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> 33d6606683SLinus Torvalds 3463551ae0SDavid Gibson #include <asm/page.h> 35ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3624669e58SAneesh Kumar K.V #include <asm/tlb.h> 3763551ae0SDavid Gibson 3824669e58SAneesh Kumar K.V #include <linux/io.h> 3963551ae0SDavid Gibson #include <linux/hugetlb.h> 409dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 419a305230SLee Schermerhorn #include <linux/node.h> 42ab5ac90aSMichal Hocko #include <linux/page_owner.h> 437835e98bSNick Piggin #include "internal.h" 441da177e4SLinus Torvalds 45c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 46e5ff2159SAndi Kleen unsigned int default_hstate_idx; 47e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 48cf11e85fSRoman Gushchin 49dbda8feaSBarry Song #ifdef CONFIG_CMA 50cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 51dbda8feaSBarry Song #endif 52dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 53cf11e85fSRoman Gushchin 54641844f5SNaoya Horiguchi /* 55641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 56641844f5SNaoya Horiguchi * at boot time. 57641844f5SNaoya Horiguchi */ 58641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 59e5ff2159SAndi Kleen 6053ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6153ba51d2SJon Tollefson 62e5ff2159SAndi Kleen /* for command line parsing */ 63e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 64e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 659fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 66282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 67e5ff2159SAndi Kleen 683935baa9SDavid Gibson /* 6931caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7031caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 713935baa9SDavid Gibson */ 72c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 730bd0f9fbSEric Paris 748382d914SDavidlohr Bueso /* 758382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 768382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 778382d914SDavidlohr Bueso */ 788382d914SDavidlohr Bueso static int num_fault_mutexes; 79c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 808382d914SDavidlohr Bueso 817ca02d0aSMike Kravetz /* Forward declaration */ 827ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 837ca02d0aSMike Kravetz 841d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 851d88433bSMiaohe Lin { 861d88433bSMiaohe Lin if (spool->count) 871d88433bSMiaohe Lin return false; 881d88433bSMiaohe Lin if (spool->max_hpages != -1) 891d88433bSMiaohe Lin return spool->used_hpages == 0; 901d88433bSMiaohe Lin if (spool->min_hpages != -1) 911d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 921d88433bSMiaohe Lin 931d88433bSMiaohe Lin return true; 941d88433bSMiaohe Lin } 951d88433bSMiaohe Lin 96db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 97db71ef79SMike Kravetz unsigned long irq_flags) 9890481622SDavid Gibson { 99db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 10090481622SDavid Gibson 10190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1027c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1037ca02d0aSMike Kravetz * free the subpool */ 1041d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1057ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1067ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1077ca02d0aSMike Kravetz -spool->min_hpages); 10890481622SDavid Gibson kfree(spool); 10990481622SDavid Gibson } 1107ca02d0aSMike Kravetz } 11190481622SDavid Gibson 1127ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1137ca02d0aSMike Kravetz long min_hpages) 11490481622SDavid Gibson { 11590481622SDavid Gibson struct hugepage_subpool *spool; 11690481622SDavid Gibson 117c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 11890481622SDavid Gibson if (!spool) 11990481622SDavid Gibson return NULL; 12090481622SDavid Gibson 12190481622SDavid Gibson spin_lock_init(&spool->lock); 12290481622SDavid Gibson spool->count = 1; 1237ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1247ca02d0aSMike Kravetz spool->hstate = h; 1257ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1267ca02d0aSMike Kravetz 1277ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1287ca02d0aSMike Kravetz kfree(spool); 1297ca02d0aSMike Kravetz return NULL; 1307ca02d0aSMike Kravetz } 1317ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 13290481622SDavid Gibson 13390481622SDavid Gibson return spool; 13490481622SDavid Gibson } 13590481622SDavid Gibson 13690481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 13790481622SDavid Gibson { 138db71ef79SMike Kravetz unsigned long flags; 139db71ef79SMike Kravetz 140db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 14190481622SDavid Gibson BUG_ON(!spool->count); 14290481622SDavid Gibson spool->count--; 143db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 14490481622SDavid Gibson } 14590481622SDavid Gibson 1461c5ecae3SMike Kravetz /* 1471c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1481c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1499e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1501c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1511c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1527c8de358SEthon Paul * a subpool minimum size must be maintained. 1531c5ecae3SMike Kravetz */ 1541c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 15590481622SDavid Gibson long delta) 15690481622SDavid Gibson { 1571c5ecae3SMike Kravetz long ret = delta; 15890481622SDavid Gibson 15990481622SDavid Gibson if (!spool) 1601c5ecae3SMike Kravetz return ret; 16190481622SDavid Gibson 162db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 16390481622SDavid Gibson 1641c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1651c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1661c5ecae3SMike Kravetz spool->used_hpages += delta; 1671c5ecae3SMike Kravetz else { 1681c5ecae3SMike Kravetz ret = -ENOMEM; 1691c5ecae3SMike Kravetz goto unlock_ret; 1701c5ecae3SMike Kravetz } 1711c5ecae3SMike Kravetz } 1721c5ecae3SMike Kravetz 17309a95e29SMike Kravetz /* minimum size accounting */ 17409a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1751c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1761c5ecae3SMike Kravetz /* 1771c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1781c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1791c5ecae3SMike Kravetz */ 1801c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1811c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1821c5ecae3SMike Kravetz } else { 1831c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1841c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1851c5ecae3SMike Kravetz } 1861c5ecae3SMike Kravetz } 1871c5ecae3SMike Kravetz 1881c5ecae3SMike Kravetz unlock_ret: 189db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 19090481622SDavid Gibson return ret; 19190481622SDavid Gibson } 19290481622SDavid Gibson 1931c5ecae3SMike Kravetz /* 1941c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1951c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1961c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1971c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1981c5ecae3SMike Kravetz */ 1991c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 20090481622SDavid Gibson long delta) 20190481622SDavid Gibson { 2021c5ecae3SMike Kravetz long ret = delta; 203db71ef79SMike Kravetz unsigned long flags; 2041c5ecae3SMike Kravetz 20590481622SDavid Gibson if (!spool) 2061c5ecae3SMike Kravetz return delta; 20790481622SDavid Gibson 208db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2091c5ecae3SMike Kravetz 2101c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 21190481622SDavid Gibson spool->used_hpages -= delta; 2121c5ecae3SMike Kravetz 21309a95e29SMike Kravetz /* minimum size accounting */ 21409a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2151c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2161c5ecae3SMike Kravetz ret = 0; 2171c5ecae3SMike Kravetz else 2181c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2191c5ecae3SMike Kravetz 2201c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2211c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2221c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2231c5ecae3SMike Kravetz } 2241c5ecae3SMike Kravetz 2251c5ecae3SMike Kravetz /* 2261c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2271c5ecae3SMike Kravetz * quota reference, free it now. 2281c5ecae3SMike Kravetz */ 229db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2301c5ecae3SMike Kravetz 2311c5ecae3SMike Kravetz return ret; 23290481622SDavid Gibson } 23390481622SDavid Gibson 23490481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 23590481622SDavid Gibson { 23690481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 23790481622SDavid Gibson } 23890481622SDavid Gibson 23990481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 24090481622SDavid Gibson { 241496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 24290481622SDavid Gibson } 24390481622SDavid Gibson 2440db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2450db9d74eSMina Almasry * it for use. 2460db9d74eSMina Almasry */ 2470db9d74eSMina Almasry static struct file_region * 2480db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2490db9d74eSMina Almasry { 2500db9d74eSMina Almasry struct file_region *nrg = NULL; 2510db9d74eSMina Almasry 2520db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2530db9d74eSMina Almasry 2540db9d74eSMina Almasry resv->region_cache_count--; 2550db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2560db9d74eSMina Almasry list_del(&nrg->link); 2570db9d74eSMina Almasry 2580db9d74eSMina Almasry nrg->from = from; 2590db9d74eSMina Almasry nrg->to = to; 2600db9d74eSMina Almasry 2610db9d74eSMina Almasry return nrg; 2620db9d74eSMina Almasry } 2630db9d74eSMina Almasry 264075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 265075a61d0SMina Almasry struct file_region *rg) 266075a61d0SMina Almasry { 267075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 268075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 269075a61d0SMina Almasry nrg->css = rg->css; 270075a61d0SMina Almasry if (rg->css) 271075a61d0SMina Almasry css_get(rg->css); 272075a61d0SMina Almasry #endif 273075a61d0SMina Almasry } 274075a61d0SMina Almasry 275075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 276075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 277075a61d0SMina Almasry struct hstate *h, 278075a61d0SMina Almasry struct resv_map *resv, 279075a61d0SMina Almasry struct file_region *nrg) 280075a61d0SMina Almasry { 281075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 282075a61d0SMina Almasry if (h_cg) { 283075a61d0SMina Almasry nrg->reservation_counter = 284075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 285075a61d0SMina Almasry nrg->css = &h_cg->css; 286d85aecf2SMiaohe Lin /* 287d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 288d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 289d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 290d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 291d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 292d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 293d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 294d85aecf2SMiaohe Lin * untouched. 295d85aecf2SMiaohe Lin */ 296d85aecf2SMiaohe Lin css_get(&h_cg->css); 297075a61d0SMina Almasry if (!resv->pages_per_hpage) 298075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 299075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 300075a61d0SMina Almasry * a resv_map. 301075a61d0SMina Almasry */ 302075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 303075a61d0SMina Almasry } else { 304075a61d0SMina Almasry nrg->reservation_counter = NULL; 305075a61d0SMina Almasry nrg->css = NULL; 306075a61d0SMina Almasry } 307075a61d0SMina Almasry #endif 308075a61d0SMina Almasry } 309075a61d0SMina Almasry 310d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 311d85aecf2SMiaohe Lin { 312d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 313d85aecf2SMiaohe Lin if (rg->css) 314d85aecf2SMiaohe Lin css_put(rg->css); 315d85aecf2SMiaohe Lin #endif 316d85aecf2SMiaohe Lin } 317d85aecf2SMiaohe Lin 318a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 319a9b3f867SMina Almasry struct file_region *org) 320a9b3f867SMina Almasry { 321a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 322a9b3f867SMina Almasry return rg && org && 323a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 324a9b3f867SMina Almasry rg->css == org->css; 325a9b3f867SMina Almasry 326a9b3f867SMina Almasry #else 327a9b3f867SMina Almasry return true; 328a9b3f867SMina Almasry #endif 329a9b3f867SMina Almasry } 330a9b3f867SMina Almasry 331a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 332a9b3f867SMina Almasry { 333a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 334a9b3f867SMina Almasry 335a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 336a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 337a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 338a9b3f867SMina Almasry prg->to = rg->to; 339a9b3f867SMina Almasry 340a9b3f867SMina Almasry list_del(&rg->link); 341d85aecf2SMiaohe Lin put_uncharge_info(rg); 342a9b3f867SMina Almasry kfree(rg); 343a9b3f867SMina Almasry 3447db5e7b6SWei Yang rg = prg; 345a9b3f867SMina Almasry } 346a9b3f867SMina Almasry 347a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 348a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 349a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 350a9b3f867SMina Almasry nrg->from = rg->from; 351a9b3f867SMina Almasry 352a9b3f867SMina Almasry list_del(&rg->link); 353d85aecf2SMiaohe Lin put_uncharge_info(rg); 354a9b3f867SMina Almasry kfree(rg); 355a9b3f867SMina Almasry } 356a9b3f867SMina Almasry } 357a9b3f867SMina Almasry 3582103cf9cSPeter Xu static inline long 3592103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3602103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3612103cf9cSPeter Xu long *regions_needed) 3622103cf9cSPeter Xu { 3632103cf9cSPeter Xu struct file_region *nrg; 3642103cf9cSPeter Xu 3652103cf9cSPeter Xu if (!regions_needed) { 3662103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3672103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3682103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3692103cf9cSPeter Xu coalesce_file_region(map, nrg); 3702103cf9cSPeter Xu } else 3712103cf9cSPeter Xu *regions_needed += 1; 3722103cf9cSPeter Xu 3732103cf9cSPeter Xu return to - from; 3742103cf9cSPeter Xu } 3752103cf9cSPeter Xu 376972a3da3SWei Yang /* 377972a3da3SWei Yang * Must be called with resv->lock held. 378972a3da3SWei Yang * 379972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 380972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 381972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 382972a3da3SWei Yang * the regions for this range. 383d75c6af9SMina Almasry */ 384d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 385075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 386972a3da3SWei Yang struct hstate *h, long *regions_needed) 387d75c6af9SMina Almasry { 3880db9d74eSMina Almasry long add = 0; 389d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3900db9d74eSMina Almasry long last_accounted_offset = f; 3912103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 392d75c6af9SMina Almasry 3930db9d74eSMina Almasry if (regions_needed) 3940db9d74eSMina Almasry *regions_needed = 0; 395d75c6af9SMina Almasry 3960db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3970db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3980db9d74eSMina Almasry * bounds checking. 3990db9d74eSMina Almasry */ 4000db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4010db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4020db9d74eSMina Almasry if (rg->from < f) { 4030db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4040db9d74eSMina Almasry * then we need to update last_accounted_offset. 4050db9d74eSMina Almasry */ 4060db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4070db9d74eSMina Almasry last_accounted_offset = rg->to; 4080db9d74eSMina Almasry continue; 4090db9d74eSMina Almasry } 410d75c6af9SMina Almasry 4110db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4120db9d74eSMina Almasry * finished. 4130db9d74eSMina Almasry */ 414ca7e0457SPeter Xu if (rg->from >= t) 415d75c6af9SMina Almasry break; 416d75c6af9SMina Almasry 4170db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4180db9d74eSMina Almasry * update last_accounted_offset. 419d75c6af9SMina Almasry */ 4202103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4212103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4222103cf9cSPeter Xu last_accounted_offset, 4232103cf9cSPeter Xu rg->from, h, h_cg, 4242103cf9cSPeter Xu regions_needed); 425d75c6af9SMina Almasry 4260db9d74eSMina Almasry last_accounted_offset = rg->to; 4270db9d74eSMina Almasry } 4280db9d74eSMina Almasry 4290db9d74eSMina Almasry /* Handle the case where our range extends beyond 4300db9d74eSMina Almasry * last_accounted_offset. 4310db9d74eSMina Almasry */ 4322103cf9cSPeter Xu if (last_accounted_offset < t) 4332103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4342103cf9cSPeter Xu t, h, h_cg, regions_needed); 4350db9d74eSMina Almasry 4360db9d74eSMina Almasry VM_BUG_ON(add < 0); 4370db9d74eSMina Almasry return add; 4380db9d74eSMina Almasry } 4390db9d74eSMina Almasry 4400db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4410db9d74eSMina Almasry */ 4420db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4430db9d74eSMina Almasry int regions_needed) 4440db9d74eSMina Almasry __must_hold(&resv->lock) 4450db9d74eSMina Almasry { 4460db9d74eSMina Almasry struct list_head allocated_regions; 4470db9d74eSMina Almasry int to_allocate = 0, i = 0; 4480db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4490db9d74eSMina Almasry 4500db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4510db9d74eSMina Almasry 4520db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4530db9d74eSMina Almasry 4540db9d74eSMina Almasry /* 4550db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4560db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4570db9d74eSMina Almasry * 4580db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4590db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4600db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4610db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4620db9d74eSMina Almasry */ 4630db9d74eSMina Almasry while (resv->region_cache_count < 4640db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4650db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4660db9d74eSMina Almasry resv->region_cache_count; 4670db9d74eSMina Almasry 4680db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4690db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4700db9d74eSMina Almasry * needing to allocate for regions_needed. 4710db9d74eSMina Almasry */ 4720db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4730db9d74eSMina Almasry 4740db9d74eSMina Almasry spin_unlock(&resv->lock); 4750db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4760db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4770db9d74eSMina Almasry if (!trg) 4780db9d74eSMina Almasry goto out_of_memory; 4790db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4800db9d74eSMina Almasry } 4810db9d74eSMina Almasry 4820db9d74eSMina Almasry spin_lock(&resv->lock); 4830db9d74eSMina Almasry 484d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 485d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4860db9d74eSMina Almasry } 4870db9d74eSMina Almasry 4880db9d74eSMina Almasry return 0; 4890db9d74eSMina Almasry 4900db9d74eSMina Almasry out_of_memory: 4910db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 492d75c6af9SMina Almasry list_del(&rg->link); 493d75c6af9SMina Almasry kfree(rg); 494d75c6af9SMina Almasry } 4950db9d74eSMina Almasry return -ENOMEM; 496d75c6af9SMina Almasry } 497d75c6af9SMina Almasry 4981dd308a7SMike Kravetz /* 4991dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5000db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5010db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5020db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5030db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5040db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 505cf3ad20bSMike Kravetz * 5060db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5070db9d74eSMina Almasry * 5080db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5090db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5107c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5110db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5120db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5130db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5141dd308a7SMike Kravetz */ 5150db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 516075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 517075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 51896822904SAndy Whitcroft { 5190db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 52096822904SAndy Whitcroft 5217b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5220db9d74eSMina Almasry retry: 5230db9d74eSMina Almasry 5240db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 525972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 526972a3da3SWei Yang &actual_regions_needed); 52796822904SAndy Whitcroft 5285e911373SMike Kravetz /* 5290db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5300db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5310db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5320db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5330db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5340db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5350db9d74eSMina Almasry * operation. 5365e911373SMike Kravetz */ 5370db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5380db9d74eSMina Almasry resv->region_cache_count < 5390db9d74eSMina Almasry resv->adds_in_progress + 5400db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5410db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5420db9d74eSMina Almasry * allocate file_region entries. 5430db9d74eSMina Almasry */ 5440db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5455e911373SMike Kravetz 5460db9d74eSMina Almasry if (allocate_file_region_entries( 5470db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5480db9d74eSMina Almasry return -ENOMEM; 5495e911373SMike Kravetz } 5505e911373SMike Kravetz 5510db9d74eSMina Almasry goto retry; 5520db9d74eSMina Almasry } 553cf3ad20bSMike Kravetz 554972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5550db9d74eSMina Almasry 5560db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5570db9d74eSMina Almasry 5587b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 559cf3ad20bSMike Kravetz return add; 56096822904SAndy Whitcroft } 56196822904SAndy Whitcroft 5621dd308a7SMike Kravetz /* 5631dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5641dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5651dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5661dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5671dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5681dd308a7SMike Kravetz * not change the number of huge pages represented by the 5690db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5700db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5710db9d74eSMina Almasry * file_region structure is added. 5720db9d74eSMina Almasry * 5730db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5740db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5750db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5765e911373SMike Kravetz * 5775e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5785e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5795e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5805e911373SMike Kravetz * is needed and can not be allocated. 5811dd308a7SMike Kravetz */ 5820db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5830db9d74eSMina Almasry long *out_regions_needed) 58496822904SAndy Whitcroft { 58596822904SAndy Whitcroft long chg = 0; 58696822904SAndy Whitcroft 5877b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5885e911373SMike Kravetz 589972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 590075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 591972a3da3SWei Yang out_regions_needed); 5925e911373SMike Kravetz 5930db9d74eSMina Almasry if (*out_regions_needed == 0) 5940db9d74eSMina Almasry *out_regions_needed = 1; 5955e911373SMike Kravetz 5960db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5975e911373SMike Kravetz return -ENOMEM; 5985e911373SMike Kravetz 5990db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 60096822904SAndy Whitcroft 6017b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 60296822904SAndy Whitcroft return chg; 60396822904SAndy Whitcroft } 60496822904SAndy Whitcroft 6051dd308a7SMike Kravetz /* 6065e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6075e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6085e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6095e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6100db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6110db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6120db9d74eSMina Almasry * the adds_in_progress counter. 6135e911373SMike Kravetz * 6145e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6155e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6165e911373SMike Kravetz * arguments will match the associated region_chg call. 6175e911373SMike Kravetz */ 6180db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6190db9d74eSMina Almasry long regions_needed) 6205e911373SMike Kravetz { 6215e911373SMike Kravetz spin_lock(&resv->lock); 6225e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6230db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6245e911373SMike Kravetz spin_unlock(&resv->lock); 6255e911373SMike Kravetz } 6265e911373SMike Kravetz 6275e911373SMike Kravetz /* 628feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 629feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 630feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 631feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 632feba16e2SMike Kravetz * 633feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 634feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 635feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 636feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 637feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 638feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 639feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6401dd308a7SMike Kravetz */ 641feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 64296822904SAndy Whitcroft { 6431406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 64496822904SAndy Whitcroft struct file_region *rg, *trg; 645feba16e2SMike Kravetz struct file_region *nrg = NULL; 646feba16e2SMike Kravetz long del = 0; 64796822904SAndy Whitcroft 648feba16e2SMike Kravetz retry: 6497b24d861SDavidlohr Bueso spin_lock(&resv->lock); 650feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 651dbe409e4SMike Kravetz /* 652dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 653dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 654dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 655dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 656dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 657dbe409e4SMike Kravetz */ 658dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 659feba16e2SMike Kravetz continue; 660dbe409e4SMike Kravetz 661feba16e2SMike Kravetz if (rg->from >= t) 66296822904SAndy Whitcroft break; 66396822904SAndy Whitcroft 664feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 665feba16e2SMike Kravetz /* 666feba16e2SMike Kravetz * Check for an entry in the cache before dropping 667feba16e2SMike Kravetz * lock and attempting allocation. 668feba16e2SMike Kravetz */ 669feba16e2SMike Kravetz if (!nrg && 670feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 671feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 672feba16e2SMike Kravetz struct file_region, 673feba16e2SMike Kravetz link); 674feba16e2SMike Kravetz list_del(&nrg->link); 675feba16e2SMike Kravetz resv->region_cache_count--; 67696822904SAndy Whitcroft } 67796822904SAndy Whitcroft 678feba16e2SMike Kravetz if (!nrg) { 679feba16e2SMike Kravetz spin_unlock(&resv->lock); 680feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 681feba16e2SMike Kravetz if (!nrg) 682feba16e2SMike Kravetz return -ENOMEM; 683feba16e2SMike Kravetz goto retry; 684feba16e2SMike Kravetz } 685feba16e2SMike Kravetz 686feba16e2SMike Kravetz del += t - f; 68779aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 688d85aecf2SMiaohe Lin resv, rg, t - f, false); 689feba16e2SMike Kravetz 690feba16e2SMike Kravetz /* New entry for end of split region */ 691feba16e2SMike Kravetz nrg->from = t; 692feba16e2SMike Kravetz nrg->to = rg->to; 693075a61d0SMina Almasry 694075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 695075a61d0SMina Almasry 696feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 697feba16e2SMike Kravetz 698feba16e2SMike Kravetz /* Original entry is trimmed */ 699feba16e2SMike Kravetz rg->to = f; 700feba16e2SMike Kravetz 701feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 702feba16e2SMike Kravetz nrg = NULL; 70396822904SAndy Whitcroft break; 704feba16e2SMike Kravetz } 705feba16e2SMike Kravetz 706feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 707feba16e2SMike Kravetz del += rg->to - rg->from; 708075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 709d85aecf2SMiaohe Lin rg->to - rg->from, true); 71096822904SAndy Whitcroft list_del(&rg->link); 71196822904SAndy Whitcroft kfree(rg); 712feba16e2SMike Kravetz continue; 71396822904SAndy Whitcroft } 7147b24d861SDavidlohr Bueso 715feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 716075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 717d85aecf2SMiaohe Lin t - rg->from, false); 718075a61d0SMina Almasry 71979aa925bSMike Kravetz del += t - rg->from; 72079aa925bSMike Kravetz rg->from = t; 72179aa925bSMike Kravetz } else { /* Trim end of region */ 722075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 723d85aecf2SMiaohe Lin rg->to - f, false); 72479aa925bSMike Kravetz 72579aa925bSMike Kravetz del += rg->to - f; 72679aa925bSMike Kravetz rg->to = f; 727feba16e2SMike Kravetz } 728feba16e2SMike Kravetz } 729feba16e2SMike Kravetz 7307b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 731feba16e2SMike Kravetz kfree(nrg); 732feba16e2SMike Kravetz return del; 73396822904SAndy Whitcroft } 73496822904SAndy Whitcroft 7351dd308a7SMike Kravetz /* 736b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 737b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 738b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 739b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 740b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 741b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 742b5cec28dSMike Kravetz * counts. 743b5cec28dSMike Kravetz */ 74472e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 745b5cec28dSMike Kravetz { 746b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 747b5cec28dSMike Kravetz long rsv_adjust; 748da56388cSMiaohe Lin bool reserved = false; 749b5cec28dSMike Kravetz 750b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 751da56388cSMiaohe Lin if (rsv_adjust > 0) { 752b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 753b5cec28dSMike Kravetz 754da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 755da56388cSMiaohe Lin reserved = true; 756da56388cSMiaohe Lin } else if (!rsv_adjust) { 757da56388cSMiaohe Lin reserved = true; 758b5cec28dSMike Kravetz } 759da56388cSMiaohe Lin 760da56388cSMiaohe Lin if (!reserved) 761da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 762b5cec28dSMike Kravetz } 763b5cec28dSMike Kravetz 764b5cec28dSMike Kravetz /* 7651dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7661dd308a7SMike Kravetz * that intersect with the range [f, t). 7671dd308a7SMike Kravetz */ 7681406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 76984afd99bSAndy Whitcroft { 7701406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 77184afd99bSAndy Whitcroft struct file_region *rg; 77284afd99bSAndy Whitcroft long chg = 0; 77384afd99bSAndy Whitcroft 7747b24d861SDavidlohr Bueso spin_lock(&resv->lock); 77584afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 77684afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 777f2135a4aSWang Sheng-Hui long seg_from; 778f2135a4aSWang Sheng-Hui long seg_to; 77984afd99bSAndy Whitcroft 78084afd99bSAndy Whitcroft if (rg->to <= f) 78184afd99bSAndy Whitcroft continue; 78284afd99bSAndy Whitcroft if (rg->from >= t) 78384afd99bSAndy Whitcroft break; 78484afd99bSAndy Whitcroft 78584afd99bSAndy Whitcroft seg_from = max(rg->from, f); 78684afd99bSAndy Whitcroft seg_to = min(rg->to, t); 78784afd99bSAndy Whitcroft 78884afd99bSAndy Whitcroft chg += seg_to - seg_from; 78984afd99bSAndy Whitcroft } 7907b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 79184afd99bSAndy Whitcroft 79284afd99bSAndy Whitcroft return chg; 79384afd99bSAndy Whitcroft } 79484afd99bSAndy Whitcroft 79596822904SAndy Whitcroft /* 796e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 797e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 798e7c4b0bfSAndy Whitcroft */ 799a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 800a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 801e7c4b0bfSAndy Whitcroft { 802a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 803a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 804e7c4b0bfSAndy Whitcroft } 805e7c4b0bfSAndy Whitcroft 8060fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8070fe6e20bSNaoya Horiguchi unsigned long address) 8080fe6e20bSNaoya Horiguchi { 8090fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8100fe6e20bSNaoya Horiguchi } 811dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8120fe6e20bSNaoya Horiguchi 81384afd99bSAndy Whitcroft /* 81408fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 81508fba699SMel Gorman * cases this will be same size as used by the page table entries. 81608fba699SMel Gorman */ 81708fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 81808fba699SMel Gorman { 81905ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 82005ea8860SDan Williams return vma->vm_ops->pagesize(vma); 82108fba699SMel Gorman return PAGE_SIZE; 82208fba699SMel Gorman } 823f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 82408fba699SMel Gorman 82508fba699SMel Gorman /* 8263340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8273340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 82809135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 82909135cc5SDan Williams * version of this symbol is required. 8303340289dSMel Gorman */ 83109135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8323340289dSMel Gorman { 8333340289dSMel Gorman return vma_kernel_pagesize(vma); 8343340289dSMel Gorman } 8353340289dSMel Gorman 8363340289dSMel Gorman /* 83784afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 83884afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 83984afd99bSAndy Whitcroft * alignment. 84084afd99bSAndy Whitcroft */ 84184afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 84284afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 84304f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 84484afd99bSAndy Whitcroft 845a1e78772SMel Gorman /* 846a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 847a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 848a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 849a1e78772SMel Gorman * 850a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 851a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 852a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 853a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 85484afd99bSAndy Whitcroft * 85584afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 85684afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 85784afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 85884afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 85984afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 86084afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 86184afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 86284afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 863a1e78772SMel Gorman */ 864e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 865e7c4b0bfSAndy Whitcroft { 866e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 867e7c4b0bfSAndy Whitcroft } 868e7c4b0bfSAndy Whitcroft 869e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 870e7c4b0bfSAndy Whitcroft unsigned long value) 871e7c4b0bfSAndy Whitcroft { 872e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 873e7c4b0bfSAndy Whitcroft } 874e7c4b0bfSAndy Whitcroft 875e9fe92aeSMina Almasry static void 876e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 877e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 878e9fe92aeSMina Almasry struct hstate *h) 879e9fe92aeSMina Almasry { 880e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 881e9fe92aeSMina Almasry if (!h_cg || !h) { 882e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 883e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 884e9fe92aeSMina Almasry resv_map->css = NULL; 885e9fe92aeSMina Almasry } else { 886e9fe92aeSMina Almasry resv_map->reservation_counter = 887e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 888e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 889e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 890e9fe92aeSMina Almasry } 891e9fe92aeSMina Almasry #endif 892e9fe92aeSMina Almasry } 893e9fe92aeSMina Almasry 8949119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 89584afd99bSAndy Whitcroft { 89684afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8975e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8985e911373SMike Kravetz 8995e911373SMike Kravetz if (!resv_map || !rg) { 9005e911373SMike Kravetz kfree(resv_map); 9015e911373SMike Kravetz kfree(rg); 90284afd99bSAndy Whitcroft return NULL; 9035e911373SMike Kravetz } 90484afd99bSAndy Whitcroft 90584afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9067b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 90784afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 90884afd99bSAndy Whitcroft 9095e911373SMike Kravetz resv_map->adds_in_progress = 0; 910e9fe92aeSMina Almasry /* 911e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 912e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 913e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 914e9fe92aeSMina Almasry * reservations are to be un-charged from here. 915e9fe92aeSMina Almasry */ 916e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9175e911373SMike Kravetz 9185e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9195e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9205e911373SMike Kravetz resv_map->region_cache_count = 1; 9215e911373SMike Kravetz 92284afd99bSAndy Whitcroft return resv_map; 92384afd99bSAndy Whitcroft } 92484afd99bSAndy Whitcroft 9259119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 92684afd99bSAndy Whitcroft { 92784afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9285e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9295e911373SMike Kravetz struct file_region *rg, *trg; 93084afd99bSAndy Whitcroft 93184afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 932feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9335e911373SMike Kravetz 9345e911373SMike Kravetz /* ... and any entries left in the cache */ 9355e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9365e911373SMike Kravetz list_del(&rg->link); 9375e911373SMike Kravetz kfree(rg); 9385e911373SMike Kravetz } 9395e911373SMike Kravetz 9405e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9415e911373SMike Kravetz 94284afd99bSAndy Whitcroft kfree(resv_map); 94384afd99bSAndy Whitcroft } 94484afd99bSAndy Whitcroft 9454e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9464e35f483SJoonsoo Kim { 947f27a5136SMike Kravetz /* 948f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 949f27a5136SMike Kravetz * address space within the inode. This original address space 950f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 951f27a5136SMike Kravetz * address space embedded within the inode. 952f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 953f27a5136SMike Kravetz * this may not be true for device special inodes. 954f27a5136SMike Kravetz */ 955f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9564e35f483SJoonsoo Kim } 9574e35f483SJoonsoo Kim 95884afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 959a1e78772SMel Gorman { 96081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9614e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9624e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9634e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9644e35f483SJoonsoo Kim 9654e35f483SJoonsoo Kim return inode_resv_map(inode); 9664e35f483SJoonsoo Kim 9674e35f483SJoonsoo Kim } else { 96884afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 96984afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9704e35f483SJoonsoo Kim } 971a1e78772SMel Gorman } 972a1e78772SMel Gorman 97384afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 974a1e78772SMel Gorman { 97581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 97681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 977a1e78772SMel Gorman 97884afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 97984afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 98004f2cbe3SMel Gorman } 98104f2cbe3SMel Gorman 98204f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 98304f2cbe3SMel Gorman { 98481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98581d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 986e7c4b0bfSAndy Whitcroft 987e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 98804f2cbe3SMel Gorman } 98904f2cbe3SMel Gorman 99004f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 99104f2cbe3SMel Gorman { 99281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 993e7c4b0bfSAndy Whitcroft 994e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 995a1e78772SMel Gorman } 996a1e78772SMel Gorman 99704f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 998a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 999a1e78772SMel Gorman { 100081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1001f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1002a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1003a1e78772SMel Gorman } 1004a1e78772SMel Gorman 1005a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1006559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1007a1e78772SMel Gorman { 1008af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1009af0ed73eSJoonsoo Kim /* 1010af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1011af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1012af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1013af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1014af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1015af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1016af0ed73eSJoonsoo Kim * properly, so add work-around here. 1017af0ed73eSJoonsoo Kim */ 1018af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1019559ec2f8SNicholas Krause return true; 1020af0ed73eSJoonsoo Kim else 1021559ec2f8SNicholas Krause return false; 1022af0ed73eSJoonsoo Kim } 1023a63884e9SJoonsoo Kim 1024a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10251fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10261fb1b0e9SMike Kravetz /* 10271fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10281fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10291fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10307c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10311fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10321fb1b0e9SMike Kravetz */ 10331fb1b0e9SMike Kravetz if (chg) 10341fb1b0e9SMike Kravetz return false; 10351fb1b0e9SMike Kravetz else 1036559ec2f8SNicholas Krause return true; 10371fb1b0e9SMike Kravetz } 1038a63884e9SJoonsoo Kim 1039a63884e9SJoonsoo Kim /* 1040a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1041a63884e9SJoonsoo Kim * private mappings. 1042a63884e9SJoonsoo Kim */ 104367961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 104467961f9dSMike Kravetz /* 104567961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 104667961f9dSMike Kravetz * could have been performed on the private mapping. 104767961f9dSMike Kravetz * Examine the value of chg to determine if reserves 104867961f9dSMike Kravetz * actually exist or were previously consumed. 104967961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 105067961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 105167961f9dSMike Kravetz * private mappings has different (opposite) semantics 105267961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 105367961f9dSMike Kravetz * has already taken this difference in semantics into 105467961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 105567961f9dSMike Kravetz * as in the shared case above. Code could easily be 105667961f9dSMike Kravetz * combined, but keeping it separate draws attention to 105767961f9dSMike Kravetz * subtle differences. 105867961f9dSMike Kravetz */ 105967961f9dSMike Kravetz if (chg) 106067961f9dSMike Kravetz return false; 106167961f9dSMike Kravetz else 1062559ec2f8SNicholas Krause return true; 106367961f9dSMike Kravetz } 1064a63884e9SJoonsoo Kim 1065559ec2f8SNicholas Krause return false; 1066a1e78772SMel Gorman } 1067a1e78772SMel Gorman 1068a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10691da177e4SLinus Torvalds { 10701da177e4SLinus Torvalds int nid = page_to_nid(page); 10719487ca60SMike Kravetz 10729487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 10730edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1074a5516438SAndi Kleen h->free_huge_pages++; 1075a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10766c037149SMike Kravetz SetHPageFreed(page); 10771da177e4SLinus Torvalds } 10781da177e4SLinus Torvalds 107994310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1080bf50bab2SNaoya Horiguchi { 1081bf50bab2SNaoya Horiguchi struct page *page; 1082bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1083bf50bab2SNaoya Horiguchi 10849487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1085bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1086bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1087bbe88753SJoonsoo Kim continue; 1088bbe88753SJoonsoo Kim 10896664bfc8SWei Yang if (PageHWPoison(page)) 10906664bfc8SWei Yang continue; 1091bbe88753SJoonsoo Kim 10920edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1093a9869b83SNaoya Horiguchi set_page_refcounted(page); 10946c037149SMike Kravetz ClearHPageFreed(page); 1095bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1096bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1097bf50bab2SNaoya Horiguchi return page; 1098bf50bab2SNaoya Horiguchi } 1099bf50bab2SNaoya Horiguchi 11006664bfc8SWei Yang return NULL; 11016664bfc8SWei Yang } 11026664bfc8SWei Yang 11033e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11043e59fcb0SMichal Hocko nodemask_t *nmask) 110594310cbcSAnshuman Khandual { 11063e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11073e59fcb0SMichal Hocko struct zonelist *zonelist; 11083e59fcb0SMichal Hocko struct zone *zone; 11093e59fcb0SMichal Hocko struct zoneref *z; 111098fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11113e59fcb0SMichal Hocko 11123e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11133e59fcb0SMichal Hocko 11143e59fcb0SMichal Hocko retry_cpuset: 11153e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11163e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 111794310cbcSAnshuman Khandual struct page *page; 111894310cbcSAnshuman Khandual 11193e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11203e59fcb0SMichal Hocko continue; 11213e59fcb0SMichal Hocko /* 11223e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11233e59fcb0SMichal Hocko * zone aware 11243e59fcb0SMichal Hocko */ 11253e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11263e59fcb0SMichal Hocko continue; 11273e59fcb0SMichal Hocko node = zone_to_nid(zone); 112894310cbcSAnshuman Khandual 112994310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 113094310cbcSAnshuman Khandual if (page) 113194310cbcSAnshuman Khandual return page; 113294310cbcSAnshuman Khandual } 11333e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11343e59fcb0SMichal Hocko goto retry_cpuset; 11353e59fcb0SMichal Hocko 113694310cbcSAnshuman Khandual return NULL; 113794310cbcSAnshuman Khandual } 113894310cbcSAnshuman Khandual 1139a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1140a5516438SAndi Kleen struct vm_area_struct *vma, 1141af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1142af0ed73eSJoonsoo Kim long chg) 11431da177e4SLinus Torvalds { 11443e59fcb0SMichal Hocko struct page *page; 1145480eccf9SLee Schermerhorn struct mempolicy *mpol; 114604ec6264SVlastimil Babka gfp_t gfp_mask; 11473e59fcb0SMichal Hocko nodemask_t *nodemask; 114804ec6264SVlastimil Babka int nid; 11491da177e4SLinus Torvalds 1150a1e78772SMel Gorman /* 1151a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1152a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1153a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1154a1e78772SMel Gorman */ 1155af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1156a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1157c0ff7453SMiao Xie goto err; 1158a1e78772SMel Gorman 115904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1160a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11616eab04a8SJustin P. Mattock goto err; 116204f2cbe3SMel Gorman 116304ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 116404ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11653e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11663e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1167d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1168a63884e9SJoonsoo Kim h->resv_huge_pages--; 1169bf50bab2SNaoya Horiguchi } 1170cc9a6c87SMel Gorman 1171cc9a6c87SMel Gorman mpol_cond_put(mpol); 1172cc9a6c87SMel Gorman return page; 1173cc9a6c87SMel Gorman 1174c0ff7453SMiao Xie err: 1175cc9a6c87SMel Gorman return NULL; 11761da177e4SLinus Torvalds } 11771da177e4SLinus Torvalds 11781cac6f2cSLuiz Capitulino /* 11791cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11801cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11811cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11821cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11831cac6f2cSLuiz Capitulino * node for alloc or free. 11841cac6f2cSLuiz Capitulino */ 11851cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11861cac6f2cSLuiz Capitulino { 11870edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11881cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11891cac6f2cSLuiz Capitulino 11901cac6f2cSLuiz Capitulino return nid; 11911cac6f2cSLuiz Capitulino } 11921cac6f2cSLuiz Capitulino 11931cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11941cac6f2cSLuiz Capitulino { 11951cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11961cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11971cac6f2cSLuiz Capitulino return nid; 11981cac6f2cSLuiz Capitulino } 11991cac6f2cSLuiz Capitulino 12001cac6f2cSLuiz Capitulino /* 12011cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12021cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12031cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12041cac6f2cSLuiz Capitulino * mask. 12051cac6f2cSLuiz Capitulino */ 12061cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12071cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12081cac6f2cSLuiz Capitulino { 12091cac6f2cSLuiz Capitulino int nid; 12101cac6f2cSLuiz Capitulino 12111cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12121cac6f2cSLuiz Capitulino 12131cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12141cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12151cac6f2cSLuiz Capitulino 12161cac6f2cSLuiz Capitulino return nid; 12171cac6f2cSLuiz Capitulino } 12181cac6f2cSLuiz Capitulino 12191cac6f2cSLuiz Capitulino /* 122010c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12211cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12221cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12231cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12241cac6f2cSLuiz Capitulino */ 12251cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12261cac6f2cSLuiz Capitulino { 12271cac6f2cSLuiz Capitulino int nid; 12281cac6f2cSLuiz Capitulino 12291cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12301cac6f2cSLuiz Capitulino 12311cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12321cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12331cac6f2cSLuiz Capitulino 12341cac6f2cSLuiz Capitulino return nid; 12351cac6f2cSLuiz Capitulino } 12361cac6f2cSLuiz Capitulino 12371cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12381cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12391cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12401cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12411cac6f2cSLuiz Capitulino nr_nodes--) 12421cac6f2cSLuiz Capitulino 12431cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12441cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12451cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12461cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12471cac6f2cSLuiz Capitulino nr_nodes--) 12481cac6f2cSLuiz Capitulino 1249e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1250944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1251d00181b9SKirill A. Shutemov unsigned int order) 1252944d9fecSLuiz Capitulino { 1253944d9fecSLuiz Capitulino int i; 1254944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1255944d9fecSLuiz Capitulino struct page *p = page + 1; 1256944d9fecSLuiz Capitulino 1257c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 125847e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 125947e29d32SJohn Hubbard 1260944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12611d798ca3SKirill A. Shutemov clear_compound_head(p); 1262944d9fecSLuiz Capitulino set_page_refcounted(p); 1263944d9fecSLuiz Capitulino } 1264944d9fecSLuiz Capitulino 1265944d9fecSLuiz Capitulino set_compound_order(page, 0); 1266ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1267944d9fecSLuiz Capitulino __ClearPageHead(page); 1268944d9fecSLuiz Capitulino } 1269944d9fecSLuiz Capitulino 1270d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1271944d9fecSLuiz Capitulino { 1272cf11e85fSRoman Gushchin /* 1273cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1274cf11e85fSRoman Gushchin * cma_release() returns false. 1275cf11e85fSRoman Gushchin */ 1276dbda8feaSBarry Song #ifdef CONFIG_CMA 1277dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1278cf11e85fSRoman Gushchin return; 1279dbda8feaSBarry Song #endif 1280cf11e85fSRoman Gushchin 1281944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1282944d9fecSLuiz Capitulino } 1283944d9fecSLuiz Capitulino 12844eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1285d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1286d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1287944d9fecSLuiz Capitulino { 128804adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1289953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1290953f064aSLi Xinhai nid = numa_mem_id(); 1291944d9fecSLuiz Capitulino 1292dbda8feaSBarry Song #ifdef CONFIG_CMA 1293dbda8feaSBarry Song { 1294cf11e85fSRoman Gushchin struct page *page; 1295cf11e85fSRoman Gushchin int node; 1296cf11e85fSRoman Gushchin 1297953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1298953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1299953f064aSLi Xinhai huge_page_order(h), true); 1300953f064aSLi Xinhai if (page) 1301953f064aSLi Xinhai return page; 1302953f064aSLi Xinhai } 1303953f064aSLi Xinhai 1304953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1305cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1306953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1307cf11e85fSRoman Gushchin continue; 1308cf11e85fSRoman Gushchin 1309cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1310cf11e85fSRoman Gushchin huge_page_order(h), true); 1311cf11e85fSRoman Gushchin if (page) 1312cf11e85fSRoman Gushchin return page; 1313cf11e85fSRoman Gushchin } 1314cf11e85fSRoman Gushchin } 1315953f064aSLi Xinhai } 1316dbda8feaSBarry Song #endif 1317cf11e85fSRoman Gushchin 13185e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1319944d9fecSLuiz Capitulino } 1320944d9fecSLuiz Capitulino 1321944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1322d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 13234eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13244eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13254eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13264eb0716eSAlexandre Ghiti { 13274eb0716eSAlexandre Ghiti return NULL; 13284eb0716eSAlexandre Ghiti } 13294eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1330944d9fecSLuiz Capitulino 1331e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1332d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13334eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13344eb0716eSAlexandre Ghiti { 13354eb0716eSAlexandre Ghiti return NULL; 13364eb0716eSAlexandre Ghiti } 1337d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1338944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1339d00181b9SKirill A. Shutemov unsigned int order) { } 1340944d9fecSLuiz Capitulino #endif 1341944d9fecSLuiz Capitulino 13426eb4e88aSMike Kravetz /* 13436eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 13446eb4e88aSMike Kravetz * as just a compound page. A reference is held on the page. 13456eb4e88aSMike Kravetz * 13466eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 13476eb4e88aSMike Kravetz */ 13486eb4e88aSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 13496eb4e88aSMike Kravetz bool adjust_surplus) 13506eb4e88aSMike Kravetz { 13516eb4e88aSMike Kravetz int nid = page_to_nid(page); 13526eb4e88aSMike Kravetz 13536eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13546eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 13556eb4e88aSMike Kravetz 13569487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 13576eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 13586eb4e88aSMike Kravetz return; 13596eb4e88aSMike Kravetz 13606eb4e88aSMike Kravetz list_del(&page->lru); 13616eb4e88aSMike Kravetz 13626eb4e88aSMike Kravetz if (HPageFreed(page)) { 13636eb4e88aSMike Kravetz h->free_huge_pages--; 13646eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 13656eb4e88aSMike Kravetz } 13666eb4e88aSMike Kravetz if (adjust_surplus) { 13676eb4e88aSMike Kravetz h->surplus_huge_pages--; 13686eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 13696eb4e88aSMike Kravetz } 13706eb4e88aSMike Kravetz 13716eb4e88aSMike Kravetz set_page_refcounted(page); 13726eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13736eb4e88aSMike Kravetz 13746eb4e88aSMike Kravetz h->nr_huge_pages--; 13756eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 13766eb4e88aSMike Kravetz } 13776eb4e88aSMike Kravetz 1378a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 13796af2acb6SAdam Litke { 13806af2acb6SAdam Litke int i; 1381dbfee5aeSMike Kravetz struct page *subpage = page; 1382a5516438SAndi Kleen 13834eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1384944d9fecSLuiz Capitulino return; 138518229df5SAndy Whitcroft 1386dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1387dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1388dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 138932f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1390a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1391a7407a27SLuiz Capitulino 1 << PG_writeback); 13926af2acb6SAdam Litke } 1393944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1394944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1395944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1396944d9fecSLuiz Capitulino } else { 1397a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13986af2acb6SAdam Litke } 1399944d9fecSLuiz Capitulino } 14006af2acb6SAdam Litke 140110c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 140210c6ec49SMike Kravetz { 140310c6ec49SMike Kravetz struct page *page, *t_page; 140410c6ec49SMike Kravetz 140510c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 140610c6ec49SMike Kravetz update_and_free_page(h, page); 140710c6ec49SMike Kravetz cond_resched(); 140810c6ec49SMike Kravetz } 140910c6ec49SMike Kravetz } 141010c6ec49SMike Kravetz 1411e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1412e5ff2159SAndi Kleen { 1413e5ff2159SAndi Kleen struct hstate *h; 1414e5ff2159SAndi Kleen 1415e5ff2159SAndi Kleen for_each_hstate(h) { 1416e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1417e5ff2159SAndi Kleen return h; 1418e5ff2159SAndi Kleen } 1419e5ff2159SAndi Kleen return NULL; 1420e5ff2159SAndi Kleen } 1421e5ff2159SAndi Kleen 1422db71ef79SMike Kravetz void free_huge_page(struct page *page) 142327a85ef1SDavid Gibson { 1424a5516438SAndi Kleen /* 1425a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1426a5516438SAndi Kleen * compound page destructor. 1427a5516438SAndi Kleen */ 1428e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 14297893d1d5SAdam Litke int nid = page_to_nid(page); 1430d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 143107443a85SJoonsoo Kim bool restore_reserve; 1432db71ef79SMike Kravetz unsigned long flags; 143327a85ef1SDavid Gibson 1434b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1435b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14368ace22bcSYongkai Wu 1437d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14388ace22bcSYongkai Wu page->mapping = NULL; 1439d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1440d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 144127a85ef1SDavid Gibson 14421c5ecae3SMike Kravetz /* 1443d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 14440919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14450919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14460919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14476c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 14480919e1b6SMike Kravetz * remove the reserved page from the subpool. 14490919e1b6SMike Kravetz */ 14500919e1b6SMike Kravetz if (!restore_reserve) { 14510919e1b6SMike Kravetz /* 14520919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14530919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14540919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14550919e1b6SMike Kravetz * operation. 14561c5ecae3SMike Kravetz */ 14571c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14581c5ecae3SMike Kravetz restore_reserve = true; 14590919e1b6SMike Kravetz } 14601c5ecae3SMike Kravetz 1461db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 14628f251a3dSMike Kravetz ClearHPageMigratable(page); 14636d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14646d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 146508cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 146608cf9fafSMina Almasry pages_per_huge_page(h), page); 146707443a85SJoonsoo Kim if (restore_reserve) 146807443a85SJoonsoo Kim h->resv_huge_pages++; 146907443a85SJoonsoo Kim 14709157c311SMike Kravetz if (HPageTemporary(page)) { 14716eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1472db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1473ab5ac90aSMichal Hocko update_and_free_page(h, page); 1474ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14750edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14766eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1477db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1478a5516438SAndi Kleen update_and_free_page(h, page); 14797893d1d5SAdam Litke } else { 14805d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1481a5516438SAndi Kleen enqueue_huge_page(h, page); 1482db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 148327a85ef1SDavid Gibson } 14841121828aSMike Kravetz } 148527a85ef1SDavid Gibson 1486d3d99fccSOscar Salvador /* 1487d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1488d3d99fccSOscar Salvador */ 1489d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1490d3d99fccSOscar Salvador { 1491d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1492d3d99fccSOscar Salvador h->nr_huge_pages++; 1493d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1494d3d99fccSOscar Salvador } 1495d3d99fccSOscar Salvador 1496d3d99fccSOscar Salvador static void __prep_new_huge_page(struct page *page) 1497b7ba30c6SAndi Kleen { 14980edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1499f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1500ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 15019dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15021adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1503d3d99fccSOscar Salvador } 1504d3d99fccSOscar Salvador 1505d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1506d3d99fccSOscar Salvador { 1507d3d99fccSOscar Salvador __prep_new_huge_page(page); 1508db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1509d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1510db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1511b7ba30c6SAndi Kleen } 1512b7ba30c6SAndi Kleen 1513d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 151420a0307cSWu Fengguang { 151520a0307cSWu Fengguang int i; 151620a0307cSWu Fengguang int nr_pages = 1 << order; 151720a0307cSWu Fengguang struct page *p = page + 1; 151820a0307cSWu Fengguang 151920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 152020a0307cSWu Fengguang set_compound_order(page, order); 1521ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1522de09d31dSKirill A. Shutemov __SetPageHead(page); 152320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1524ef5a22beSAndrea Arcangeli /* 1525ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1526ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1527ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15287c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1529ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1530ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1531ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1532ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1533ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1534ef5a22beSAndrea Arcangeli * after get_user_pages(). 1535ef5a22beSAndrea Arcangeli */ 1536ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 153758a84aa9SYouquan Song set_page_count(p, 0); 15381d798ca3SKirill A. Shutemov set_compound_head(p, page); 153920a0307cSWu Fengguang } 1540b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 154147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 154220a0307cSWu Fengguang } 154320a0307cSWu Fengguang 15447795912cSAndrew Morton /* 15457795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15467795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15477795912cSAndrew Morton * details. 15487795912cSAndrew Morton */ 154920a0307cSWu Fengguang int PageHuge(struct page *page) 155020a0307cSWu Fengguang { 155120a0307cSWu Fengguang if (!PageCompound(page)) 155220a0307cSWu Fengguang return 0; 155320a0307cSWu Fengguang 155420a0307cSWu Fengguang page = compound_head(page); 1555f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 155620a0307cSWu Fengguang } 155743131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 155843131e14SNaoya Horiguchi 155927c73ae7SAndrea Arcangeli /* 156027c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 156127c73ae7SAndrea Arcangeli * normal or transparent huge pages. 156227c73ae7SAndrea Arcangeli */ 156327c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 156427c73ae7SAndrea Arcangeli { 156527c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 156627c73ae7SAndrea Arcangeli return 0; 156727c73ae7SAndrea Arcangeli 1568d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 156927c73ae7SAndrea Arcangeli } 157027c73ae7SAndrea Arcangeli 1571c0d0381aSMike Kravetz /* 1572c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1573c0d0381aSMike Kravetz * 1574336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1575336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1576336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1577c0d0381aSMike Kravetz */ 1578c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1579c0d0381aSMike Kravetz { 1580336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1581c0d0381aSMike Kravetz 1582c0d0381aSMike Kravetz if (!mapping) 1583c0d0381aSMike Kravetz return mapping; 1584c0d0381aSMike Kravetz 1585c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1586c0d0381aSMike Kravetz return mapping; 1587c0d0381aSMike Kravetz 1588c0d0381aSMike Kravetz return NULL; 1589c0d0381aSMike Kravetz } 1590c0d0381aSMike Kravetz 159113d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 159213d60f4bSZhang Yi { 159313d60f4bSZhang Yi struct page *page_head = compound_head(page); 159413d60f4bSZhang Yi pgoff_t index = page_index(page_head); 159513d60f4bSZhang Yi unsigned long compound_idx; 159613d60f4bSZhang Yi 159713d60f4bSZhang Yi if (!PageHuge(page_head)) 159813d60f4bSZhang Yi return page_index(page); 159913d60f4bSZhang Yi 160013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 160113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 160213d60f4bSZhang Yi else 160313d60f4bSZhang Yi compound_idx = page - page_head; 160413d60f4bSZhang Yi 160513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 160613d60f4bSZhang Yi } 160713d60f4bSZhang Yi 16080c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1609f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1610f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16111da177e4SLinus Torvalds { 1612af0fb9dfSMichal Hocko int order = huge_page_order(h); 16131da177e4SLinus Torvalds struct page *page; 1614f60858f9SMike Kravetz bool alloc_try_hard = true; 1615f96efd58SJoe Jin 1616f60858f9SMike Kravetz /* 1617f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1618f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1619f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1620f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1621f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1622f60858f9SMike Kravetz */ 1623f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1624f60858f9SMike Kravetz alloc_try_hard = false; 1625f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1626f60858f9SMike Kravetz if (alloc_try_hard) 1627f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1628af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1629af0fb9dfSMichal Hocko nid = numa_mem_id(); 163084172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1631af0fb9dfSMichal Hocko if (page) 1632af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1633af0fb9dfSMichal Hocko else 1634af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 163563b4613cSNishanth Aravamudan 1636f60858f9SMike Kravetz /* 1637f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1638f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1639f60858f9SMike Kravetz * normal 'try hard' allocations. 1640f60858f9SMike Kravetz */ 1641f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1642f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1643f60858f9SMike Kravetz 1644f60858f9SMike Kravetz /* 1645f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1646f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1647f60858f9SMike Kravetz * overall state change. 1648f60858f9SMike Kravetz */ 1649f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1650f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1651f60858f9SMike Kravetz 165263b4613cSNishanth Aravamudan return page; 165363b4613cSNishanth Aravamudan } 165463b4613cSNishanth Aravamudan 1655af0fb9dfSMichal Hocko /* 16560c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16570c397daeSMichal Hocko * should use this function to get new hugetlb pages 16580c397daeSMichal Hocko */ 16590c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1660f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1661f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16620c397daeSMichal Hocko { 16630c397daeSMichal Hocko struct page *page; 16640c397daeSMichal Hocko 16650c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16660c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16670c397daeSMichal Hocko else 16680c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1669f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16700c397daeSMichal Hocko if (!page) 16710c397daeSMichal Hocko return NULL; 16720c397daeSMichal Hocko 16730c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16740c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16750c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16760c397daeSMichal Hocko 16770c397daeSMichal Hocko return page; 16780c397daeSMichal Hocko } 16790c397daeSMichal Hocko 16800c397daeSMichal Hocko /* 1681af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1682af0fb9dfSMichal Hocko * manner. 1683af0fb9dfSMichal Hocko */ 1684f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1685f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1686b2261026SJoonsoo Kim { 1687b2261026SJoonsoo Kim struct page *page; 1688b2261026SJoonsoo Kim int nr_nodes, node; 1689af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1690b2261026SJoonsoo Kim 1691b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1692f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1693f60858f9SMike Kravetz node_alloc_noretry); 1694af0fb9dfSMichal Hocko if (page) 1695b2261026SJoonsoo Kim break; 1696b2261026SJoonsoo Kim } 1697b2261026SJoonsoo Kim 1698af0fb9dfSMichal Hocko if (!page) 1699af0fb9dfSMichal Hocko return 0; 1700b2261026SJoonsoo Kim 1701af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1702af0fb9dfSMichal Hocko 1703af0fb9dfSMichal Hocko return 1; 1704b2261026SJoonsoo Kim } 1705b2261026SJoonsoo Kim 1706e8c5c824SLee Schermerhorn /* 170710c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 170810c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 170910c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 171010c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 1711e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1712e8c5c824SLee Schermerhorn */ 171310c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 171410c6ec49SMike Kravetz nodemask_t *nodes_allowed, 17156ae11b27SLee Schermerhorn bool acct_surplus) 1716e8c5c824SLee Schermerhorn { 1717b2261026SJoonsoo Kim int nr_nodes, node; 171810c6ec49SMike Kravetz struct page *page = NULL; 1719e8c5c824SLee Schermerhorn 17209487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1721b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1722685f3457SLee Schermerhorn /* 1723685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1724685f3457SLee Schermerhorn * nodes with surplus pages. 1725685f3457SLee Schermerhorn */ 1726b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1727b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 172810c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 1729e8c5c824SLee Schermerhorn struct page, lru); 17306eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 17319a76db09SLee Schermerhorn break; 1732e8c5c824SLee Schermerhorn } 1733b2261026SJoonsoo Kim } 1734e8c5c824SLee Schermerhorn 173510c6ec49SMike Kravetz return page; 1736e8c5c824SLee Schermerhorn } 1737e8c5c824SLee Schermerhorn 1738c8721bbbSNaoya Horiguchi /* 1739c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1740faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1741faf53defSNaoya Horiguchi * This function returns values like below: 1742faf53defSNaoya Horiguchi * 1743faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1744faf53defSNaoya Horiguchi * (allocated or reserved.) 1745faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1746faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1747c8721bbbSNaoya Horiguchi */ 1748c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1749c8721bbbSNaoya Horiguchi { 17506bc9b564SNaoya Horiguchi int rc = -EBUSY; 1751082d5b6bSGerald Schaefer 17527ffddd49SMuchun Song retry: 1753faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1754faf53defSNaoya Horiguchi if (!PageHuge(page)) 1755faf53defSNaoya Horiguchi return 0; 1756faf53defSNaoya Horiguchi 1757db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1758faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1759faf53defSNaoya Horiguchi rc = 0; 1760faf53defSNaoya Horiguchi goto out; 1761faf53defSNaoya Horiguchi } 1762faf53defSNaoya Horiguchi 1763faf53defSNaoya Horiguchi if (!page_count(page)) { 17642247bb33SGerald Schaefer struct page *head = compound_head(page); 17652247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17666bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1767082d5b6bSGerald Schaefer goto out; 17687ffddd49SMuchun Song 17697ffddd49SMuchun Song /* 17707ffddd49SMuchun Song * We should make sure that the page is already on the free list 17717ffddd49SMuchun Song * when it is dissolved. 17727ffddd49SMuchun Song */ 17736c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 1774db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 17757ffddd49SMuchun Song cond_resched(); 17767ffddd49SMuchun Song 17777ffddd49SMuchun Song /* 17787ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 17797ffddd49SMuchun Song * encounter this race. In fact, we have a chance 17807ffddd49SMuchun Song * to successfully dissolve the page if we do a 17817ffddd49SMuchun Song * retry. Because the race window is quite small. 17827ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 17837ffddd49SMuchun Song * for increasing the success rate of dissolving page. 17847ffddd49SMuchun Song */ 17857ffddd49SMuchun Song goto retry; 17867ffddd49SMuchun Song } 17877ffddd49SMuchun Song 1788c3114a84SAnshuman Khandual /* 1789c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1790c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1791c3114a84SAnshuman Khandual */ 1792c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1793c3114a84SAnshuman Khandual SetPageHWPoison(page); 1794c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1795c3114a84SAnshuman Khandual } 17966eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1797c1470b33Szhong jiang h->max_huge_pages--; 1798db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 17992247bb33SGerald Schaefer update_and_free_page(h, head); 18001121828aSMike Kravetz return 0; 1801c8721bbbSNaoya Horiguchi } 1802082d5b6bSGerald Schaefer out: 1803db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1804082d5b6bSGerald Schaefer return rc; 1805c8721bbbSNaoya Horiguchi } 1806c8721bbbSNaoya Horiguchi 1807c8721bbbSNaoya Horiguchi /* 1808c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1809c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18102247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18112247bb33SGerald Schaefer * part of it lies within the given range. 1812082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1813082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1814c8721bbbSNaoya Horiguchi */ 1815082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1816c8721bbbSNaoya Horiguchi { 1817c8721bbbSNaoya Horiguchi unsigned long pfn; 1818eb03aa00SGerald Schaefer struct page *page; 1819082d5b6bSGerald Schaefer int rc = 0; 1820c8721bbbSNaoya Horiguchi 1821d0177639SLi Zhong if (!hugepages_supported()) 1822082d5b6bSGerald Schaefer return rc; 1823d0177639SLi Zhong 1824eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1825eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1826eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1827eb03aa00SGerald Schaefer if (rc) 1828082d5b6bSGerald Schaefer break; 1829eb03aa00SGerald Schaefer } 1830082d5b6bSGerald Schaefer 1831082d5b6bSGerald Schaefer return rc; 1832c8721bbbSNaoya Horiguchi } 1833c8721bbbSNaoya Horiguchi 1834ab5ac90aSMichal Hocko /* 1835ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1836ab5ac90aSMichal Hocko */ 18370c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1838aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18397893d1d5SAdam Litke { 18409980d744SMichal Hocko struct page *page = NULL; 18417893d1d5SAdam Litke 1842bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1843aa888a74SAndi Kleen return NULL; 1844aa888a74SAndi Kleen 1845db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 18469980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18479980d744SMichal Hocko goto out_unlock; 1848db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1849d1c3fb1fSNishanth Aravamudan 1850f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18519980d744SMichal Hocko if (!page) 18520c397daeSMichal Hocko return NULL; 1853d1c3fb1fSNishanth Aravamudan 1854db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 18559980d744SMichal Hocko /* 18569980d744SMichal Hocko * We could have raced with the pool size change. 18579980d744SMichal Hocko * Double check that and simply deallocate the new page 18589980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18599980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18609980d744SMichal Hocko * codeflow 18619980d744SMichal Hocko */ 18629980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18639157c311SMike Kravetz SetHPageTemporary(page); 1864db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18659980d744SMichal Hocko put_page(page); 18662bf753e6SKai Shen return NULL; 18679980d744SMichal Hocko } else { 18689980d744SMichal Hocko h->surplus_huge_pages++; 18694704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18707893d1d5SAdam Litke } 18719980d744SMichal Hocko 18729980d744SMichal Hocko out_unlock: 1873db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18747893d1d5SAdam Litke 18757893d1d5SAdam Litke return page; 18767893d1d5SAdam Litke } 18777893d1d5SAdam Litke 1878bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1879ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1880ab5ac90aSMichal Hocko { 1881ab5ac90aSMichal Hocko struct page *page; 1882ab5ac90aSMichal Hocko 1883ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1884ab5ac90aSMichal Hocko return NULL; 1885ab5ac90aSMichal Hocko 1886f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1887ab5ac90aSMichal Hocko if (!page) 1888ab5ac90aSMichal Hocko return NULL; 1889ab5ac90aSMichal Hocko 1890ab5ac90aSMichal Hocko /* 1891ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1892ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1893ab5ac90aSMichal Hocko */ 18949157c311SMike Kravetz SetHPageTemporary(page); 1895ab5ac90aSMichal Hocko 1896ab5ac90aSMichal Hocko return page; 1897ab5ac90aSMichal Hocko } 1898ab5ac90aSMichal Hocko 1899e4e574b7SAdam Litke /* 1900099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1901099730d6SDave Hansen */ 1902e0ec90eeSDave Hansen static 19030c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1904099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1905099730d6SDave Hansen { 1906aaf14e40SMichal Hocko struct page *page; 1907aaf14e40SMichal Hocko struct mempolicy *mpol; 1908aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1909aaf14e40SMichal Hocko int nid; 1910aaf14e40SMichal Hocko nodemask_t *nodemask; 1911aaf14e40SMichal Hocko 1912aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19130c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1914aaf14e40SMichal Hocko mpol_cond_put(mpol); 1915aaf14e40SMichal Hocko 1916aaf14e40SMichal Hocko return page; 1917099730d6SDave Hansen } 1918099730d6SDave Hansen 1919ab5ac90aSMichal Hocko /* page migration callback function */ 19203e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1921d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19224db9b2efSMichal Hocko { 1923db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 19244db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19253e59fcb0SMichal Hocko struct page *page; 19263e59fcb0SMichal Hocko 19273e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19283e59fcb0SMichal Hocko if (page) { 1929db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19303e59fcb0SMichal Hocko return page; 19314db9b2efSMichal Hocko } 19324db9b2efSMichal Hocko } 1933db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19344db9b2efSMichal Hocko 19350c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19364db9b2efSMichal Hocko } 19374db9b2efSMichal Hocko 1938ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1939389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1940389c8178SMichal Hocko unsigned long address) 1941ebd63723SMichal Hocko { 1942ebd63723SMichal Hocko struct mempolicy *mpol; 1943ebd63723SMichal Hocko nodemask_t *nodemask; 1944ebd63723SMichal Hocko struct page *page; 1945ebd63723SMichal Hocko gfp_t gfp_mask; 1946ebd63723SMichal Hocko int node; 1947ebd63723SMichal Hocko 1948ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1949ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1950d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1951ebd63723SMichal Hocko mpol_cond_put(mpol); 1952ebd63723SMichal Hocko 1953ebd63723SMichal Hocko return page; 1954ebd63723SMichal Hocko } 1955ebd63723SMichal Hocko 1956bf50bab2SNaoya Horiguchi /* 195725985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1958e4e574b7SAdam Litke * of size 'delta'. 1959e4e574b7SAdam Litke */ 19600a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19611b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1962e4e574b7SAdam Litke { 1963e4e574b7SAdam Litke struct list_head surplus_list; 1964e4e574b7SAdam Litke struct page *page, *tmp; 19650a4f3d1bSLiu Xiang int ret; 19660a4f3d1bSLiu Xiang long i; 19670a4f3d1bSLiu Xiang long needed, allocated; 196828073b02SHillf Danton bool alloc_ok = true; 1969e4e574b7SAdam Litke 19709487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1971a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1972ac09b3a1SAdam Litke if (needed <= 0) { 1973a5516438SAndi Kleen h->resv_huge_pages += delta; 1974e4e574b7SAdam Litke return 0; 1975ac09b3a1SAdam Litke } 1976e4e574b7SAdam Litke 1977e4e574b7SAdam Litke allocated = 0; 1978e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1979e4e574b7SAdam Litke 1980e4e574b7SAdam Litke ret = -ENOMEM; 1981e4e574b7SAdam Litke retry: 1982db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1983e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19840c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1985aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 198628073b02SHillf Danton if (!page) { 198728073b02SHillf Danton alloc_ok = false; 198828073b02SHillf Danton break; 198928073b02SHillf Danton } 1990e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 199169ed779aSDavid Rientjes cond_resched(); 1992e4e574b7SAdam Litke } 199328073b02SHillf Danton allocated += i; 1994e4e574b7SAdam Litke 1995e4e574b7SAdam Litke /* 1996e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1997e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1998e4e574b7SAdam Litke */ 1999db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2000a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2001a5516438SAndi Kleen (h->free_huge_pages + allocated); 200228073b02SHillf Danton if (needed > 0) { 200328073b02SHillf Danton if (alloc_ok) 2004e4e574b7SAdam Litke goto retry; 200528073b02SHillf Danton /* 200628073b02SHillf Danton * We were not able to allocate enough pages to 200728073b02SHillf Danton * satisfy the entire reservation so we free what 200828073b02SHillf Danton * we've allocated so far. 200928073b02SHillf Danton */ 201028073b02SHillf Danton goto free; 201128073b02SHillf Danton } 2012e4e574b7SAdam Litke /* 2013e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 201425985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2015e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2016ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2017ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2018ac09b3a1SAdam Litke * before they are reserved. 2019e4e574b7SAdam Litke */ 2020e4e574b7SAdam Litke needed += allocated; 2021a5516438SAndi Kleen h->resv_huge_pages += delta; 2022e4e574b7SAdam Litke ret = 0; 2023a9869b83SNaoya Horiguchi 202419fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 202519fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2026e558464bSMuchun Song int zeroed; 2027e558464bSMuchun Song 202819fc3f0aSAdam Litke if ((--needed) < 0) 202919fc3f0aSAdam Litke break; 2030a9869b83SNaoya Horiguchi /* 2031a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2032a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2033a9869b83SNaoya Horiguchi */ 2034e558464bSMuchun Song zeroed = put_page_testzero(page); 2035e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2036a5516438SAndi Kleen enqueue_huge_page(h, page); 203719fc3f0aSAdam Litke } 203828073b02SHillf Danton free: 2039db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 204019fc3f0aSAdam Litke 204119fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2042c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2043a9869b83SNaoya Horiguchi put_page(page); 2044db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2045e4e574b7SAdam Litke 2046e4e574b7SAdam Litke return ret; 2047e4e574b7SAdam Litke } 2048e4e574b7SAdam Litke 2049e4e574b7SAdam Litke /* 2050e5bbc8a6SMike Kravetz * This routine has two main purposes: 2051e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2052e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2053e5bbc8a6SMike Kravetz * to the associated reservation map. 2054e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2055e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2056e4e574b7SAdam Litke */ 2057a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2058a5516438SAndi Kleen unsigned long unused_resv_pages) 2059e4e574b7SAdam Litke { 2060e4e574b7SAdam Litke unsigned long nr_pages; 206110c6ec49SMike Kravetz struct page *page; 206210c6ec49SMike Kravetz LIST_HEAD(page_list); 206310c6ec49SMike Kravetz 20649487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 206510c6ec49SMike Kravetz /* Uncommit the reservation */ 206610c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2067e4e574b7SAdam Litke 2068aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2069bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2070e5bbc8a6SMike Kravetz goto out; 2071aa888a74SAndi Kleen 2072e5bbc8a6SMike Kravetz /* 2073e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2074e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2075e5bbc8a6SMike Kravetz */ 2076a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2077e4e574b7SAdam Litke 2078685f3457SLee Schermerhorn /* 2079685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20809b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20819b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20829b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 208310c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 20849b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2085685f3457SLee Schermerhorn */ 2086685f3457SLee Schermerhorn while (nr_pages--) { 208710c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 208810c6ec49SMike Kravetz if (!page) 2089e5bbc8a6SMike Kravetz goto out; 209010c6ec49SMike Kravetz 209110c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2092e4e574b7SAdam Litke } 2093e5bbc8a6SMike Kravetz 2094e5bbc8a6SMike Kravetz out: 2095db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 209610c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2097db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2098e4e574b7SAdam Litke } 2099e4e574b7SAdam Litke 21005e911373SMike Kravetz 2101c37f9fb1SAndy Whitcroft /* 2102feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21035e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2104cf3ad20bSMike Kravetz * 2105cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2106cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2107cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2108cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2109cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2110feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2111feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2112feba16e2SMike Kravetz * is called in such cases. 2113cf3ad20bSMike Kravetz * 2114cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2115cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2116cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2117cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2118cf3ad20bSMike Kravetz * take appropriate action. 211996b96a96SMike Kravetz * 212096b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 212196b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 212296b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 212396b96a96SMike Kravetz * reservation exists. 2124c37f9fb1SAndy Whitcroft */ 21255e911373SMike Kravetz enum vma_resv_mode { 21265e911373SMike Kravetz VMA_NEEDS_RESV, 21275e911373SMike Kravetz VMA_COMMIT_RESV, 2128feba16e2SMike Kravetz VMA_END_RESV, 212996b96a96SMike Kravetz VMA_ADD_RESV, 21305e911373SMike Kravetz }; 2131cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2132cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21335e911373SMike Kravetz enum vma_resv_mode mode) 2134c37f9fb1SAndy Whitcroft { 21354e35f483SJoonsoo Kim struct resv_map *resv; 21364e35f483SJoonsoo Kim pgoff_t idx; 2137cf3ad20bSMike Kravetz long ret; 21380db9d74eSMina Almasry long dummy_out_regions_needed; 2139c37f9fb1SAndy Whitcroft 21404e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21414e35f483SJoonsoo Kim if (!resv) 2142c37f9fb1SAndy Whitcroft return 1; 2143c37f9fb1SAndy Whitcroft 21444e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21455e911373SMike Kravetz switch (mode) { 21465e911373SMike Kravetz case VMA_NEEDS_RESV: 21470db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21480db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21490db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21500db9d74eSMina Almasry * ever require 1 region. 21510db9d74eSMina Almasry */ 21520db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21535e911373SMike Kravetz break; 21545e911373SMike Kravetz case VMA_COMMIT_RESV: 2155075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21560db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21570db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21585e911373SMike Kravetz break; 2159feba16e2SMike Kravetz case VMA_END_RESV: 21600db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21615e911373SMike Kravetz ret = 0; 21625e911373SMike Kravetz break; 216396b96a96SMike Kravetz case VMA_ADD_RESV: 21640db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2165075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21660db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21670db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21680db9d74eSMina Almasry } else { 21690db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 217096b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 217196b96a96SMike Kravetz } 217296b96a96SMike Kravetz break; 21735e911373SMike Kravetz default: 21745e911373SMike Kravetz BUG(); 21755e911373SMike Kravetz } 217684afd99bSAndy Whitcroft 21774e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2178cf3ad20bSMike Kravetz return ret; 217967961f9dSMike Kravetz /* 2180bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2181bf3d12b9SMiaohe Lin * 218267961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 218367961f9dSMike Kravetz * However, a file associated with mapping could have been 218467961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 218567961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 218667961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 218767961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 218867961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 218967961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 219067961f9dSMike Kravetz * reservation has already been consumed. As a result, the 219167961f9dSMike Kravetz * return value of this routine is the opposite of the 219267961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 219367961f9dSMike Kravetz */ 2194bf3d12b9SMiaohe Lin if (ret > 0) 219567961f9dSMike Kravetz return 0; 2196bf3d12b9SMiaohe Lin if (ret == 0) 219767961f9dSMike Kravetz return 1; 2198bf3d12b9SMiaohe Lin return ret; 219984afd99bSAndy Whitcroft } 2200cf3ad20bSMike Kravetz 2201cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2202a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2203c37f9fb1SAndy Whitcroft { 22045e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2205cf3ad20bSMike Kravetz } 2206c37f9fb1SAndy Whitcroft 2207cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2208cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2209cf3ad20bSMike Kravetz { 22105e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22115e911373SMike Kravetz } 22125e911373SMike Kravetz 2213feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22145e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22155e911373SMike Kravetz { 2216feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2217c37f9fb1SAndy Whitcroft } 2218c37f9fb1SAndy Whitcroft 221996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 222096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 222196b96a96SMike Kravetz { 222296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 222396b96a96SMike Kravetz } 222496b96a96SMike Kravetz 222596b96a96SMike Kravetz /* 222696b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 222796b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 222896b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 2229d6995da3SMike Kravetz * alloc_huge_page would have consumed the reservation and set 2230d6995da3SMike Kravetz * HPageRestoreReserve in the newly allocated page. When the page is freed 2231d6995da3SMike Kravetz * via free_huge_page, the global reservation count will be incremented if 2232d6995da3SMike Kravetz * HPageRestoreReserve is set. However, free_huge_page can not adjust the 2233d6995da3SMike Kravetz * reserve map. Adjust the reserve map here to be consistent with global 2234d6995da3SMike Kravetz * reserve count adjustments to be made by free_huge_page. 223596b96a96SMike Kravetz */ 223696b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 223796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 223896b96a96SMike Kravetz struct page *page) 223996b96a96SMike Kravetz { 2240d6995da3SMike Kravetz if (unlikely(HPageRestoreReserve(page))) { 224196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 224296b96a96SMike Kravetz 224396b96a96SMike Kravetz if (unlikely(rc < 0)) { 224496b96a96SMike Kravetz /* 224596b96a96SMike Kravetz * Rare out of memory condition in reserve map 2246d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 224796b96a96SMike Kravetz * global reserve count will not be incremented 224896b96a96SMike Kravetz * by free_huge_page. This will make it appear 224996b96a96SMike Kravetz * as though the reservation for this page was 225096b96a96SMike Kravetz * consumed. This may prevent the task from 225196b96a96SMike Kravetz * faulting in the page at a later time. This 225296b96a96SMike Kravetz * is better than inconsistent global huge page 225396b96a96SMike Kravetz * accounting of reserve counts. 225496b96a96SMike Kravetz */ 2255d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 225696b96a96SMike Kravetz } else if (rc) { 225796b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 225896b96a96SMike Kravetz if (unlikely(rc < 0)) 225996b96a96SMike Kravetz /* 226096b96a96SMike Kravetz * See above comment about rare out of 226196b96a96SMike Kravetz * memory condition. 226296b96a96SMike Kravetz */ 2263d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 226496b96a96SMike Kravetz } else 226596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 226696b96a96SMike Kravetz } 226796b96a96SMike Kravetz } 226896b96a96SMike Kravetz 2269369fa227SOscar Salvador /* 2270369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2271369fa227SOscar Salvador * @h: struct hstate old page belongs to 2272369fa227SOscar Salvador * @old_page: Old page to dissolve 2273ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2274369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2275369fa227SOscar Salvador */ 2276ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2277ae37c7ffSOscar Salvador struct list_head *list) 2278369fa227SOscar Salvador { 2279369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2280369fa227SOscar Salvador int nid = page_to_nid(old_page); 2281369fa227SOscar Salvador struct page *new_page; 2282369fa227SOscar Salvador int ret = 0; 2283369fa227SOscar Salvador 2284369fa227SOscar Salvador /* 2285369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2286369fa227SOscar Salvador * pool to remain stable. Using alloc_buddy_huge_page() allows us to 2287369fa227SOscar Salvador * not having to deal with prep_new_huge_page() and avoids dealing of any 2288369fa227SOscar Salvador * counters. This simplifies and let us do the whole thing under the 2289369fa227SOscar Salvador * lock. 2290369fa227SOscar Salvador */ 2291369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2292369fa227SOscar Salvador if (!new_page) 2293369fa227SOscar Salvador return -ENOMEM; 2294369fa227SOscar Salvador 2295369fa227SOscar Salvador retry: 2296369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2297369fa227SOscar Salvador if (!PageHuge(old_page)) { 2298369fa227SOscar Salvador /* 2299369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2300369fa227SOscar Salvador */ 2301369fa227SOscar Salvador goto free_new; 2302369fa227SOscar Salvador } else if (page_count(old_page)) { 2303369fa227SOscar Salvador /* 2304ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2305ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2306369fa227SOscar Salvador */ 2307ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2308ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2309369fa227SOscar Salvador ret = -EBUSY; 2310ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2311369fa227SOscar Salvador goto free_new; 2312369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2313369fa227SOscar Salvador /* 2314369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2315369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2316369fa227SOscar Salvador * we retry. 2317369fa227SOscar Salvador */ 2318369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2319369fa227SOscar Salvador cond_resched(); 2320369fa227SOscar Salvador goto retry; 2321369fa227SOscar Salvador } else { 2322369fa227SOscar Salvador /* 2323369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2324369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2325369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2326369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2327369fa227SOscar Salvador * stable since this happens under the lock. 2328369fa227SOscar Salvador */ 2329369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2330369fa227SOscar Salvador 2331369fa227SOscar Salvador /* 2332369fa227SOscar Salvador * new_page needs to be initialized with the standard hugetlb 2333369fa227SOscar Salvador * state. This is normally done by prep_new_huge_page() but 2334369fa227SOscar Salvador * that takes hugetlb_lock which is already held so we need to 2335369fa227SOscar Salvador * open code it here. 2336369fa227SOscar Salvador * Reference count trick is needed because allocator gives us 2337369fa227SOscar Salvador * referenced page but the pool requires pages with 0 refcount. 2338369fa227SOscar Salvador */ 2339369fa227SOscar Salvador __prep_new_huge_page(new_page); 2340369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2341369fa227SOscar Salvador page_ref_dec(new_page); 2342369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2343369fa227SOscar Salvador 2344369fa227SOscar Salvador /* 2345369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2346369fa227SOscar Salvador */ 2347369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2348369fa227SOscar Salvador update_and_free_page(h, old_page); 2349369fa227SOscar Salvador } 2350369fa227SOscar Salvador 2351369fa227SOscar Salvador return ret; 2352369fa227SOscar Salvador 2353369fa227SOscar Salvador free_new: 2354369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2355369fa227SOscar Salvador __free_pages(new_page, huge_page_order(h)); 2356369fa227SOscar Salvador 2357369fa227SOscar Salvador return ret; 2358369fa227SOscar Salvador } 2359369fa227SOscar Salvador 2360ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2361369fa227SOscar Salvador { 2362369fa227SOscar Salvador struct hstate *h; 2363369fa227SOscar Salvador struct page *head; 2364ae37c7ffSOscar Salvador int ret = -EBUSY; 2365369fa227SOscar Salvador 2366369fa227SOscar Salvador /* 2367369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2368369fa227SOscar Salvador * to carefully check the state under the lock. 2369369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2370369fa227SOscar Salvador */ 2371369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2372369fa227SOscar Salvador if (PageHuge(page)) { 2373369fa227SOscar Salvador head = compound_head(page); 2374369fa227SOscar Salvador h = page_hstate(head); 2375369fa227SOscar Salvador } else { 2376369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2377369fa227SOscar Salvador return 0; 2378369fa227SOscar Salvador } 2379369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2380369fa227SOscar Salvador 2381369fa227SOscar Salvador /* 2382369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2383369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2384369fa227SOscar Salvador * of bailing out right away without further retrying. 2385369fa227SOscar Salvador */ 2386369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2387369fa227SOscar Salvador return -ENOMEM; 2388369fa227SOscar Salvador 2389ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2390ae37c7ffSOscar Salvador ret = 0; 2391ae37c7ffSOscar Salvador else if (!page_count(head)) 2392ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2393ae37c7ffSOscar Salvador 2394ae37c7ffSOscar Salvador return ret; 2395369fa227SOscar Salvador } 2396369fa227SOscar Salvador 239770c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 239804f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2399348ea204SAdam Litke { 240090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2401a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2402348ea204SAdam Litke struct page *page; 2403d85f69b0SMike Kravetz long map_chg, map_commit; 2404d85f69b0SMike Kravetz long gbl_chg; 24056d76dcf4SAneesh Kumar K.V int ret, idx; 24066d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 240708cf9fafSMina Almasry bool deferred_reserve; 24082fc39cecSAdam Litke 24096d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2410a1e78772SMel Gorman /* 2411d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2412d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2413d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2414a1e78772SMel Gorman */ 2415d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2416d85f69b0SMike Kravetz if (map_chg < 0) 241776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2418d85f69b0SMike Kravetz 2419d85f69b0SMike Kravetz /* 2420d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2421d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2422d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2423d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2424d85f69b0SMike Kravetz * checked against any subpool limit. 2425d85f69b0SMike Kravetz */ 2426d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2427d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2428d85f69b0SMike Kravetz if (gbl_chg < 0) { 2429feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 243076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 24315e911373SMike Kravetz } 243290d8b7e6SAdam Litke 2433d85f69b0SMike Kravetz /* 2434d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2435d85f69b0SMike Kravetz * map, there could be reservations associated with the 2436d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2437d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2438d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2439d85f69b0SMike Kravetz * the subpool reservations. 2440d85f69b0SMike Kravetz */ 2441d85f69b0SMike Kravetz if (avoid_reserve) 2442d85f69b0SMike Kravetz gbl_chg = 1; 2443d85f69b0SMike Kravetz } 2444d85f69b0SMike Kravetz 244508cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 244608cf9fafSMina Almasry */ 24476501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 244808cf9fafSMina Almasry if (deferred_reserve) { 244908cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 245008cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 24518f34af6fSJianyu Zhan if (ret) 24528f34af6fSJianyu Zhan goto out_subpool_put; 245308cf9fafSMina Almasry } 245408cf9fafSMina Almasry 245508cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 245608cf9fafSMina Almasry if (ret) 245708cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 24588f34af6fSJianyu Zhan 2459db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2460d85f69b0SMike Kravetz /* 2461d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2462d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2463d85f69b0SMike Kravetz * a reservation exists for the allocation. 2464d85f69b0SMike Kravetz */ 2465d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 246681a6fcaeSJoonsoo Kim if (!page) { 2467db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24680c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 24698f34af6fSJianyu Zhan if (!page) 24708f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2471a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2472d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2473a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2474a88c7695SNaoya Horiguchi } 2475db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 247615a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 247781a6fcaeSJoonsoo Kim /* Fall through */ 2478a1e78772SMel Gorman } 247981a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 248008cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 248108cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 248208cf9fafSMina Almasry */ 248308cf9fafSMina Almasry if (deferred_reserve) { 248408cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 248508cf9fafSMina Almasry h_cg, page); 248608cf9fafSMina Almasry } 248708cf9fafSMina Almasry 2488db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2489a1e78772SMel Gorman 2490d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2491a1e78772SMel Gorman 2492d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2493d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 249433039678SMike Kravetz /* 249533039678SMike Kravetz * The page was added to the reservation map between 249633039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 249733039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 249833039678SMike Kravetz * Adjust for the subpool count incremented above AND 249933039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 250033039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 250133039678SMike Kravetz * no longer applies. 250233039678SMike Kravetz */ 250333039678SMike Kravetz long rsv_adjust; 250433039678SMike Kravetz 250533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 250633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 250779aa925bSMike Kravetz if (deferred_reserve) 250879aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 250979aa925bSMike Kravetz pages_per_huge_page(h), page); 251033039678SMike Kravetz } 25117893d1d5SAdam Litke return page; 25128f34af6fSJianyu Zhan 25138f34af6fSJianyu Zhan out_uncharge_cgroup: 25148f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 251508cf9fafSMina Almasry out_uncharge_cgroup_reservation: 251608cf9fafSMina Almasry if (deferred_reserve) 251708cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 251808cf9fafSMina Almasry h_cg); 25198f34af6fSJianyu Zhan out_subpool_put: 2520d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 25218f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2522feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 25238f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2524b45b5bd6SDavid Gibson } 2525b45b5bd6SDavid Gibson 2526e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2527e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2528e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2529aa888a74SAndi Kleen { 2530aa888a74SAndi Kleen struct huge_bootmem_page *m; 2531b2261026SJoonsoo Kim int nr_nodes, node; 2532aa888a74SAndi Kleen 2533b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2534aa888a74SAndi Kleen void *addr; 2535aa888a74SAndi Kleen 2536eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 25378b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 253897ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2539aa888a74SAndi Kleen if (addr) { 2540aa888a74SAndi Kleen /* 2541aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2542aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2543aa888a74SAndi Kleen * puts them into the mem_map). 2544aa888a74SAndi Kleen */ 2545aa888a74SAndi Kleen m = addr; 2546aa888a74SAndi Kleen goto found; 2547aa888a74SAndi Kleen } 2548aa888a74SAndi Kleen } 2549aa888a74SAndi Kleen return 0; 2550aa888a74SAndi Kleen 2551aa888a74SAndi Kleen found: 2552df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2553aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2554330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2555aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2556aa888a74SAndi Kleen m->hstate = h; 2557aa888a74SAndi Kleen return 1; 2558aa888a74SAndi Kleen } 2559aa888a74SAndi Kleen 2560d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2561d00181b9SKirill A. Shutemov unsigned int order) 256218229df5SAndy Whitcroft { 256318229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 256418229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 256518229df5SAndy Whitcroft else 256618229df5SAndy Whitcroft prep_compound_page(page, order); 256718229df5SAndy Whitcroft } 256818229df5SAndy Whitcroft 2569aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2570aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2571aa888a74SAndi Kleen { 2572aa888a74SAndi Kleen struct huge_bootmem_page *m; 2573aa888a74SAndi Kleen 2574aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 257540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2576aa888a74SAndi Kleen struct hstate *h = m->hstate; 2577ee8f248dSBecky Bruce 2578aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2579c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2580ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2581aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2582af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2583af0fb9dfSMichal Hocko 2584b0320c7bSRafael Aquini /* 2585b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2586b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2587b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2588b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2589b0320c7bSRafael Aquini */ 2590bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2591c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2592520495feSCannon Matthews cond_resched(); 2593aa888a74SAndi Kleen } 2594aa888a74SAndi Kleen } 2595aa888a74SAndi Kleen 25968faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 25971da177e4SLinus Torvalds { 25981da177e4SLinus Torvalds unsigned long i; 2599f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2600f60858f9SMike Kravetz 2601f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2602f60858f9SMike Kravetz /* 2603f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2604f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2605f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2606f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2607f60858f9SMike Kravetz */ 2608f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2609f60858f9SMike Kravetz GFP_KERNEL); 2610f60858f9SMike Kravetz } else { 2611f60858f9SMike Kravetz /* allocations done at boot time */ 2612f60858f9SMike Kravetz node_alloc_noretry = NULL; 2613f60858f9SMike Kravetz } 2614f60858f9SMike Kravetz 2615f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2616f60858f9SMike Kravetz if (node_alloc_noretry) 2617f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 26181da177e4SLinus Torvalds 2619e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2620bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2621dbda8feaSBarry Song if (hugetlb_cma_size) { 2622cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 26237ecc9565SChen Wandun goto free; 2624cf11e85fSRoman Gushchin } 2625aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2626aa888a74SAndi Kleen break; 26270c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2628f60858f9SMike Kravetz &node_states[N_MEMORY], 2629f60858f9SMike Kravetz node_alloc_noretry)) 26301da177e4SLinus Torvalds break; 263169ed779aSDavid Rientjes cond_resched(); 26321da177e4SLinus Torvalds } 2633d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2634d715cf80SLiam R. Howlett char buf[32]; 2635d715cf80SLiam R. Howlett 2636c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2637d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2638d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 26398faa8b07SAndi Kleen h->max_huge_pages = i; 2640e5ff2159SAndi Kleen } 26417ecc9565SChen Wandun free: 2642f60858f9SMike Kravetz kfree(node_alloc_noretry); 2643d715cf80SLiam R. Howlett } 2644e5ff2159SAndi Kleen 2645e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2646e5ff2159SAndi Kleen { 2647e5ff2159SAndi Kleen struct hstate *h; 2648e5ff2159SAndi Kleen 2649e5ff2159SAndi Kleen for_each_hstate(h) { 2650641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2651641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2652641844f5SNaoya Horiguchi 26538faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2654bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 26558faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2656e5ff2159SAndi Kleen } 2657641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2658e5ff2159SAndi Kleen } 2659e5ff2159SAndi Kleen 2660e5ff2159SAndi Kleen static void __init report_hugepages(void) 2661e5ff2159SAndi Kleen { 2662e5ff2159SAndi Kleen struct hstate *h; 2663e5ff2159SAndi Kleen 2664e5ff2159SAndi Kleen for_each_hstate(h) { 26654abd32dbSAndi Kleen char buf[32]; 2666c6247f72SMatthew Wilcox 2667c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2668ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2669c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2670e5ff2159SAndi Kleen } 2671e5ff2159SAndi Kleen } 2672e5ff2159SAndi Kleen 26731da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 26746ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 26756ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26761da177e4SLinus Torvalds { 26774415cc8dSChristoph Lameter int i; 26781121828aSMike Kravetz LIST_HEAD(page_list); 26794415cc8dSChristoph Lameter 26809487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2681bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2682aa888a74SAndi Kleen return; 2683aa888a74SAndi Kleen 26841121828aSMike Kravetz /* 26851121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 26861121828aSMike Kravetz */ 26876ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 268810c6ec49SMike Kravetz struct page *page, *next; 2689a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2690a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2691a5516438SAndi Kleen if (count >= h->nr_huge_pages) 26921121828aSMike Kravetz goto out; 26931da177e4SLinus Torvalds if (PageHighMem(page)) 26941da177e4SLinus Torvalds continue; 26956eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 26961121828aSMike Kravetz list_add(&page->lru, &page_list); 26971121828aSMike Kravetz } 26981121828aSMike Kravetz } 26991121828aSMike Kravetz 27001121828aSMike Kravetz out: 2701db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 270210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2703db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 27041da177e4SLinus Torvalds } 27051da177e4SLinus Torvalds #else 27066ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 27076ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 27081da177e4SLinus Torvalds { 27091da177e4SLinus Torvalds } 27101da177e4SLinus Torvalds #endif 27111da177e4SLinus Torvalds 271220a0307cSWu Fengguang /* 271320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 271420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 271520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 271620a0307cSWu Fengguang */ 27176ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 27186ae11b27SLee Schermerhorn int delta) 271920a0307cSWu Fengguang { 2720b2261026SJoonsoo Kim int nr_nodes, node; 272120a0307cSWu Fengguang 27229487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 272320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 272420a0307cSWu Fengguang 2725e8c5c824SLee Schermerhorn if (delta < 0) { 2726b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2727b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2728b2261026SJoonsoo Kim goto found; 2729b2261026SJoonsoo Kim } 2730b2261026SJoonsoo Kim } else { 2731b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2732b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2733b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2734b2261026SJoonsoo Kim goto found; 2735e8c5c824SLee Schermerhorn } 27369a76db09SLee Schermerhorn } 2737b2261026SJoonsoo Kim return 0; 273820a0307cSWu Fengguang 2739b2261026SJoonsoo Kim found: 274020a0307cSWu Fengguang h->surplus_huge_pages += delta; 2741b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2742b2261026SJoonsoo Kim return 1; 274320a0307cSWu Fengguang } 274420a0307cSWu Fengguang 2745a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2746fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 27476ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 27481da177e4SLinus Torvalds { 27497893d1d5SAdam Litke unsigned long min_count, ret; 275010c6ec49SMike Kravetz struct page *page; 275110c6ec49SMike Kravetz LIST_HEAD(page_list); 2752f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2753f60858f9SMike Kravetz 2754f60858f9SMike Kravetz /* 2755f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2756f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2757f60858f9SMike Kravetz * the requested huge pages. 2758f60858f9SMike Kravetz */ 2759f60858f9SMike Kravetz if (node_alloc_noretry) 2760f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2761f60858f9SMike Kravetz else 2762f60858f9SMike Kravetz return -ENOMEM; 27631da177e4SLinus Torvalds 276429383967SMike Kravetz /* 276529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 276629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 276729383967SMike Kravetz */ 276829383967SMike Kravetz mutex_lock(&h->resize_lock); 2769db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 27704eb0716eSAlexandre Ghiti 27714eb0716eSAlexandre Ghiti /* 2772fd875dcaSMike Kravetz * Check for a node specific request. 2773fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2774fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2775fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2776fd875dcaSMike Kravetz */ 2777fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2778fd875dcaSMike Kravetz unsigned long old_count = count; 2779fd875dcaSMike Kravetz 2780fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2781fd875dcaSMike Kravetz /* 2782fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2783fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2784fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2785fd875dcaSMike Kravetz * largest possible value to align with their intention. 2786fd875dcaSMike Kravetz */ 2787fd875dcaSMike Kravetz if (count < old_count) 2788fd875dcaSMike Kravetz count = ULONG_MAX; 2789fd875dcaSMike Kravetz } 2790fd875dcaSMike Kravetz 2791fd875dcaSMike Kravetz /* 27924eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 27934eb0716eSAlexandre Ghiti * page range allocation. 27944eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 27954eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 27964eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 27974eb0716eSAlexandre Ghiti */ 27984eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 27994eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 2800db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 280129383967SMike Kravetz mutex_unlock(&h->resize_lock); 2802f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 28034eb0716eSAlexandre Ghiti return -EINVAL; 28044eb0716eSAlexandre Ghiti } 28054eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 28064eb0716eSAlexandre Ghiti } 2807aa888a74SAndi Kleen 28087893d1d5SAdam Litke /* 28097893d1d5SAdam Litke * Increase the pool size 28107893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 28117893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2812d1c3fb1fSNishanth Aravamudan * 28130c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2814d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2815d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2816d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2817d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 28187893d1d5SAdam Litke */ 2819a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 28206ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 28217893d1d5SAdam Litke break; 28227893d1d5SAdam Litke } 28237893d1d5SAdam Litke 2824a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 28257893d1d5SAdam Litke /* 28267893d1d5SAdam Litke * If this allocation races such that we no longer need the 28277893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 28287893d1d5SAdam Litke * and reducing the surplus. 28297893d1d5SAdam Litke */ 2830db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2831649920c6SJia He 2832649920c6SJia He /* yield cpu to avoid soft lockup */ 2833649920c6SJia He cond_resched(); 2834649920c6SJia He 2835f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2836f60858f9SMike Kravetz node_alloc_noretry); 2837db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 28387893d1d5SAdam Litke if (!ret) 28397893d1d5SAdam Litke goto out; 28407893d1d5SAdam Litke 2841536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2842536240f2SMel Gorman if (signal_pending(current)) 2843536240f2SMel Gorman goto out; 28447893d1d5SAdam Litke } 28457893d1d5SAdam Litke 28467893d1d5SAdam Litke /* 28477893d1d5SAdam Litke * Decrease the pool size 28487893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 28497893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 28507893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 28517893d1d5SAdam Litke * to the desired size as pages become free. 2852d1c3fb1fSNishanth Aravamudan * 2853d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2854d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2855d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 28560c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2857d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2858d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2859d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 28607893d1d5SAdam Litke */ 2861a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 28626b0c880dSAdam Litke min_count = max(count, min_count); 28636ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 286410c6ec49SMike Kravetz 286510c6ec49SMike Kravetz /* 286610c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 286710c6ec49SMike Kravetz */ 2868a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 286910c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 287010c6ec49SMike Kravetz if (!page) 28711da177e4SLinus Torvalds break; 287210c6ec49SMike Kravetz 287310c6ec49SMike Kravetz list_add(&page->lru, &page_list); 28741da177e4SLinus Torvalds } 287510c6ec49SMike Kravetz /* free the pages after dropping lock */ 2876db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 287710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2878db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 287910c6ec49SMike Kravetz 2880a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 28816ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 28827893d1d5SAdam Litke break; 28837893d1d5SAdam Litke } 28847893d1d5SAdam Litke out: 28854eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 2886db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 288729383967SMike Kravetz mutex_unlock(&h->resize_lock); 28884eb0716eSAlexandre Ghiti 2889f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2890f60858f9SMike Kravetz 28914eb0716eSAlexandre Ghiti return 0; 28921da177e4SLinus Torvalds } 28931da177e4SLinus Torvalds 2894a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2895a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2896a3437870SNishanth Aravamudan 2897a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2898a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2899a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2900a3437870SNishanth Aravamudan 2901a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2902a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2903a3437870SNishanth Aravamudan 29049a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 29059a305230SLee Schermerhorn 29069a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2907a3437870SNishanth Aravamudan { 2908a3437870SNishanth Aravamudan int i; 29099a305230SLee Schermerhorn 2910a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 29119a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 29129a305230SLee Schermerhorn if (nidp) 29139a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2914a3437870SNishanth Aravamudan return &hstates[i]; 29159a305230SLee Schermerhorn } 29169a305230SLee Schermerhorn 29179a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2918a3437870SNishanth Aravamudan } 2919a3437870SNishanth Aravamudan 292006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2921a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2922a3437870SNishanth Aravamudan { 29239a305230SLee Schermerhorn struct hstate *h; 29249a305230SLee Schermerhorn unsigned long nr_huge_pages; 29259a305230SLee Schermerhorn int nid; 29269a305230SLee Schermerhorn 29279a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29289a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29299a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 29309a305230SLee Schermerhorn else 29319a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 29329a305230SLee Schermerhorn 2933ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2934a3437870SNishanth Aravamudan } 2935adbe8726SEric B Munson 2936238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2937238d3c13SDavid Rientjes struct hstate *h, int nid, 2938238d3c13SDavid Rientjes unsigned long count, size_t len) 2939a3437870SNishanth Aravamudan { 2940a3437870SNishanth Aravamudan int err; 29412d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2942a3437870SNishanth Aravamudan 29432d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 29442d0adf7eSOscar Salvador return -EINVAL; 2945adbe8726SEric B Munson 29469a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 29479a305230SLee Schermerhorn /* 29489a305230SLee Schermerhorn * global hstate attribute 29499a305230SLee Schermerhorn */ 29509a305230SLee Schermerhorn if (!(obey_mempolicy && 29512d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 29522d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 29532d0adf7eSOscar Salvador else 29542d0adf7eSOscar Salvador n_mask = &nodes_allowed; 29552d0adf7eSOscar Salvador } else { 29569a305230SLee Schermerhorn /* 2957fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2958fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 29599a305230SLee Schermerhorn */ 29602d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 29612d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2962fd875dcaSMike Kravetz } 29639a305230SLee Schermerhorn 29642d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 296506808b08SLee Schermerhorn 29664eb0716eSAlexandre Ghiti return err ? err : len; 296706808b08SLee Schermerhorn } 296806808b08SLee Schermerhorn 2969238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2970238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2971238d3c13SDavid Rientjes size_t len) 2972238d3c13SDavid Rientjes { 2973238d3c13SDavid Rientjes struct hstate *h; 2974238d3c13SDavid Rientjes unsigned long count; 2975238d3c13SDavid Rientjes int nid; 2976238d3c13SDavid Rientjes int err; 2977238d3c13SDavid Rientjes 2978238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2979238d3c13SDavid Rientjes if (err) 2980238d3c13SDavid Rientjes return err; 2981238d3c13SDavid Rientjes 2982238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2983238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2984238d3c13SDavid Rientjes } 2985238d3c13SDavid Rientjes 298606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 298706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 298806808b08SLee Schermerhorn { 298906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 299006808b08SLee Schermerhorn } 299106808b08SLee Schermerhorn 299206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 299306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 299406808b08SLee Schermerhorn { 2995238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2996a3437870SNishanth Aravamudan } 2997a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2998a3437870SNishanth Aravamudan 299906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 300006808b08SLee Schermerhorn 300106808b08SLee Schermerhorn /* 300206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 300306808b08SLee Schermerhorn * huge page alloc/free. 300406808b08SLee Schermerhorn */ 300506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3006ae7a927dSJoe Perches struct kobj_attribute *attr, 3007ae7a927dSJoe Perches char *buf) 300806808b08SLee Schermerhorn { 300906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 301006808b08SLee Schermerhorn } 301106808b08SLee Schermerhorn 301206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 301306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 301406808b08SLee Schermerhorn { 3015238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 301606808b08SLee Schermerhorn } 301706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 301806808b08SLee Schermerhorn #endif 301906808b08SLee Schermerhorn 302006808b08SLee Schermerhorn 3021a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3022a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3023a3437870SNishanth Aravamudan { 30249a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3025ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3026a3437870SNishanth Aravamudan } 3027adbe8726SEric B Munson 3028a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3029a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3030a3437870SNishanth Aravamudan { 3031a3437870SNishanth Aravamudan int err; 3032a3437870SNishanth Aravamudan unsigned long input; 30339a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3034a3437870SNishanth Aravamudan 3035bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3036adbe8726SEric B Munson return -EINVAL; 3037adbe8726SEric B Munson 30383dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3039a3437870SNishanth Aravamudan if (err) 304073ae31e5SEric B Munson return err; 3041a3437870SNishanth Aravamudan 3042db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3043a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3044db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3045a3437870SNishanth Aravamudan 3046a3437870SNishanth Aravamudan return count; 3047a3437870SNishanth Aravamudan } 3048a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3049a3437870SNishanth Aravamudan 3050a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3051a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3052a3437870SNishanth Aravamudan { 30539a305230SLee Schermerhorn struct hstate *h; 30549a305230SLee Schermerhorn unsigned long free_huge_pages; 30559a305230SLee Schermerhorn int nid; 30569a305230SLee Schermerhorn 30579a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 30589a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 30599a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 30609a305230SLee Schermerhorn else 30619a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 30629a305230SLee Schermerhorn 3063ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3064a3437870SNishanth Aravamudan } 3065a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3066a3437870SNishanth Aravamudan 3067a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3068a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3069a3437870SNishanth Aravamudan { 30709a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3071ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3072a3437870SNishanth Aravamudan } 3073a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3074a3437870SNishanth Aravamudan 3075a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3076a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3077a3437870SNishanth Aravamudan { 30789a305230SLee Schermerhorn struct hstate *h; 30799a305230SLee Schermerhorn unsigned long surplus_huge_pages; 30809a305230SLee Schermerhorn int nid; 30819a305230SLee Schermerhorn 30829a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 30839a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 30849a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 30859a305230SLee Schermerhorn else 30869a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 30879a305230SLee Schermerhorn 3088ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3089a3437870SNishanth Aravamudan } 3090a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3091a3437870SNishanth Aravamudan 3092a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3093a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3094a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3095a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3096a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3097a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 309806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 309906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 310006808b08SLee Schermerhorn #endif 3101a3437870SNishanth Aravamudan NULL, 3102a3437870SNishanth Aravamudan }; 3103a3437870SNishanth Aravamudan 310467e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3105a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3106a3437870SNishanth Aravamudan }; 3107a3437870SNishanth Aravamudan 3108094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 31099a305230SLee Schermerhorn struct kobject **hstate_kobjs, 311067e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3111a3437870SNishanth Aravamudan { 3112a3437870SNishanth Aravamudan int retval; 3113972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3114a3437870SNishanth Aravamudan 31159a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 31169a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3117a3437870SNishanth Aravamudan return -ENOMEM; 3118a3437870SNishanth Aravamudan 31199a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3120cc2205a6SMiaohe Lin if (retval) { 31219a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3122cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3123cc2205a6SMiaohe Lin } 3124a3437870SNishanth Aravamudan 3125a3437870SNishanth Aravamudan return retval; 3126a3437870SNishanth Aravamudan } 3127a3437870SNishanth Aravamudan 3128a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3129a3437870SNishanth Aravamudan { 3130a3437870SNishanth Aravamudan struct hstate *h; 3131a3437870SNishanth Aravamudan int err; 3132a3437870SNishanth Aravamudan 3133a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3134a3437870SNishanth Aravamudan if (!hugepages_kobj) 3135a3437870SNishanth Aravamudan return; 3136a3437870SNishanth Aravamudan 3137a3437870SNishanth Aravamudan for_each_hstate(h) { 31389a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 31399a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3140a3437870SNishanth Aravamudan if (err) 3141282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3142a3437870SNishanth Aravamudan } 3143a3437870SNishanth Aravamudan } 3144a3437870SNishanth Aravamudan 31459a305230SLee Schermerhorn #ifdef CONFIG_NUMA 31469a305230SLee Schermerhorn 31479a305230SLee Schermerhorn /* 31489a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 314910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 315010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 315110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 31529a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 31539a305230SLee Schermerhorn */ 31549a305230SLee Schermerhorn struct node_hstate { 31559a305230SLee Schermerhorn struct kobject *hugepages_kobj; 31569a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 31579a305230SLee Schermerhorn }; 3158b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 31599a305230SLee Schermerhorn 31609a305230SLee Schermerhorn /* 316110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 31629a305230SLee Schermerhorn */ 31639a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 31649a305230SLee Schermerhorn &nr_hugepages_attr.attr, 31659a305230SLee Schermerhorn &free_hugepages_attr.attr, 31669a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 31679a305230SLee Schermerhorn NULL, 31689a305230SLee Schermerhorn }; 31699a305230SLee Schermerhorn 317067e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 31719a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 31729a305230SLee Schermerhorn }; 31739a305230SLee Schermerhorn 31749a305230SLee Schermerhorn /* 317510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 31769a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 31779a305230SLee Schermerhorn */ 31789a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31799a305230SLee Schermerhorn { 31809a305230SLee Schermerhorn int nid; 31819a305230SLee Schermerhorn 31829a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 31839a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 31849a305230SLee Schermerhorn int i; 31859a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 31869a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 31879a305230SLee Schermerhorn if (nidp) 31889a305230SLee Schermerhorn *nidp = nid; 31899a305230SLee Schermerhorn return &hstates[i]; 31909a305230SLee Schermerhorn } 31919a305230SLee Schermerhorn } 31929a305230SLee Schermerhorn 31939a305230SLee Schermerhorn BUG(); 31949a305230SLee Schermerhorn return NULL; 31959a305230SLee Schermerhorn } 31969a305230SLee Schermerhorn 31979a305230SLee Schermerhorn /* 319810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 31999a305230SLee Schermerhorn * No-op if no hstate attributes attached. 32009a305230SLee Schermerhorn */ 32013cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 32029a305230SLee Schermerhorn { 32039a305230SLee Schermerhorn struct hstate *h; 320410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 32059a305230SLee Schermerhorn 32069a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 32079b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 32089a305230SLee Schermerhorn 3209972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3210972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3211972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3212972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3213972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3214972dc4deSAneesh Kumar K.V } 32159a305230SLee Schermerhorn } 32169a305230SLee Schermerhorn 32179a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 32189a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 32199a305230SLee Schermerhorn } 32209a305230SLee Schermerhorn 32219a305230SLee Schermerhorn 32229a305230SLee Schermerhorn /* 322310fbcf4cSKay Sievers * Register hstate attributes for a single node device. 32249a305230SLee Schermerhorn * No-op if attributes already registered. 32259a305230SLee Schermerhorn */ 32263cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 32279a305230SLee Schermerhorn { 32289a305230SLee Schermerhorn struct hstate *h; 322910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 32309a305230SLee Schermerhorn int err; 32319a305230SLee Schermerhorn 32329a305230SLee Schermerhorn if (nhs->hugepages_kobj) 32339a305230SLee Schermerhorn return; /* already allocated */ 32349a305230SLee Schermerhorn 32359a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 323610fbcf4cSKay Sievers &node->dev.kobj); 32379a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 32389a305230SLee Schermerhorn return; 32399a305230SLee Schermerhorn 32409a305230SLee Schermerhorn for_each_hstate(h) { 32419a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 32429a305230SLee Schermerhorn nhs->hstate_kobjs, 32439a305230SLee Schermerhorn &per_node_hstate_attr_group); 32449a305230SLee Schermerhorn if (err) { 3245282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 324610fbcf4cSKay Sievers h->name, node->dev.id); 32479a305230SLee Schermerhorn hugetlb_unregister_node(node); 32489a305230SLee Schermerhorn break; 32499a305230SLee Schermerhorn } 32509a305230SLee Schermerhorn } 32519a305230SLee Schermerhorn } 32529a305230SLee Schermerhorn 32539a305230SLee Schermerhorn /* 32549b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 325510fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 325610fbcf4cSKay Sievers * registered their associated device by this time. 32579a305230SLee Schermerhorn */ 32587d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 32599a305230SLee Schermerhorn { 32609a305230SLee Schermerhorn int nid; 32619a305230SLee Schermerhorn 32628cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 32638732794bSWen Congyang struct node *node = node_devices[nid]; 326410fbcf4cSKay Sievers if (node->dev.id == nid) 32659a305230SLee Schermerhorn hugetlb_register_node(node); 32669a305230SLee Schermerhorn } 32679a305230SLee Schermerhorn 32689a305230SLee Schermerhorn /* 326910fbcf4cSKay Sievers * Let the node device driver know we're here so it can 32709a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 32719a305230SLee Schermerhorn */ 32729a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 32739a305230SLee Schermerhorn hugetlb_unregister_node); 32749a305230SLee Schermerhorn } 32759a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 32769a305230SLee Schermerhorn 32779a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 32789a305230SLee Schermerhorn { 32799a305230SLee Schermerhorn BUG(); 32809a305230SLee Schermerhorn if (nidp) 32819a305230SLee Schermerhorn *nidp = -1; 32829a305230SLee Schermerhorn return NULL; 32839a305230SLee Schermerhorn } 32849a305230SLee Schermerhorn 32859a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 32869a305230SLee Schermerhorn 32879a305230SLee Schermerhorn #endif 32889a305230SLee Schermerhorn 3289a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3290a3437870SNishanth Aravamudan { 32918382d914SDavidlohr Bueso int i; 32928382d914SDavidlohr Bueso 3293d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3294d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3295d6995da3SMike Kravetz 3296c2833a5bSMike Kravetz if (!hugepages_supported()) { 3297c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3298c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 32990ef89d25SBenjamin Herrenschmidt return 0; 3300d715cf80SLiam R. Howlett } 3301d715cf80SLiam R. Howlett 3302282f4214SMike Kravetz /* 3303282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3304282f4214SMike Kravetz * architectures depend on setup being done here. 3305282f4214SMike Kravetz */ 3306a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3307282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3308282f4214SMike Kravetz /* 3309282f4214SMike Kravetz * If we did not parse a default huge page size, set 3310282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3311282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3312282f4214SMike Kravetz * specified, set that here as well. 3313282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3314282f4214SMike Kravetz * setting. A warning will be printed in this case. 3315282f4214SMike Kravetz */ 3316282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3317f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3318282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3319282f4214SMike Kravetz char buf[32]; 3320282f4214SMike Kravetz 3321282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3322282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3323282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3324282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3325282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3326282f4214SMike Kravetz default_hstate_max_huge_pages); 3327282f4214SMike Kravetz } 3328282f4214SMike Kravetz default_hstate.max_huge_pages = 3329282f4214SMike Kravetz default_hstate_max_huge_pages; 3330282f4214SMike Kravetz } 3331f8b74815SVaishali Thakkar } 3332a3437870SNishanth Aravamudan 3333cf11e85fSRoman Gushchin hugetlb_cma_check(); 3334a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3335aa888a74SAndi Kleen gather_bootmem_prealloc(); 3336a3437870SNishanth Aravamudan report_hugepages(); 3337a3437870SNishanth Aravamudan 3338a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 33399a305230SLee Schermerhorn hugetlb_register_all_nodes(); 33407179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 33419a305230SLee Schermerhorn 33428382d914SDavidlohr Bueso #ifdef CONFIG_SMP 33438382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 33448382d914SDavidlohr Bueso #else 33458382d914SDavidlohr Bueso num_fault_mutexes = 1; 33468382d914SDavidlohr Bueso #endif 3347c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 33486da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 33496da2ec56SKees Cook GFP_KERNEL); 3350c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 33518382d914SDavidlohr Bueso 33528382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3353c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3354a3437870SNishanth Aravamudan return 0; 3355a3437870SNishanth Aravamudan } 33563e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3357a3437870SNishanth Aravamudan 3358ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3359ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 33609fee021dSVaishali Thakkar { 3361ae94da89SMike Kravetz return size == HPAGE_SIZE; 33629fee021dSVaishali Thakkar } 33639fee021dSVaishali Thakkar 3364d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3365a3437870SNishanth Aravamudan { 3366a3437870SNishanth Aravamudan struct hstate *h; 33678faa8b07SAndi Kleen unsigned long i; 33688faa8b07SAndi Kleen 3369a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3370a3437870SNishanth Aravamudan return; 3371a3437870SNishanth Aravamudan } 337247d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3373a3437870SNishanth Aravamudan BUG_ON(order == 0); 337447d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 337529383967SMike Kravetz mutex_init(&h->resize_lock); 3376a3437870SNishanth Aravamudan h->order = order; 3377aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 33788faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 33798faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 33800edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 338154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 338254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3383a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3384a3437870SNishanth Aravamudan huge_page_size(h)/1024); 33858faa8b07SAndi Kleen 3386a3437870SNishanth Aravamudan parsed_hstate = h; 3387a3437870SNishanth Aravamudan } 3388a3437870SNishanth Aravamudan 3389282f4214SMike Kravetz /* 3390282f4214SMike Kravetz * hugepages command line processing 3391282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3392282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3393282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3394282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3395282f4214SMike Kravetz */ 3396282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3397a3437870SNishanth Aravamudan { 3398a3437870SNishanth Aravamudan unsigned long *mhp; 33998faa8b07SAndi Kleen static unsigned long *last_mhp; 3400a3437870SNishanth Aravamudan 34019fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3402282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 34039fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3404282f4214SMike Kravetz return 0; 34059fee021dSVaishali Thakkar } 3406282f4214SMike Kravetz 3407a3437870SNishanth Aravamudan /* 3408282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3409282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3410282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3411282f4214SMike Kravetz * default_hugepagesz. 3412a3437870SNishanth Aravamudan */ 34139fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3414a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3415a3437870SNishanth Aravamudan else 3416a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3417a3437870SNishanth Aravamudan 34188faa8b07SAndi Kleen if (mhp == last_mhp) { 3419282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3420282f4214SMike Kravetz return 0; 34218faa8b07SAndi Kleen } 34228faa8b07SAndi Kleen 3423a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3424a3437870SNishanth Aravamudan *mhp = 0; 3425a3437870SNishanth Aravamudan 34268faa8b07SAndi Kleen /* 34278faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 342804adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 34298faa8b07SAndi Kleen * use the bootmem allocator. 34308faa8b07SAndi Kleen */ 343104adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 34328faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 34338faa8b07SAndi Kleen 34348faa8b07SAndi Kleen last_mhp = mhp; 34358faa8b07SAndi Kleen 3436a3437870SNishanth Aravamudan return 1; 3437a3437870SNishanth Aravamudan } 3438282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3439e11bfbfcSNick Piggin 3440282f4214SMike Kravetz /* 3441282f4214SMike Kravetz * hugepagesz command line processing 3442282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3443282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3444282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3445282f4214SMike Kravetz * hugepagesz argument was valid. 3446282f4214SMike Kravetz */ 3447359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3448e11bfbfcSNick Piggin { 3449359f2544SMike Kravetz unsigned long size; 3450282f4214SMike Kravetz struct hstate *h; 3451282f4214SMike Kravetz 3452282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3453359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3454359f2544SMike Kravetz 3455359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3456282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3457359f2544SMike Kravetz return 0; 3458359f2544SMike Kravetz } 3459359f2544SMike Kravetz 3460282f4214SMike Kravetz h = size_to_hstate(size); 3461282f4214SMike Kravetz if (h) { 3462282f4214SMike Kravetz /* 3463282f4214SMike Kravetz * hstate for this size already exists. This is normally 3464282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3465282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3466282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3467282f4214SMike Kravetz * previously specified. 3468282f4214SMike Kravetz */ 3469282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3470282f4214SMike Kravetz default_hstate.max_huge_pages) { 3471282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 347238237830SMike Kravetz return 0; 347338237830SMike Kravetz } 347438237830SMike Kravetz 3475282f4214SMike Kravetz /* 3476282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3477282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3478282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3479282f4214SMike Kravetz */ 3480282f4214SMike Kravetz parsed_hstate = h; 3481282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3482e11bfbfcSNick Piggin return 1; 3483e11bfbfcSNick Piggin } 3484282f4214SMike Kravetz 3485359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3486282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3487359f2544SMike Kravetz return 1; 3488359f2544SMike Kravetz } 3489359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3490359f2544SMike Kravetz 3491282f4214SMike Kravetz /* 3492282f4214SMike Kravetz * default_hugepagesz command line input 3493282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3494282f4214SMike Kravetz */ 3495ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3496e11bfbfcSNick Piggin { 3497ae94da89SMike Kravetz unsigned long size; 3498ae94da89SMike Kravetz 3499282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3500282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3501282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3502282f4214SMike Kravetz return 0; 3503282f4214SMike Kravetz } 3504282f4214SMike Kravetz 3505282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3506282f4214SMike Kravetz 3507282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3508282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3509282f4214SMike Kravetz return 0; 3510282f4214SMike Kravetz } 3511282f4214SMike Kravetz 3512282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3513282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3514282f4214SMike Kravetz parsed_default_hugepagesz = true; 3515282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3516282f4214SMike Kravetz 3517282f4214SMike Kravetz /* 3518282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3519282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3520282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3521282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3522282f4214SMike Kravetz * allocated here from bootmem allocator. 3523282f4214SMike Kravetz */ 3524282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3525282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3526282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3527282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3528282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3529282f4214SMike Kravetz } 3530282f4214SMike Kravetz 3531e11bfbfcSNick Piggin return 1; 3532e11bfbfcSNick Piggin } 3533ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3534a3437870SNishanth Aravamudan 35358ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 35368a213460SNishanth Aravamudan { 35378a213460SNishanth Aravamudan int node; 35388a213460SNishanth Aravamudan unsigned int nr = 0; 35398ca39e68SMuchun Song nodemask_t *mpol_allowed; 35408ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 35418ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 35428a213460SNishanth Aravamudan 35438ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 35448ca39e68SMuchun Song 35458ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3546c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 35478a213460SNishanth Aravamudan nr += array[node]; 35488ca39e68SMuchun Song } 35498a213460SNishanth Aravamudan 35508a213460SNishanth Aravamudan return nr; 35518a213460SNishanth Aravamudan } 35528a213460SNishanth Aravamudan 35538a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 355417743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 355517743798SMuchun Song void *buffer, size_t *length, 355617743798SMuchun Song loff_t *ppos, unsigned long *out) 355717743798SMuchun Song { 355817743798SMuchun Song struct ctl_table dup_table; 355917743798SMuchun Song 356017743798SMuchun Song /* 356117743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 356217743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 356317743798SMuchun Song */ 356417743798SMuchun Song dup_table = *table; 356517743798SMuchun Song dup_table.data = out; 356617743798SMuchun Song 356717743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 356817743798SMuchun Song } 356917743798SMuchun Song 357006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 357106808b08SLee Schermerhorn struct ctl_table *table, int write, 357232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 35731da177e4SLinus Torvalds { 3574e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3575238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 357608d4a246SMichal Hocko int ret; 3577e5ff2159SAndi Kleen 3578457c1b27SNishanth Aravamudan if (!hugepages_supported()) 357986613628SJan Stancek return -EOPNOTSUPP; 3580457c1b27SNishanth Aravamudan 358117743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 358217743798SMuchun Song &tmp); 358308d4a246SMichal Hocko if (ret) 358408d4a246SMichal Hocko goto out; 3585e5ff2159SAndi Kleen 3586238d3c13SDavid Rientjes if (write) 3587238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3588238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 358908d4a246SMichal Hocko out: 359008d4a246SMichal Hocko return ret; 35911da177e4SLinus Torvalds } 3592396faf03SMel Gorman 359306808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 359432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 359506808b08SLee Schermerhorn { 359606808b08SLee Schermerhorn 359706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 359806808b08SLee Schermerhorn buffer, length, ppos); 359906808b08SLee Schermerhorn } 360006808b08SLee Schermerhorn 360106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 360206808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 360332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 360406808b08SLee Schermerhorn { 360506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 360606808b08SLee Schermerhorn buffer, length, ppos); 360706808b08SLee Schermerhorn } 360806808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 360906808b08SLee Schermerhorn 3610a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 361132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3612a3d0c6aaSNishanth Aravamudan { 3613a5516438SAndi Kleen struct hstate *h = &default_hstate; 3614e5ff2159SAndi Kleen unsigned long tmp; 361508d4a246SMichal Hocko int ret; 3616e5ff2159SAndi Kleen 3617457c1b27SNishanth Aravamudan if (!hugepages_supported()) 361886613628SJan Stancek return -EOPNOTSUPP; 3619457c1b27SNishanth Aravamudan 3620e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3621e5ff2159SAndi Kleen 3622bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3623adbe8726SEric B Munson return -EINVAL; 3624adbe8726SEric B Munson 362517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 362617743798SMuchun Song &tmp); 362708d4a246SMichal Hocko if (ret) 362808d4a246SMichal Hocko goto out; 3629e5ff2159SAndi Kleen 3630e5ff2159SAndi Kleen if (write) { 3631db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3632e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3633db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3634e5ff2159SAndi Kleen } 363508d4a246SMichal Hocko out: 363608d4a246SMichal Hocko return ret; 3637a3d0c6aaSNishanth Aravamudan } 3638a3d0c6aaSNishanth Aravamudan 36391da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 36401da177e4SLinus Torvalds 3641e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 36421da177e4SLinus Torvalds { 3643fcb2b0c5SRoman Gushchin struct hstate *h; 3644fcb2b0c5SRoman Gushchin unsigned long total = 0; 3645fcb2b0c5SRoman Gushchin 3646457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3647457c1b27SNishanth Aravamudan return; 3648fcb2b0c5SRoman Gushchin 3649fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3650fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3651fcb2b0c5SRoman Gushchin 3652aca78307SMiaohe Lin total += huge_page_size(h) * count; 3653fcb2b0c5SRoman Gushchin 3654fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3655e1759c21SAlexey Dobriyan seq_printf(m, 36561da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 36571da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3658b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 36597893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 36604f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3661fcb2b0c5SRoman Gushchin count, 3662a5516438SAndi Kleen h->free_huge_pages, 3663a5516438SAndi Kleen h->resv_huge_pages, 3664a5516438SAndi Kleen h->surplus_huge_pages, 3665aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3666fcb2b0c5SRoman Gushchin } 3667fcb2b0c5SRoman Gushchin 3668aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 36691da177e4SLinus Torvalds } 36701da177e4SLinus Torvalds 36717981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 36721da177e4SLinus Torvalds { 3673a5516438SAndi Kleen struct hstate *h = &default_hstate; 36747981593bSJoe Perches 3675457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3676457c1b27SNishanth Aravamudan return 0; 36777981593bSJoe Perches 36787981593bSJoe Perches return sysfs_emit_at(buf, len, 36791da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3680a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3681a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3682a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3683a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3684a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 36851da177e4SLinus Torvalds } 36861da177e4SLinus Torvalds 3687949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3688949f7ec5SDavid Rientjes { 3689949f7ec5SDavid Rientjes struct hstate *h; 3690949f7ec5SDavid Rientjes int nid; 3691949f7ec5SDavid Rientjes 3692457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3693457c1b27SNishanth Aravamudan return; 3694457c1b27SNishanth Aravamudan 3695949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3696949f7ec5SDavid Rientjes for_each_hstate(h) 3697949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3698949f7ec5SDavid Rientjes nid, 3699949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3700949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3701949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3702aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3703949f7ec5SDavid Rientjes } 3704949f7ec5SDavid Rientjes 37055d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 37065d317b2bSNaoya Horiguchi { 37075d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 37085d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 37095d317b2bSNaoya Horiguchi } 37105d317b2bSNaoya Horiguchi 37111da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 37121da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 37131da177e4SLinus Torvalds { 3714d0028588SWanpeng Li struct hstate *h; 3715d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3716d0028588SWanpeng Li 3717d0028588SWanpeng Li for_each_hstate(h) 3718d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3719d0028588SWanpeng Li return nr_total_pages; 37201da177e4SLinus Torvalds } 37211da177e4SLinus Torvalds 3722a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3723fc1b8a73SMel Gorman { 3724fc1b8a73SMel Gorman int ret = -ENOMEM; 3725fc1b8a73SMel Gorman 37260aa7f354SMiaohe Lin if (!delta) 37270aa7f354SMiaohe Lin return 0; 37280aa7f354SMiaohe Lin 3729db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3730fc1b8a73SMel Gorman /* 3731fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3732fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3733fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3734fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3735fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3736fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3737fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3738fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3739fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3740fc1b8a73SMel Gorman * 3741fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3742fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3743fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3744fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3745fc1b8a73SMel Gorman * semantics that cpuset has. 37468ca39e68SMuchun Song * 37478ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 37488ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 37498ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 37508ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 37518ca39e68SMuchun Song * above. 3752fc1b8a73SMel Gorman */ 3753fc1b8a73SMel Gorman if (delta > 0) { 3754a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3755fc1b8a73SMel Gorman goto out; 3756fc1b8a73SMel Gorman 37578ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3758a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3759fc1b8a73SMel Gorman goto out; 3760fc1b8a73SMel Gorman } 3761fc1b8a73SMel Gorman } 3762fc1b8a73SMel Gorman 3763fc1b8a73SMel Gorman ret = 0; 3764fc1b8a73SMel Gorman if (delta < 0) 3765a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3766fc1b8a73SMel Gorman 3767fc1b8a73SMel Gorman out: 3768db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3769fc1b8a73SMel Gorman return ret; 3770fc1b8a73SMel Gorman } 3771fc1b8a73SMel Gorman 377284afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 377384afd99bSAndy Whitcroft { 3774f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 377584afd99bSAndy Whitcroft 377684afd99bSAndy Whitcroft /* 377784afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 377884afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 377984afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 378025985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 378184afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 378284afd99bSAndy Whitcroft * new reference here without additional locking. 378384afd99bSAndy Whitcroft */ 37844e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3785f522c3acSJoonsoo Kim kref_get(&resv->refs); 378684afd99bSAndy Whitcroft } 378784afd99bSAndy Whitcroft 3788a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3789a1e78772SMel Gorman { 3790a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3791f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 379290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 37934e35f483SJoonsoo Kim unsigned long reserve, start, end; 37941c5ecae3SMike Kravetz long gbl_reserve; 379584afd99bSAndy Whitcroft 37964e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 37974e35f483SJoonsoo Kim return; 37984e35f483SJoonsoo Kim 3799a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3800a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 380184afd99bSAndy Whitcroft 38024e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3803e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 38047251ff78SAdam Litke if (reserve) { 38051c5ecae3SMike Kravetz /* 38061c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 38071c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 38081c5ecae3SMike Kravetz */ 38091c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 38101c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 38117251ff78SAdam Litke } 3812e9fe92aeSMina Almasry 3813e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3814a1e78772SMel Gorman } 3815a1e78772SMel Gorman 381631383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 381731383c68SDan Williams { 381831383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 381931383c68SDan Williams return -EINVAL; 382031383c68SDan Williams return 0; 382131383c68SDan Williams } 382231383c68SDan Williams 382305ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 382405ea8860SDan Williams { 3825aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 382605ea8860SDan Williams } 382705ea8860SDan Williams 38281da177e4SLinus Torvalds /* 38291da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 38301da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 38316c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 38321da177e4SLinus Torvalds * this far. 38331da177e4SLinus Torvalds */ 3834b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 38351da177e4SLinus Torvalds { 38361da177e4SLinus Torvalds BUG(); 3837d0217ac0SNick Piggin return 0; 38381da177e4SLinus Torvalds } 38391da177e4SLinus Torvalds 3840eec3636aSJane Chu /* 3841eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3842eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3843eec3636aSJane Chu * This is because under System V memory model, mappings created via 3844eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3845eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3846eec3636aSJane Chu */ 3847f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3848d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 384984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3850a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3851dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 385205ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 38531da177e4SLinus Torvalds }; 38541da177e4SLinus Torvalds 38551e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 38561e8f889bSDavid Gibson int writable) 385763551ae0SDavid Gibson { 385863551ae0SDavid Gibson pte_t entry; 385963551ae0SDavid Gibson 38601e8f889bSDavid Gibson if (writable) { 3861106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3862106c992aSGerald Schaefer vma->vm_page_prot))); 386363551ae0SDavid Gibson } else { 3864106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3865106c992aSGerald Schaefer vma->vm_page_prot)); 386663551ae0SDavid Gibson } 386763551ae0SDavid Gibson entry = pte_mkyoung(entry); 386863551ae0SDavid Gibson entry = pte_mkhuge(entry); 3869d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 387063551ae0SDavid Gibson 387163551ae0SDavid Gibson return entry; 387263551ae0SDavid Gibson } 387363551ae0SDavid Gibson 38741e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 38751e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 38761e8f889bSDavid Gibson { 38771e8f889bSDavid Gibson pte_t entry; 38781e8f889bSDavid Gibson 3879106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 388032f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 38814b3073e1SRussell King update_mmu_cache(vma, address, ptep); 38821e8f889bSDavid Gibson } 38831e8f889bSDavid Gibson 3884d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 38854a705fefSNaoya Horiguchi { 38864a705fefSNaoya Horiguchi swp_entry_t swp; 38874a705fefSNaoya Horiguchi 38884a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3889d5ed7444SAneesh Kumar K.V return false; 38904a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3891d79d176aSBaoquan He if (is_migration_entry(swp)) 3892d5ed7444SAneesh Kumar K.V return true; 38934a705fefSNaoya Horiguchi else 3894d5ed7444SAneesh Kumar K.V return false; 38954a705fefSNaoya Horiguchi } 38964a705fefSNaoya Horiguchi 38973e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 38984a705fefSNaoya Horiguchi { 38994a705fefSNaoya Horiguchi swp_entry_t swp; 39004a705fefSNaoya Horiguchi 39014a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 39023e5c3600SBaoquan He return false; 39034a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3904d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 39053e5c3600SBaoquan He return true; 39064a705fefSNaoya Horiguchi else 39073e5c3600SBaoquan He return false; 39084a705fefSNaoya Horiguchi } 39091e8f889bSDavid Gibson 39104eae4efaSPeter Xu static void 39114eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 39124eae4efaSPeter Xu struct page *new_page) 39134eae4efaSPeter Xu { 39144eae4efaSPeter Xu __SetPageUptodate(new_page); 39154eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 39164eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 39174eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 39184eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 39194eae4efaSPeter Xu SetHPageMigratable(new_page); 39204eae4efaSPeter Xu } 39214eae4efaSPeter Xu 392263551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 392363551ae0SDavid Gibson struct vm_area_struct *vma) 392463551ae0SDavid Gibson { 39255e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 392663551ae0SDavid Gibson struct page *ptepage; 39271c59827dSHugh Dickins unsigned long addr; 3928ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 3929a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3930a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 39314eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 3932c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3933ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3934e8569dd2SAndreas Sandberg int ret = 0; 39351e8f889bSDavid Gibson 3936ac46d4f3SJérôme Glisse if (cow) { 39377269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 39386f4f13e8SJérôme Glisse vma->vm_start, 3939ac46d4f3SJérôme Glisse vma->vm_end); 3940ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3941c0d0381aSMike Kravetz } else { 3942c0d0381aSMike Kravetz /* 3943c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3944c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3945c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3946c0d0381aSMike Kravetz * huge_pmd_unshare. 3947c0d0381aSMike Kravetz */ 3948c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3949ac46d4f3SJérôme Glisse } 3950e8569dd2SAndreas Sandberg 3951a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3952cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 39537868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3954c74df32cSHugh Dickins if (!src_pte) 3955c74df32cSHugh Dickins continue; 3956aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 3957e8569dd2SAndreas Sandberg if (!dst_pte) { 3958e8569dd2SAndreas Sandberg ret = -ENOMEM; 3959e8569dd2SAndreas Sandberg break; 3960e8569dd2SAndreas Sandberg } 3961c5c99429SLarry Woodman 39625e41540cSMike Kravetz /* 39635e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 39645e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 39655e41540cSMike Kravetz * 39665e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 39675e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 39685e41540cSMike Kravetz * Check here before taking page table lock, and once again 39695e41540cSMike Kravetz * after taking the lock below. 39705e41540cSMike Kravetz */ 39715e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 39725e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3973c5c99429SLarry Woodman continue; 3974c5c99429SLarry Woodman 3975cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3976cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3977cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 39784a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 39795e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 39804eae4efaSPeter Xu again: 39815e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 39825e41540cSMike Kravetz /* 39835e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 39845e41540cSMike Kravetz * unlikely case dst entry !none as this implies 39855e41540cSMike Kravetz * sharing with another vma. 39865e41540cSMike Kravetz */ 39874a705fefSNaoya Horiguchi ; 39884a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 39894a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 39904a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 39914a705fefSNaoya Horiguchi 39924a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 39934a705fefSNaoya Horiguchi /* 39944a705fefSNaoya Horiguchi * COW mappings require pages in both 39954a705fefSNaoya Horiguchi * parent and child to be set to read. 39964a705fefSNaoya Horiguchi */ 39974a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 39984a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3999e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4000e5251fd4SPunit Agrawal entry, sz); 40014a705fefSNaoya Horiguchi } 4002e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 40034a705fefSNaoya Horiguchi } else { 40044eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 40054eae4efaSPeter Xu ptepage = pte_page(entry); 40064eae4efaSPeter Xu get_page(ptepage); 40074eae4efaSPeter Xu 40084eae4efaSPeter Xu /* 40094eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 40104eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 40114eae4efaSPeter Xu * COW earlier during fork. 40124eae4efaSPeter Xu * 40134eae4efaSPeter Xu * When pre-allocating the page or copying data, we 40144eae4efaSPeter Xu * need to be without the pgtable locks since we could 40154eae4efaSPeter Xu * sleep during the process. 40164eae4efaSPeter Xu */ 40174eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 40184eae4efaSPeter Xu pte_t src_pte_old = entry; 40194eae4efaSPeter Xu struct page *new; 40204eae4efaSPeter Xu 40214eae4efaSPeter Xu spin_unlock(src_ptl); 40224eae4efaSPeter Xu spin_unlock(dst_ptl); 40234eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 40244eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 40254eae4efaSPeter Xu if (IS_ERR(new)) { 40264eae4efaSPeter Xu put_page(ptepage); 40274eae4efaSPeter Xu ret = PTR_ERR(new); 40284eae4efaSPeter Xu break; 40294eae4efaSPeter Xu } 40304eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 40314eae4efaSPeter Xu npages); 40324eae4efaSPeter Xu put_page(ptepage); 40334eae4efaSPeter Xu 40344eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 40354eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 40364eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 40374eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 40384eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 40394eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 40404eae4efaSPeter Xu put_page(new); 40414eae4efaSPeter Xu /* dst_entry won't change as in child */ 40424eae4efaSPeter Xu goto again; 40434eae4efaSPeter Xu } 40444eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 40454eae4efaSPeter Xu spin_unlock(src_ptl); 40464eae4efaSPeter Xu spin_unlock(dst_ptl); 40474eae4efaSPeter Xu continue; 40484eae4efaSPeter Xu } 40494eae4efaSPeter Xu 405034ee645eSJoerg Roedel if (cow) { 40510f10851eSJérôme Glisse /* 40520f10851eSJérôme Glisse * No need to notify as we are downgrading page 40530f10851eSJérôme Glisse * table protection not changing it to point 40540f10851eSJérôme Glisse * to a new page. 40550f10851eSJérôme Glisse * 4056ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 40570f10851eSJérôme Glisse */ 40587f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 405934ee645eSJoerg Roedel } 40604eae4efaSPeter Xu 406153f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 406263551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 40634eae4efaSPeter Xu hugetlb_count_add(npages, dst); 40641c59827dSHugh Dickins } 4065cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4066cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 406763551ae0SDavid Gibson } 406863551ae0SDavid Gibson 4069e8569dd2SAndreas Sandberg if (cow) 4070ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4071c0d0381aSMike Kravetz else 4072c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4073e8569dd2SAndreas Sandberg 4074e8569dd2SAndreas Sandberg return ret; 407563551ae0SDavid Gibson } 407663551ae0SDavid Gibson 407724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 407824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 407924669e58SAneesh Kumar K.V struct page *ref_page) 408063551ae0SDavid Gibson { 408163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 408263551ae0SDavid Gibson unsigned long address; 4083c7546f8fSDavid Gibson pte_t *ptep; 408463551ae0SDavid Gibson pte_t pte; 4085cb900f41SKirill A. Shutemov spinlock_t *ptl; 408663551ae0SDavid Gibson struct page *page; 4087a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4088a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4089ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4090a5516438SAndi Kleen 409163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4092a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4093a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 409463551ae0SDavid Gibson 409507e32661SAneesh Kumar K.V /* 409607e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 409707e32661SAneesh Kumar K.V * to huge page. 409807e32661SAneesh Kumar K.V */ 4099ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 410024669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4101dff11abeSMike Kravetz 4102dff11abeSMike Kravetz /* 4103dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4104dff11abeSMike Kravetz */ 41056f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 41066f4f13e8SJérôme Glisse end); 4107ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4108ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4109569f48b8SHillf Danton address = start; 4110569f48b8SHillf Danton for (; address < end; address += sz) { 41117868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4112c7546f8fSDavid Gibson if (!ptep) 4113c7546f8fSDavid Gibson continue; 4114c7546f8fSDavid Gibson 4115cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 411634ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 411731d49da5SAneesh Kumar K.V spin_unlock(ptl); 4118dff11abeSMike Kravetz /* 4119dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 4120dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 4121dff11abeSMike Kravetz */ 412231d49da5SAneesh Kumar K.V continue; 412331d49da5SAneesh Kumar K.V } 412439dde65cSChen, Kenneth W 41256629326bSHillf Danton pte = huge_ptep_get(ptep); 412631d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 412731d49da5SAneesh Kumar K.V spin_unlock(ptl); 412831d49da5SAneesh Kumar K.V continue; 412931d49da5SAneesh Kumar K.V } 41306629326bSHillf Danton 41316629326bSHillf Danton /* 41329fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 41339fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 41346629326bSHillf Danton */ 41359fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 41369386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 413731d49da5SAneesh Kumar K.V spin_unlock(ptl); 413831d49da5SAneesh Kumar K.V continue; 41398c4894c6SNaoya Horiguchi } 41406629326bSHillf Danton 41416629326bSHillf Danton page = pte_page(pte); 414204f2cbe3SMel Gorman /* 414304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 414404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 414504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 414604f2cbe3SMel Gorman */ 414704f2cbe3SMel Gorman if (ref_page) { 414831d49da5SAneesh Kumar K.V if (page != ref_page) { 414931d49da5SAneesh Kumar K.V spin_unlock(ptl); 415031d49da5SAneesh Kumar K.V continue; 415131d49da5SAneesh Kumar K.V } 415204f2cbe3SMel Gorman /* 415304f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 415404f2cbe3SMel Gorman * future faults in this VMA will fail rather than 415504f2cbe3SMel Gorman * looking like data was lost 415604f2cbe3SMel Gorman */ 415704f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 415804f2cbe3SMel Gorman } 415904f2cbe3SMel Gorman 4160c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4161b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4162106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 41636649a386SKen Chen set_page_dirty(page); 41649e81130bSHillf Danton 41655d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4166d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 416731d49da5SAneesh Kumar K.V 4168cb900f41SKirill A. Shutemov spin_unlock(ptl); 4169e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 417024669e58SAneesh Kumar K.V /* 417131d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 417224669e58SAneesh Kumar K.V */ 417331d49da5SAneesh Kumar K.V if (ref_page) 417431d49da5SAneesh Kumar K.V break; 4175fe1668aeSChen, Kenneth W } 4176ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 417724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 41781da177e4SLinus Torvalds } 417963551ae0SDavid Gibson 4180d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4181d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4182d833352aSMel Gorman unsigned long end, struct page *ref_page) 4183d833352aSMel Gorman { 4184d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4185d833352aSMel Gorman 4186d833352aSMel Gorman /* 4187d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4188d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4189d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4190d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4191d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4192c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4193d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4194c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4195d833352aSMel Gorman */ 4196d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4197d833352aSMel Gorman } 4198d833352aSMel Gorman 4199502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 420004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4201502717f4SChen, Kenneth W { 420224669e58SAneesh Kumar K.V struct mmu_gather tlb; 4203dff11abeSMike Kravetz 4204a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 420524669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4206ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4207502717f4SChen, Kenneth W } 4208502717f4SChen, Kenneth W 420904f2cbe3SMel Gorman /* 421004f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4211578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 421204f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 421304f2cbe3SMel Gorman * same region. 421404f2cbe3SMel Gorman */ 42152f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 42162a4b3dedSHarvey Harrison struct page *page, unsigned long address) 421704f2cbe3SMel Gorman { 42187526674dSAdam Litke struct hstate *h = hstate_vma(vma); 421904f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 422004f2cbe3SMel Gorman struct address_space *mapping; 422104f2cbe3SMel Gorman pgoff_t pgoff; 422204f2cbe3SMel Gorman 422304f2cbe3SMel Gorman /* 422404f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 422504f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 422604f2cbe3SMel Gorman */ 42277526674dSAdam Litke address = address & huge_page_mask(h); 422836e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 422936e4f20aSMichal Hocko vma->vm_pgoff; 423093c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 423104f2cbe3SMel Gorman 42324eb2b1dcSMel Gorman /* 42334eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 42344eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 42354eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 42364eb2b1dcSMel Gorman */ 423783cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 42386b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 423904f2cbe3SMel Gorman /* Do not unmap the current VMA */ 424004f2cbe3SMel Gorman if (iter_vma == vma) 424104f2cbe3SMel Gorman continue; 424204f2cbe3SMel Gorman 424304f2cbe3SMel Gorman /* 42442f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 42452f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 42462f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 42472f84a899SMel Gorman */ 42482f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 42492f84a899SMel Gorman continue; 42502f84a899SMel Gorman 42512f84a899SMel Gorman /* 425204f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 425304f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 425404f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 425504f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 425604f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 425704f2cbe3SMel Gorman */ 425804f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 425924669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 426024669e58SAneesh Kumar K.V address + huge_page_size(h), page); 426104f2cbe3SMel Gorman } 426283cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 426304f2cbe3SMel Gorman } 426404f2cbe3SMel Gorman 42650fe6e20bSNaoya Horiguchi /* 42660fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4267ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4268ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4269ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 42700fe6e20bSNaoya Horiguchi */ 42712b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4272974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4273cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 42741e8f889bSDavid Gibson { 42753999f52eSAneesh Kumar K.V pte_t pte; 4276a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42771e8f889bSDavid Gibson struct page *old_page, *new_page; 42782b740303SSouptick Joarder int outside_reserve = 0; 42792b740303SSouptick Joarder vm_fault_t ret = 0; 4280974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4281ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 42821e8f889bSDavid Gibson 42833999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 42841e8f889bSDavid Gibson old_page = pte_page(pte); 42851e8f889bSDavid Gibson 428604f2cbe3SMel Gorman retry_avoidcopy: 42871e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 42881e8f889bSDavid Gibson * and just make the page writable */ 428937a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 42905a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 42915b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 429283c54070SNick Piggin return 0; 42931e8f889bSDavid Gibson } 42941e8f889bSDavid Gibson 429504f2cbe3SMel Gorman /* 429604f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 429704f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 429804f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 429904f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 430004f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 430104f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 430204f2cbe3SMel Gorman * of the full address range. 430304f2cbe3SMel Gorman */ 43045944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 430504f2cbe3SMel Gorman old_page != pagecache_page) 430604f2cbe3SMel Gorman outside_reserve = 1; 430704f2cbe3SMel Gorman 430809cbfeafSKirill A. Shutemov get_page(old_page); 4309b76c8cfbSLarry Woodman 4310ad4404a2SDavidlohr Bueso /* 4311ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4312ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4313ad4404a2SDavidlohr Bueso */ 4314cb900f41SKirill A. Shutemov spin_unlock(ptl); 43155b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 43161e8f889bSDavid Gibson 43172fc39cecSAdam Litke if (IS_ERR(new_page)) { 431804f2cbe3SMel Gorman /* 431904f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 432004f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 432104f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 432204f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 432304f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 432404f2cbe3SMel Gorman */ 432504f2cbe3SMel Gorman if (outside_reserve) { 4326e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4327e7dd91c4SMike Kravetz pgoff_t idx; 4328e7dd91c4SMike Kravetz u32 hash; 4329e7dd91c4SMike Kravetz 433009cbfeafSKirill A. Shutemov put_page(old_page); 433104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4332e7dd91c4SMike Kravetz /* 4333e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4334e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4335e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4336e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4337e7dd91c4SMike Kravetz * PMD sharing. 4338e7dd91c4SMike Kravetz * 4339e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4340e7dd91c4SMike Kravetz */ 4341e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4342e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4343e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4344e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4345e7dd91c4SMike Kravetz 43465b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4347e7dd91c4SMike Kravetz 4348e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4349e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4350cb900f41SKirill A. Shutemov spin_lock(ptl); 43515b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4352a9af0c5dSNaoya Horiguchi if (likely(ptep && 4353a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 435404f2cbe3SMel Gorman goto retry_avoidcopy; 4355a734bcc8SHillf Danton /* 4356cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4357cb900f41SKirill A. Shutemov * lock, and our job is done. 4358a734bcc8SHillf Danton */ 4359a734bcc8SHillf Danton return 0; 436004f2cbe3SMel Gorman } 436104f2cbe3SMel Gorman 43622b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4363ad4404a2SDavidlohr Bueso goto out_release_old; 43641e8f889bSDavid Gibson } 43651e8f889bSDavid Gibson 43660fe6e20bSNaoya Horiguchi /* 43670fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 43680fe6e20bSNaoya Horiguchi * anon_vma prepared. 43690fe6e20bSNaoya Horiguchi */ 437044e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4371ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4372ad4404a2SDavidlohr Bueso goto out_release_all; 437344e2aa93SDean Nelson } 43740fe6e20bSNaoya Horiguchi 4375974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 437647ad8475SAndrea Arcangeli pages_per_huge_page(h)); 43770ed361deSNick Piggin __SetPageUptodate(new_page); 43781e8f889bSDavid Gibson 43797269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 43806f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4381ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4382ad4404a2SDavidlohr Bueso 4383b76c8cfbSLarry Woodman /* 4384cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4385b76c8cfbSLarry Woodman * before the page tables are altered 4386b76c8cfbSLarry Woodman */ 4387cb900f41SKirill A. Shutemov spin_lock(ptl); 43885b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4389a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4390d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 439107443a85SJoonsoo Kim 43921e8f889bSDavid Gibson /* Break COW */ 43935b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4394ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 43955b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 43961e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4397d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 43985b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 43998f251a3dSMike Kravetz SetHPageMigratable(new_page); 44001e8f889bSDavid Gibson /* Make the old page be freed below */ 44011e8f889bSDavid Gibson new_page = old_page; 44021e8f889bSDavid Gibson } 4403cb900f41SKirill A. Shutemov spin_unlock(ptl); 4404ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4405ad4404a2SDavidlohr Bueso out_release_all: 44065b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 440709cbfeafSKirill A. Shutemov put_page(new_page); 4408ad4404a2SDavidlohr Bueso out_release_old: 440909cbfeafSKirill A. Shutemov put_page(old_page); 44108312034fSJoonsoo Kim 4411ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4412ad4404a2SDavidlohr Bueso return ret; 44131e8f889bSDavid Gibson } 44141e8f889bSDavid Gibson 441504f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4416a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4417a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 441804f2cbe3SMel Gorman { 441904f2cbe3SMel Gorman struct address_space *mapping; 4420e7c4b0bfSAndy Whitcroft pgoff_t idx; 442104f2cbe3SMel Gorman 442204f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4423a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 442404f2cbe3SMel Gorman 442504f2cbe3SMel Gorman return find_lock_page(mapping, idx); 442604f2cbe3SMel Gorman } 442704f2cbe3SMel Gorman 44283ae77f43SHugh Dickins /* 44293ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 44303ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 44313ae77f43SHugh Dickins */ 44323ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 44332a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 44342a15efc9SHugh Dickins { 44352a15efc9SHugh Dickins struct address_space *mapping; 44362a15efc9SHugh Dickins pgoff_t idx; 44372a15efc9SHugh Dickins struct page *page; 44382a15efc9SHugh Dickins 44392a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 44402a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 44412a15efc9SHugh Dickins 44422a15efc9SHugh Dickins page = find_get_page(mapping, idx); 44432a15efc9SHugh Dickins if (page) 44442a15efc9SHugh Dickins put_page(page); 44452a15efc9SHugh Dickins return page != NULL; 44462a15efc9SHugh Dickins } 44472a15efc9SHugh Dickins 4448ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4449ab76ad54SMike Kravetz pgoff_t idx) 4450ab76ad54SMike Kravetz { 4451ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4452ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4453ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4454ab76ad54SMike Kravetz 4455ab76ad54SMike Kravetz if (err) 4456ab76ad54SMike Kravetz return err; 4457d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4458ab76ad54SMike Kravetz 445922146c3cSMike Kravetz /* 446022146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 446122146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 446222146c3cSMike Kravetz */ 446322146c3cSMike Kravetz set_page_dirty(page); 446422146c3cSMike Kravetz 4465ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4466ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4467ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4468ab76ad54SMike Kravetz return 0; 4469ab76ad54SMike Kravetz } 4470ab76ad54SMike Kravetz 44717677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 44727677f7fdSAxel Rasmussen struct address_space *mapping, 44737677f7fdSAxel Rasmussen pgoff_t idx, 44747677f7fdSAxel Rasmussen unsigned int flags, 44757677f7fdSAxel Rasmussen unsigned long haddr, 44767677f7fdSAxel Rasmussen unsigned long reason) 44777677f7fdSAxel Rasmussen { 44787677f7fdSAxel Rasmussen vm_fault_t ret; 44797677f7fdSAxel Rasmussen u32 hash; 44807677f7fdSAxel Rasmussen struct vm_fault vmf = { 44817677f7fdSAxel Rasmussen .vma = vma, 44827677f7fdSAxel Rasmussen .address = haddr, 44837677f7fdSAxel Rasmussen .flags = flags, 44847677f7fdSAxel Rasmussen 44857677f7fdSAxel Rasmussen /* 44867677f7fdSAxel Rasmussen * Hard to debug if it ends up being 44877677f7fdSAxel Rasmussen * used by a callee that assumes 44887677f7fdSAxel Rasmussen * something about the other 44897677f7fdSAxel Rasmussen * uninitialized fields... same as in 44907677f7fdSAxel Rasmussen * memory.c 44917677f7fdSAxel Rasmussen */ 44927677f7fdSAxel Rasmussen }; 44937677f7fdSAxel Rasmussen 44947677f7fdSAxel Rasmussen /* 44957677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 44967677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 44977677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 44987677f7fdSAxel Rasmussen */ 44997677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 45007677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 45017677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 45027677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 45037677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 45047677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 45057677f7fdSAxel Rasmussen 45067677f7fdSAxel Rasmussen return ret; 45077677f7fdSAxel Rasmussen } 45087677f7fdSAxel Rasmussen 45092b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 45102b740303SSouptick Joarder struct vm_area_struct *vma, 45118382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4512788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4513ac9b9c66SHugh Dickins { 4514a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45152b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4516409eb8c2SHillf Danton int anon_rmap = 0; 45174c887265SAdam Litke unsigned long size; 45184c887265SAdam Litke struct page *page; 45191e8f889bSDavid Gibson pte_t new_pte; 4520cb900f41SKirill A. Shutemov spinlock_t *ptl; 4521285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4522cb6acd01SMike Kravetz bool new_page = false; 45234c887265SAdam Litke 452404f2cbe3SMel Gorman /* 452504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 452604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 452725985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 452804f2cbe3SMel Gorman */ 452904f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4530910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 453104f2cbe3SMel Gorman current->pid); 453204f2cbe3SMel Gorman return ret; 453304f2cbe3SMel Gorman } 453404f2cbe3SMel Gorman 45354c887265SAdam Litke /* 453687bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 453787bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 453887bf91d3SMike Kravetz * once for faults beyond end of file. 45394c887265SAdam Litke */ 4540a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4541ebed4bfcSHugh Dickins if (idx >= size) 4542ebed4bfcSHugh Dickins goto out; 45431a1aad8aSMike Kravetz 454487bf91d3SMike Kravetz retry: 454587bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 454687bf91d3SMike Kravetz if (!page) { 45477677f7fdSAxel Rasmussen /* Check for page in userfault range */ 45481a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 45497677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 45507677f7fdSAxel Rasmussen flags, haddr, 45517677f7fdSAxel Rasmussen VM_UFFD_MISSING); 45521a1aad8aSMike Kravetz goto out; 45531a1aad8aSMike Kravetz } 45541a1aad8aSMike Kravetz 4555285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 45562fc39cecSAdam Litke if (IS_ERR(page)) { 45574643d67eSMike Kravetz /* 45584643d67eSMike Kravetz * Returning error will result in faulting task being 45594643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 45604643d67eSMike Kravetz * tasks from racing to fault in the same page which 45614643d67eSMike Kravetz * could result in false unable to allocate errors. 45624643d67eSMike Kravetz * Page migration does not take the fault mutex, but 45634643d67eSMike Kravetz * does a clear then write of pte's under page table 45644643d67eSMike Kravetz * lock. Page fault code could race with migration, 45654643d67eSMike Kravetz * notice the clear pte and try to allocate a page 45664643d67eSMike Kravetz * here. Before returning error, get ptl and make 45674643d67eSMike Kravetz * sure there really is no pte entry. 45684643d67eSMike Kravetz */ 45694643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 45704643d67eSMike Kravetz ret = 0; 4571d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 45722b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 4573d83e6c8aSMiaohe Lin spin_unlock(ptl); 45746bda666aSChristoph Lameter goto out; 45756bda666aSChristoph Lameter } 457647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 45770ed361deSNick Piggin __SetPageUptodate(page); 4578cb6acd01SMike Kravetz new_page = true; 4579ac9b9c66SHugh Dickins 4580f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4581ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 45826bda666aSChristoph Lameter if (err) { 45836bda666aSChristoph Lameter put_page(page); 45846bda666aSChristoph Lameter if (err == -EEXIST) 45856bda666aSChristoph Lameter goto retry; 45866bda666aSChristoph Lameter goto out; 45876bda666aSChristoph Lameter } 458823be7468SMel Gorman } else { 45896bda666aSChristoph Lameter lock_page(page); 45900fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 45910fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 45920fe6e20bSNaoya Horiguchi goto backout_unlocked; 459323be7468SMel Gorman } 4594409eb8c2SHillf Danton anon_rmap = 1; 45950fe6e20bSNaoya Horiguchi } 45960fe6e20bSNaoya Horiguchi } else { 459757303d80SAndy Whitcroft /* 4598998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4599998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4600998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4601fd6a03edSNaoya Horiguchi */ 4602fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 46030eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4604972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4605fd6a03edSNaoya Horiguchi goto backout_unlocked; 46066bda666aSChristoph Lameter } 46077677f7fdSAxel Rasmussen 46087677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 46097677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 46107677f7fdSAxel Rasmussen unlock_page(page); 46117677f7fdSAxel Rasmussen put_page(page); 46127677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 46137677f7fdSAxel Rasmussen flags, haddr, 46147677f7fdSAxel Rasmussen VM_UFFD_MINOR); 46157677f7fdSAxel Rasmussen goto out; 46167677f7fdSAxel Rasmussen } 4617998b4382SNaoya Horiguchi } 46181e8f889bSDavid Gibson 461957303d80SAndy Whitcroft /* 462057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 462157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 462257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 462357303d80SAndy Whitcroft * the spinlock. 462457303d80SAndy Whitcroft */ 46255e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4626285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 46272b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 46282b26736cSAndy Whitcroft goto backout_unlocked; 46292b26736cSAndy Whitcroft } 46305e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4631285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 46325e911373SMike Kravetz } 463357303d80SAndy Whitcroft 46348bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 463583c54070SNick Piggin ret = 0; 46367f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 46374c887265SAdam Litke goto backout; 46384c887265SAdam Litke 463907443a85SJoonsoo Kim if (anon_rmap) { 4640d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4641285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4642ac714904SChoi Gi-yong } else 464353f9263bSKirill A. Shutemov page_dup_rmap(page, true); 46441e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 46451e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4646285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 46471e8f889bSDavid Gibson 46485d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4649788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 46501e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4651974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 46521e8f889bSDavid Gibson } 46531e8f889bSDavid Gibson 4654cb900f41SKirill A. Shutemov spin_unlock(ptl); 4655cb6acd01SMike Kravetz 4656cb6acd01SMike Kravetz /* 46578f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 46588f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 46598f251a3dSMike Kravetz * been isolated for migration. 4660cb6acd01SMike Kravetz */ 4661cb6acd01SMike Kravetz if (new_page) 46628f251a3dSMike Kravetz SetHPageMigratable(page); 4663cb6acd01SMike Kravetz 46644c887265SAdam Litke unlock_page(page); 46654c887265SAdam Litke out: 4666ac9b9c66SHugh Dickins return ret; 46674c887265SAdam Litke 46684c887265SAdam Litke backout: 4669cb900f41SKirill A. Shutemov spin_unlock(ptl); 46702b26736cSAndy Whitcroft backout_unlocked: 46714c887265SAdam Litke unlock_page(page); 4672285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 46734c887265SAdam Litke put_page(page); 46744c887265SAdam Litke goto out; 4675ac9b9c66SHugh Dickins } 4676ac9b9c66SHugh Dickins 46778382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4678188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 46798382d914SDavidlohr Bueso { 46808382d914SDavidlohr Bueso unsigned long key[2]; 46818382d914SDavidlohr Bueso u32 hash; 46828382d914SDavidlohr Bueso 46838382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 46848382d914SDavidlohr Bueso key[1] = idx; 46858382d914SDavidlohr Bueso 468655254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 46878382d914SDavidlohr Bueso 46888382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 46898382d914SDavidlohr Bueso } 46908382d914SDavidlohr Bueso #else 46918382d914SDavidlohr Bueso /* 46926c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 46938382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 46948382d914SDavidlohr Bueso */ 4695188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 46968382d914SDavidlohr Bueso { 46978382d914SDavidlohr Bueso return 0; 46988382d914SDavidlohr Bueso } 46998382d914SDavidlohr Bueso #endif 47008382d914SDavidlohr Bueso 47012b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4702788c7df4SHugh Dickins unsigned long address, unsigned int flags) 470386e5216fSAdam Litke { 47048382d914SDavidlohr Bueso pte_t *ptep, entry; 4705cb900f41SKirill A. Shutemov spinlock_t *ptl; 47062b740303SSouptick Joarder vm_fault_t ret; 47078382d914SDavidlohr Bueso u32 hash; 47088382d914SDavidlohr Bueso pgoff_t idx; 47090fe6e20bSNaoya Horiguchi struct page *page = NULL; 471057303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4711a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 47128382d914SDavidlohr Bueso struct address_space *mapping; 47130f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4714285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 471586e5216fSAdam Litke 4716285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4717fd6a03edSNaoya Horiguchi if (ptep) { 4718c0d0381aSMike Kravetz /* 4719c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4720c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4721c0d0381aSMike Kravetz * not actually modifying content here. 4722c0d0381aSMike Kravetz */ 4723fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4724290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4725cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4726290408d4SNaoya Horiguchi return 0; 4727290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4728aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4729972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4730b43a9990SMike Kravetz } 47318382d914SDavidlohr Bueso 4732c0d0381aSMike Kravetz /* 4733c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 473487bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 473587bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 473687bf91d3SMike Kravetz * and making the ptep no longer valid. 473787bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4738c0d0381aSMike Kravetz * 4739c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4740c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4741c0d0381aSMike Kravetz * something has changed. 4742c0d0381aSMike Kravetz */ 4743ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4744c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4745aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 4746c0d0381aSMike Kravetz if (!ptep) { 4747c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4748c0d0381aSMike Kravetz return VM_FAULT_OOM; 4749c0d0381aSMike Kravetz } 4750ddeaab32SMike Kravetz 47513935baa9SDavid Gibson /* 47523935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 47533935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 47543935baa9SDavid Gibson * the same page in the page cache. 47553935baa9SDavid Gibson */ 4756c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4757188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4758c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 47598382d914SDavidlohr Bueso 47607f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 47617f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 47628382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4763b4d1d99fSDavid Gibson goto out_mutex; 47643935baa9SDavid Gibson } 476586e5216fSAdam Litke 476683c54070SNick Piggin ret = 0; 47671e8f889bSDavid Gibson 476857303d80SAndy Whitcroft /* 47690f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 47700f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 47717c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 47727c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 47737c8de358SEthon Paul * properly handle it. 47740f792cf9SNaoya Horiguchi */ 47750f792cf9SNaoya Horiguchi if (!pte_present(entry)) 47760f792cf9SNaoya Horiguchi goto out_mutex; 47770f792cf9SNaoya Horiguchi 47780f792cf9SNaoya Horiguchi /* 477957303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 478057303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 478157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 478257303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 478357303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 478457303d80SAndy Whitcroft * consumed. 478557303d80SAndy Whitcroft */ 4786106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4787285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 47882b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4789b4d1d99fSDavid Gibson goto out_mutex; 47902b26736cSAndy Whitcroft } 47915e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4792285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 479357303d80SAndy Whitcroft 4794f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 479557303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4796285b8dcaSHuang Ying vma, haddr); 479757303d80SAndy Whitcroft } 479857303d80SAndy Whitcroft 47990f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 48000fe6e20bSNaoya Horiguchi 48011e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4802b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4803cb900f41SKirill A. Shutemov goto out_ptl; 4804b4d1d99fSDavid Gibson 48050f792cf9SNaoya Horiguchi /* 48060f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 48070f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 48080f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 48090f792cf9SNaoya Horiguchi */ 48100f792cf9SNaoya Horiguchi page = pte_page(entry); 48110f792cf9SNaoya Horiguchi if (page != pagecache_page) 48120f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 48130f792cf9SNaoya Horiguchi need_wait_lock = 1; 48140f792cf9SNaoya Horiguchi goto out_ptl; 48150f792cf9SNaoya Horiguchi } 48160f792cf9SNaoya Horiguchi 48170f792cf9SNaoya Horiguchi get_page(page); 4818b4d1d99fSDavid Gibson 4819788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4820106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4821974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4822cb900f41SKirill A. Shutemov pagecache_page, ptl); 48230f792cf9SNaoya Horiguchi goto out_put_page; 4824b4d1d99fSDavid Gibson } 4825106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4826b4d1d99fSDavid Gibson } 4827b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4828285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4829788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4830285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 48310f792cf9SNaoya Horiguchi out_put_page: 48320f792cf9SNaoya Horiguchi if (page != pagecache_page) 48330f792cf9SNaoya Horiguchi unlock_page(page); 48340f792cf9SNaoya Horiguchi put_page(page); 4835cb900f41SKirill A. Shutemov out_ptl: 4836cb900f41SKirill A. Shutemov spin_unlock(ptl); 483757303d80SAndy Whitcroft 483857303d80SAndy Whitcroft if (pagecache_page) { 483957303d80SAndy Whitcroft unlock_page(pagecache_page); 484057303d80SAndy Whitcroft put_page(pagecache_page); 484157303d80SAndy Whitcroft } 4842b4d1d99fSDavid Gibson out_mutex: 4843c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4844c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 48450f792cf9SNaoya Horiguchi /* 48460f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 48470f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 48480f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 48490f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 48500f792cf9SNaoya Horiguchi * here without taking refcount. 48510f792cf9SNaoya Horiguchi */ 48520f792cf9SNaoya Horiguchi if (need_wait_lock) 48530f792cf9SNaoya Horiguchi wait_on_page_locked(page); 48541e8f889bSDavid Gibson return ret; 485586e5216fSAdam Litke } 485686e5216fSAdam Litke 4857714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 48588fb5debcSMike Kravetz /* 48598fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 48608fb5debcSMike Kravetz * modifications for huge pages. 48618fb5debcSMike Kravetz */ 48628fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 48638fb5debcSMike Kravetz pte_t *dst_pte, 48648fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 48658fb5debcSMike Kravetz unsigned long dst_addr, 48668fb5debcSMike Kravetz unsigned long src_addr, 4867*f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 48688fb5debcSMike Kravetz struct page **pagep) 48698fb5debcSMike Kravetz { 4870*f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 48711e392147SAndrea Arcangeli struct address_space *mapping; 48721e392147SAndrea Arcangeli pgoff_t idx; 48731e392147SAndrea Arcangeli unsigned long size; 48741c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 48758fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 48768fb5debcSMike Kravetz pte_t _dst_pte; 48778fb5debcSMike Kravetz spinlock_t *ptl; 48788fb5debcSMike Kravetz int ret; 48798fb5debcSMike Kravetz struct page *page; 4880*f6191471SAxel Rasmussen int writable; 48818fb5debcSMike Kravetz 4882*f6191471SAxel Rasmussen mapping = dst_vma->vm_file->f_mapping; 4883*f6191471SAxel Rasmussen idx = vma_hugecache_offset(h, dst_vma, dst_addr); 4884*f6191471SAxel Rasmussen 4885*f6191471SAxel Rasmussen if (is_continue) { 4886*f6191471SAxel Rasmussen ret = -EFAULT; 4887*f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 4888*f6191471SAxel Rasmussen if (!page) 4889*f6191471SAxel Rasmussen goto out; 4890*f6191471SAxel Rasmussen } else if (!*pagep) { 48918fb5debcSMike Kravetz ret = -ENOMEM; 48928fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 48938fb5debcSMike Kravetz if (IS_ERR(page)) 48948fb5debcSMike Kravetz goto out; 48958fb5debcSMike Kravetz 48968fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 48978fb5debcSMike Kravetz (const void __user *) src_addr, 4898810a56b9SMike Kravetz pages_per_huge_page(h), false); 48998fb5debcSMike Kravetz 4900c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 49018fb5debcSMike Kravetz if (unlikely(ret)) { 49029e368259SAndrea Arcangeli ret = -ENOENT; 49038fb5debcSMike Kravetz *pagep = page; 49048fb5debcSMike Kravetz /* don't free the page */ 49058fb5debcSMike Kravetz goto out; 49068fb5debcSMike Kravetz } 49078fb5debcSMike Kravetz } else { 49088fb5debcSMike Kravetz page = *pagep; 49098fb5debcSMike Kravetz *pagep = NULL; 49108fb5debcSMike Kravetz } 49118fb5debcSMike Kravetz 49128fb5debcSMike Kravetz /* 49138fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 49148fb5debcSMike Kravetz * preceding stores to the page contents become visible before 49158fb5debcSMike Kravetz * the set_pte_at() write. 49168fb5debcSMike Kravetz */ 49178fb5debcSMike Kravetz __SetPageUptodate(page); 49188fb5debcSMike Kravetz 4919*f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 4920*f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 49211e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 49221e392147SAndrea Arcangeli ret = -EFAULT; 49231e392147SAndrea Arcangeli if (idx >= size) 49241e392147SAndrea Arcangeli goto out_release_nounlock; 49251c9e8defSMike Kravetz 49261e392147SAndrea Arcangeli /* 49271e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 49281e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 49291e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 49301e392147SAndrea Arcangeli * the caller. 49311e392147SAndrea Arcangeli */ 49321c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 49331c9e8defSMike Kravetz if (ret) 49341c9e8defSMike Kravetz goto out_release_nounlock; 49351c9e8defSMike Kravetz } 49361c9e8defSMike Kravetz 49378fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 49388fb5debcSMike Kravetz spin_lock(ptl); 49398fb5debcSMike Kravetz 49401e392147SAndrea Arcangeli /* 49411e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 49421e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 49431e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 49441e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 49451e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 49461e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 49471e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 49481e392147SAndrea Arcangeli */ 49491e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 49501e392147SAndrea Arcangeli ret = -EFAULT; 49511e392147SAndrea Arcangeli if (idx >= size) 49521e392147SAndrea Arcangeli goto out_release_unlock; 49531e392147SAndrea Arcangeli 49548fb5debcSMike Kravetz ret = -EEXIST; 49558fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 49568fb5debcSMike Kravetz goto out_release_unlock; 49578fb5debcSMike Kravetz 49581c9e8defSMike Kravetz if (vm_shared) { 49591c9e8defSMike Kravetz page_dup_rmap(page, true); 49601c9e8defSMike Kravetz } else { 4961d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 49628fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 49631c9e8defSMike Kravetz } 49648fb5debcSMike Kravetz 4965*f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 4966*f6191471SAxel Rasmussen if (is_continue && !vm_shared) 4967*f6191471SAxel Rasmussen writable = 0; 4968*f6191471SAxel Rasmussen else 4969*f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 4970*f6191471SAxel Rasmussen 4971*f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 4972*f6191471SAxel Rasmussen if (writable) 49738fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 49748fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 49758fb5debcSMike Kravetz 49768fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 49778fb5debcSMike Kravetz 49788fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 49798fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 49808fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 49818fb5debcSMike Kravetz 49828fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 49838fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 49848fb5debcSMike Kravetz 49858fb5debcSMike Kravetz spin_unlock(ptl); 4986*f6191471SAxel Rasmussen if (!is_continue) 49878f251a3dSMike Kravetz SetHPageMigratable(page); 4988*f6191471SAxel Rasmussen if (vm_shared || is_continue) 49891c9e8defSMike Kravetz unlock_page(page); 49908fb5debcSMike Kravetz ret = 0; 49918fb5debcSMike Kravetz out: 49928fb5debcSMike Kravetz return ret; 49938fb5debcSMike Kravetz out_release_unlock: 49948fb5debcSMike Kravetz spin_unlock(ptl); 4995*f6191471SAxel Rasmussen if (vm_shared || is_continue) 49961c9e8defSMike Kravetz unlock_page(page); 49975af10dfdSAndrea Arcangeli out_release_nounlock: 49988fb5debcSMike Kravetz put_page(page); 49998fb5debcSMike Kravetz goto out; 50008fb5debcSMike Kravetz } 5001714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 50028fb5debcSMike Kravetz 500382e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 500482e5d378SJoao Martins int refs, struct page **pages, 500582e5d378SJoao Martins struct vm_area_struct **vmas) 500682e5d378SJoao Martins { 500782e5d378SJoao Martins int nr; 500882e5d378SJoao Martins 500982e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 501082e5d378SJoao Martins if (likely(pages)) 501182e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 501282e5d378SJoao Martins if (vmas) 501382e5d378SJoao Martins vmas[nr] = vma; 501482e5d378SJoao Martins } 501582e5d378SJoao Martins } 501682e5d378SJoao Martins 501728a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 501863551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 501928a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 50204f6da934SPeter Xu long i, unsigned int flags, int *locked) 502163551ae0SDavid Gibson { 5022d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5023d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 502428a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5025a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 50260fa5bc40SJoao Martins int err = -EFAULT, refs; 502763551ae0SDavid Gibson 502863551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 502963551ae0SDavid Gibson pte_t *pte; 5030cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 50312a15efc9SHugh Dickins int absent; 503263551ae0SDavid Gibson struct page *page; 503363551ae0SDavid Gibson 50344c887265SAdam Litke /* 503502057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 503602057967SDavid Rientjes * potentially allocating memory. 503702057967SDavid Rientjes */ 5038fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 503902057967SDavid Rientjes remainder = 0; 504002057967SDavid Rientjes break; 504102057967SDavid Rientjes } 504202057967SDavid Rientjes 504302057967SDavid Rientjes /* 50444c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 50452a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 50464c887265SAdam Litke * first, for the page indexing below to work. 5047cb900f41SKirill A. Shutemov * 5048cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 50494c887265SAdam Litke */ 50507868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 50517868a208SPunit Agrawal huge_page_size(h)); 5052cb900f41SKirill A. Shutemov if (pte) 5053cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 50542a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 505563551ae0SDavid Gibson 50562a15efc9SHugh Dickins /* 50572a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 50583ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 50593ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 50603ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 50613ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 50622a15efc9SHugh Dickins */ 50633ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 50643ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5065cb900f41SKirill A. Shutemov if (pte) 5066cb900f41SKirill A. Shutemov spin_unlock(ptl); 50672a15efc9SHugh Dickins remainder = 0; 50682a15efc9SHugh Dickins break; 50692a15efc9SHugh Dickins } 50702a15efc9SHugh Dickins 50719cc3a5bdSNaoya Horiguchi /* 50729cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 50739cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 50749cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 50759cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 50769cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 50779cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 50789cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 50799cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 50809cc3a5bdSNaoya Horiguchi */ 50819cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5082106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5083106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 50842b740303SSouptick Joarder vm_fault_t ret; 508587ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 50864c887265SAdam Litke 5087cb900f41SKirill A. Shutemov if (pte) 5088cb900f41SKirill A. Shutemov spin_unlock(ptl); 508987ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 509087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 50914f6da934SPeter Xu if (locked) 509271335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 509371335f37SPeter Xu FAULT_FLAG_KILLABLE; 509487ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 509587ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 509687ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 509787ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 50984426e945SPeter Xu /* 50994426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 51004426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 51014426e945SPeter Xu */ 510287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 510387ffc118SAndrea Arcangeli } 510487ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 510587ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 51062be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 51071c59827dSHugh Dickins remainder = 0; 51081c59827dSHugh Dickins break; 51091c59827dSHugh Dickins } 511087ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 51114f6da934SPeter Xu if (locked && 51121ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 51134f6da934SPeter Xu *locked = 0; 511487ffc118SAndrea Arcangeli *nr_pages = 0; 511587ffc118SAndrea Arcangeli /* 511687ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 511787ffc118SAndrea Arcangeli * error, it will return zero 511887ffc118SAndrea Arcangeli * instead. 511987ffc118SAndrea Arcangeli * 512087ffc118SAndrea Arcangeli * No need to update "position" as the 512187ffc118SAndrea Arcangeli * caller will not check it after 512287ffc118SAndrea Arcangeli * *nr_pages is set to 0. 512387ffc118SAndrea Arcangeli */ 512487ffc118SAndrea Arcangeli return i; 512587ffc118SAndrea Arcangeli } 512687ffc118SAndrea Arcangeli continue; 512787ffc118SAndrea Arcangeli } 512863551ae0SDavid Gibson 5129a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 51307f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 51318fde12caSLinus Torvalds 51328fde12caSLinus Torvalds /* 5133acbfb087SZhigang Lu * If subpage information not requested, update counters 5134acbfb087SZhigang Lu * and skip the same_page loop below. 5135acbfb087SZhigang Lu */ 5136acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 5137acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 5138acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 5139acbfb087SZhigang Lu vaddr += huge_page_size(h); 5140acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 5141acbfb087SZhigang Lu i += pages_per_huge_page(h); 5142acbfb087SZhigang Lu spin_unlock(ptl); 5143acbfb087SZhigang Lu continue; 5144acbfb087SZhigang Lu } 5145acbfb087SZhigang Lu 514682e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 514782e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 51480fa5bc40SJoao Martins 514982e5d378SJoao Martins if (pages || vmas) 515082e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 515182e5d378SJoao Martins vma, refs, 515282e5d378SJoao Martins likely(pages) ? pages + i : NULL, 515382e5d378SJoao Martins vmas ? vmas + i : NULL); 515463551ae0SDavid Gibson 515582e5d378SJoao Martins if (pages) { 51560fa5bc40SJoao Martins /* 51570fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 51580fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 51590fa5bc40SJoao Martins * checked that the huge page is present in the page 51600fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 51610fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 51620fa5bc40SJoao Martins * head page and tail pages from being rearranged in 51630fa5bc40SJoao Martins * any way. So this page must be available at this 51640fa5bc40SJoao Martins * point, unless the page refcount overflowed: 51650fa5bc40SJoao Martins */ 516682e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 51670fa5bc40SJoao Martins refs, 51680fa5bc40SJoao Martins flags))) { 51690fa5bc40SJoao Martins spin_unlock(ptl); 51700fa5bc40SJoao Martins remainder = 0; 51710fa5bc40SJoao Martins err = -ENOMEM; 51720fa5bc40SJoao Martins break; 51730fa5bc40SJoao Martins } 5174d5d4b0aaSChen, Kenneth W } 517582e5d378SJoao Martins 517682e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 517782e5d378SJoao Martins remainder -= refs; 517882e5d378SJoao Martins i += refs; 517982e5d378SJoao Martins 5180cb900f41SKirill A. Shutemov spin_unlock(ptl); 518163551ae0SDavid Gibson } 518228a35716SMichel Lespinasse *nr_pages = remainder; 518387ffc118SAndrea Arcangeli /* 518487ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 518587ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 518687ffc118SAndrea Arcangeli * branch. 518787ffc118SAndrea Arcangeli */ 518863551ae0SDavid Gibson *position = vaddr; 518963551ae0SDavid Gibson 51902be7cfedSDaniel Jordan return i ? i : err; 519163551ae0SDavid Gibson } 51928f860591SZhang, Yanmin 51937da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 51948f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 51958f860591SZhang, Yanmin { 51968f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 51978f860591SZhang, Yanmin unsigned long start = address; 51988f860591SZhang, Yanmin pte_t *ptep; 51998f860591SZhang, Yanmin pte_t pte; 5200a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52017da4d641SPeter Zijlstra unsigned long pages = 0; 5202dff11abeSMike Kravetz bool shared_pmd = false; 5203ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5204dff11abeSMike Kravetz 5205dff11abeSMike Kravetz /* 5206dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5207ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5208dff11abeSMike Kravetz * range if PMD sharing is possible. 5209dff11abeSMike Kravetz */ 52107269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 52117269f999SJérôme Glisse 0, vma, mm, start, end); 5212ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52138f860591SZhang, Yanmin 52148f860591SZhang, Yanmin BUG_ON(address >= end); 5215ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 52168f860591SZhang, Yanmin 5217ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 521883cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5219a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5220cb900f41SKirill A. Shutemov spinlock_t *ptl; 52217868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 52228f860591SZhang, Yanmin if (!ptep) 52238f860591SZhang, Yanmin continue; 5224cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 522534ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 52267da4d641SPeter Zijlstra pages++; 5227cb900f41SKirill A. Shutemov spin_unlock(ptl); 5228dff11abeSMike Kravetz shared_pmd = true; 522939dde65cSChen, Kenneth W continue; 52307da4d641SPeter Zijlstra } 5231a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5232a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5233a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5234a8bda28dSNaoya Horiguchi continue; 5235a8bda28dSNaoya Horiguchi } 5236a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5237a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5238a8bda28dSNaoya Horiguchi 5239a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5240a8bda28dSNaoya Horiguchi pte_t newpte; 5241a8bda28dSNaoya Horiguchi 5242a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5243a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5244e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5245e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5246a8bda28dSNaoya Horiguchi pages++; 5247a8bda28dSNaoya Horiguchi } 5248a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5249a8bda28dSNaoya Horiguchi continue; 5250a8bda28dSNaoya Horiguchi } 5251a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5252023bdd00SAneesh Kumar K.V pte_t old_pte; 5253023bdd00SAneesh Kumar K.V 5254023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5255023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5256be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5257023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 52587da4d641SPeter Zijlstra pages++; 52598f860591SZhang, Yanmin } 5260cb900f41SKirill A. Shutemov spin_unlock(ptl); 52618f860591SZhang, Yanmin } 5262d833352aSMel Gorman /* 5263c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5264d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5265c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5266dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5267dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5268d833352aSMel Gorman */ 5269dff11abeSMike Kravetz if (shared_pmd) 5270ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5271dff11abeSMike Kravetz else 52725491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 52730f10851eSJérôme Glisse /* 52740f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 52750f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 52760f10851eSJérôme Glisse * 5277ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 52780f10851eSJérôme Glisse */ 527983cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5280ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 52817da4d641SPeter Zijlstra 52827da4d641SPeter Zijlstra return pages << h->order; 52838f860591SZhang, Yanmin } 52848f860591SZhang, Yanmin 528533b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 528633b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5287a1e78772SMel Gorman long from, long to, 52885a6fe125SMel Gorman struct vm_area_struct *vma, 5289ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5290e4e574b7SAdam Litke { 529133b8f84aSMike Kravetz long chg, add = -1; 5292a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 529390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52949119a41eSJoonsoo Kim struct resv_map *resv_map; 5295075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 52960db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5297e4e574b7SAdam Litke 529863489f8eSMike Kravetz /* This should never happen */ 529963489f8eSMike Kravetz if (from > to) { 530063489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 530133b8f84aSMike Kravetz return false; 530263489f8eSMike Kravetz } 530363489f8eSMike Kravetz 5304a1e78772SMel Gorman /* 530517c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 530617c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 530790481622SDavid Gibson * without using reserves 530817c9d12eSMel Gorman */ 5309ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 531033b8f84aSMike Kravetz return true; 531117c9d12eSMel Gorman 531217c9d12eSMel Gorman /* 5313a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5314a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5315a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5316a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5317a1e78772SMel Gorman */ 53189119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5319f27a5136SMike Kravetz /* 5320f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5321f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5322f27a5136SMike Kravetz * hugetlbfs_get_inode). 5323f27a5136SMike Kravetz */ 53244e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 53259119a41eSJoonsoo Kim 53260db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 53279119a41eSJoonsoo Kim 53289119a41eSJoonsoo Kim } else { 5329e9fe92aeSMina Almasry /* Private mapping. */ 53309119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 53315a6fe125SMel Gorman if (!resv_map) 533233b8f84aSMike Kravetz return false; 53335a6fe125SMel Gorman 533417c9d12eSMel Gorman chg = to - from; 533517c9d12eSMel Gorman 53365a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 53375a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 53385a6fe125SMel Gorman } 53395a6fe125SMel Gorman 534033b8f84aSMike Kravetz if (chg < 0) 5341c50ac050SDave Hansen goto out_err; 534217c9d12eSMel Gorman 534333b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 534433b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5345075a61d0SMina Almasry goto out_err; 5346075a61d0SMina Almasry 5347075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5348075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5349075a61d0SMina Almasry * of the resv_map. 5350075a61d0SMina Almasry */ 5351075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5352075a61d0SMina Almasry } 5353075a61d0SMina Almasry 53541c5ecae3SMike Kravetz /* 53551c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 53561c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 53571c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 53581c5ecae3SMike Kravetz */ 53591c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 536033b8f84aSMike Kravetz if (gbl_reserve < 0) 5361075a61d0SMina Almasry goto out_uncharge_cgroup; 536217c9d12eSMel Gorman 536317c9d12eSMel Gorman /* 536417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 536590481622SDavid Gibson * Hand the pages back to the subpool if there are not 536617c9d12eSMel Gorman */ 536733b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5368075a61d0SMina Almasry goto out_put_pages; 536917c9d12eSMel Gorman 537017c9d12eSMel Gorman /* 537117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 537217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 537317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 537417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 537517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 537617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 537717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 537817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 537917c9d12eSMel Gorman * else has to be done for private mappings here 538017c9d12eSMel Gorman */ 538133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5382075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 538333039678SMike Kravetz 53840db9d74eSMina Almasry if (unlikely(add < 0)) { 53850db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5386075a61d0SMina Almasry goto out_put_pages; 53870db9d74eSMina Almasry } else if (unlikely(chg > add)) { 538833039678SMike Kravetz /* 538933039678SMike Kravetz * pages in this range were added to the reserve 539033039678SMike Kravetz * map between region_chg and region_add. This 539133039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 539233039678SMike Kravetz * the subpool and reserve counts modified above 539333039678SMike Kravetz * based on the difference. 539433039678SMike Kravetz */ 539533039678SMike Kravetz long rsv_adjust; 539633039678SMike Kravetz 5397d85aecf2SMiaohe Lin /* 5398d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5399d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5400d85aecf2SMiaohe Lin */ 5401075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5402075a61d0SMina Almasry hstate_index(h), 5403075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5404075a61d0SMina Almasry 540533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 540633039678SMike Kravetz chg - add); 540733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5408d85aecf2SMiaohe Lin } else if (h_cg) { 5409d85aecf2SMiaohe Lin /* 5410d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5411d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5412d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5413d85aecf2SMiaohe Lin * done. 5414d85aecf2SMiaohe Lin */ 5415d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 541633039678SMike Kravetz } 541733039678SMike Kravetz } 541833b8f84aSMike Kravetz return true; 541933b8f84aSMike Kravetz 5420075a61d0SMina Almasry out_put_pages: 5421075a61d0SMina Almasry /* put back original number of pages, chg */ 5422075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5423075a61d0SMina Almasry out_uncharge_cgroup: 5424075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5425075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5426c50ac050SDave Hansen out_err: 54275e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 54280db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 54290db9d74eSMina Almasry * region_add failed or didn't run. 54300db9d74eSMina Almasry */ 54310db9d74eSMina Almasry if (chg >= 0 && add < 0) 54320db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5433f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5434f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 543533b8f84aSMike Kravetz return false; 5436a43a8c39SChen, Kenneth W } 5437a43a8c39SChen, Kenneth W 5438b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5439b5cec28dSMike Kravetz long freed) 5440a43a8c39SChen, Kenneth W { 5441a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 54424e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 54439119a41eSJoonsoo Kim long chg = 0; 544490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 54451c5ecae3SMike Kravetz long gbl_reserve; 544645c682a6SKen Chen 5447f27a5136SMike Kravetz /* 5448f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5449f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5450f27a5136SMike Kravetz */ 5451b5cec28dSMike Kravetz if (resv_map) { 5452b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5453b5cec28dSMike Kravetz /* 5454b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5455b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5456b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5457b5cec28dSMike Kravetz */ 5458b5cec28dSMike Kravetz if (chg < 0) 5459b5cec28dSMike Kravetz return chg; 5460b5cec28dSMike Kravetz } 5461b5cec28dSMike Kravetz 546245c682a6SKen Chen spin_lock(&inode->i_lock); 5463e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 546445c682a6SKen Chen spin_unlock(&inode->i_lock); 546545c682a6SKen Chen 54661c5ecae3SMike Kravetz /* 54671c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 54681c5ecae3SMike Kravetz * reservations to be released may be adjusted. 5469dddf31a4SMiaohe Lin * 5470dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 5471dddf31a4SMiaohe Lin * won't go negative. 54721c5ecae3SMike Kravetz */ 54731c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 54741c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5475b5cec28dSMike Kravetz 5476b5cec28dSMike Kravetz return 0; 5477a43a8c39SChen, Kenneth W } 547893f70f90SNaoya Horiguchi 54793212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 54803212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 54813212b535SSteve Capper struct vm_area_struct *vma, 54823212b535SSteve Capper unsigned long addr, pgoff_t idx) 54833212b535SSteve Capper { 54843212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 54853212b535SSteve Capper svma->vm_start; 54863212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 54873212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 54883212b535SSteve Capper 54893212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5490de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5491de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 54923212b535SSteve Capper 54933212b535SSteve Capper /* 54943212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 54953212b535SSteve Capper * page table page. 54963212b535SSteve Capper */ 54973212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 54983212b535SSteve Capper vm_flags != svm_flags || 549907e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 55003212b535SSteve Capper return 0; 55013212b535SSteve Capper 55023212b535SSteve Capper return saddr; 55033212b535SSteve Capper } 55043212b535SSteve Capper 550531aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 55063212b535SSteve Capper { 55073212b535SSteve Capper unsigned long base = addr & PUD_MASK; 55083212b535SSteve Capper unsigned long end = base + PUD_SIZE; 55093212b535SSteve Capper 55103212b535SSteve Capper /* 55113212b535SSteve Capper * check on proper vm_flags and page table alignment 55123212b535SSteve Capper */ 5513017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 551431aafb45SNicholas Krause return true; 551531aafb45SNicholas Krause return false; 55163212b535SSteve Capper } 55173212b535SSteve Capper 5518c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5519c1991e07SPeter Xu { 5520c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5521c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5522c1991e07SPeter Xu return false; 5523c1991e07SPeter Xu #endif 5524c1991e07SPeter Xu return vma_shareable(vma, addr); 5525c1991e07SPeter Xu } 5526c1991e07SPeter Xu 55273212b535SSteve Capper /* 5528017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5529017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5530017b1660SMike Kravetz * shared pmd mappings. 5531017b1660SMike Kravetz */ 5532017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5533017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5534017b1660SMike Kravetz { 5535a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5536a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5537017b1660SMike Kravetz 5538a1ba9da8SLi Xinhai /* 5539a1ba9da8SLi Xinhai * vma need span at least one aligned PUD size and the start,end range 5540a1ba9da8SLi Xinhai * must at least partialy within it. 5541a1ba9da8SLi Xinhai */ 5542a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5543a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5544017b1660SMike Kravetz return; 5545017b1660SMike Kravetz 554675802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5547a1ba9da8SLi Xinhai if (*start > v_start) 5548a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5549017b1660SMike Kravetz 5550a1ba9da8SLi Xinhai if (*end < v_end) 5551a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5552017b1660SMike Kravetz } 5553017b1660SMike Kravetz 5554017b1660SMike Kravetz /* 55553212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 55563212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 55573212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5558c0d0381aSMike Kravetz * code much cleaner. 5559c0d0381aSMike Kravetz * 55600bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 55610bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 55620bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 55630bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 55640bf7b64eSMike Kravetz * 55650bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 55660bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 55670bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 55680bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 55690bf7b64eSMike Kravetz * only required for subsequent processing. 55703212b535SSteve Capper */ 5571aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5572aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 55733212b535SSteve Capper { 55743212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 55753212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 55763212b535SSteve Capper vma->vm_pgoff; 55773212b535SSteve Capper struct vm_area_struct *svma; 55783212b535SSteve Capper unsigned long saddr; 55793212b535SSteve Capper pte_t *spte = NULL; 55803212b535SSteve Capper pte_t *pte; 5581cb900f41SKirill A. Shutemov spinlock_t *ptl; 55823212b535SSteve Capper 55830bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 55843212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 55853212b535SSteve Capper if (svma == vma) 55863212b535SSteve Capper continue; 55873212b535SSteve Capper 55883212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 55893212b535SSteve Capper if (saddr) { 55907868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 55917868a208SPunit Agrawal vma_mmu_pagesize(svma)); 55923212b535SSteve Capper if (spte) { 55933212b535SSteve Capper get_page(virt_to_page(spte)); 55943212b535SSteve Capper break; 55953212b535SSteve Capper } 55963212b535SSteve Capper } 55973212b535SSteve Capper } 55983212b535SSteve Capper 55993212b535SSteve Capper if (!spte) 56003212b535SSteve Capper goto out; 56013212b535SSteve Capper 56028bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5603dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 56043212b535SSteve Capper pud_populate(mm, pud, 56053212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5606c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5607dc6c9a35SKirill A. Shutemov } else { 56083212b535SSteve Capper put_page(virt_to_page(spte)); 5609dc6c9a35SKirill A. Shutemov } 5610cb900f41SKirill A. Shutemov spin_unlock(ptl); 56113212b535SSteve Capper out: 56123212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 56133212b535SSteve Capper return pte; 56143212b535SSteve Capper } 56153212b535SSteve Capper 56163212b535SSteve Capper /* 56173212b535SSteve Capper * unmap huge page backed by shared pte. 56183212b535SSteve Capper * 56193212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 56203212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 56213212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 56223212b535SSteve Capper * 5623c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 56243212b535SSteve Capper * 56253212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 56263212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 56273212b535SSteve Capper */ 562834ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 562934ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 56303212b535SSteve Capper { 56313212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5632c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5633c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 56343212b535SSteve Capper 563534ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 56363212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 56373212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 56383212b535SSteve Capper return 0; 56393212b535SSteve Capper 56403212b535SSteve Capper pud_clear(pud); 56413212b535SSteve Capper put_page(virt_to_page(ptep)); 5642dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 56433212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 56443212b535SSteve Capper return 1; 56453212b535SSteve Capper } 5646c1991e07SPeter Xu 56479e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5648aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5649aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 56509e5fc74cSSteve Capper { 56519e5fc74cSSteve Capper return NULL; 56529e5fc74cSSteve Capper } 5653e81f2d22SZhang Zhen 565434ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 565534ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5656e81f2d22SZhang Zhen { 5657e81f2d22SZhang Zhen return 0; 5658e81f2d22SZhang Zhen } 5659017b1660SMike Kravetz 5660017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5661017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5662017b1660SMike Kravetz { 5663017b1660SMike Kravetz } 5664c1991e07SPeter Xu 5665c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5666c1991e07SPeter Xu { 5667c1991e07SPeter Xu return false; 5668c1991e07SPeter Xu } 56693212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 56703212b535SSteve Capper 56719e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5672aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 56739e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 56749e5fc74cSSteve Capper { 56759e5fc74cSSteve Capper pgd_t *pgd; 5676c2febafcSKirill A. Shutemov p4d_t *p4d; 56779e5fc74cSSteve Capper pud_t *pud; 56789e5fc74cSSteve Capper pte_t *pte = NULL; 56799e5fc74cSSteve Capper 56809e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5681f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5682f4f0a3d8SKirill A. Shutemov if (!p4d) 5683f4f0a3d8SKirill A. Shutemov return NULL; 5684c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 56859e5fc74cSSteve Capper if (pud) { 56869e5fc74cSSteve Capper if (sz == PUD_SIZE) { 56879e5fc74cSSteve Capper pte = (pte_t *)pud; 56889e5fc74cSSteve Capper } else { 56899e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5690c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5691aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 56929e5fc74cSSteve Capper else 56939e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 56949e5fc74cSSteve Capper } 56959e5fc74cSSteve Capper } 56964e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 56979e5fc74cSSteve Capper 56989e5fc74cSSteve Capper return pte; 56999e5fc74cSSteve Capper } 57009e5fc74cSSteve Capper 57019b19df29SPunit Agrawal /* 57029b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 57039b19df29SPunit Agrawal * entry at address @addr 57049b19df29SPunit Agrawal * 57058ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 57068ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 57079b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 57089b19df29SPunit Agrawal * table. 57099b19df29SPunit Agrawal */ 57107868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 57117868a208SPunit Agrawal unsigned long addr, unsigned long sz) 57129e5fc74cSSteve Capper { 57139e5fc74cSSteve Capper pgd_t *pgd; 5714c2febafcSKirill A. Shutemov p4d_t *p4d; 57158ac0b81aSLi Xinhai pud_t *pud; 57168ac0b81aSLi Xinhai pmd_t *pmd; 57179e5fc74cSSteve Capper 57189e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5719c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5720c2febafcSKirill A. Shutemov return NULL; 5721c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5722c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5723c2febafcSKirill A. Shutemov return NULL; 57249b19df29SPunit Agrawal 5725c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 57268ac0b81aSLi Xinhai if (sz == PUD_SIZE) 57278ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 57289e5fc74cSSteve Capper return (pte_t *)pud; 57298ac0b81aSLi Xinhai if (!pud_present(*pud)) 57308ac0b81aSLi Xinhai return NULL; 57318ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 57329b19df29SPunit Agrawal 57339e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 57348ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 57359e5fc74cSSteve Capper return (pte_t *)pmd; 57369e5fc74cSSteve Capper } 57379e5fc74cSSteve Capper 573861f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 573961f77edaSNaoya Horiguchi 574061f77edaSNaoya Horiguchi /* 574161f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 574261f77edaSNaoya Horiguchi * behavior. 574361f77edaSNaoya Horiguchi */ 574461f77edaSNaoya Horiguchi struct page * __weak 574561f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 574661f77edaSNaoya Horiguchi int write) 574761f77edaSNaoya Horiguchi { 574861f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 574961f77edaSNaoya Horiguchi } 575061f77edaSNaoya Horiguchi 575161f77edaSNaoya Horiguchi struct page * __weak 57524dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 57534dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 57544dc71451SAneesh Kumar K.V { 57554dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 57564dc71451SAneesh Kumar K.V return NULL; 57574dc71451SAneesh Kumar K.V } 57584dc71451SAneesh Kumar K.V 57594dc71451SAneesh Kumar K.V struct page * __weak 57609e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5761e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 57629e5fc74cSSteve Capper { 5763e66f17ffSNaoya Horiguchi struct page *page = NULL; 5764e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5765c9d398faSNaoya Horiguchi pte_t pte; 57663faa52c0SJohn Hubbard 57673faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 57683faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 57693faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 57703faa52c0SJohn Hubbard return NULL; 57713faa52c0SJohn Hubbard 5772e66f17ffSNaoya Horiguchi retry: 5773e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5774e66f17ffSNaoya Horiguchi spin_lock(ptl); 5775e66f17ffSNaoya Horiguchi /* 5776e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5777e66f17ffSNaoya Horiguchi * unmapped from other threads. 5778e66f17ffSNaoya Horiguchi */ 5779e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5780e66f17ffSNaoya Horiguchi goto out; 5781c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5782c9d398faSNaoya Horiguchi if (pte_present(pte)) { 578397534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 57843faa52c0SJohn Hubbard /* 57853faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 57863faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 57873faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 57883faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 57893faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 57903faa52c0SJohn Hubbard * unless the page refcount overflowed: 57913faa52c0SJohn Hubbard */ 57923faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 57933faa52c0SJohn Hubbard page = NULL; 57943faa52c0SJohn Hubbard goto out; 57953faa52c0SJohn Hubbard } 5796e66f17ffSNaoya Horiguchi } else { 5797c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5798e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5799e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5800e66f17ffSNaoya Horiguchi goto retry; 5801e66f17ffSNaoya Horiguchi } 5802e66f17ffSNaoya Horiguchi /* 5803e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5804e66f17ffSNaoya Horiguchi * follow_page_mask(). 5805e66f17ffSNaoya Horiguchi */ 5806e66f17ffSNaoya Horiguchi } 5807e66f17ffSNaoya Horiguchi out: 5808e66f17ffSNaoya Horiguchi spin_unlock(ptl); 58099e5fc74cSSteve Capper return page; 58109e5fc74cSSteve Capper } 58119e5fc74cSSteve Capper 581261f77edaSNaoya Horiguchi struct page * __weak 58139e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5814e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 58159e5fc74cSSteve Capper { 58163faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5817e66f17ffSNaoya Horiguchi return NULL; 58189e5fc74cSSteve Capper 5819e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 58209e5fc74cSSteve Capper } 58219e5fc74cSSteve Capper 5822faaa5b62SAnshuman Khandual struct page * __weak 5823faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5824faaa5b62SAnshuman Khandual { 58253faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5826faaa5b62SAnshuman Khandual return NULL; 5827faaa5b62SAnshuman Khandual 5828faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5829faaa5b62SAnshuman Khandual } 5830faaa5b62SAnshuman Khandual 583131caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 583231caf665SNaoya Horiguchi { 5833bcc54222SNaoya Horiguchi bool ret = true; 5834bcc54222SNaoya Horiguchi 5835db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 58368f251a3dSMike Kravetz if (!PageHeadHuge(page) || 58378f251a3dSMike Kravetz !HPageMigratable(page) || 58380eb2df2bSMuchun Song !get_page_unless_zero(page)) { 5839bcc54222SNaoya Horiguchi ret = false; 5840bcc54222SNaoya Horiguchi goto unlock; 5841bcc54222SNaoya Horiguchi } 58428f251a3dSMike Kravetz ClearHPageMigratable(page); 584331caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5844bcc54222SNaoya Horiguchi unlock: 5845db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 5846bcc54222SNaoya Horiguchi return ret; 584731caf665SNaoya Horiguchi } 584831caf665SNaoya Horiguchi 584931caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 585031caf665SNaoya Horiguchi { 5851db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 58528f251a3dSMike Kravetz SetHPageMigratable(page); 585331caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 5854db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 585531caf665SNaoya Horiguchi put_page(page); 585631caf665SNaoya Horiguchi } 5857ab5ac90aSMichal Hocko 5858ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5859ab5ac90aSMichal Hocko { 5860ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5861ab5ac90aSMichal Hocko 5862ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5863ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5864ab5ac90aSMichal Hocko 5865ab5ac90aSMichal Hocko /* 5866ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5867ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5868ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5869ab5ac90aSMichal Hocko * the temporary status. 5870ab5ac90aSMichal Hocko * 5871ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5872ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5873ab5ac90aSMichal Hocko * the per-node's. 5874ab5ac90aSMichal Hocko */ 58759157c311SMike Kravetz if (HPageTemporary(newpage)) { 5876ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5877ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5878ab5ac90aSMichal Hocko 58799157c311SMike Kravetz SetHPageTemporary(oldpage); 58809157c311SMike Kravetz ClearHPageTemporary(newpage); 5881ab5ac90aSMichal Hocko 58825af1ab1dSMiaohe Lin /* 58835af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 58845af1ab1dSMiaohe Lin * when we do not cross the node. 58855af1ab1dSMiaohe Lin */ 58865af1ab1dSMiaohe Lin if (new_nid == old_nid) 58875af1ab1dSMiaohe Lin return; 5888db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 5889ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5890ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5891ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5892ab5ac90aSMichal Hocko } 5893db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 5894ab5ac90aSMichal Hocko } 5895ab5ac90aSMichal Hocko } 5896cf11e85fSRoman Gushchin 58976dfeaff9SPeter Xu /* 58986dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 58996dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 59006dfeaff9SPeter Xu */ 59016dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 59026dfeaff9SPeter Xu { 59036dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 59046dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 59056dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 59066dfeaff9SPeter Xu struct mmu_notifier_range range; 59076dfeaff9SPeter Xu unsigned long address, start, end; 59086dfeaff9SPeter Xu spinlock_t *ptl; 59096dfeaff9SPeter Xu pte_t *ptep; 59106dfeaff9SPeter Xu 59116dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 59126dfeaff9SPeter Xu return; 59136dfeaff9SPeter Xu 59146dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 59156dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 59166dfeaff9SPeter Xu 59176dfeaff9SPeter Xu if (start >= end) 59186dfeaff9SPeter Xu return; 59196dfeaff9SPeter Xu 59206dfeaff9SPeter Xu /* 59216dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 59226dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 59236dfeaff9SPeter Xu */ 59246dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 59256dfeaff9SPeter Xu start, end); 59266dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 59276dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 59286dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 59296dfeaff9SPeter Xu unsigned long tmp = address; 59306dfeaff9SPeter Xu 59316dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 59326dfeaff9SPeter Xu if (!ptep) 59336dfeaff9SPeter Xu continue; 59346dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 59356dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 59366dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 59376dfeaff9SPeter Xu spin_unlock(ptl); 59386dfeaff9SPeter Xu } 59396dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 59406dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 59416dfeaff9SPeter Xu /* 59426dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 59436dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 59446dfeaff9SPeter Xu */ 59456dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 59466dfeaff9SPeter Xu } 59476dfeaff9SPeter Xu 5948cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5949cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5950cf11e85fSRoman Gushchin 5951cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5952cf11e85fSRoman Gushchin { 5953cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5954cf11e85fSRoman Gushchin return 0; 5955cf11e85fSRoman Gushchin } 5956cf11e85fSRoman Gushchin 5957cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5958cf11e85fSRoman Gushchin 5959cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5960cf11e85fSRoman Gushchin { 5961cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5962cf11e85fSRoman Gushchin int nid; 5963cf11e85fSRoman Gushchin 5964cf11e85fSRoman Gushchin cma_reserve_called = true; 5965cf11e85fSRoman Gushchin 5966cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5967cf11e85fSRoman Gushchin return; 5968cf11e85fSRoman Gushchin 5969cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5970cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5971cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5972cf11e85fSRoman Gushchin return; 5973cf11e85fSRoman Gushchin } 5974cf11e85fSRoman Gushchin 5975cf11e85fSRoman Gushchin /* 5976cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5977cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5978cf11e85fSRoman Gushchin */ 5979cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5980cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5981cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5982cf11e85fSRoman Gushchin 5983cf11e85fSRoman Gushchin reserved = 0; 5984cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5985cf11e85fSRoman Gushchin int res; 59862281f797SBarry Song char name[CMA_MAX_NAME]; 5987cf11e85fSRoman Gushchin 5988cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5989cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5990cf11e85fSRoman Gushchin 59912281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5992cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 599329d0f41dSBarry Song 0, false, name, 5994cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5995cf11e85fSRoman Gushchin if (res) { 5996cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5997cf11e85fSRoman Gushchin res, nid); 5998cf11e85fSRoman Gushchin continue; 5999cf11e85fSRoman Gushchin } 6000cf11e85fSRoman Gushchin 6001cf11e85fSRoman Gushchin reserved += size; 6002cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 6003cf11e85fSRoman Gushchin size / SZ_1M, nid); 6004cf11e85fSRoman Gushchin 6005cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 6006cf11e85fSRoman Gushchin break; 6007cf11e85fSRoman Gushchin } 6008cf11e85fSRoman Gushchin } 6009cf11e85fSRoman Gushchin 6010cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 6011cf11e85fSRoman Gushchin { 6012cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 6013cf11e85fSRoman Gushchin return; 6014cf11e85fSRoman Gushchin 6015cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 6016cf11e85fSRoman Gushchin } 6017cf11e85fSRoman Gushchin 6018cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 6019