAxel Rasmussen */ 55217677f7fdSAxel Rasmussen }; 55227677f7fdSAxel Rasmussen 55237677f7fdSAxel Rasmussen /* 55247677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 55257677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 55267677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 55277677f7fdSAxel Rasmussen */ 55287677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 55297677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 55307677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 55317677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 55327677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 55337677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 55347677f7fdSAxel Rasmussen 55357677f7fdSAxel Rasmussen return ret; 55367677f7fdSAxel Rasmussen } 55377677f7fdSAxel Rasmussen 55382b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 55392b740303SSouptick Joarder struct vm_area_struct *vma, 55408382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5541c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5542c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5543ac9b9c66SHugh Dickins { 5544a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55452b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5546409eb8c2SHillf Danton int anon_rmap = 0; 55474c887265SAdam Litke unsigned long size; 55484c887265SAdam Litke struct page *page; 55491e8f889bSDavid Gibson pte_t new_pte; 5550cb900f41SKirill A. Shutemov spinlock_t *ptl; 5551285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5552c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 55534c887265SAdam Litke 555404f2cbe3SMel Gorman /* 555504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 555604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5557c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5558c89357e2SDavid Hildenbrand * be obvious. 555904f2cbe3SMel Gorman */ 556004f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5561910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 556204f2cbe3SMel Gorman current->pid); 556304f2cbe3SMel Gorman return ret; 556404f2cbe3SMel Gorman } 556504f2cbe3SMel Gorman 55664c887265SAdam Litke /* 556787bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 556887bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 556987bf91d3SMike Kravetz * once for faults beyond end of file. 55704c887265SAdam Litke */ 5571a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 5572ebed4bfcSHugh Dickins if (idx >= size) 5573ebed4bfcSHugh Dickins goto out; 55741a1aad8aSMike Kravetz 5575c7b1850dSMike Kravetz new_page = false; 557687bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 557787bf91d3SMike Kravetz if (!page) { 55787677f7fdSAxel Rasmussen /* Check for page in userfault range */ 55791a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 55807677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5581824ddc60SNadav Amit flags, haddr, address, 55827677f7fdSAxel Rasmussen VM_UFFD_MISSING); 55831a1aad8aSMike Kravetz goto out; 55841a1aad8aSMike Kravetz } 55851a1aad8aSMike Kravetz 5586285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 55872fc39cecSAdam Litke if (IS_ERR(page)) { 55884643d67eSMike Kravetz /* 55894643d67eSMike Kravetz * Returning error will result in faulting task being 55904643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 55914643d67eSMike Kravetz * tasks from racing to fault in the same page which 55924643d67eSMike Kravetz * could result in false unable to allocate errors. 55934643d67eSMike Kravetz * Page migration does not take the fault mutex, but 55944643d67eSMike Kravetz * does a clear then write of pte's under page table 55954643d67eSMike Kravetz * lock. Page fault code could race with migration, 55964643d67eSMike Kravetz * notice the clear pte and try to allocate a page 55974643d67eSMike Kravetz * here. Before returning error, get ptl and make 55984643d67eSMike Kravetz * sure there really is no pte entry. 55994643d67eSMike Kravetz */ 56004643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 56014643d67eSMike Kravetz ret = 0; 5602d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 56032b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5604d83e6c8aSMiaohe Lin spin_unlock(ptl); 56056bda666aSChristoph Lameter goto out; 56066bda666aSChristoph Lameter } 560747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 56080ed361deSNick Piggin __SetPageUptodate(page); 5609cb6acd01SMike Kravetz new_page = true; 5610ac9b9c66SHugh Dickins 5611f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 5612ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 56136bda666aSChristoph Lameter if (err) { 56143a5497a2SMiaohe Lin /* 56153a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 56163a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 56173a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 56183a5497a2SMiaohe Lin * to the page cache. So it's safe to call 56193a5497a2SMiaohe Lin * restore_reserve_on_error() here. 56203a5497a2SMiaohe Lin */ 56213a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 56226bda666aSChristoph Lameter put_page(page); 56236bda666aSChristoph Lameter goto out; 56246bda666aSChristoph Lameter } 5625c7b1850dSMike Kravetz new_pagecache_page = true; 562623be7468SMel Gorman } else { 56276bda666aSChristoph Lameter lock_page(page); 56280fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 56290fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 56300fe6e20bSNaoya Horiguchi goto backout_unlocked; 563123be7468SMel Gorman } 5632409eb8c2SHillf Danton anon_rmap = 1; 56330fe6e20bSNaoya Horiguchi } 56340fe6e20bSNaoya Horiguchi } else { 563557303d80SAndy Whitcroft /* 5636998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5637998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5638998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5639fd6a03edSNaoya Horiguchi */ 5640fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 56410eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5642972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5643fd6a03edSNaoya Horiguchi goto backout_unlocked; 56446bda666aSChristoph Lameter } 56457677f7fdSAxel Rasmussen 56467677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 56477677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 56487677f7fdSAxel Rasmussen unlock_page(page); 56497677f7fdSAxel Rasmussen put_page(page); 56507677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5651824ddc60SNadav Amit flags, haddr, address, 56527677f7fdSAxel Rasmussen VM_UFFD_MINOR); 56537677f7fdSAxel Rasmussen goto out; 56547677f7fdSAxel Rasmussen } 5655998b4382SNaoya Horiguchi } 56561e8f889bSDavid Gibson 565757303d80SAndy Whitcroft /* 565857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 565957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 566057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 566157303d80SAndy Whitcroft * the spinlock. 566257303d80SAndy Whitcroft */ 56635e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5664285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 56652b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 56662b26736cSAndy Whitcroft goto backout_unlocked; 56672b26736cSAndy Whitcroft } 56685e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5669285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 56705e911373SMike Kravetz } 567157303d80SAndy Whitcroft 56728bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 567383c54070SNick Piggin ret = 0; 5674c64e912cSPeter Xu /* If pte changed from under us, retry */ 5675c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 56764c887265SAdam Litke goto backout; 56774c887265SAdam Litke 567807443a85SJoonsoo Kim if (anon_rmap) { 5679d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5680285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5681ac714904SChoi Gi-yong } else 5682fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 56831e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 56841e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5685c64e912cSPeter Xu /* 5686c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5687c64e912cSPeter Xu * if populated. 5688c64e912cSPeter Xu */ 5689c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5690c64e912cSPeter Xu new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); 5691285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 56921e8f889bSDavid Gibson 56935d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5694788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 56951e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5696c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 56971e8f889bSDavid Gibson } 56981e8f889bSDavid Gibson 5699cb900f41SKirill A. Shutemov spin_unlock(ptl); 5700cb6acd01SMike Kravetz 5701cb6acd01SMike Kravetz /* 57028f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 57038f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 57048f251a3dSMike Kravetz * been isolated for migration. 5705cb6acd01SMike Kravetz */ 5706cb6acd01SMike Kravetz if (new_page) 57078f251a3dSMike Kravetz SetHPageMigratable(page); 5708cb6acd01SMike Kravetz 57094c887265SAdam Litke unlock_page(page); 57104c887265SAdam Litke out: 5711ac9b9c66SHugh Dickins return ret; 57124c887265SAdam Litke 57134c887265SAdam Litke backout: 5714cb900f41SKirill A. Shutemov spin_unlock(ptl); 57152b26736cSAndy Whitcroft backout_unlocked: 57164c887265SAdam Litke unlock_page(page); 5717c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5718c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5719285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 57204c887265SAdam Litke put_page(page); 57214c887265SAdam Litke goto out; 5722ac9b9c66SHugh Dickins } 5723ac9b9c66SHugh Dickins 57248382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5725188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57268382d914SDavidlohr Bueso { 57278382d914SDavidlohr Bueso unsigned long key[2]; 57288382d914SDavidlohr Bueso u32 hash; 57298382d914SDavidlohr Bueso 57308382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 57318382d914SDavidlohr Bueso key[1] = idx; 57328382d914SDavidlohr Bueso 573355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 57348382d914SDavidlohr Bueso 57358382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 57368382d914SDavidlohr Bueso } 57378382d914SDavidlohr Bueso #else 57388382d914SDavidlohr Bueso /* 57396c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 57408382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 57418382d914SDavidlohr Bueso */ 5742188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57438382d914SDavidlohr Bueso { 57448382d914SDavidlohr Bueso return 0; 57458382d914SDavidlohr Bueso } 57468382d914SDavidlohr Bueso #endif 57478382d914SDavidlohr Bueso 57482b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5749788c7df4SHugh Dickins unsigned long address, unsigned int flags) 575086e5216fSAdam Litke { 57518382d914SDavidlohr Bueso pte_t *ptep, entry; 5752cb900f41SKirill A. Shutemov spinlock_t *ptl; 57532b740303SSouptick Joarder vm_fault_t ret; 57548382d914SDavidlohr Bueso u32 hash; 57558382d914SDavidlohr Bueso pgoff_t idx; 57560fe6e20bSNaoya Horiguchi struct page *page = NULL; 575757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5758a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57598382d914SDavidlohr Bueso struct address_space *mapping; 57600f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5761285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 576286e5216fSAdam Litke 5763285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5764fd6a03edSNaoya Horiguchi if (ptep) { 5765c0d0381aSMike Kravetz /* 5766c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 5767c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 5768c0d0381aSMike Kravetz * not actually modifying content here. 5769c0d0381aSMike Kravetz */ 5770fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5771290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5772ad1ac596SMiaohe Lin migration_entry_wait_huge(vma, ptep); 5773290408d4SNaoya Horiguchi return 0; 5774290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5775aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5776972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5777b43a9990SMike Kravetz } 57788382d914SDavidlohr Bueso 5779c0d0381aSMike Kravetz /* 5780c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 578187bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 578287bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 578387bf91d3SMike Kravetz * and making the ptep no longer valid. 578487bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 5785c0d0381aSMike Kravetz * 5786c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5787c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5788c0d0381aSMike Kravetz * something has changed. 5789c0d0381aSMike Kravetz */ 5790ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5791c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5792aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5793c0d0381aSMike Kravetz if (!ptep) { 5794c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5795c0d0381aSMike Kravetz return VM_FAULT_OOM; 5796c0d0381aSMike Kravetz } 5797ddeaab32SMike Kravetz 57983935baa9SDavid Gibson /* 57993935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 58003935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 58013935baa9SDavid Gibson * the same page in the page cache. 58023935baa9SDavid Gibson */ 5803c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5804188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5805c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 58068382d914SDavidlohr Bueso 58077f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 5808c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 5809c64e912cSPeter Xu if (huge_pte_none_mostly(entry)) { 5810c64e912cSPeter Xu ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 5811c64e912cSPeter Xu entry, flags); 5812b4d1d99fSDavid Gibson goto out_mutex; 58133935baa9SDavid Gibson } 581486e5216fSAdam Litke 581583c54070SNick Piggin ret = 0; 58161e8f889bSDavid Gibson 581757303d80SAndy Whitcroft /* 58180f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 58190f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 58207c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 58217c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 58227c8de358SEthon Paul * properly handle it. 58230f792cf9SNaoya Horiguchi */ 58240f792cf9SNaoya Horiguchi if (!pte_present(entry)) 58250f792cf9SNaoya Horiguchi goto out_mutex; 58260f792cf9SNaoya Horiguchi 58270f792cf9SNaoya Horiguchi /* 5828c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 5829c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 583057303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 58311d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 58321d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 583357303d80SAndy Whitcroft */ 5834c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 58351d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 5836285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58372b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5838b4d1d99fSDavid Gibson goto out_mutex; 58392b26736cSAndy Whitcroft } 58405e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5841285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 584257303d80SAndy Whitcroft 58431d8d1464SDavid Hildenbrand pagecache_page = hugetlbfs_pagecache_page(h, vma, haddr); 584457303d80SAndy Whitcroft } 584557303d80SAndy Whitcroft 58460f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 58470fe6e20bSNaoya Horiguchi 5848c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 5849b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5850cb900f41SKirill A. Shutemov goto out_ptl; 5851b4d1d99fSDavid Gibson 5852166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 5853166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 5854166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5855166f3eccSPeter Xu struct vm_fault vmf = { 5856166f3eccSPeter Xu .vma = vma, 5857166f3eccSPeter Xu .address = haddr, 5858166f3eccSPeter Xu .real_address = address, 5859166f3eccSPeter Xu .flags = flags, 5860166f3eccSPeter Xu }; 5861166f3eccSPeter Xu 5862166f3eccSPeter Xu spin_unlock(ptl); 5863166f3eccSPeter Xu if (pagecache_page) { 5864166f3eccSPeter Xu unlock_page(pagecache_page); 5865166f3eccSPeter Xu put_page(pagecache_page); 5866166f3eccSPeter Xu } 5867166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5868166f3eccSPeter Xu i_mmap_unlock_read(mapping); 5869166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 5870166f3eccSPeter Xu } 5871166f3eccSPeter Xu 58720f792cf9SNaoya Horiguchi /* 5873c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 58740f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 58750f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 58760f792cf9SNaoya Horiguchi */ 58770f792cf9SNaoya Horiguchi page = pte_page(entry); 58780f792cf9SNaoya Horiguchi if (page != pagecache_page) 58790f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 58800f792cf9SNaoya Horiguchi need_wait_lock = 1; 58810f792cf9SNaoya Horiguchi goto out_ptl; 58820f792cf9SNaoya Horiguchi } 58830f792cf9SNaoya Horiguchi 58840f792cf9SNaoya Horiguchi get_page(page); 5885b4d1d99fSDavid Gibson 5886c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 5887106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5888c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 5889cb900f41SKirill A. Shutemov pagecache_page, ptl); 58900f792cf9SNaoya Horiguchi goto out_put_page; 5891c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 5892106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5893b4d1d99fSDavid Gibson } 5894c89357e2SDavid Hildenbrand } 5895b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5896285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5897788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5898285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 58990f792cf9SNaoya Horiguchi out_put_page: 59000f792cf9SNaoya Horiguchi if (page != pagecache_page) 59010f792cf9SNaoya Horiguchi unlock_page(page); 59020f792cf9SNaoya Horiguchi put_page(page); 5903cb900f41SKirill A. Shutemov out_ptl: 5904cb900f41SKirill A. Shutemov spin_unlock(ptl); 590557303d80SAndy Whitcroft 590657303d80SAndy Whitcroft if (pagecache_page) { 590757303d80SAndy Whitcroft unlock_page(pagecache_page); 590857303d80SAndy Whitcroft put_page(pagecache_page); 590957303d80SAndy Whitcroft } 5910b4d1d99fSDavid Gibson out_mutex: 5911c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5912c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 59130f792cf9SNaoya Horiguchi /* 59140f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 59150f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 59160f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 59170f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 59180f792cf9SNaoya Horiguchi * here without taking refcount. 59190f792cf9SNaoya Horiguchi */ 59200f792cf9SNaoya Horiguchi if (need_wait_lock) 59210f792cf9SNaoya Horiguchi wait_on_page_locked(page); 59221e8f889bSDavid Gibson return ret; 592386e5216fSAdam Litke } 592486e5216fSAdam Litke 5925714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 59268fb5debcSMike Kravetz /* 59278fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 59288fb5debcSMike Kravetz * modifications for huge pages. 59298fb5debcSMike Kravetz */ 59308fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 59318fb5debcSMike Kravetz pte_t *dst_pte, 59328fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 59338fb5debcSMike Kravetz unsigned long dst_addr, 59348fb5debcSMike Kravetz unsigned long src_addr, 5935f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 59366041c691SPeter Xu struct page **pagep, 59376041c691SPeter Xu bool wp_copy) 59388fb5debcSMike Kravetz { 5939f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 59408cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 59418cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 59428cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 59431e392147SAndrea Arcangeli unsigned long size; 59441c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 59458fb5debcSMike Kravetz pte_t _dst_pte; 59468fb5debcSMike Kravetz spinlock_t *ptl; 59478cc5fcbbSMina Almasry int ret = -ENOMEM; 59488fb5debcSMike Kravetz struct page *page; 5949f6191471SAxel Rasmussen int writable; 5950cc30042dSMina Almasry bool page_in_pagecache = false; 59518fb5debcSMike Kravetz 5952f6191471SAxel Rasmussen if (is_continue) { 5953f6191471SAxel Rasmussen ret = -EFAULT; 5954f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5955f6191471SAxel Rasmussen if (!page) 5956f6191471SAxel Rasmussen goto out; 5957cc30042dSMina Almasry page_in_pagecache = true; 5958f6191471SAxel Rasmussen } else if (!*pagep) { 5959d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5960d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5961d84cf06eSMina Almasry */ 5962d84cf06eSMina Almasry if (vm_shared && 5963d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5964d84cf06eSMina Almasry ret = -EEXIST; 59658fb5debcSMike Kravetz goto out; 5966d84cf06eSMina Almasry } 5967d84cf06eSMina Almasry 5968d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5969d84cf06eSMina Almasry if (IS_ERR(page)) { 5970d84cf06eSMina Almasry ret = -ENOMEM; 5971d84cf06eSMina Almasry goto out; 5972d84cf06eSMina Almasry } 59738fb5debcSMike Kravetz 59748fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 59758fb5debcSMike Kravetz (const void __user *) src_addr, 5976810a56b9SMike Kravetz pages_per_huge_page(h), false); 59778fb5debcSMike Kravetz 5978c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 59798fb5debcSMike Kravetz if (unlikely(ret)) { 59809e368259SAndrea Arcangeli ret = -ENOENT; 59818cc5fcbbSMina Almasry /* Free the allocated page which may have 59828cc5fcbbSMina Almasry * consumed a reservation. 59838cc5fcbbSMina Almasry */ 59848cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 59858cc5fcbbSMina Almasry put_page(page); 59868cc5fcbbSMina Almasry 59878cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 59888cc5fcbbSMina Almasry * contents. 59898cc5fcbbSMina Almasry */ 59908cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 59918cc5fcbbSMina Almasry if (!page) { 59928cc5fcbbSMina Almasry ret = -ENOMEM; 59938cc5fcbbSMina Almasry goto out; 59948cc5fcbbSMina Almasry } 59958fb5debcSMike Kravetz *pagep = page; 59968cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 59978cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 59988cc5fcbbSMina Almasry * page. 59998cc5fcbbSMina Almasry */ 60008fb5debcSMike Kravetz goto out; 60018fb5debcSMike Kravetz } 60028fb5debcSMike Kravetz } else { 60038cc5fcbbSMina Almasry if (vm_shared && 60048cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 60058cc5fcbbSMina Almasry put_page(*pagep); 60068cc5fcbbSMina Almasry ret = -EEXIST; 60078cc5fcbbSMina Almasry *pagep = NULL; 60088cc5fcbbSMina Almasry goto out; 60098cc5fcbbSMina Almasry } 60108cc5fcbbSMina Almasry 60118cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 60128cc5fcbbSMina Almasry if (IS_ERR(page)) { 6013da9a298fSMiaohe Lin put_page(*pagep); 60148cc5fcbbSMina Almasry ret = -ENOMEM; 60158cc5fcbbSMina Almasry *pagep = NULL; 60168cc5fcbbSMina Almasry goto out; 60178cc5fcbbSMina Almasry } 601834892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 601934892366SMuchun Song pages_per_huge_page(h)); 60208cc5fcbbSMina Almasry put_page(*pagep); 60218fb5debcSMike Kravetz *pagep = NULL; 60228fb5debcSMike Kravetz } 60238fb5debcSMike Kravetz 60248fb5debcSMike Kravetz /* 60258fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 60268fb5debcSMike Kravetz * preceding stores to the page contents become visible before 60278fb5debcSMike Kravetz * the set_pte_at() write. 60288fb5debcSMike Kravetz */ 60298fb5debcSMike Kravetz __SetPageUptodate(page); 60308fb5debcSMike Kravetz 6031f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6032f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 60331e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60341e392147SAndrea Arcangeli ret = -EFAULT; 60351e392147SAndrea Arcangeli if (idx >= size) 60361e392147SAndrea Arcangeli goto out_release_nounlock; 60371c9e8defSMike Kravetz 60381e392147SAndrea Arcangeli /* 60391e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 60401e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 60411e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 60421e392147SAndrea Arcangeli * the caller. 60431e392147SAndrea Arcangeli */ 60441c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 60451c9e8defSMike Kravetz if (ret) 60461c9e8defSMike Kravetz goto out_release_nounlock; 6047cc30042dSMina Almasry page_in_pagecache = true; 60481c9e8defSMike Kravetz } 60491c9e8defSMike Kravetz 6050bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 60518fb5debcSMike Kravetz 60521e392147SAndrea Arcangeli /* 60531e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 60541e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 60551e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 60561e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 60571e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 60581e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 60591e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 60601e392147SAndrea Arcangeli */ 60611e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60621e392147SAndrea Arcangeli ret = -EFAULT; 60631e392147SAndrea Arcangeli if (idx >= size) 60641e392147SAndrea Arcangeli goto out_release_unlock; 60651e392147SAndrea Arcangeli 60668fb5debcSMike Kravetz ret = -EEXIST; 60676041c691SPeter Xu /* 60686041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 60696041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 60706041c691SPeter Xu * page backing it, then access the page. 60716041c691SPeter Xu */ 60726041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 60738fb5debcSMike Kravetz goto out_release_unlock; 60748fb5debcSMike Kravetz 6075ab74ef70SMiaohe Lin if (page_in_pagecache) { 6076fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 60771c9e8defSMike Kravetz } else { 6078d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 60798fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 60801c9e8defSMike Kravetz } 60818fb5debcSMike Kravetz 60826041c691SPeter Xu /* 60836041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 60846041c691SPeter Xu * with wp flag set, don't set pte write bit. 60856041c691SPeter Xu */ 60866041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6087f6191471SAxel Rasmussen writable = 0; 6088f6191471SAxel Rasmussen else 6089f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6090f6191471SAxel Rasmussen 6091f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 60926041c691SPeter Xu /* 60936041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 60946041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 60956041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 60966041c691SPeter Xu * thrown away at will, even if write bit not set. 60976041c691SPeter Xu */ 60988fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 60998fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 61008fb5debcSMike Kravetz 61016041c691SPeter Xu if (wp_copy) 61026041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 61036041c691SPeter Xu 61048fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 61058fb5debcSMike Kravetz 61068fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 61078fb5debcSMike Kravetz 61088fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 61098fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 61108fb5debcSMike Kravetz 61118fb5debcSMike Kravetz spin_unlock(ptl); 6112f6191471SAxel Rasmussen if (!is_continue) 61138f251a3dSMike Kravetz SetHPageMigratable(page); 6114f6191471SAxel Rasmussen if (vm_shared || is_continue) 61151c9e8defSMike Kravetz unlock_page(page); 61168fb5debcSMike Kravetz ret = 0; 61178fb5debcSMike Kravetz out: 61188fb5debcSMike Kravetz return ret; 61198fb5debcSMike Kravetz out_release_unlock: 61208fb5debcSMike Kravetz spin_unlock(ptl); 6121f6191471SAxel Rasmussen if (vm_shared || is_continue) 61221c9e8defSMike Kravetz unlock_page(page); 61235af10dfdSAndrea Arcangeli out_release_nounlock: 6124cc30042dSMina Almasry if (!page_in_pagecache) 6125846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 61268fb5debcSMike Kravetz put_page(page); 61278fb5debcSMike Kravetz goto out; 61288fb5debcSMike Kravetz } 6129714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 61308fb5debcSMike Kravetz 613182e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 613282e5d378SJoao Martins int refs, struct page **pages, 613382e5d378SJoao Martins struct vm_area_struct **vmas) 613482e5d378SJoao Martins { 613582e5d378SJoao Martins int nr; 613682e5d378SJoao Martins 613782e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 613882e5d378SJoao Martins if (likely(pages)) 613982e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 614082e5d378SJoao Martins if (vmas) 614182e5d378SJoao Martins vmas[nr] = vma; 614282e5d378SJoao Martins } 614382e5d378SJoao Martins } 614482e5d378SJoao Martins 6145a7f22660SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, 6146a7f22660SDavid Hildenbrand bool *unshare) 6147a7f22660SDavid Hildenbrand { 6148a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6149a7f22660SDavid Hildenbrand 6150a7f22660SDavid Hildenbrand *unshare = false; 6151a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6152a7f22660SDavid Hildenbrand return true; 6153a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6154a7f22660SDavid Hildenbrand return false; 6155a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6156a7f22660SDavid Hildenbrand return true; 6157a7f22660SDavid Hildenbrand if (gup_must_unshare(flags, pte_page(pteval))) { 6158a7f22660SDavid Hildenbrand *unshare = true; 6159a7f22660SDavid Hildenbrand return true; 6160a7f22660SDavid Hildenbrand } 6161a7f22660SDavid Hildenbrand return false; 6162a7f22660SDavid Hildenbrand } 6163a7f22660SDavid Hildenbrand 616428a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 616563551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 616628a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 61674f6da934SPeter Xu long i, unsigned int flags, int *locked) 616863551ae0SDavid Gibson { 6169d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6170d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 617128a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6172a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 61730fa5bc40SJoao Martins int err = -EFAULT, refs; 617463551ae0SDavid Gibson 617563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 617663551ae0SDavid Gibson pte_t *pte; 6177cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6178a7f22660SDavid Hildenbrand bool unshare = false; 61792a15efc9SHugh Dickins int absent; 618063551ae0SDavid Gibson struct page *page; 618163551ae0SDavid Gibson 61824c887265SAdam Litke /* 618302057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 618402057967SDavid Rientjes * potentially allocating memory. 618502057967SDavid Rientjes */ 6186fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 618702057967SDavid Rientjes remainder = 0; 618802057967SDavid Rientjes break; 618902057967SDavid Rientjes } 619002057967SDavid Rientjes 619102057967SDavid Rientjes /* 61924c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 61932a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 61944c887265SAdam Litke * first, for the page indexing below to work. 6195cb900f41SKirill A. Shutemov * 6196cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 61974c887265SAdam Litke */ 61987868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 61997868a208SPunit Agrawal huge_page_size(h)); 6200cb900f41SKirill A. Shutemov if (pte) 6201cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 62022a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 620363551ae0SDavid Gibson 62042a15efc9SHugh Dickins /* 62052a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 62063ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 62073ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 62083ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 62093ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 62102a15efc9SHugh Dickins */ 62113ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 62123ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6213cb900f41SKirill A. Shutemov if (pte) 6214cb900f41SKirill A. Shutemov spin_unlock(ptl); 62152a15efc9SHugh Dickins remainder = 0; 62162a15efc9SHugh Dickins break; 62172a15efc9SHugh Dickins } 62182a15efc9SHugh Dickins 62199cc3a5bdSNaoya Horiguchi /* 62209cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 62219cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 62229cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 62239cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 62249cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 62259cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 62269cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 62279cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 62289cc3a5bdSNaoya Horiguchi */ 6229a7f22660SDavid Hildenbrand if (absent || 6230a7f22660SDavid Hildenbrand __follow_hugetlb_must_fault(flags, pte, &unshare)) { 62312b740303SSouptick Joarder vm_fault_t ret; 623287ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 62334c887265SAdam Litke 6234cb900f41SKirill A. Shutemov if (pte) 6235cb900f41SKirill A. Shutemov spin_unlock(ptl); 623687ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 623787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6238a7f22660SDavid Hildenbrand else if (unshare) 6239a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 62404f6da934SPeter Xu if (locked) 624171335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 624271335f37SPeter Xu FAULT_FLAG_KILLABLE; 624387ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 624487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 624587ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 624687ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 62474426e945SPeter Xu /* 62484426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 62494426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 62504426e945SPeter Xu */ 625187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 625287ffc118SAndrea Arcangeli } 625387ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 625487ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 62552be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 62561c59827dSHugh Dickins remainder = 0; 62571c59827dSHugh Dickins break; 62581c59827dSHugh Dickins } 625987ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 62604f6da934SPeter Xu if (locked && 62611ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 62624f6da934SPeter Xu *locked = 0; 626387ffc118SAndrea Arcangeli *nr_pages = 0; 626487ffc118SAndrea Arcangeli /* 626587ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 626687ffc118SAndrea Arcangeli * error, it will return zero 626787ffc118SAndrea Arcangeli * instead. 626887ffc118SAndrea Arcangeli * 626987ffc118SAndrea Arcangeli * No need to update "position" as the 627087ffc118SAndrea Arcangeli * caller will not check it after 627187ffc118SAndrea Arcangeli * *nr_pages is set to 0. 627287ffc118SAndrea Arcangeli */ 627387ffc118SAndrea Arcangeli return i; 627487ffc118SAndrea Arcangeli } 627587ffc118SAndrea Arcangeli continue; 627687ffc118SAndrea Arcangeli } 627763551ae0SDavid Gibson 6278a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 62797f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 62808fde12caSLinus Torvalds 6281b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6282b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6283b6a2619cSDavid Hildenbrand 62848fde12caSLinus Torvalds /* 6285acbfb087SZhigang Lu * If subpage information not requested, update counters 6286acbfb087SZhigang Lu * and skip the same_page loop below. 6287acbfb087SZhigang Lu */ 6288acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6289acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6290acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6291acbfb087SZhigang Lu vaddr += huge_page_size(h); 6292acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6293acbfb087SZhigang Lu i += pages_per_huge_page(h); 6294acbfb087SZhigang Lu spin_unlock(ptl); 6295acbfb087SZhigang Lu continue; 6296acbfb087SZhigang Lu } 6297acbfb087SZhigang Lu 6298d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6299d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6300d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 63010fa5bc40SJoao Martins 630282e5d378SJoao Martins if (pages || vmas) 630382e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 630482e5d378SJoao Martins vma, refs, 630582e5d378SJoao Martins likely(pages) ? pages + i : NULL, 630682e5d378SJoao Martins vmas ? vmas + i : NULL); 630763551ae0SDavid Gibson 630882e5d378SJoao Martins if (pages) { 63090fa5bc40SJoao Martins /* 6310822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 63110fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 63120fa5bc40SJoao Martins * checked that the huge page is present in the page 63130fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 63140fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 63150fa5bc40SJoao Martins * head page and tail pages from being rearranged in 63160fa5bc40SJoao Martins * any way. So this page must be available at this 63170fa5bc40SJoao Martins * point, unless the page refcount overflowed: 63180fa5bc40SJoao Martins */ 6319822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 63200fa5bc40SJoao Martins flags))) { 63210fa5bc40SJoao Martins spin_unlock(ptl); 63220fa5bc40SJoao Martins remainder = 0; 63230fa5bc40SJoao Martins err = -ENOMEM; 63240fa5bc40SJoao Martins break; 63250fa5bc40SJoao Martins } 6326d5d4b0aaSChen, Kenneth W } 632782e5d378SJoao Martins 632882e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 632982e5d378SJoao Martins remainder -= refs; 633082e5d378SJoao Martins i += refs; 633182e5d378SJoao Martins 6332cb900f41SKirill A. Shutemov spin_unlock(ptl); 633363551ae0SDavid Gibson } 633428a35716SMichel Lespinasse *nr_pages = remainder; 633587ffc118SAndrea Arcangeli /* 633687ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 633787ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 633887ffc118SAndrea Arcangeli * branch. 633987ffc118SAndrea Arcangeli */ 634063551ae0SDavid Gibson *position = vaddr; 634163551ae0SDavid Gibson 63422be7cfedSDaniel Jordan return i ? i : err; 634363551ae0SDavid Gibson } 63448f860591SZhang, Yanmin 63457da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 63465a90d5a1SPeter Xu unsigned long address, unsigned long end, 63475a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 63488f860591SZhang, Yanmin { 63498f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 63508f860591SZhang, Yanmin unsigned long start = address; 63518f860591SZhang, Yanmin pte_t *ptep; 63528f860591SZhang, Yanmin pte_t pte; 6353a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 635460dfaad6SPeter Xu unsigned long pages = 0, psize = huge_page_size(h); 6355dff11abeSMike Kravetz bool shared_pmd = false; 6356ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6357e95a9851SMike Kravetz unsigned long last_addr_mask; 63585a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 63595a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6360dff11abeSMike Kravetz 6361dff11abeSMike Kravetz /* 6362dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6363ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6364dff11abeSMike Kravetz * range if PMD sharing is possible. 6365dff11abeSMike Kravetz */ 63667269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 63677269f999SJérôme Glisse 0, vma, mm, start, end); 6368ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 63698f860591SZhang, Yanmin 63708f860591SZhang, Yanmin BUG_ON(address >= end); 6371ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 63728f860591SZhang, Yanmin 6373ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 6374e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 637583cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 637660dfaad6SPeter Xu for (; address < end; address += psize) { 6377cb900f41SKirill A. Shutemov spinlock_t *ptl; 637860dfaad6SPeter Xu ptep = huge_pte_offset(mm, address, psize); 6379e95a9851SMike Kravetz if (!ptep) { 6380e95a9851SMike Kravetz address |= last_addr_mask; 63818f860591SZhang, Yanmin continue; 6382e95a9851SMike Kravetz } 6383cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 63844ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 638560dfaad6SPeter Xu /* 638660dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 638760dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 638860dfaad6SPeter Xu * happened due to some reason. 638960dfaad6SPeter Xu */ 639060dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 63917da4d641SPeter Zijlstra pages++; 6392cb900f41SKirill A. Shutemov spin_unlock(ptl); 6393dff11abeSMike Kravetz shared_pmd = true; 63944ddb4d91SMike Kravetz address |= last_addr_mask; 639539dde65cSChen, Kenneth W continue; 63967da4d641SPeter Zijlstra } 6397a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6398a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6399a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6400a8bda28dSNaoya Horiguchi continue; 6401a8bda28dSNaoya Horiguchi } 6402a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6403a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 64046c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 6405a8bda28dSNaoya Horiguchi 64066c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 6407a8bda28dSNaoya Horiguchi pte_t newpte; 6408a8bda28dSNaoya Horiguchi 64096c287605SDavid Hildenbrand if (PageAnon(page)) 64106c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 64116c287605SDavid Hildenbrand swp_offset(entry)); 64126c287605SDavid Hildenbrand else 64134dd845b5SAlistair Popple entry = make_readable_migration_entry( 64144dd845b5SAlistair Popple swp_offset(entry)); 6415a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 64165a90d5a1SPeter Xu if (uffd_wp) 64175a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 64185a90d5a1SPeter Xu else if (uffd_wp_resolve) 64195a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 642018f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 6421a8bda28dSNaoya Horiguchi pages++; 6422a8bda28dSNaoya Horiguchi } 6423a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6424a8bda28dSNaoya Horiguchi continue; 6425a8bda28dSNaoya Horiguchi } 642660dfaad6SPeter Xu if (unlikely(pte_marker_uffd_wp(pte))) { 642760dfaad6SPeter Xu /* 642860dfaad6SPeter Xu * This is changing a non-present pte into a none pte, 642960dfaad6SPeter Xu * no need for huge_ptep_modify_prot_start/commit(). 643060dfaad6SPeter Xu */ 643160dfaad6SPeter Xu if (uffd_wp_resolve) 643260dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 643360dfaad6SPeter Xu } 6434a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6435023bdd00SAneesh Kumar K.V pte_t old_pte; 643679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6437023bdd00SAneesh Kumar K.V 6438023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 643916785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 644079c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 64415a90d5a1SPeter Xu if (uffd_wp) 64425a90d5a1SPeter Xu pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); 64435a90d5a1SPeter Xu else if (uffd_wp_resolve) 64445a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6445023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 64467da4d641SPeter Zijlstra pages++; 644760dfaad6SPeter Xu } else { 644860dfaad6SPeter Xu /* None pte */ 644960dfaad6SPeter Xu if (unlikely(uffd_wp)) 645060dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 645160dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 645260dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 64538f860591SZhang, Yanmin } 6454cb900f41SKirill A. Shutemov spin_unlock(ptl); 64558f860591SZhang, Yanmin } 6456d833352aSMel Gorman /* 6457c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6458d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6459c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6460dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6461dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6462d833352aSMel Gorman */ 6463dff11abeSMike Kravetz if (shared_pmd) 6464ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6465dff11abeSMike Kravetz else 64665491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 64670f10851eSJérôme Glisse /* 64680f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 64690f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 64700f10851eSJérôme Glisse * 6471ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 64720f10851eSJérôme Glisse */ 647383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 6474ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 64757da4d641SPeter Zijlstra 64767da4d641SPeter Zijlstra return pages << h->order; 64778f860591SZhang, Yanmin } 64788f860591SZhang, Yanmin 647933b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 648033b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6481a1e78772SMel Gorman long from, long to, 64825a6fe125SMel Gorman struct vm_area_struct *vma, 6483ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6484e4e574b7SAdam Litke { 648533b8f84aSMike Kravetz long chg, add = -1; 6486a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 648790481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 64889119a41eSJoonsoo Kim struct resv_map *resv_map; 6489075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 64900db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6491e4e574b7SAdam Litke 649263489f8eSMike Kravetz /* This should never happen */ 649363489f8eSMike Kravetz if (from > to) { 649463489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 649533b8f84aSMike Kravetz return false; 649663489f8eSMike Kravetz } 649763489f8eSMike Kravetz 6498a1e78772SMel Gorman /* 649917c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 650017c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 650190481622SDavid Gibson * without using reserves 650217c9d12eSMel Gorman */ 6503ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 650433b8f84aSMike Kravetz return true; 650517c9d12eSMel Gorman 650617c9d12eSMel Gorman /* 6507a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6508a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6509a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6510a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6511a1e78772SMel Gorman */ 65129119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6513f27a5136SMike Kravetz /* 6514f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6515f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6516f27a5136SMike Kravetz * hugetlbfs_get_inode). 6517f27a5136SMike Kravetz */ 65184e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 65199119a41eSJoonsoo Kim 65200db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 65219119a41eSJoonsoo Kim 65229119a41eSJoonsoo Kim } else { 6523e9fe92aeSMina Almasry /* Private mapping. */ 65249119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 65255a6fe125SMel Gorman if (!resv_map) 652633b8f84aSMike Kravetz return false; 65275a6fe125SMel Gorman 652817c9d12eSMel Gorman chg = to - from; 652917c9d12eSMel Gorman 65305a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 65315a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 65325a6fe125SMel Gorman } 65335a6fe125SMel Gorman 653433b8f84aSMike Kravetz if (chg < 0) 6535c50ac050SDave Hansen goto out_err; 653617c9d12eSMel Gorman 653733b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 653833b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6539075a61d0SMina Almasry goto out_err; 6540075a61d0SMina Almasry 6541075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6542075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6543075a61d0SMina Almasry * of the resv_map. 6544075a61d0SMina Almasry */ 6545075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6546075a61d0SMina Almasry } 6547075a61d0SMina Almasry 65481c5ecae3SMike Kravetz /* 65491c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 65501c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 65511c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 65521c5ecae3SMike Kravetz */ 65531c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 655433b8f84aSMike Kravetz if (gbl_reserve < 0) 6555075a61d0SMina Almasry goto out_uncharge_cgroup; 655617c9d12eSMel Gorman 655717c9d12eSMel Gorman /* 655817c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 655990481622SDavid Gibson * Hand the pages back to the subpool if there are not 656017c9d12eSMel Gorman */ 656133b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6562075a61d0SMina Almasry goto out_put_pages; 656317c9d12eSMel Gorman 656417c9d12eSMel Gorman /* 656517c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 656617c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 656717c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 656817c9d12eSMel Gorman * the reservation was and the page cache tells how much of 656917c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 657017c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 657117c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 657217c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 657317c9d12eSMel Gorman * else has to be done for private mappings here 657417c9d12eSMel Gorman */ 657533039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6576075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 657733039678SMike Kravetz 65780db9d74eSMina Almasry if (unlikely(add < 0)) { 65790db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6580075a61d0SMina Almasry goto out_put_pages; 65810db9d74eSMina Almasry } else if (unlikely(chg > add)) { 658233039678SMike Kravetz /* 658333039678SMike Kravetz * pages in this range were added to the reserve 658433039678SMike Kravetz * map between region_chg and region_add. This 658533039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 658633039678SMike Kravetz * the subpool and reserve counts modified above 658733039678SMike Kravetz * based on the difference. 658833039678SMike Kravetz */ 658933039678SMike Kravetz long rsv_adjust; 659033039678SMike Kravetz 6591d85aecf2SMiaohe Lin /* 6592d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6593d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6594d85aecf2SMiaohe Lin */ 6595075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6596075a61d0SMina Almasry hstate_index(h), 6597075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6598075a61d0SMina Almasry 659933039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 660033039678SMike Kravetz chg - add); 660133039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6602d85aecf2SMiaohe Lin } else if (h_cg) { 6603d85aecf2SMiaohe Lin /* 6604d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6605d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6606d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6607d85aecf2SMiaohe Lin * done. 6608d85aecf2SMiaohe Lin */ 6609d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 661033039678SMike Kravetz } 661133039678SMike Kravetz } 661233b8f84aSMike Kravetz return true; 661333b8f84aSMike Kravetz 6614075a61d0SMina Almasry out_put_pages: 6615075a61d0SMina Almasry /* put back original number of pages, chg */ 6616075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6617075a61d0SMina Almasry out_uncharge_cgroup: 6618075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6619075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6620c50ac050SDave Hansen out_err: 66215e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 66220db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 66230db9d74eSMina Almasry * region_add failed or didn't run. 66240db9d74eSMina Almasry */ 66250db9d74eSMina Almasry if (chg >= 0 && add < 0) 66260db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6627f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6628f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 662933b8f84aSMike Kravetz return false; 6630a43a8c39SChen, Kenneth W } 6631a43a8c39SChen, Kenneth W 6632b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6633b5cec28dSMike Kravetz long freed) 6634a43a8c39SChen, Kenneth W { 6635a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 66364e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 66379119a41eSJoonsoo Kim long chg = 0; 663890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66391c5ecae3SMike Kravetz long gbl_reserve; 664045c682a6SKen Chen 6641f27a5136SMike Kravetz /* 6642f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6643f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6644f27a5136SMike Kravetz */ 6645b5cec28dSMike Kravetz if (resv_map) { 6646b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6647b5cec28dSMike Kravetz /* 6648b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6649b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6650b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6651b5cec28dSMike Kravetz */ 6652b5cec28dSMike Kravetz if (chg < 0) 6653b5cec28dSMike Kravetz return chg; 6654b5cec28dSMike Kravetz } 6655b5cec28dSMike Kravetz 665645c682a6SKen Chen spin_lock(&inode->i_lock); 6657e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 665845c682a6SKen Chen spin_unlock(&inode->i_lock); 665945c682a6SKen Chen 66601c5ecae3SMike Kravetz /* 66611c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 66621c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6663dddf31a4SMiaohe Lin * 6664dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6665dddf31a4SMiaohe Lin * won't go negative. 66661c5ecae3SMike Kravetz */ 66671c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 66681c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6669b5cec28dSMike Kravetz 6670b5cec28dSMike Kravetz return 0; 6671a43a8c39SChen, Kenneth W } 667293f70f90SNaoya Horiguchi 66733212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 66743212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 66753212b535SSteve Capper struct vm_area_struct *vma, 66763212b535SSteve Capper unsigned long addr, pgoff_t idx) 66773212b535SSteve Capper { 66783212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 66793212b535SSteve Capper svma->vm_start; 66803212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 66813212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 66823212b535SSteve Capper 66833212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6684de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6685de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 66863212b535SSteve Capper 66873212b535SSteve Capper /* 66883212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 66893212b535SSteve Capper * page table page. 66903212b535SSteve Capper */ 66913212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 66923212b535SSteve Capper vm_flags != svm_flags || 669307e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 66943212b535SSteve Capper return 0; 66953212b535SSteve Capper 66963212b535SSteve Capper return saddr; 66973212b535SSteve Capper } 66983212b535SSteve Capper 669931aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 67003212b535SSteve Capper { 67013212b535SSteve Capper unsigned long base = addr & PUD_MASK; 67023212b535SSteve Capper unsigned long end = base + PUD_SIZE; 67033212b535SSteve Capper 67043212b535SSteve Capper /* 67053212b535SSteve Capper * check on proper vm_flags and page table alignment 67063212b535SSteve Capper */ 6707017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 670831aafb45SNicholas Krause return true; 670931aafb45SNicholas Krause return false; 67103212b535SSteve Capper } 67113212b535SSteve Capper 6712c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6713c1991e07SPeter Xu { 6714c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 6715c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 6716c1991e07SPeter Xu return false; 6717c1991e07SPeter Xu #endif 6718c1991e07SPeter Xu return vma_shareable(vma, addr); 6719c1991e07SPeter Xu } 6720c1991e07SPeter Xu 67213212b535SSteve Capper /* 6722017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6723017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6724017b1660SMike Kravetz * shared pmd mappings. 6725017b1660SMike Kravetz */ 6726017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6727017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6728017b1660SMike Kravetz { 6729a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6730a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6731017b1660SMike Kravetz 6732a1ba9da8SLi Xinhai /* 6733f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6734f0953a1bSIngo Molnar * must be at least partially within in. 6735a1ba9da8SLi Xinhai */ 6736a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6737a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6738017b1660SMike Kravetz return; 6739017b1660SMike Kravetz 674075802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6741a1ba9da8SLi Xinhai if (*start > v_start) 6742a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6743017b1660SMike Kravetz 6744a1ba9da8SLi Xinhai if (*end < v_end) 6745a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6746017b1660SMike Kravetz } 6747017b1660SMike Kravetz 6748017b1660SMike Kravetz /* 67493212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 67503212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 67513212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 6752c0d0381aSMike Kravetz * code much cleaner. 6753c0d0381aSMike Kravetz * 67540bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 67550bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 67560bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 67570bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 67583212b535SSteve Capper */ 6759aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6760aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 67613212b535SSteve Capper { 67623212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 67633212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 67643212b535SSteve Capper vma->vm_pgoff; 67653212b535SSteve Capper struct vm_area_struct *svma; 67663212b535SSteve Capper unsigned long saddr; 67673212b535SSteve Capper pte_t *spte = NULL; 67683212b535SSteve Capper pte_t *pte; 6769cb900f41SKirill A. Shutemov spinlock_t *ptl; 67703212b535SSteve Capper 67710bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 67723212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 67733212b535SSteve Capper if (svma == vma) 67743212b535SSteve Capper continue; 67753212b535SSteve Capper 67763212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 67773212b535SSteve Capper if (saddr) { 67787868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 67797868a208SPunit Agrawal vma_mmu_pagesize(svma)); 67803212b535SSteve Capper if (spte) { 67813212b535SSteve Capper get_page(virt_to_page(spte)); 67823212b535SSteve Capper break; 67833212b535SSteve Capper } 67843212b535SSteve Capper } 67853212b535SSteve Capper } 67863212b535SSteve Capper 67873212b535SSteve Capper if (!spte) 67883212b535SSteve Capper goto out; 67893212b535SSteve Capper 67908bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6791dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 67923212b535SSteve Capper pud_populate(mm, pud, 67933212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6794c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6795dc6c9a35SKirill A. Shutemov } else { 67963212b535SSteve Capper put_page(virt_to_page(spte)); 6797dc6c9a35SKirill A. Shutemov } 6798cb900f41SKirill A. Shutemov spin_unlock(ptl); 67993212b535SSteve Capper out: 68003212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 68013212b535SSteve Capper return pte; 68023212b535SSteve Capper } 68033212b535SSteve Capper 68043212b535SSteve Capper /* 68053212b535SSteve Capper * unmap huge page backed by shared pte. 68063212b535SSteve Capper * 68073212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 68083212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 68093212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 68103212b535SSteve Capper * 6811c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 68123212b535SSteve Capper * 68133212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 68143212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 68153212b535SSteve Capper */ 681634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 68174ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 68183212b535SSteve Capper { 68194ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 68204ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 68214ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 68223212b535SSteve Capper 682334ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 68243212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 68253212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 68263212b535SSteve Capper return 0; 68273212b535SSteve Capper 68283212b535SSteve Capper pud_clear(pud); 68293212b535SSteve Capper put_page(virt_to_page(ptep)); 6830dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 68313212b535SSteve Capper return 1; 68323212b535SSteve Capper } 6833c1991e07SPeter Xu 68349e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 6835aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6836aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 68379e5fc74cSSteve Capper { 68389e5fc74cSSteve Capper return NULL; 68399e5fc74cSSteve Capper } 6840e81f2d22SZhang Zhen 684134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 68424ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 6843e81f2d22SZhang Zhen { 6844e81f2d22SZhang Zhen return 0; 6845e81f2d22SZhang Zhen } 6846017b1660SMike Kravetz 6847017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6848017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6849017b1660SMike Kravetz { 6850017b1660SMike Kravetz } 6851c1991e07SPeter Xu 6852c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6853c1991e07SPeter Xu { 6854c1991e07SPeter Xu return false; 6855c1991e07SPeter Xu } 68563212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 68573212b535SSteve Capper 68589e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 6859aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 68609e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 68619e5fc74cSSteve Capper { 68629e5fc74cSSteve Capper pgd_t *pgd; 6863c2febafcSKirill A. Shutemov p4d_t *p4d; 68649e5fc74cSSteve Capper pud_t *pud; 68659e5fc74cSSteve Capper pte_t *pte = NULL; 68669e5fc74cSSteve Capper 68679e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6868f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 6869f4f0a3d8SKirill A. Shutemov if (!p4d) 6870f4f0a3d8SKirill A. Shutemov return NULL; 6871c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 68729e5fc74cSSteve Capper if (pud) { 68739e5fc74cSSteve Capper if (sz == PUD_SIZE) { 68749e5fc74cSSteve Capper pte = (pte_t *)pud; 68759e5fc74cSSteve Capper } else { 68769e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 6877c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 6878aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 68799e5fc74cSSteve Capper else 68809e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 68819e5fc74cSSteve Capper } 68829e5fc74cSSteve Capper } 68834e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 68849e5fc74cSSteve Capper 68859e5fc74cSSteve Capper return pte; 68869e5fc74cSSteve Capper } 68879e5fc74cSSteve Capper 68889b19df29SPunit Agrawal /* 68899b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 68909b19df29SPunit Agrawal * entry at address @addr 68919b19df29SPunit Agrawal * 68928ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 68938ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 68949b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 68959b19df29SPunit Agrawal * table. 68969b19df29SPunit Agrawal */ 68977868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 68987868a208SPunit Agrawal unsigned long addr, unsigned long sz) 68999e5fc74cSSteve Capper { 69009e5fc74cSSteve Capper pgd_t *pgd; 6901c2febafcSKirill A. Shutemov p4d_t *p4d; 69028ac0b81aSLi Xinhai pud_t *pud; 69038ac0b81aSLi Xinhai pmd_t *pmd; 69049e5fc74cSSteve Capper 69059e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6906c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6907c2febafcSKirill A. Shutemov return NULL; 6908c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6909c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6910c2febafcSKirill A. Shutemov return NULL; 69119b19df29SPunit Agrawal 6912c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 69138ac0b81aSLi Xinhai if (sz == PUD_SIZE) 69148ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 69159e5fc74cSSteve Capper return (pte_t *)pud; 69168ac0b81aSLi Xinhai if (!pud_present(*pud)) 69178ac0b81aSLi Xinhai return NULL; 69188ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 69199b19df29SPunit Agrawal 69209e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 69218ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 69229e5fc74cSSteve Capper return (pte_t *)pmd; 69239e5fc74cSSteve Capper } 69249e5fc74cSSteve Capper 6925e95a9851SMike Kravetz /* 6926e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 6927e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 6928e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 6929e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 6930e95a9851SMike Kravetz * version of this routine. 6931e95a9851SMike Kravetz */ 6932e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 6933e95a9851SMike Kravetz { 6934e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 6935e95a9851SMike Kravetz 6936e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 6937e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 6938e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 6939e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 6940e95a9851SMike Kravetz else 6941e95a9851SMike Kravetz return 0UL; 6942e95a9851SMike Kravetz } 6943e95a9851SMike Kravetz 6944e95a9851SMike Kravetz #else 6945e95a9851SMike Kravetz 6946e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 6947e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 6948e95a9851SMike Kravetz { 69494ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 69504ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 69514ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 69524ddb4d91SMike Kravetz #endif 6953e95a9851SMike Kravetz return 0UL; 6954e95a9851SMike Kravetz } 6955e95a9851SMike Kravetz 695661f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 695761f77edaSNaoya Horiguchi 695861f77edaSNaoya Horiguchi /* 695961f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 696061f77edaSNaoya Horiguchi * behavior. 696161f77edaSNaoya Horiguchi */ 696261f77edaSNaoya Horiguchi struct page * __weak 696361f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 696461f77edaSNaoya Horiguchi int write) 696561f77edaSNaoya Horiguchi { 696661f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 696761f77edaSNaoya Horiguchi } 696861f77edaSNaoya Horiguchi 696961f77edaSNaoya Horiguchi struct page * __weak 69704dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 69714dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 69724dc71451SAneesh Kumar K.V { 69734dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 69744dc71451SAneesh Kumar K.V return NULL; 69754dc71451SAneesh Kumar K.V } 69764dc71451SAneesh Kumar K.V 69774dc71451SAneesh Kumar K.V struct page * __weak 69789e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6979e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 69809e5fc74cSSteve Capper { 6981e66f17ffSNaoya Horiguchi struct page *page = NULL; 6982e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6983c9d398faSNaoya Horiguchi pte_t pte; 69843faa52c0SJohn Hubbard 69858909691bSDavid Hildenbrand /* 69868909691bSDavid Hildenbrand * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 69878909691bSDavid Hildenbrand * follow_hugetlb_page(). 69888909691bSDavid Hildenbrand */ 69898909691bSDavid Hildenbrand if (WARN_ON_ONCE(flags & FOLL_PIN)) 69903faa52c0SJohn Hubbard return NULL; 69913faa52c0SJohn Hubbard 6992e66f17ffSNaoya Horiguchi retry: 6993e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 6994e66f17ffSNaoya Horiguchi spin_lock(ptl); 6995e66f17ffSNaoya Horiguchi /* 6996e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 6997e66f17ffSNaoya Horiguchi * unmapped from other threads. 6998e66f17ffSNaoya Horiguchi */ 6999e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 7000e66f17ffSNaoya Horiguchi goto out; 7001c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 7002c9d398faSNaoya Horiguchi if (pte_present(pte)) { 700397534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 70043faa52c0SJohn Hubbard /* 70053faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 70063faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 70073faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 70083faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 70093faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 70103faa52c0SJohn Hubbard * unless the page refcount overflowed: 70113faa52c0SJohn Hubbard */ 70123faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 70133faa52c0SJohn Hubbard page = NULL; 70143faa52c0SJohn Hubbard goto out; 70153faa52c0SJohn Hubbard } 7016e66f17ffSNaoya Horiguchi } else { 7017c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 7018e66f17ffSNaoya Horiguchi spin_unlock(ptl); 7019ad1ac596SMiaohe Lin __migration_entry_wait_huge((pte_t *)pmd, ptl); 7020e66f17ffSNaoya Horiguchi goto retry; 7021e66f17ffSNaoya Horiguchi } 7022e66f17ffSNaoya Horiguchi /* 7023e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 7024e66f17ffSNaoya Horiguchi * follow_page_mask(). 7025e66f17ffSNaoya Horiguchi */ 7026e66f17ffSNaoya Horiguchi } 7027e66f17ffSNaoya Horiguchi out: 7028e66f17ffSNaoya Horiguchi spin_unlock(ptl); 70299e5fc74cSSteve Capper return page; 70309e5fc74cSSteve Capper } 70319e5fc74cSSteve Capper 703261f77edaSNaoya Horiguchi struct page * __weak 70339e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 7034e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 70359e5fc74cSSteve Capper { 70363a194f3fSNaoya Horiguchi struct page *page = NULL; 70373a194f3fSNaoya Horiguchi spinlock_t *ptl; 70383a194f3fSNaoya Horiguchi pte_t pte; 70393a194f3fSNaoya Horiguchi 70403a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(flags & FOLL_PIN)) 7041e66f17ffSNaoya Horiguchi return NULL; 70429e5fc74cSSteve Capper 70433a194f3fSNaoya Horiguchi retry: 70443a194f3fSNaoya Horiguchi ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud); 70453a194f3fSNaoya Horiguchi if (!pud_huge(*pud)) 70463a194f3fSNaoya Horiguchi goto out; 70473a194f3fSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pud); 70483a194f3fSNaoya Horiguchi if (pte_present(pte)) { 70493a194f3fSNaoya Horiguchi page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 70503a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 70513a194f3fSNaoya Horiguchi page = NULL; 70523a194f3fSNaoya Horiguchi goto out; 70533a194f3fSNaoya Horiguchi } 70543a194f3fSNaoya Horiguchi } else { 70553a194f3fSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 70563a194f3fSNaoya Horiguchi spin_unlock(ptl); 70573a194f3fSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pud, ptl); 70583a194f3fSNaoya Horiguchi goto retry; 70593a194f3fSNaoya Horiguchi } 70603a194f3fSNaoya Horiguchi /* 70613a194f3fSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 70623a194f3fSNaoya Horiguchi * follow_page_mask(). 70633a194f3fSNaoya Horiguchi */ 70643a194f3fSNaoya Horiguchi } 70653a194f3fSNaoya Horiguchi out: 70663a194f3fSNaoya Horiguchi spin_unlock(ptl); 70673a194f3fSNaoya Horiguchi return page; 70689e5fc74cSSteve Capper } 70699e5fc74cSSteve Capper 7070faaa5b62SAnshuman Khandual struct page * __weak 7071faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 7072faaa5b62SAnshuman Khandual { 70733faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 7074faaa5b62SAnshuman Khandual return NULL; 7075faaa5b62SAnshuman Khandual 7076faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 7077faaa5b62SAnshuman Khandual } 7078faaa5b62SAnshuman Khandual 70797ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 708031caf665SNaoya Horiguchi { 70817ce82f4cSMiaohe Lin int ret = 0; 7082bcc54222SNaoya Horiguchi 7083db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 70848f251a3dSMike Kravetz if (!PageHeadHuge(page) || 70858f251a3dSMike Kravetz !HPageMigratable(page) || 70860eb2df2bSMuchun Song !get_page_unless_zero(page)) { 70877ce82f4cSMiaohe Lin ret = -EBUSY; 7088bcc54222SNaoya Horiguchi goto unlock; 7089bcc54222SNaoya Horiguchi } 70908f251a3dSMike Kravetz ClearHPageMigratable(page); 709131caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7092bcc54222SNaoya Horiguchi unlock: 7093db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7094bcc54222SNaoya Horiguchi return ret; 709531caf665SNaoya Horiguchi } 709631caf665SNaoya Horiguchi 709725182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 709825182f05SNaoya Horiguchi { 709925182f05SNaoya Horiguchi int ret = 0; 710025182f05SNaoya Horiguchi 710125182f05SNaoya Horiguchi *hugetlb = false; 710225182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 710325182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 710425182f05SNaoya Horiguchi *hugetlb = true; 7105b283d983SNaoya Horiguchi if (HPageFreed(page)) 7106b283d983SNaoya Horiguchi ret = 0; 7107b283d983SNaoya Horiguchi else if (HPageMigratable(page)) 710825182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 71090ed950d1SNaoya Horiguchi else 71100ed950d1SNaoya Horiguchi ret = -EBUSY; 711125182f05SNaoya Horiguchi } 711225182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 711325182f05SNaoya Horiguchi return ret; 711425182f05SNaoya Horiguchi } 711525182f05SNaoya Horiguchi 7116405ce051SNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags) 7117405ce051SNaoya Horiguchi { 7118405ce051SNaoya Horiguchi int ret; 7119405ce051SNaoya Horiguchi 7120405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7121405ce051SNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags); 7122405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7123405ce051SNaoya Horiguchi return ret; 7124405ce051SNaoya Horiguchi } 7125405ce051SNaoya Horiguchi 712631caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 712731caf665SNaoya Horiguchi { 7128db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 71298f251a3dSMike Kravetz SetHPageMigratable(page); 713031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7131db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 713231caf665SNaoya Horiguchi put_page(page); 713331caf665SNaoya Horiguchi } 7134ab5ac90aSMichal Hocko 7135ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 7136ab5ac90aSMichal Hocko { 7137ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 7138ab5ac90aSMichal Hocko 7139ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 7140ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 7141ab5ac90aSMichal Hocko 7142ab5ac90aSMichal Hocko /* 7143ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 7144ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7145ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7146ab5ac90aSMichal Hocko * the temporary status. 7147ab5ac90aSMichal Hocko * 7148ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7149ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7150ab5ac90aSMichal Hocko * the per-node's. 7151ab5ac90aSMichal Hocko */ 71529157c311SMike Kravetz if (HPageTemporary(newpage)) { 7153ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 7154ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 7155ab5ac90aSMichal Hocko 71569157c311SMike Kravetz SetHPageTemporary(oldpage); 71579157c311SMike Kravetz ClearHPageTemporary(newpage); 7158ab5ac90aSMichal Hocko 71595af1ab1dSMiaohe Lin /* 71605af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 71615af1ab1dSMiaohe Lin * when we do not cross the node. 71625af1ab1dSMiaohe Lin */ 71635af1ab1dSMiaohe Lin if (new_nid == old_nid) 71645af1ab1dSMiaohe Lin return; 7165db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7166ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7167ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7168ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7169ab5ac90aSMichal Hocko } 7170db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7171ab5ac90aSMichal Hocko } 7172ab5ac90aSMichal Hocko } 7173cf11e85fSRoman Gushchin 71746dfeaff9SPeter Xu /* 71756dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 71766dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 71776dfeaff9SPeter Xu */ 71786dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 71796dfeaff9SPeter Xu { 71806dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 71816dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 71826dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 71836dfeaff9SPeter Xu struct mmu_notifier_range range; 71846dfeaff9SPeter Xu unsigned long address, start, end; 71856dfeaff9SPeter Xu spinlock_t *ptl; 71866dfeaff9SPeter Xu pte_t *ptep; 71876dfeaff9SPeter Xu 71886dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 71896dfeaff9SPeter Xu return; 71906dfeaff9SPeter Xu 71916dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 71926dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 71936dfeaff9SPeter Xu 71946dfeaff9SPeter Xu if (start >= end) 71956dfeaff9SPeter Xu return; 71966dfeaff9SPeter Xu 71979c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 71986dfeaff9SPeter Xu /* 71996dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 72006dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 72016dfeaff9SPeter Xu */ 72026dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 72036dfeaff9SPeter Xu start, end); 72046dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 72056dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 72066dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 72076dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 72086dfeaff9SPeter Xu if (!ptep) 72096dfeaff9SPeter Xu continue; 72106dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 72114ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 72126dfeaff9SPeter Xu spin_unlock(ptl); 72136dfeaff9SPeter Xu } 72146dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 72156dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 72166dfeaff9SPeter Xu /* 72176dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7218ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 72196dfeaff9SPeter Xu */ 72206dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 72216dfeaff9SPeter Xu } 72226dfeaff9SPeter Xu 7223cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7224cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7225cf11e85fSRoman Gushchin 7226cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7227cf11e85fSRoman Gushchin { 722838e719abSBaolin Wang int nid, count = 0; 722938e719abSBaolin Wang unsigned long tmp; 723038e719abSBaolin Wang char *s = p; 723138e719abSBaolin Wang 723238e719abSBaolin Wang while (*s) { 723338e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 723438e719abSBaolin Wang break; 723538e719abSBaolin Wang 723638e719abSBaolin Wang if (s[count] == ':') { 7237f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 723838e719abSBaolin Wang break; 7239f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 724038e719abSBaolin Wang 724138e719abSBaolin Wang s += count + 1; 724238e719abSBaolin Wang tmp = memparse(s, &s); 724338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 724438e719abSBaolin Wang hugetlb_cma_size += tmp; 724538e719abSBaolin Wang 724638e719abSBaolin Wang /* 724738e719abSBaolin Wang * Skip the separator if have one, otherwise 724838e719abSBaolin Wang * break the parsing. 724938e719abSBaolin Wang */ 725038e719abSBaolin Wang if (*s == ',') 725138e719abSBaolin Wang s++; 725238e719abSBaolin Wang else 725338e719abSBaolin Wang break; 725438e719abSBaolin Wang } else { 7255cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 725638e719abSBaolin Wang break; 725738e719abSBaolin Wang } 725838e719abSBaolin Wang } 725938e719abSBaolin Wang 7260cf11e85fSRoman Gushchin return 0; 7261cf11e85fSRoman Gushchin } 7262cf11e85fSRoman Gushchin 7263cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7264cf11e85fSRoman Gushchin 7265cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7266cf11e85fSRoman Gushchin { 7267cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 726838e719abSBaolin Wang bool node_specific_cma_alloc = false; 7269cf11e85fSRoman Gushchin int nid; 7270cf11e85fSRoman Gushchin 7271cf11e85fSRoman Gushchin cma_reserve_called = true; 7272cf11e85fSRoman Gushchin 7273cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7274cf11e85fSRoman Gushchin return; 7275cf11e85fSRoman Gushchin 727638e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 727738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 727838e719abSBaolin Wang continue; 727938e719abSBaolin Wang 728030a51400SPeng Liu if (!node_online(nid)) { 728138e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 728238e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 728338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 728438e719abSBaolin Wang continue; 728538e719abSBaolin Wang } 728638e719abSBaolin Wang 728738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 728838e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 728938e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 729038e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 729138e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 729238e719abSBaolin Wang } else { 729338e719abSBaolin Wang node_specific_cma_alloc = true; 729438e719abSBaolin Wang } 729538e719abSBaolin Wang } 729638e719abSBaolin Wang 729738e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 729838e719abSBaolin Wang if (!hugetlb_cma_size) 729938e719abSBaolin Wang return; 730038e719abSBaolin Wang 7301cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7302cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7303cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7304a01f4390SMike Kravetz hugetlb_cma_size = 0; 7305cf11e85fSRoman Gushchin return; 7306cf11e85fSRoman Gushchin } 7307cf11e85fSRoman Gushchin 730838e719abSBaolin Wang if (!node_specific_cma_alloc) { 7309cf11e85fSRoman Gushchin /* 7310cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7311cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7312cf11e85fSRoman Gushchin */ 7313cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7314cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7315cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 731638e719abSBaolin Wang } 7317cf11e85fSRoman Gushchin 7318cf11e85fSRoman Gushchin reserved = 0; 731930a51400SPeng Liu for_each_online_node(nid) { 7320cf11e85fSRoman Gushchin int res; 73212281f797SBarry Song char name[CMA_MAX_NAME]; 7322cf11e85fSRoman Gushchin 732338e719abSBaolin Wang if (node_specific_cma_alloc) { 732438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 732538e719abSBaolin Wang continue; 732638e719abSBaolin Wang 732738e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 732838e719abSBaolin Wang } else { 7329cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 733038e719abSBaolin Wang } 733138e719abSBaolin Wang 7332cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7333cf11e85fSRoman Gushchin 73342281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7335a01f4390SMike Kravetz /* 7336a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7337a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7338a01f4390SMike Kravetz * huge page demotion. 7339a01f4390SMike Kravetz */ 7340a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7341a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 734229d0f41dSBarry Song 0, false, name, 7343cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7344cf11e85fSRoman Gushchin if (res) { 7345cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7346cf11e85fSRoman Gushchin res, nid); 7347cf11e85fSRoman Gushchin continue; 7348cf11e85fSRoman Gushchin } 7349cf11e85fSRoman Gushchin 7350cf11e85fSRoman Gushchin reserved += size; 7351cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7352cf11e85fSRoman Gushchin size / SZ_1M, nid); 7353cf11e85fSRoman Gushchin 7354cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7355cf11e85fSRoman Gushchin break; 7356cf11e85fSRoman Gushchin } 7357a01f4390SMike Kravetz 7358a01f4390SMike Kravetz if (!reserved) 7359a01f4390SMike Kravetz /* 7360a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7361a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7362a01f4390SMike Kravetz */ 7363a01f4390SMike Kravetz hugetlb_cma_size = 0; 7364cf11e85fSRoman Gushchin } 7365cf11e85fSRoman Gushchin 7366263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7367cf11e85fSRoman Gushchin { 7368cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7369cf11e85fSRoman Gushchin return; 7370cf11e85fSRoman Gushchin 7371cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7372cf11e85fSRoman Gushchin } 7373cf11e85fSRoman Gushchin 7374cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7375