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> 421a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 451da177e4SLinus Torvalds 46c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 47e5ff2159SAndi Kleen unsigned int default_hstate_idx; 48e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 49cf11e85fSRoman Gushchin 50dbda8feaSBarry Song #ifdef CONFIG_CMA 51cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 52dbda8feaSBarry Song #endif 53dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 54cf11e85fSRoman Gushchin 55641844f5SNaoya Horiguchi /* 56641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 57641844f5SNaoya Horiguchi * at boot time. 58641844f5SNaoya Horiguchi */ 59641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 60e5ff2159SAndi Kleen 6153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6253ba51d2SJon Tollefson 63e5ff2159SAndi Kleen /* for command line parsing */ 64e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 65e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 669fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 67282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 68e5ff2159SAndi Kleen 693935baa9SDavid Gibson /* 7031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 723935baa9SDavid Gibson */ 73c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 740bd0f9fbSEric Paris 758382d914SDavidlohr Bueso /* 768382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 778382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 788382d914SDavidlohr Bueso */ 798382d914SDavidlohr Bueso static int num_fault_mutexes; 80c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 818382d914SDavidlohr Bueso 827ca02d0aSMike Kravetz /* Forward declaration */ 837ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 847ca02d0aSMike Kravetz 851d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 861d88433bSMiaohe Lin { 871d88433bSMiaohe Lin if (spool->count) 881d88433bSMiaohe Lin return false; 891d88433bSMiaohe Lin if (spool->max_hpages != -1) 901d88433bSMiaohe Lin return spool->used_hpages == 0; 911d88433bSMiaohe Lin if (spool->min_hpages != -1) 921d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 931d88433bSMiaohe Lin 941d88433bSMiaohe Lin return true; 951d88433bSMiaohe Lin } 961d88433bSMiaohe Lin 9790481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 9890481622SDavid Gibson { 9990481622SDavid Gibson spin_unlock(&spool->lock); 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 { 13890481622SDavid Gibson spin_lock(&spool->lock); 13990481622SDavid Gibson BUG_ON(!spool->count); 14090481622SDavid Gibson spool->count--; 14190481622SDavid Gibson unlock_or_release_subpool(spool); 14290481622SDavid Gibson } 14390481622SDavid Gibson 1441c5ecae3SMike Kravetz /* 1451c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1461c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1479e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1481c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1491c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1507c8de358SEthon Paul * a subpool minimum size must be maintained. 1511c5ecae3SMike Kravetz */ 1521c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 15390481622SDavid Gibson long delta) 15490481622SDavid Gibson { 1551c5ecae3SMike Kravetz long ret = delta; 15690481622SDavid Gibson 15790481622SDavid Gibson if (!spool) 1581c5ecae3SMike Kravetz return ret; 15990481622SDavid Gibson 16090481622SDavid Gibson spin_lock(&spool->lock); 16190481622SDavid Gibson 1621c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1631c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1641c5ecae3SMike Kravetz spool->used_hpages += delta; 1651c5ecae3SMike Kravetz else { 1661c5ecae3SMike Kravetz ret = -ENOMEM; 1671c5ecae3SMike Kravetz goto unlock_ret; 1681c5ecae3SMike Kravetz } 1691c5ecae3SMike Kravetz } 1701c5ecae3SMike Kravetz 17109a95e29SMike Kravetz /* minimum size accounting */ 17209a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1731c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1741c5ecae3SMike Kravetz /* 1751c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1761c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1771c5ecae3SMike Kravetz */ 1781c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1791c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1801c5ecae3SMike Kravetz } else { 1811c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1821c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1831c5ecae3SMike Kravetz } 1841c5ecae3SMike Kravetz } 1851c5ecae3SMike Kravetz 1861c5ecae3SMike Kravetz unlock_ret: 1871c5ecae3SMike Kravetz spin_unlock(&spool->lock); 18890481622SDavid Gibson return ret; 18990481622SDavid Gibson } 19090481622SDavid Gibson 1911c5ecae3SMike Kravetz /* 1921c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1931c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1941c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1951c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1961c5ecae3SMike Kravetz */ 1971c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 19890481622SDavid Gibson long delta) 19990481622SDavid Gibson { 2001c5ecae3SMike Kravetz long ret = delta; 2011c5ecae3SMike Kravetz 20290481622SDavid Gibson if (!spool) 2031c5ecae3SMike Kravetz return delta; 20490481622SDavid Gibson 20590481622SDavid Gibson spin_lock(&spool->lock); 2061c5ecae3SMike Kravetz 2071c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 20890481622SDavid Gibson spool->used_hpages -= delta; 2091c5ecae3SMike Kravetz 21009a95e29SMike Kravetz /* minimum size accounting */ 21109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2121c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2131c5ecae3SMike Kravetz ret = 0; 2141c5ecae3SMike Kravetz else 2151c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2161c5ecae3SMike Kravetz 2171c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2181c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2191c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2201c5ecae3SMike Kravetz } 2211c5ecae3SMike Kravetz 2221c5ecae3SMike Kravetz /* 2231c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2241c5ecae3SMike Kravetz * quota reference, free it now. 2251c5ecae3SMike Kravetz */ 22690481622SDavid Gibson unlock_or_release_subpool(spool); 2271c5ecae3SMike Kravetz 2281c5ecae3SMike Kravetz return ret; 22990481622SDavid Gibson } 23090481622SDavid Gibson 23190481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 23290481622SDavid Gibson { 23390481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 23490481622SDavid Gibson } 23590481622SDavid Gibson 23690481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 23790481622SDavid Gibson { 238496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 23990481622SDavid Gibson } 24090481622SDavid Gibson 2410db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2420db9d74eSMina Almasry * it for use. 2430db9d74eSMina Almasry */ 2440db9d74eSMina Almasry static struct file_region * 2450db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2460db9d74eSMina Almasry { 2470db9d74eSMina Almasry struct file_region *nrg = NULL; 2480db9d74eSMina Almasry 2490db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2500db9d74eSMina Almasry 2510db9d74eSMina Almasry resv->region_cache_count--; 2520db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2530db9d74eSMina Almasry list_del(&nrg->link); 2540db9d74eSMina Almasry 2550db9d74eSMina Almasry nrg->from = from; 2560db9d74eSMina Almasry nrg->to = to; 2570db9d74eSMina Almasry 2580db9d74eSMina Almasry return nrg; 2590db9d74eSMina Almasry } 2600db9d74eSMina Almasry 261075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 262075a61d0SMina Almasry struct file_region *rg) 263075a61d0SMina Almasry { 264075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 265075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 266075a61d0SMina Almasry nrg->css = rg->css; 267075a61d0SMina Almasry if (rg->css) 268075a61d0SMina Almasry css_get(rg->css); 269075a61d0SMina Almasry #endif 270075a61d0SMina Almasry } 271075a61d0SMina Almasry 272075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 273075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 274075a61d0SMina Almasry struct hstate *h, 275075a61d0SMina Almasry struct resv_map *resv, 276075a61d0SMina Almasry struct file_region *nrg) 277075a61d0SMina Almasry { 278075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 279075a61d0SMina Almasry if (h_cg) { 280075a61d0SMina Almasry nrg->reservation_counter = 281075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 282075a61d0SMina Almasry nrg->css = &h_cg->css; 283d85aecf2SMiaohe Lin /* 284d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 285d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 286d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 287d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 288d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 289d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 290d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 291d85aecf2SMiaohe Lin * untouched. 292d85aecf2SMiaohe Lin */ 293d85aecf2SMiaohe Lin css_get(&h_cg->css); 294075a61d0SMina Almasry if (!resv->pages_per_hpage) 295075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 296075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 297075a61d0SMina Almasry * a resv_map. 298075a61d0SMina Almasry */ 299075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 300075a61d0SMina Almasry } else { 301075a61d0SMina Almasry nrg->reservation_counter = NULL; 302075a61d0SMina Almasry nrg->css = NULL; 303075a61d0SMina Almasry } 304075a61d0SMina Almasry #endif 305075a61d0SMina Almasry } 306075a61d0SMina Almasry 307d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 308d85aecf2SMiaohe Lin { 309d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 310d85aecf2SMiaohe Lin if (rg->css) 311d85aecf2SMiaohe Lin css_put(rg->css); 312d85aecf2SMiaohe Lin #endif 313d85aecf2SMiaohe Lin } 314d85aecf2SMiaohe Lin 315a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 316a9b3f867SMina Almasry struct file_region *org) 317a9b3f867SMina Almasry { 318a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 319a9b3f867SMina Almasry return rg && org && 320a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 321a9b3f867SMina Almasry rg->css == org->css; 322a9b3f867SMina Almasry 323a9b3f867SMina Almasry #else 324a9b3f867SMina Almasry return true; 325a9b3f867SMina Almasry #endif 326a9b3f867SMina Almasry } 327a9b3f867SMina Almasry 328a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 329a9b3f867SMina Almasry { 330a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 331a9b3f867SMina Almasry 332a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 333a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 334a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 335a9b3f867SMina Almasry prg->to = rg->to; 336a9b3f867SMina Almasry 337a9b3f867SMina Almasry list_del(&rg->link); 338d85aecf2SMiaohe Lin put_uncharge_info(rg); 339a9b3f867SMina Almasry kfree(rg); 340a9b3f867SMina Almasry 3417db5e7b6SWei Yang rg = prg; 342a9b3f867SMina Almasry } 343a9b3f867SMina Almasry 344a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 345a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 346a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 347a9b3f867SMina Almasry nrg->from = rg->from; 348a9b3f867SMina Almasry 349a9b3f867SMina Almasry list_del(&rg->link); 350d85aecf2SMiaohe Lin put_uncharge_info(rg); 351a9b3f867SMina Almasry kfree(rg); 352a9b3f867SMina Almasry } 353a9b3f867SMina Almasry } 354a9b3f867SMina Almasry 3552103cf9cSPeter Xu static inline long 3562103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3572103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3582103cf9cSPeter Xu long *regions_needed) 3592103cf9cSPeter Xu { 3602103cf9cSPeter Xu struct file_region *nrg; 3612103cf9cSPeter Xu 3622103cf9cSPeter Xu if (!regions_needed) { 3632103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3642103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3652103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3662103cf9cSPeter Xu coalesce_file_region(map, nrg); 3672103cf9cSPeter Xu } else 3682103cf9cSPeter Xu *regions_needed += 1; 3692103cf9cSPeter Xu 3702103cf9cSPeter Xu return to - from; 3712103cf9cSPeter Xu } 3722103cf9cSPeter Xu 373972a3da3SWei Yang /* 374972a3da3SWei Yang * Must be called with resv->lock held. 375972a3da3SWei Yang * 376972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 377972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 378972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 379972a3da3SWei Yang * the regions for this range. 380d75c6af9SMina Almasry */ 381d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 382075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 383972a3da3SWei Yang struct hstate *h, long *regions_needed) 384d75c6af9SMina Almasry { 3850db9d74eSMina Almasry long add = 0; 386d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3870db9d74eSMina Almasry long last_accounted_offset = f; 3882103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 389d75c6af9SMina Almasry 3900db9d74eSMina Almasry if (regions_needed) 3910db9d74eSMina Almasry *regions_needed = 0; 392d75c6af9SMina Almasry 3930db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3940db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3950db9d74eSMina Almasry * bounds checking. 3960db9d74eSMina Almasry */ 3970db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 3980db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 3990db9d74eSMina Almasry if (rg->from < f) { 4000db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4010db9d74eSMina Almasry * then we need to update last_accounted_offset. 4020db9d74eSMina Almasry */ 4030db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4040db9d74eSMina Almasry last_accounted_offset = rg->to; 4050db9d74eSMina Almasry continue; 4060db9d74eSMina Almasry } 407d75c6af9SMina Almasry 4080db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4090db9d74eSMina Almasry * finished. 4100db9d74eSMina Almasry */ 411ca7e0457SPeter Xu if (rg->from >= t) 412d75c6af9SMina Almasry break; 413d75c6af9SMina Almasry 4140db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4150db9d74eSMina Almasry * update last_accounted_offset. 416d75c6af9SMina Almasry */ 4172103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4182103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4192103cf9cSPeter Xu last_accounted_offset, 4202103cf9cSPeter Xu rg->from, h, h_cg, 4212103cf9cSPeter Xu regions_needed); 422d75c6af9SMina Almasry 4230db9d74eSMina Almasry last_accounted_offset = rg->to; 4240db9d74eSMina Almasry } 4250db9d74eSMina Almasry 4260db9d74eSMina Almasry /* Handle the case where our range extends beyond 4270db9d74eSMina Almasry * last_accounted_offset. 4280db9d74eSMina Almasry */ 4292103cf9cSPeter Xu if (last_accounted_offset < t) 4302103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4312103cf9cSPeter Xu t, h, h_cg, regions_needed); 4320db9d74eSMina Almasry 4330db9d74eSMina Almasry VM_BUG_ON(add < 0); 4340db9d74eSMina Almasry return add; 4350db9d74eSMina Almasry } 4360db9d74eSMina Almasry 4370db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4380db9d74eSMina Almasry */ 4390db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4400db9d74eSMina Almasry int regions_needed) 4410db9d74eSMina Almasry __must_hold(&resv->lock) 4420db9d74eSMina Almasry { 4430db9d74eSMina Almasry struct list_head allocated_regions; 4440db9d74eSMina Almasry int to_allocate = 0, i = 0; 4450db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4460db9d74eSMina Almasry 4470db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4480db9d74eSMina Almasry 4490db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4500db9d74eSMina Almasry 4510db9d74eSMina Almasry /* 4520db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4530db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4540db9d74eSMina Almasry * 4550db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4560db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4570db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4580db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4590db9d74eSMina Almasry */ 4600db9d74eSMina Almasry while (resv->region_cache_count < 4610db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4620db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4630db9d74eSMina Almasry resv->region_cache_count; 4640db9d74eSMina Almasry 4650db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4660db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4670db9d74eSMina Almasry * needing to allocate for regions_needed. 4680db9d74eSMina Almasry */ 4690db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4700db9d74eSMina Almasry 4710db9d74eSMina Almasry spin_unlock(&resv->lock); 4720db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4730db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4740db9d74eSMina Almasry if (!trg) 4750db9d74eSMina Almasry goto out_of_memory; 4760db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4770db9d74eSMina Almasry } 4780db9d74eSMina Almasry 4790db9d74eSMina Almasry spin_lock(&resv->lock); 4800db9d74eSMina Almasry 481d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 482d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4830db9d74eSMina Almasry } 4840db9d74eSMina Almasry 4850db9d74eSMina Almasry return 0; 4860db9d74eSMina Almasry 4870db9d74eSMina Almasry out_of_memory: 4880db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 489d75c6af9SMina Almasry list_del(&rg->link); 490d75c6af9SMina Almasry kfree(rg); 491d75c6af9SMina Almasry } 4920db9d74eSMina Almasry return -ENOMEM; 493d75c6af9SMina Almasry } 494d75c6af9SMina Almasry 4951dd308a7SMike Kravetz /* 4961dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 4970db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 4980db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 4990db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5000db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5010db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 502cf3ad20bSMike Kravetz * 5030db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5040db9d74eSMina Almasry * 5050db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5060db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5077c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5080db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5090db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5100db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5111dd308a7SMike Kravetz */ 5120db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 513075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 514075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 51596822904SAndy Whitcroft { 5160db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 51796822904SAndy Whitcroft 5187b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5190db9d74eSMina Almasry retry: 5200db9d74eSMina Almasry 5210db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 522972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 523972a3da3SWei Yang &actual_regions_needed); 52496822904SAndy Whitcroft 5255e911373SMike Kravetz /* 5260db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5270db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5280db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5290db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5300db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5310db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5320db9d74eSMina Almasry * operation. 5335e911373SMike Kravetz */ 5340db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5350db9d74eSMina Almasry resv->region_cache_count < 5360db9d74eSMina Almasry resv->adds_in_progress + 5370db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5380db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5390db9d74eSMina Almasry * allocate file_region entries. 5400db9d74eSMina Almasry */ 5410db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5425e911373SMike Kravetz 5430db9d74eSMina Almasry if (allocate_file_region_entries( 5440db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5450db9d74eSMina Almasry return -ENOMEM; 5465e911373SMike Kravetz } 5475e911373SMike Kravetz 5480db9d74eSMina Almasry goto retry; 5490db9d74eSMina Almasry } 550cf3ad20bSMike Kravetz 551972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5520db9d74eSMina Almasry 5530db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5540db9d74eSMina Almasry 5557b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 556cf3ad20bSMike Kravetz return add; 55796822904SAndy Whitcroft } 55896822904SAndy Whitcroft 5591dd308a7SMike Kravetz /* 5601dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5611dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5621dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5631dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5641dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5651dd308a7SMike Kravetz * not change the number of huge pages represented by the 5660db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5670db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5680db9d74eSMina Almasry * file_region structure is added. 5690db9d74eSMina Almasry * 5700db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5710db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5720db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5735e911373SMike Kravetz * 5745e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5755e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5765e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5775e911373SMike Kravetz * is needed and can not be allocated. 5781dd308a7SMike Kravetz */ 5790db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5800db9d74eSMina Almasry long *out_regions_needed) 58196822904SAndy Whitcroft { 58296822904SAndy Whitcroft long chg = 0; 58396822904SAndy Whitcroft 5847b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5855e911373SMike Kravetz 586972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 587075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 588972a3da3SWei Yang out_regions_needed); 5895e911373SMike Kravetz 5900db9d74eSMina Almasry if (*out_regions_needed == 0) 5910db9d74eSMina Almasry *out_regions_needed = 1; 5925e911373SMike Kravetz 5930db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5945e911373SMike Kravetz return -ENOMEM; 5955e911373SMike Kravetz 5960db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 59796822904SAndy Whitcroft 5987b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 59996822904SAndy Whitcroft return chg; 60096822904SAndy Whitcroft } 60196822904SAndy Whitcroft 6021dd308a7SMike Kravetz /* 6035e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6045e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6055e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6065e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6070db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6080db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6090db9d74eSMina Almasry * the adds_in_progress counter. 6105e911373SMike Kravetz * 6115e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6125e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6135e911373SMike Kravetz * arguments will match the associated region_chg call. 6145e911373SMike Kravetz */ 6150db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6160db9d74eSMina Almasry long regions_needed) 6175e911373SMike Kravetz { 6185e911373SMike Kravetz spin_lock(&resv->lock); 6195e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6200db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6215e911373SMike Kravetz spin_unlock(&resv->lock); 6225e911373SMike Kravetz } 6235e911373SMike Kravetz 6245e911373SMike Kravetz /* 625feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 626feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 627feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 628feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 629feba16e2SMike Kravetz * 630feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 631feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 632feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 633feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 634feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 635feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 636feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6371dd308a7SMike Kravetz */ 638feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 63996822904SAndy Whitcroft { 6401406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 64196822904SAndy Whitcroft struct file_region *rg, *trg; 642feba16e2SMike Kravetz struct file_region *nrg = NULL; 643feba16e2SMike Kravetz long del = 0; 64496822904SAndy Whitcroft 645feba16e2SMike Kravetz retry: 6467b24d861SDavidlohr Bueso spin_lock(&resv->lock); 647feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 648dbe409e4SMike Kravetz /* 649dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 650dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 651dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 652dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 653dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 654dbe409e4SMike Kravetz */ 655dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 656feba16e2SMike Kravetz continue; 657dbe409e4SMike Kravetz 658feba16e2SMike Kravetz if (rg->from >= t) 65996822904SAndy Whitcroft break; 66096822904SAndy Whitcroft 661feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 662feba16e2SMike Kravetz /* 663feba16e2SMike Kravetz * Check for an entry in the cache before dropping 664feba16e2SMike Kravetz * lock and attempting allocation. 665feba16e2SMike Kravetz */ 666feba16e2SMike Kravetz if (!nrg && 667feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 668feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 669feba16e2SMike Kravetz struct file_region, 670feba16e2SMike Kravetz link); 671feba16e2SMike Kravetz list_del(&nrg->link); 672feba16e2SMike Kravetz resv->region_cache_count--; 67396822904SAndy Whitcroft } 67496822904SAndy Whitcroft 675feba16e2SMike Kravetz if (!nrg) { 676feba16e2SMike Kravetz spin_unlock(&resv->lock); 677feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 678feba16e2SMike Kravetz if (!nrg) 679feba16e2SMike Kravetz return -ENOMEM; 680feba16e2SMike Kravetz goto retry; 681feba16e2SMike Kravetz } 682feba16e2SMike Kravetz 683feba16e2SMike Kravetz del += t - f; 68479aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 685d85aecf2SMiaohe Lin resv, rg, t - f, false); 686feba16e2SMike Kravetz 687feba16e2SMike Kravetz /* New entry for end of split region */ 688feba16e2SMike Kravetz nrg->from = t; 689feba16e2SMike Kravetz nrg->to = rg->to; 690075a61d0SMina Almasry 691075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 692075a61d0SMina Almasry 693feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 694feba16e2SMike Kravetz 695feba16e2SMike Kravetz /* Original entry is trimmed */ 696feba16e2SMike Kravetz rg->to = f; 697feba16e2SMike Kravetz 698feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 699feba16e2SMike Kravetz nrg = NULL; 70096822904SAndy Whitcroft break; 701feba16e2SMike Kravetz } 702feba16e2SMike Kravetz 703feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 704feba16e2SMike Kravetz del += rg->to - rg->from; 705075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 706d85aecf2SMiaohe Lin rg->to - rg->from, true); 70796822904SAndy Whitcroft list_del(&rg->link); 70896822904SAndy Whitcroft kfree(rg); 709feba16e2SMike Kravetz continue; 71096822904SAndy Whitcroft } 7117b24d861SDavidlohr Bueso 712feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 713075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 714d85aecf2SMiaohe Lin t - rg->from, false); 715075a61d0SMina Almasry 71679aa925bSMike Kravetz del += t - rg->from; 71779aa925bSMike Kravetz rg->from = t; 71879aa925bSMike Kravetz } else { /* Trim end of region */ 719075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 720d85aecf2SMiaohe Lin rg->to - f, false); 72179aa925bSMike Kravetz 72279aa925bSMike Kravetz del += rg->to - f; 72379aa925bSMike Kravetz rg->to = f; 724feba16e2SMike Kravetz } 725feba16e2SMike Kravetz } 726feba16e2SMike Kravetz 7277b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 728feba16e2SMike Kravetz kfree(nrg); 729feba16e2SMike Kravetz return del; 73096822904SAndy Whitcroft } 73196822904SAndy Whitcroft 7321dd308a7SMike Kravetz /* 733b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 734b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 735b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 736b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 737b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 738b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 739b5cec28dSMike Kravetz * counts. 740b5cec28dSMike Kravetz */ 74172e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 742b5cec28dSMike Kravetz { 743b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 744b5cec28dSMike Kravetz long rsv_adjust; 745da56388cSMiaohe Lin bool reserved = false; 746b5cec28dSMike Kravetz 747b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 748da56388cSMiaohe Lin if (rsv_adjust > 0) { 749b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 750b5cec28dSMike Kravetz 751da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 752da56388cSMiaohe Lin reserved = true; 753da56388cSMiaohe Lin } else if (!rsv_adjust) { 754da56388cSMiaohe Lin reserved = true; 755b5cec28dSMike Kravetz } 756da56388cSMiaohe Lin 757da56388cSMiaohe Lin if (!reserved) 758da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 759b5cec28dSMike Kravetz } 760b5cec28dSMike Kravetz 761b5cec28dSMike Kravetz /* 7621dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7631dd308a7SMike Kravetz * that intersect with the range [f, t). 7641dd308a7SMike Kravetz */ 7651406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 76684afd99bSAndy Whitcroft { 7671406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 76884afd99bSAndy Whitcroft struct file_region *rg; 76984afd99bSAndy Whitcroft long chg = 0; 77084afd99bSAndy Whitcroft 7717b24d861SDavidlohr Bueso spin_lock(&resv->lock); 77284afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 77384afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 774f2135a4aSWang Sheng-Hui long seg_from; 775f2135a4aSWang Sheng-Hui long seg_to; 77684afd99bSAndy Whitcroft 77784afd99bSAndy Whitcroft if (rg->to <= f) 77884afd99bSAndy Whitcroft continue; 77984afd99bSAndy Whitcroft if (rg->from >= t) 78084afd99bSAndy Whitcroft break; 78184afd99bSAndy Whitcroft 78284afd99bSAndy Whitcroft seg_from = max(rg->from, f); 78384afd99bSAndy Whitcroft seg_to = min(rg->to, t); 78484afd99bSAndy Whitcroft 78584afd99bSAndy Whitcroft chg += seg_to - seg_from; 78684afd99bSAndy Whitcroft } 7877b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 78884afd99bSAndy Whitcroft 78984afd99bSAndy Whitcroft return chg; 79084afd99bSAndy Whitcroft } 79184afd99bSAndy Whitcroft 79296822904SAndy Whitcroft /* 793e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 794e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 795e7c4b0bfSAndy Whitcroft */ 796a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 797a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 798e7c4b0bfSAndy Whitcroft { 799a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 800a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 801e7c4b0bfSAndy Whitcroft } 802e7c4b0bfSAndy Whitcroft 8030fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8040fe6e20bSNaoya Horiguchi unsigned long address) 8050fe6e20bSNaoya Horiguchi { 8060fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8070fe6e20bSNaoya Horiguchi } 808dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8090fe6e20bSNaoya Horiguchi 81084afd99bSAndy Whitcroft /* 81108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 81208fba699SMel Gorman * cases this will be same size as used by the page table entries. 81308fba699SMel Gorman */ 81408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 81508fba699SMel Gorman { 81605ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 81705ea8860SDan Williams return vma->vm_ops->pagesize(vma); 81808fba699SMel Gorman return PAGE_SIZE; 81908fba699SMel Gorman } 820f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 82108fba699SMel Gorman 82208fba699SMel Gorman /* 8233340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8243340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 82509135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 82609135cc5SDan Williams * version of this symbol is required. 8273340289dSMel Gorman */ 82809135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8293340289dSMel Gorman { 8303340289dSMel Gorman return vma_kernel_pagesize(vma); 8313340289dSMel Gorman } 8323340289dSMel Gorman 8333340289dSMel Gorman /* 83484afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 83584afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 83684afd99bSAndy Whitcroft * alignment. 83784afd99bSAndy Whitcroft */ 83884afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 83984afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 84004f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 84184afd99bSAndy Whitcroft 842a1e78772SMel Gorman /* 843a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 844a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 845a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 846a1e78772SMel Gorman * 847a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 848a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 849a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 850a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 85184afd99bSAndy Whitcroft * 85284afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 85384afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 85484afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 85584afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 85684afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 85784afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 85884afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 85984afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 860a1e78772SMel Gorman */ 861e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 862e7c4b0bfSAndy Whitcroft { 863e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 864e7c4b0bfSAndy Whitcroft } 865e7c4b0bfSAndy Whitcroft 866e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 867e7c4b0bfSAndy Whitcroft unsigned long value) 868e7c4b0bfSAndy Whitcroft { 869e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 870e7c4b0bfSAndy Whitcroft } 871e7c4b0bfSAndy Whitcroft 872e9fe92aeSMina Almasry static void 873e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 874e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 875e9fe92aeSMina Almasry struct hstate *h) 876e9fe92aeSMina Almasry { 877e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 878e9fe92aeSMina Almasry if (!h_cg || !h) { 879e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 880e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 881e9fe92aeSMina Almasry resv_map->css = NULL; 882e9fe92aeSMina Almasry } else { 883e9fe92aeSMina Almasry resv_map->reservation_counter = 884e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 885e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 886e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 887e9fe92aeSMina Almasry } 888e9fe92aeSMina Almasry #endif 889e9fe92aeSMina Almasry } 890e9fe92aeSMina Almasry 8919119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 89284afd99bSAndy Whitcroft { 89384afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8945e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8955e911373SMike Kravetz 8965e911373SMike Kravetz if (!resv_map || !rg) { 8975e911373SMike Kravetz kfree(resv_map); 8985e911373SMike Kravetz kfree(rg); 89984afd99bSAndy Whitcroft return NULL; 9005e911373SMike Kravetz } 90184afd99bSAndy Whitcroft 90284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9037b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 90484afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 90584afd99bSAndy Whitcroft 9065e911373SMike Kravetz resv_map->adds_in_progress = 0; 907e9fe92aeSMina Almasry /* 908e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 909e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 910e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 911e9fe92aeSMina Almasry * reservations are to be un-charged from here. 912e9fe92aeSMina Almasry */ 913e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9145e911373SMike Kravetz 9155e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9165e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9175e911373SMike Kravetz resv_map->region_cache_count = 1; 9185e911373SMike Kravetz 91984afd99bSAndy Whitcroft return resv_map; 92084afd99bSAndy Whitcroft } 92184afd99bSAndy Whitcroft 9229119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 92384afd99bSAndy Whitcroft { 92484afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9255e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9265e911373SMike Kravetz struct file_region *rg, *trg; 92784afd99bSAndy Whitcroft 92884afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 929feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9305e911373SMike Kravetz 9315e911373SMike Kravetz /* ... and any entries left in the cache */ 9325e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9335e911373SMike Kravetz list_del(&rg->link); 9345e911373SMike Kravetz kfree(rg); 9355e911373SMike Kravetz } 9365e911373SMike Kravetz 9375e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9385e911373SMike Kravetz 93984afd99bSAndy Whitcroft kfree(resv_map); 94084afd99bSAndy Whitcroft } 94184afd99bSAndy Whitcroft 9424e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9434e35f483SJoonsoo Kim { 944f27a5136SMike Kravetz /* 945f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 946f27a5136SMike Kravetz * address space within the inode. This original address space 947f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 948f27a5136SMike Kravetz * address space embedded within the inode. 949f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 950f27a5136SMike Kravetz * this may not be true for device special inodes. 951f27a5136SMike Kravetz */ 952f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9534e35f483SJoonsoo Kim } 9544e35f483SJoonsoo Kim 95584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 956a1e78772SMel Gorman { 95781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9584e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9594e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9604e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9614e35f483SJoonsoo Kim 9624e35f483SJoonsoo Kim return inode_resv_map(inode); 9634e35f483SJoonsoo Kim 9644e35f483SJoonsoo Kim } else { 96584afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 96684afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9674e35f483SJoonsoo Kim } 968a1e78772SMel Gorman } 969a1e78772SMel Gorman 97084afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 971a1e78772SMel Gorman { 97281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 97381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 974a1e78772SMel Gorman 97584afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 97684afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 97704f2cbe3SMel Gorman } 97804f2cbe3SMel Gorman 97904f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 98004f2cbe3SMel Gorman { 98181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98281d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 983e7c4b0bfSAndy Whitcroft 984e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 98504f2cbe3SMel Gorman } 98604f2cbe3SMel Gorman 98704f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 98804f2cbe3SMel Gorman { 98981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 990e7c4b0bfSAndy Whitcroft 991e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 992a1e78772SMel Gorman } 993a1e78772SMel Gorman 99404f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 995a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 996a1e78772SMel Gorman { 99781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 998f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 999a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1000a1e78772SMel Gorman } 1001a1e78772SMel Gorman 1002a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1003559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1004a1e78772SMel Gorman { 1005af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1006af0ed73eSJoonsoo Kim /* 1007af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1008af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1009af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1010af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1011af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1012af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1013af0ed73eSJoonsoo Kim * properly, so add work-around here. 1014af0ed73eSJoonsoo Kim */ 1015af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1016559ec2f8SNicholas Krause return true; 1017af0ed73eSJoonsoo Kim else 1018559ec2f8SNicholas Krause return false; 1019af0ed73eSJoonsoo Kim } 1020a63884e9SJoonsoo Kim 1021a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10221fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10231fb1b0e9SMike Kravetz /* 10241fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10251fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10261fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10277c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10281fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10291fb1b0e9SMike Kravetz */ 10301fb1b0e9SMike Kravetz if (chg) 10311fb1b0e9SMike Kravetz return false; 10321fb1b0e9SMike Kravetz else 1033559ec2f8SNicholas Krause return true; 10341fb1b0e9SMike Kravetz } 1035a63884e9SJoonsoo Kim 1036a63884e9SJoonsoo Kim /* 1037a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1038a63884e9SJoonsoo Kim * private mappings. 1039a63884e9SJoonsoo Kim */ 104067961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 104167961f9dSMike Kravetz /* 104267961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 104367961f9dSMike Kravetz * could have been performed on the private mapping. 104467961f9dSMike Kravetz * Examine the value of chg to determine if reserves 104567961f9dSMike Kravetz * actually exist or were previously consumed. 104667961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 104767961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 104867961f9dSMike Kravetz * private mappings has different (opposite) semantics 104967961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 105067961f9dSMike Kravetz * has already taken this difference in semantics into 105167961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 105267961f9dSMike Kravetz * as in the shared case above. Code could easily be 105367961f9dSMike Kravetz * combined, but keeping it separate draws attention to 105467961f9dSMike Kravetz * subtle differences. 105567961f9dSMike Kravetz */ 105667961f9dSMike Kravetz if (chg) 105767961f9dSMike Kravetz return false; 105867961f9dSMike Kravetz else 1059559ec2f8SNicholas Krause return true; 106067961f9dSMike Kravetz } 1061a63884e9SJoonsoo Kim 1062559ec2f8SNicholas Krause return false; 1063a1e78772SMel Gorman } 1064a1e78772SMel Gorman 1065a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10661da177e4SLinus Torvalds { 10671da177e4SLinus Torvalds int nid = page_to_nid(page); 10680edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1069a5516438SAndi Kleen h->free_huge_pages++; 1070a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10716c037149SMike Kravetz SetHPageFreed(page); 10721da177e4SLinus Torvalds } 10731da177e4SLinus Torvalds 107494310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1075bf50bab2SNaoya Horiguchi { 1076bf50bab2SNaoya Horiguchi struct page *page; 1077bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1078bf50bab2SNaoya Horiguchi 1079bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1080bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1081bbe88753SJoonsoo Kim continue; 1082bbe88753SJoonsoo Kim 10836664bfc8SWei Yang if (PageHWPoison(page)) 10846664bfc8SWei Yang continue; 1085bbe88753SJoonsoo Kim 10860edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1087a9869b83SNaoya Horiguchi set_page_refcounted(page); 10886c037149SMike Kravetz ClearHPageFreed(page); 1089bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1090bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1091bf50bab2SNaoya Horiguchi return page; 1092bf50bab2SNaoya Horiguchi } 1093bf50bab2SNaoya Horiguchi 10946664bfc8SWei Yang return NULL; 10956664bfc8SWei Yang } 10966664bfc8SWei Yang 10973e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10983e59fcb0SMichal Hocko nodemask_t *nmask) 109994310cbcSAnshuman Khandual { 11003e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11013e59fcb0SMichal Hocko struct zonelist *zonelist; 11023e59fcb0SMichal Hocko struct zone *zone; 11033e59fcb0SMichal Hocko struct zoneref *z; 110498fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11053e59fcb0SMichal Hocko 11063e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11073e59fcb0SMichal Hocko 11083e59fcb0SMichal Hocko retry_cpuset: 11093e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11103e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 111194310cbcSAnshuman Khandual struct page *page; 111294310cbcSAnshuman Khandual 11133e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11143e59fcb0SMichal Hocko continue; 11153e59fcb0SMichal Hocko /* 11163e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11173e59fcb0SMichal Hocko * zone aware 11183e59fcb0SMichal Hocko */ 11193e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11203e59fcb0SMichal Hocko continue; 11213e59fcb0SMichal Hocko node = zone_to_nid(zone); 112294310cbcSAnshuman Khandual 112394310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 112494310cbcSAnshuman Khandual if (page) 112594310cbcSAnshuman Khandual return page; 112694310cbcSAnshuman Khandual } 11273e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11283e59fcb0SMichal Hocko goto retry_cpuset; 11293e59fcb0SMichal Hocko 113094310cbcSAnshuman Khandual return NULL; 113194310cbcSAnshuman Khandual } 113294310cbcSAnshuman Khandual 1133a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1134a5516438SAndi Kleen struct vm_area_struct *vma, 1135af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1136af0ed73eSJoonsoo Kim long chg) 11371da177e4SLinus Torvalds { 11383e59fcb0SMichal Hocko struct page *page; 1139480eccf9SLee Schermerhorn struct mempolicy *mpol; 114004ec6264SVlastimil Babka gfp_t gfp_mask; 11413e59fcb0SMichal Hocko nodemask_t *nodemask; 114204ec6264SVlastimil Babka int nid; 11431da177e4SLinus Torvalds 1144a1e78772SMel Gorman /* 1145a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1146a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1147a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1148a1e78772SMel Gorman */ 1149af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1150a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1151c0ff7453SMiao Xie goto err; 1152a1e78772SMel Gorman 115304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1154a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11556eab04a8SJustin P. Mattock goto err; 115604f2cbe3SMel Gorman 115704ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 115804ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11593e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11603e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1161d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1162a63884e9SJoonsoo Kim h->resv_huge_pages--; 1163bf50bab2SNaoya Horiguchi } 1164cc9a6c87SMel Gorman 1165cc9a6c87SMel Gorman mpol_cond_put(mpol); 1166cc9a6c87SMel Gorman return page; 1167cc9a6c87SMel Gorman 1168c0ff7453SMiao Xie err: 1169cc9a6c87SMel Gorman return NULL; 11701da177e4SLinus Torvalds } 11711da177e4SLinus Torvalds 11721cac6f2cSLuiz Capitulino /* 11731cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11741cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11751cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11761cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11771cac6f2cSLuiz Capitulino * node for alloc or free. 11781cac6f2cSLuiz Capitulino */ 11791cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11801cac6f2cSLuiz Capitulino { 11810edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11821cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11831cac6f2cSLuiz Capitulino 11841cac6f2cSLuiz Capitulino return nid; 11851cac6f2cSLuiz Capitulino } 11861cac6f2cSLuiz Capitulino 11871cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11881cac6f2cSLuiz Capitulino { 11891cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11901cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11911cac6f2cSLuiz Capitulino return nid; 11921cac6f2cSLuiz Capitulino } 11931cac6f2cSLuiz Capitulino 11941cac6f2cSLuiz Capitulino /* 11951cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11961cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11971cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11981cac6f2cSLuiz Capitulino * mask. 11991cac6f2cSLuiz Capitulino */ 12001cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12011cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12021cac6f2cSLuiz Capitulino { 12031cac6f2cSLuiz Capitulino int nid; 12041cac6f2cSLuiz Capitulino 12051cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12061cac6f2cSLuiz Capitulino 12071cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12081cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12091cac6f2cSLuiz Capitulino 12101cac6f2cSLuiz Capitulino return nid; 12111cac6f2cSLuiz Capitulino } 12121cac6f2cSLuiz Capitulino 12131cac6f2cSLuiz Capitulino /* 12141cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 12151cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12161cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12171cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12181cac6f2cSLuiz Capitulino */ 12191cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12201cac6f2cSLuiz Capitulino { 12211cac6f2cSLuiz Capitulino int nid; 12221cac6f2cSLuiz Capitulino 12231cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12241cac6f2cSLuiz Capitulino 12251cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12261cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12271cac6f2cSLuiz Capitulino 12281cac6f2cSLuiz Capitulino return nid; 12291cac6f2cSLuiz Capitulino } 12301cac6f2cSLuiz Capitulino 12311cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12321cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12331cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12341cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12351cac6f2cSLuiz Capitulino nr_nodes--) 12361cac6f2cSLuiz Capitulino 12371cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(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_free(hs, mask)) || 1); \ 12411cac6f2cSLuiz Capitulino nr_nodes--) 12421cac6f2cSLuiz Capitulino 1243e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1244944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1245d00181b9SKirill A. Shutemov unsigned int order) 1246944d9fecSLuiz Capitulino { 1247944d9fecSLuiz Capitulino int i; 1248944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1249944d9fecSLuiz Capitulino struct page *p = page + 1; 1250944d9fecSLuiz Capitulino 1251c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 125247e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 125347e29d32SJohn Hubbard 1254944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12551d798ca3SKirill A. Shutemov clear_compound_head(p); 1256944d9fecSLuiz Capitulino set_page_refcounted(p); 1257944d9fecSLuiz Capitulino } 1258944d9fecSLuiz Capitulino 1259944d9fecSLuiz Capitulino set_compound_order(page, 0); 1260ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1261944d9fecSLuiz Capitulino __ClearPageHead(page); 1262944d9fecSLuiz Capitulino } 1263944d9fecSLuiz Capitulino 1264d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1265944d9fecSLuiz Capitulino { 1266cf11e85fSRoman Gushchin /* 1267cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1268cf11e85fSRoman Gushchin * cma_release() returns false. 1269cf11e85fSRoman Gushchin */ 1270dbda8feaSBarry Song #ifdef CONFIG_CMA 1271dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1272cf11e85fSRoman Gushchin return; 1273dbda8feaSBarry Song #endif 1274cf11e85fSRoman Gushchin 1275944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1276944d9fecSLuiz Capitulino } 1277944d9fecSLuiz Capitulino 12784eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1279d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1280d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1281944d9fecSLuiz Capitulino { 128204adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1283953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1284953f064aSLi Xinhai nid = numa_mem_id(); 1285944d9fecSLuiz Capitulino 1286dbda8feaSBarry Song #ifdef CONFIG_CMA 1287dbda8feaSBarry Song { 1288cf11e85fSRoman Gushchin struct page *page; 1289cf11e85fSRoman Gushchin int node; 1290cf11e85fSRoman Gushchin 1291953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1292953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1293953f064aSLi Xinhai huge_page_order(h), true); 1294953f064aSLi Xinhai if (page) 1295953f064aSLi Xinhai return page; 1296953f064aSLi Xinhai } 1297953f064aSLi Xinhai 1298953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1299cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1300953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1301cf11e85fSRoman Gushchin continue; 1302cf11e85fSRoman Gushchin 1303cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1304cf11e85fSRoman Gushchin huge_page_order(h), true); 1305cf11e85fSRoman Gushchin if (page) 1306cf11e85fSRoman Gushchin return page; 1307cf11e85fSRoman Gushchin } 1308cf11e85fSRoman Gushchin } 1309953f064aSLi Xinhai } 1310dbda8feaSBarry Song #endif 1311cf11e85fSRoman Gushchin 13125e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1313944d9fecSLuiz Capitulino } 1314944d9fecSLuiz Capitulino 1315944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1316d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 13174eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13184eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13194eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13204eb0716eSAlexandre Ghiti { 13214eb0716eSAlexandre Ghiti return NULL; 13224eb0716eSAlexandre Ghiti } 13234eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1324944d9fecSLuiz Capitulino 1325e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1326d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13274eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13284eb0716eSAlexandre Ghiti { 13294eb0716eSAlexandre Ghiti return NULL; 13304eb0716eSAlexandre Ghiti } 1331d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1332944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1333d00181b9SKirill A. Shutemov unsigned int order) { } 1334944d9fecSLuiz Capitulino #endif 1335944d9fecSLuiz Capitulino 1336*6eb4e88aSMike Kravetz /* 1337*6eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 1338*6eb4e88aSMike Kravetz * as just a compound page. A reference is held on the page. 1339*6eb4e88aSMike Kravetz * 1340*6eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 1341*6eb4e88aSMike Kravetz */ 1342*6eb4e88aSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 1343*6eb4e88aSMike Kravetz bool adjust_surplus) 1344*6eb4e88aSMike Kravetz { 1345*6eb4e88aSMike Kravetz int nid = page_to_nid(page); 1346*6eb4e88aSMike Kravetz 1347*6eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1348*6eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1349*6eb4e88aSMike Kravetz 1350*6eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1351*6eb4e88aSMike Kravetz return; 1352*6eb4e88aSMike Kravetz 1353*6eb4e88aSMike Kravetz list_del(&page->lru); 1354*6eb4e88aSMike Kravetz 1355*6eb4e88aSMike Kravetz if (HPageFreed(page)) { 1356*6eb4e88aSMike Kravetz h->free_huge_pages--; 1357*6eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 1358*6eb4e88aSMike Kravetz } 1359*6eb4e88aSMike Kravetz if (adjust_surplus) { 1360*6eb4e88aSMike Kravetz h->surplus_huge_pages--; 1361*6eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 1362*6eb4e88aSMike Kravetz } 1363*6eb4e88aSMike Kravetz 1364*6eb4e88aSMike Kravetz set_page_refcounted(page); 1365*6eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1366*6eb4e88aSMike Kravetz 1367*6eb4e88aSMike Kravetz h->nr_huge_pages--; 1368*6eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 1369*6eb4e88aSMike Kravetz } 1370*6eb4e88aSMike Kravetz 1371a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 13726af2acb6SAdam Litke { 13736af2acb6SAdam Litke int i; 1374dbfee5aeSMike Kravetz struct page *subpage = page; 1375a5516438SAndi Kleen 13764eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1377944d9fecSLuiz Capitulino return; 137818229df5SAndy Whitcroft 1379dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1380dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1381dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 138232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1383a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1384a7407a27SLuiz Capitulino 1 << PG_writeback); 13856af2acb6SAdam Litke } 1386944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1387944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1388944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1389944d9fecSLuiz Capitulino } else { 1390a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13916af2acb6SAdam Litke } 1392944d9fecSLuiz Capitulino } 13936af2acb6SAdam Litke 1394e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1395e5ff2159SAndi Kleen { 1396e5ff2159SAndi Kleen struct hstate *h; 1397e5ff2159SAndi Kleen 1398e5ff2159SAndi Kleen for_each_hstate(h) { 1399e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1400e5ff2159SAndi Kleen return h; 1401e5ff2159SAndi Kleen } 1402e5ff2159SAndi Kleen return NULL; 1403e5ff2159SAndi Kleen } 1404e5ff2159SAndi Kleen 1405c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 140627a85ef1SDavid Gibson { 1407a5516438SAndi Kleen /* 1408a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1409a5516438SAndi Kleen * compound page destructor. 1410a5516438SAndi Kleen */ 1411e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 14127893d1d5SAdam Litke int nid = page_to_nid(page); 1413d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 141407443a85SJoonsoo Kim bool restore_reserve; 141527a85ef1SDavid Gibson 1416b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1417b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14188ace22bcSYongkai Wu 1419d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14208ace22bcSYongkai Wu page->mapping = NULL; 1421d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1422d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 142327a85ef1SDavid Gibson 14241c5ecae3SMike Kravetz /* 1425d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 14260919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14270919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14280919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14296c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 14300919e1b6SMike Kravetz * remove the reserved page from the subpool. 14310919e1b6SMike Kravetz */ 14320919e1b6SMike Kravetz if (!restore_reserve) { 14330919e1b6SMike Kravetz /* 14340919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14350919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14360919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14370919e1b6SMike Kravetz * operation. 14381c5ecae3SMike Kravetz */ 14391c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14401c5ecae3SMike Kravetz restore_reserve = true; 14410919e1b6SMike Kravetz } 14421c5ecae3SMike Kravetz 144327a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 14448f251a3dSMike Kravetz ClearHPageMigratable(page); 14456d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14466d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 144708cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 144808cf9fafSMina Almasry pages_per_huge_page(h), page); 144907443a85SJoonsoo Kim if (restore_reserve) 145007443a85SJoonsoo Kim h->resv_huge_pages++; 145107443a85SJoonsoo Kim 14529157c311SMike Kravetz if (HPageTemporary(page)) { 1453*6eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1454ab5ac90aSMichal Hocko update_and_free_page(h, page); 1455ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14560edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1457*6eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1458a5516438SAndi Kleen update_and_free_page(h, page); 14597893d1d5SAdam Litke } else { 14605d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1461a5516438SAndi Kleen enqueue_huge_page(h, page); 14627893d1d5SAdam Litke } 146327a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 146427a85ef1SDavid Gibson } 146527a85ef1SDavid Gibson 1466c77c0a8aSWaiman Long /* 1467c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1468c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1469c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1470c77c0a8aSWaiman Long * 1471c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1472c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1473c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1474c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1475c77c0a8aSWaiman Long */ 1476c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1477c77c0a8aSWaiman Long 1478c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1479c77c0a8aSWaiman Long { 1480c77c0a8aSWaiman Long struct llist_node *node; 1481c77c0a8aSWaiman Long struct page *page; 1482c77c0a8aSWaiman Long 1483c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1484c77c0a8aSWaiman Long 1485c77c0a8aSWaiman Long while (node) { 1486c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1487c77c0a8aSWaiman Long struct page, mapping); 1488c77c0a8aSWaiman Long node = node->next; 1489c77c0a8aSWaiman Long __free_huge_page(page); 1490c77c0a8aSWaiman Long } 1491c77c0a8aSWaiman Long } 1492c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1493c77c0a8aSWaiman Long 1494c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1495c77c0a8aSWaiman Long { 1496c77c0a8aSWaiman Long /* 1497c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1498c77c0a8aSWaiman Long */ 1499c77c0a8aSWaiman Long if (!in_task()) { 1500c77c0a8aSWaiman Long /* 1501c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1502c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1503c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1504c77c0a8aSWaiman Long */ 1505c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1506c77c0a8aSWaiman Long &hpage_freelist)) 1507c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1508c77c0a8aSWaiman Long return; 1509c77c0a8aSWaiman Long } 1510c77c0a8aSWaiman Long 1511c77c0a8aSWaiman Long __free_huge_page(page); 1512c77c0a8aSWaiman Long } 1513c77c0a8aSWaiman Long 1514a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1515b7ba30c6SAndi Kleen { 15160edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1517f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1518ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 15199dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15201adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 15212f37511cSWei Yang spin_lock(&hugetlb_lock); 1522a5516438SAndi Kleen h->nr_huge_pages++; 1523a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 15246c037149SMike Kravetz ClearHPageFreed(page); 1525b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1526b7ba30c6SAndi Kleen } 1527b7ba30c6SAndi Kleen 1528d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 152920a0307cSWu Fengguang { 153020a0307cSWu Fengguang int i; 153120a0307cSWu Fengguang int nr_pages = 1 << order; 153220a0307cSWu Fengguang struct page *p = page + 1; 153320a0307cSWu Fengguang 153420a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 153520a0307cSWu Fengguang set_compound_order(page, order); 1536ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1537de09d31dSKirill A. Shutemov __SetPageHead(page); 153820a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1539ef5a22beSAndrea Arcangeli /* 1540ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1541ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1542ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15437c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1544ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1545ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1546ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1547ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1548ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1549ef5a22beSAndrea Arcangeli * after get_user_pages(). 1550ef5a22beSAndrea Arcangeli */ 1551ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 155258a84aa9SYouquan Song set_page_count(p, 0); 15531d798ca3SKirill A. Shutemov set_compound_head(p, page); 155420a0307cSWu Fengguang } 1555b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 155647e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 155720a0307cSWu Fengguang } 155820a0307cSWu Fengguang 15597795912cSAndrew Morton /* 15607795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15617795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15627795912cSAndrew Morton * details. 15637795912cSAndrew Morton */ 156420a0307cSWu Fengguang int PageHuge(struct page *page) 156520a0307cSWu Fengguang { 156620a0307cSWu Fengguang if (!PageCompound(page)) 156720a0307cSWu Fengguang return 0; 156820a0307cSWu Fengguang 156920a0307cSWu Fengguang page = compound_head(page); 1570f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 157120a0307cSWu Fengguang } 157243131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 157343131e14SNaoya Horiguchi 157427c73ae7SAndrea Arcangeli /* 157527c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 157627c73ae7SAndrea Arcangeli * normal or transparent huge pages. 157727c73ae7SAndrea Arcangeli */ 157827c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 157927c73ae7SAndrea Arcangeli { 158027c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 158127c73ae7SAndrea Arcangeli return 0; 158227c73ae7SAndrea Arcangeli 1583d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 158427c73ae7SAndrea Arcangeli } 158527c73ae7SAndrea Arcangeli 1586c0d0381aSMike Kravetz /* 1587c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1588c0d0381aSMike Kravetz * 1589336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1590336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1591336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1592c0d0381aSMike Kravetz */ 1593c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1594c0d0381aSMike Kravetz { 1595336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1596c0d0381aSMike Kravetz 1597c0d0381aSMike Kravetz if (!mapping) 1598c0d0381aSMike Kravetz return mapping; 1599c0d0381aSMike Kravetz 1600c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1601c0d0381aSMike Kravetz return mapping; 1602c0d0381aSMike Kravetz 1603c0d0381aSMike Kravetz return NULL; 1604c0d0381aSMike Kravetz } 1605c0d0381aSMike Kravetz 160613d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 160713d60f4bSZhang Yi { 160813d60f4bSZhang Yi struct page *page_head = compound_head(page); 160913d60f4bSZhang Yi pgoff_t index = page_index(page_head); 161013d60f4bSZhang Yi unsigned long compound_idx; 161113d60f4bSZhang Yi 161213d60f4bSZhang Yi if (!PageHuge(page_head)) 161313d60f4bSZhang Yi return page_index(page); 161413d60f4bSZhang Yi 161513d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 161613d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 161713d60f4bSZhang Yi else 161813d60f4bSZhang Yi compound_idx = page - page_head; 161913d60f4bSZhang Yi 162013d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 162113d60f4bSZhang Yi } 162213d60f4bSZhang Yi 16230c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1624f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1625f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16261da177e4SLinus Torvalds { 1627af0fb9dfSMichal Hocko int order = huge_page_order(h); 16281da177e4SLinus Torvalds struct page *page; 1629f60858f9SMike Kravetz bool alloc_try_hard = true; 1630f96efd58SJoe Jin 1631f60858f9SMike Kravetz /* 1632f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1633f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1634f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1635f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1636f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1637f60858f9SMike Kravetz */ 1638f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1639f60858f9SMike Kravetz alloc_try_hard = false; 1640f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1641f60858f9SMike Kravetz if (alloc_try_hard) 1642f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1643af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1644af0fb9dfSMichal Hocko nid = numa_mem_id(); 164584172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1646af0fb9dfSMichal Hocko if (page) 1647af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1648af0fb9dfSMichal Hocko else 1649af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 165063b4613cSNishanth Aravamudan 1651f60858f9SMike Kravetz /* 1652f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1653f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1654f60858f9SMike Kravetz * normal 'try hard' allocations. 1655f60858f9SMike Kravetz */ 1656f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1657f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1658f60858f9SMike Kravetz 1659f60858f9SMike Kravetz /* 1660f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1661f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1662f60858f9SMike Kravetz * overall state change. 1663f60858f9SMike Kravetz */ 1664f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1665f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1666f60858f9SMike Kravetz 166763b4613cSNishanth Aravamudan return page; 166863b4613cSNishanth Aravamudan } 166963b4613cSNishanth Aravamudan 1670af0fb9dfSMichal Hocko /* 16710c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16720c397daeSMichal Hocko * should use this function to get new hugetlb pages 16730c397daeSMichal Hocko */ 16740c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1675f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1676f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16770c397daeSMichal Hocko { 16780c397daeSMichal Hocko struct page *page; 16790c397daeSMichal Hocko 16800c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16810c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16820c397daeSMichal Hocko else 16830c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1684f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16850c397daeSMichal Hocko if (!page) 16860c397daeSMichal Hocko return NULL; 16870c397daeSMichal Hocko 16880c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16890c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16900c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16910c397daeSMichal Hocko 16920c397daeSMichal Hocko return page; 16930c397daeSMichal Hocko } 16940c397daeSMichal Hocko 16950c397daeSMichal Hocko /* 1696af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1697af0fb9dfSMichal Hocko * manner. 1698af0fb9dfSMichal Hocko */ 1699f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1700f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1701b2261026SJoonsoo Kim { 1702b2261026SJoonsoo Kim struct page *page; 1703b2261026SJoonsoo Kim int nr_nodes, node; 1704af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1705b2261026SJoonsoo Kim 1706b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1707f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1708f60858f9SMike Kravetz node_alloc_noretry); 1709af0fb9dfSMichal Hocko if (page) 1710b2261026SJoonsoo Kim break; 1711b2261026SJoonsoo Kim } 1712b2261026SJoonsoo Kim 1713af0fb9dfSMichal Hocko if (!page) 1714af0fb9dfSMichal Hocko return 0; 1715b2261026SJoonsoo Kim 1716af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1717af0fb9dfSMichal Hocko 1718af0fb9dfSMichal Hocko return 1; 1719b2261026SJoonsoo Kim } 1720b2261026SJoonsoo Kim 1721e8c5c824SLee Schermerhorn /* 1722e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1723e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1724e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1725e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1726e8c5c824SLee Schermerhorn */ 17276ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17286ae11b27SLee Schermerhorn bool acct_surplus) 1729e8c5c824SLee Schermerhorn { 1730b2261026SJoonsoo Kim int nr_nodes, node; 1731e8c5c824SLee Schermerhorn int ret = 0; 1732e8c5c824SLee Schermerhorn 1733b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1734685f3457SLee Schermerhorn /* 1735685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1736685f3457SLee Schermerhorn * nodes with surplus pages. 1737685f3457SLee Schermerhorn */ 1738b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1739b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1740e8c5c824SLee Schermerhorn struct page *page = 1741b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1742e8c5c824SLee Schermerhorn struct page, lru); 1743*6eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 1744e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1745e8c5c824SLee Schermerhorn ret = 1; 17469a76db09SLee Schermerhorn break; 1747e8c5c824SLee Schermerhorn } 1748b2261026SJoonsoo Kim } 1749e8c5c824SLee Schermerhorn 1750e8c5c824SLee Schermerhorn return ret; 1751e8c5c824SLee Schermerhorn } 1752e8c5c824SLee Schermerhorn 1753c8721bbbSNaoya Horiguchi /* 1754c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1755faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1756faf53defSNaoya Horiguchi * This function returns values like below: 1757faf53defSNaoya Horiguchi * 1758faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1759faf53defSNaoya Horiguchi * (allocated or reserved.) 1760faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1761faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1762c8721bbbSNaoya Horiguchi */ 1763c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1764c8721bbbSNaoya Horiguchi { 17656bc9b564SNaoya Horiguchi int rc = -EBUSY; 1766082d5b6bSGerald Schaefer 17677ffddd49SMuchun Song retry: 1768faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1769faf53defSNaoya Horiguchi if (!PageHuge(page)) 1770faf53defSNaoya Horiguchi return 0; 1771faf53defSNaoya Horiguchi 1772c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1773faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1774faf53defSNaoya Horiguchi rc = 0; 1775faf53defSNaoya Horiguchi goto out; 1776faf53defSNaoya Horiguchi } 1777faf53defSNaoya Horiguchi 1778faf53defSNaoya Horiguchi if (!page_count(page)) { 17792247bb33SGerald Schaefer struct page *head = compound_head(page); 17802247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17816bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1782082d5b6bSGerald Schaefer goto out; 17837ffddd49SMuchun Song 17847ffddd49SMuchun Song /* 17857ffddd49SMuchun Song * We should make sure that the page is already on the free list 17867ffddd49SMuchun Song * when it is dissolved. 17877ffddd49SMuchun Song */ 17886c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 17897ffddd49SMuchun Song spin_unlock(&hugetlb_lock); 17907ffddd49SMuchun Song cond_resched(); 17917ffddd49SMuchun Song 17927ffddd49SMuchun Song /* 17937ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 17947ffddd49SMuchun Song * encounter this race. In fact, we have a chance 17957ffddd49SMuchun Song * to successfully dissolve the page if we do a 17967ffddd49SMuchun Song * retry. Because the race window is quite small. 17977ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 17987ffddd49SMuchun Song * for increasing the success rate of dissolving page. 17997ffddd49SMuchun Song */ 18007ffddd49SMuchun Song goto retry; 18017ffddd49SMuchun Song } 18027ffddd49SMuchun Song 1803c3114a84SAnshuman Khandual /* 1804c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1805c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1806c3114a84SAnshuman Khandual */ 1807c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1808c3114a84SAnshuman Khandual SetPageHWPoison(page); 1809c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1810c3114a84SAnshuman Khandual } 1811*6eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1812c1470b33Szhong jiang h->max_huge_pages--; 18132247bb33SGerald Schaefer update_and_free_page(h, head); 18146bc9b564SNaoya Horiguchi rc = 0; 1815c8721bbbSNaoya Horiguchi } 1816082d5b6bSGerald Schaefer out: 1817c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1818082d5b6bSGerald Schaefer return rc; 1819c8721bbbSNaoya Horiguchi } 1820c8721bbbSNaoya Horiguchi 1821c8721bbbSNaoya Horiguchi /* 1822c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1823c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18242247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18252247bb33SGerald Schaefer * part of it lies within the given range. 1826082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1827082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1828c8721bbbSNaoya Horiguchi */ 1829082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1830c8721bbbSNaoya Horiguchi { 1831c8721bbbSNaoya Horiguchi unsigned long pfn; 1832eb03aa00SGerald Schaefer struct page *page; 1833082d5b6bSGerald Schaefer int rc = 0; 1834c8721bbbSNaoya Horiguchi 1835d0177639SLi Zhong if (!hugepages_supported()) 1836082d5b6bSGerald Schaefer return rc; 1837d0177639SLi Zhong 1838eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1839eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1840eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1841eb03aa00SGerald Schaefer if (rc) 1842082d5b6bSGerald Schaefer break; 1843eb03aa00SGerald Schaefer } 1844082d5b6bSGerald Schaefer 1845082d5b6bSGerald Schaefer return rc; 1846c8721bbbSNaoya Horiguchi } 1847c8721bbbSNaoya Horiguchi 1848ab5ac90aSMichal Hocko /* 1849ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1850ab5ac90aSMichal Hocko */ 18510c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1852aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18537893d1d5SAdam Litke { 18549980d744SMichal Hocko struct page *page = NULL; 18557893d1d5SAdam Litke 1856bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1857aa888a74SAndi Kleen return NULL; 1858aa888a74SAndi Kleen 1859d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 18609980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18619980d744SMichal Hocko goto out_unlock; 1862d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1863d1c3fb1fSNishanth Aravamudan 1864f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18659980d744SMichal Hocko if (!page) 18660c397daeSMichal Hocko return NULL; 1867d1c3fb1fSNishanth Aravamudan 18687893d1d5SAdam Litke spin_lock(&hugetlb_lock); 18699980d744SMichal Hocko /* 18709980d744SMichal Hocko * We could have raced with the pool size change. 18719980d744SMichal Hocko * Double check that and simply deallocate the new page 18729980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18739980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18749980d744SMichal Hocko * codeflow 18759980d744SMichal Hocko */ 18769980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18779157c311SMike Kravetz SetHPageTemporary(page); 18782bf753e6SKai Shen spin_unlock(&hugetlb_lock); 18799980d744SMichal Hocko put_page(page); 18802bf753e6SKai Shen return NULL; 18819980d744SMichal Hocko } else { 18829980d744SMichal Hocko h->surplus_huge_pages++; 18834704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18847893d1d5SAdam Litke } 18859980d744SMichal Hocko 18869980d744SMichal Hocko out_unlock: 1887d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 18887893d1d5SAdam Litke 18897893d1d5SAdam Litke return page; 18907893d1d5SAdam Litke } 18917893d1d5SAdam Litke 1892bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1893ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1894ab5ac90aSMichal Hocko { 1895ab5ac90aSMichal Hocko struct page *page; 1896ab5ac90aSMichal Hocko 1897ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1898ab5ac90aSMichal Hocko return NULL; 1899ab5ac90aSMichal Hocko 1900f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1901ab5ac90aSMichal Hocko if (!page) 1902ab5ac90aSMichal Hocko return NULL; 1903ab5ac90aSMichal Hocko 1904ab5ac90aSMichal Hocko /* 1905ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1906ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1907ab5ac90aSMichal Hocko */ 19089157c311SMike Kravetz SetHPageTemporary(page); 1909ab5ac90aSMichal Hocko 1910ab5ac90aSMichal Hocko return page; 1911ab5ac90aSMichal Hocko } 1912ab5ac90aSMichal Hocko 1913e4e574b7SAdam Litke /* 1914099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1915099730d6SDave Hansen */ 1916e0ec90eeSDave Hansen static 19170c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1918099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1919099730d6SDave Hansen { 1920aaf14e40SMichal Hocko struct page *page; 1921aaf14e40SMichal Hocko struct mempolicy *mpol; 1922aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1923aaf14e40SMichal Hocko int nid; 1924aaf14e40SMichal Hocko nodemask_t *nodemask; 1925aaf14e40SMichal Hocko 1926aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19270c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1928aaf14e40SMichal Hocko mpol_cond_put(mpol); 1929aaf14e40SMichal Hocko 1930aaf14e40SMichal Hocko return page; 1931099730d6SDave Hansen } 1932099730d6SDave Hansen 1933ab5ac90aSMichal Hocko /* page migration callback function */ 19343e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1935d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19364db9b2efSMichal Hocko { 19374db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 19384db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19393e59fcb0SMichal Hocko struct page *page; 19403e59fcb0SMichal Hocko 19413e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19423e59fcb0SMichal Hocko if (page) { 19433e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 19443e59fcb0SMichal Hocko return page; 19454db9b2efSMichal Hocko } 19464db9b2efSMichal Hocko } 19474db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 19484db9b2efSMichal Hocko 19490c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19504db9b2efSMichal Hocko } 19514db9b2efSMichal Hocko 1952ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1953389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1954389c8178SMichal Hocko unsigned long address) 1955ebd63723SMichal Hocko { 1956ebd63723SMichal Hocko struct mempolicy *mpol; 1957ebd63723SMichal Hocko nodemask_t *nodemask; 1958ebd63723SMichal Hocko struct page *page; 1959ebd63723SMichal Hocko gfp_t gfp_mask; 1960ebd63723SMichal Hocko int node; 1961ebd63723SMichal Hocko 1962ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1963ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1964d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1965ebd63723SMichal Hocko mpol_cond_put(mpol); 1966ebd63723SMichal Hocko 1967ebd63723SMichal Hocko return page; 1968ebd63723SMichal Hocko } 1969ebd63723SMichal Hocko 1970bf50bab2SNaoya Horiguchi /* 197125985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1972e4e574b7SAdam Litke * of size 'delta'. 1973e4e574b7SAdam Litke */ 19740a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19751b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1976e4e574b7SAdam Litke { 1977e4e574b7SAdam Litke struct list_head surplus_list; 1978e4e574b7SAdam Litke struct page *page, *tmp; 19790a4f3d1bSLiu Xiang int ret; 19800a4f3d1bSLiu Xiang long i; 19810a4f3d1bSLiu Xiang long needed, allocated; 198228073b02SHillf Danton bool alloc_ok = true; 1983e4e574b7SAdam Litke 1984a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1985ac09b3a1SAdam Litke if (needed <= 0) { 1986a5516438SAndi Kleen h->resv_huge_pages += delta; 1987e4e574b7SAdam Litke return 0; 1988ac09b3a1SAdam Litke } 1989e4e574b7SAdam Litke 1990e4e574b7SAdam Litke allocated = 0; 1991e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1992e4e574b7SAdam Litke 1993e4e574b7SAdam Litke ret = -ENOMEM; 1994e4e574b7SAdam Litke retry: 1995e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1996e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19970c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1998aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 199928073b02SHillf Danton if (!page) { 200028073b02SHillf Danton alloc_ok = false; 200128073b02SHillf Danton break; 200228073b02SHillf Danton } 2003e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 200469ed779aSDavid Rientjes cond_resched(); 2005e4e574b7SAdam Litke } 200628073b02SHillf Danton allocated += i; 2007e4e574b7SAdam Litke 2008e4e574b7SAdam Litke /* 2009e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2010e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2011e4e574b7SAdam Litke */ 2012e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 2013a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2014a5516438SAndi Kleen (h->free_huge_pages + allocated); 201528073b02SHillf Danton if (needed > 0) { 201628073b02SHillf Danton if (alloc_ok) 2017e4e574b7SAdam Litke goto retry; 201828073b02SHillf Danton /* 201928073b02SHillf Danton * We were not able to allocate enough pages to 202028073b02SHillf Danton * satisfy the entire reservation so we free what 202128073b02SHillf Danton * we've allocated so far. 202228073b02SHillf Danton */ 202328073b02SHillf Danton goto free; 202428073b02SHillf Danton } 2025e4e574b7SAdam Litke /* 2026e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 202725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2028e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2029ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2030ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2031ac09b3a1SAdam Litke * before they are reserved. 2032e4e574b7SAdam Litke */ 2033e4e574b7SAdam Litke needed += allocated; 2034a5516438SAndi Kleen h->resv_huge_pages += delta; 2035e4e574b7SAdam Litke ret = 0; 2036a9869b83SNaoya Horiguchi 203719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 203819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2039e558464bSMuchun Song int zeroed; 2040e558464bSMuchun Song 204119fc3f0aSAdam Litke if ((--needed) < 0) 204219fc3f0aSAdam Litke break; 2043a9869b83SNaoya Horiguchi /* 2044a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2045a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2046a9869b83SNaoya Horiguchi */ 2047e558464bSMuchun Song zeroed = put_page_testzero(page); 2048e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2049a5516438SAndi Kleen enqueue_huge_page(h, page); 205019fc3f0aSAdam Litke } 205128073b02SHillf Danton free: 2052b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 205319fc3f0aSAdam Litke 205419fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2055c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2056a9869b83SNaoya Horiguchi put_page(page); 205719fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2058e4e574b7SAdam Litke 2059e4e574b7SAdam Litke return ret; 2060e4e574b7SAdam Litke } 2061e4e574b7SAdam Litke 2062e4e574b7SAdam Litke /* 2063e5bbc8a6SMike Kravetz * This routine has two main purposes: 2064e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2065e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2066e5bbc8a6SMike Kravetz * to the associated reservation map. 2067e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2068e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2069e5bbc8a6SMike Kravetz * 2070e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2071e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2072e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2073e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2074e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2075e4e574b7SAdam Litke */ 2076a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2077a5516438SAndi Kleen unsigned long unused_resv_pages) 2078e4e574b7SAdam Litke { 2079e4e574b7SAdam Litke unsigned long nr_pages; 2080e4e574b7SAdam Litke 2081aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2082bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2083e5bbc8a6SMike Kravetz goto out; 2084aa888a74SAndi Kleen 2085e5bbc8a6SMike Kravetz /* 2086e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2087e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2088e5bbc8a6SMike Kravetz */ 2089a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2090e4e574b7SAdam Litke 2091685f3457SLee Schermerhorn /* 2092685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20939b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20949b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20959b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 20969e7ee400SRandy Dunlap * free_pool_huge_page() will balance the freed pages across the 20979b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2098e5bbc8a6SMike Kravetz * 2099e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2100e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2101e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2102685f3457SLee Schermerhorn */ 2103685f3457SLee Schermerhorn while (nr_pages--) { 2104e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2105e5bbc8a6SMike Kravetz unused_resv_pages--; 21068cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2107e5bbc8a6SMike Kravetz goto out; 21087848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2109e4e574b7SAdam Litke } 2110e5bbc8a6SMike Kravetz 2111e5bbc8a6SMike Kravetz out: 2112e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2113e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2114e4e574b7SAdam Litke } 2115e4e574b7SAdam Litke 21165e911373SMike Kravetz 2117c37f9fb1SAndy Whitcroft /* 2118feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21195e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2120cf3ad20bSMike Kravetz * 2121cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2122cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2123cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2124cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2125cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2126feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2127feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2128feba16e2SMike Kravetz * is called in such cases. 2129cf3ad20bSMike Kravetz * 2130cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2131cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2132cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2133cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2134cf3ad20bSMike Kravetz * take appropriate action. 213596b96a96SMike Kravetz * 213696b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 213796b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 213896b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 213996b96a96SMike Kravetz * reservation exists. 2140c37f9fb1SAndy Whitcroft */ 21415e911373SMike Kravetz enum vma_resv_mode { 21425e911373SMike Kravetz VMA_NEEDS_RESV, 21435e911373SMike Kravetz VMA_COMMIT_RESV, 2144feba16e2SMike Kravetz VMA_END_RESV, 214596b96a96SMike Kravetz VMA_ADD_RESV, 21465e911373SMike Kravetz }; 2147cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2148cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21495e911373SMike Kravetz enum vma_resv_mode mode) 2150c37f9fb1SAndy Whitcroft { 21514e35f483SJoonsoo Kim struct resv_map *resv; 21524e35f483SJoonsoo Kim pgoff_t idx; 2153cf3ad20bSMike Kravetz long ret; 21540db9d74eSMina Almasry long dummy_out_regions_needed; 2155c37f9fb1SAndy Whitcroft 21564e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21574e35f483SJoonsoo Kim if (!resv) 2158c37f9fb1SAndy Whitcroft return 1; 2159c37f9fb1SAndy Whitcroft 21604e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21615e911373SMike Kravetz switch (mode) { 21625e911373SMike Kravetz case VMA_NEEDS_RESV: 21630db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21640db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21650db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21660db9d74eSMina Almasry * ever require 1 region. 21670db9d74eSMina Almasry */ 21680db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21695e911373SMike Kravetz break; 21705e911373SMike Kravetz case VMA_COMMIT_RESV: 2171075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21720db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21730db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21745e911373SMike Kravetz break; 2175feba16e2SMike Kravetz case VMA_END_RESV: 21760db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21775e911373SMike Kravetz ret = 0; 21785e911373SMike Kravetz break; 217996b96a96SMike Kravetz case VMA_ADD_RESV: 21800db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2181075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21820db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21830db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21840db9d74eSMina Almasry } else { 21850db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 218696b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 218796b96a96SMike Kravetz } 218896b96a96SMike Kravetz break; 21895e911373SMike Kravetz default: 21905e911373SMike Kravetz BUG(); 21915e911373SMike Kravetz } 219284afd99bSAndy Whitcroft 21934e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2194cf3ad20bSMike Kravetz return ret; 219567961f9dSMike Kravetz /* 2196bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2197bf3d12b9SMiaohe Lin * 219867961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 219967961f9dSMike Kravetz * However, a file associated with mapping could have been 220067961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 220167961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 220267961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 220367961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 220467961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 220567961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 220667961f9dSMike Kravetz * reservation has already been consumed. As a result, the 220767961f9dSMike Kravetz * return value of this routine is the opposite of the 220867961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 220967961f9dSMike Kravetz */ 2210bf3d12b9SMiaohe Lin if (ret > 0) 221167961f9dSMike Kravetz return 0; 2212bf3d12b9SMiaohe Lin if (ret == 0) 221367961f9dSMike Kravetz return 1; 2214bf3d12b9SMiaohe Lin return ret; 221584afd99bSAndy Whitcroft } 2216cf3ad20bSMike Kravetz 2217cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2218a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2219c37f9fb1SAndy Whitcroft { 22205e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2221cf3ad20bSMike Kravetz } 2222c37f9fb1SAndy Whitcroft 2223cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2224cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2225cf3ad20bSMike Kravetz { 22265e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22275e911373SMike Kravetz } 22285e911373SMike Kravetz 2229feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22305e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22315e911373SMike Kravetz { 2232feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2233c37f9fb1SAndy Whitcroft } 2234c37f9fb1SAndy Whitcroft 223596b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 223696b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 223796b96a96SMike Kravetz { 223896b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 223996b96a96SMike Kravetz } 224096b96a96SMike Kravetz 224196b96a96SMike Kravetz /* 224296b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 224396b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 224496b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 2245d6995da3SMike Kravetz * alloc_huge_page would have consumed the reservation and set 2246d6995da3SMike Kravetz * HPageRestoreReserve in the newly allocated page. When the page is freed 2247d6995da3SMike Kravetz * via free_huge_page, the global reservation count will be incremented if 2248d6995da3SMike Kravetz * HPageRestoreReserve is set. However, free_huge_page can not adjust the 2249d6995da3SMike Kravetz * reserve map. Adjust the reserve map here to be consistent with global 2250d6995da3SMike Kravetz * reserve count adjustments to be made by free_huge_page. 225196b96a96SMike Kravetz */ 225296b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 225396b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 225496b96a96SMike Kravetz struct page *page) 225596b96a96SMike Kravetz { 2256d6995da3SMike Kravetz if (unlikely(HPageRestoreReserve(page))) { 225796b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 225896b96a96SMike Kravetz 225996b96a96SMike Kravetz if (unlikely(rc < 0)) { 226096b96a96SMike Kravetz /* 226196b96a96SMike Kravetz * Rare out of memory condition in reserve map 2262d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 226396b96a96SMike Kravetz * global reserve count will not be incremented 226496b96a96SMike Kravetz * by free_huge_page. This will make it appear 226596b96a96SMike Kravetz * as though the reservation for this page was 226696b96a96SMike Kravetz * consumed. This may prevent the task from 226796b96a96SMike Kravetz * faulting in the page at a later time. This 226896b96a96SMike Kravetz * is better than inconsistent global huge page 226996b96a96SMike Kravetz * accounting of reserve counts. 227096b96a96SMike Kravetz */ 2271d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 227296b96a96SMike Kravetz } else if (rc) { 227396b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 227496b96a96SMike Kravetz if (unlikely(rc < 0)) 227596b96a96SMike Kravetz /* 227696b96a96SMike Kravetz * See above comment about rare out of 227796b96a96SMike Kravetz * memory condition. 227896b96a96SMike Kravetz */ 2279d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 228096b96a96SMike Kravetz } else 228196b96a96SMike Kravetz vma_end_reservation(h, vma, address); 228296b96a96SMike Kravetz } 228396b96a96SMike Kravetz } 228496b96a96SMike Kravetz 228570c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 228604f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2287348ea204SAdam Litke { 228890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2289a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2290348ea204SAdam Litke struct page *page; 2291d85f69b0SMike Kravetz long map_chg, map_commit; 2292d85f69b0SMike Kravetz long gbl_chg; 22936d76dcf4SAneesh Kumar K.V int ret, idx; 22946d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 229508cf9fafSMina Almasry bool deferred_reserve; 22962fc39cecSAdam Litke 22976d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2298a1e78772SMel Gorman /* 2299d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2300d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2301d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2302a1e78772SMel Gorman */ 2303d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2304d85f69b0SMike Kravetz if (map_chg < 0) 230576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2306d85f69b0SMike Kravetz 2307d85f69b0SMike Kravetz /* 2308d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2309d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2310d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2311d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2312d85f69b0SMike Kravetz * checked against any subpool limit. 2313d85f69b0SMike Kravetz */ 2314d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2315d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2316d85f69b0SMike Kravetz if (gbl_chg < 0) { 2317feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 231876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 23195e911373SMike Kravetz } 232090d8b7e6SAdam Litke 2321d85f69b0SMike Kravetz /* 2322d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2323d85f69b0SMike Kravetz * map, there could be reservations associated with the 2324d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2325d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2326d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2327d85f69b0SMike Kravetz * the subpool reservations. 2328d85f69b0SMike Kravetz */ 2329d85f69b0SMike Kravetz if (avoid_reserve) 2330d85f69b0SMike Kravetz gbl_chg = 1; 2331d85f69b0SMike Kravetz } 2332d85f69b0SMike Kravetz 233308cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 233408cf9fafSMina Almasry */ 23356501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 233608cf9fafSMina Almasry if (deferred_reserve) { 233708cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 233808cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 23398f34af6fSJianyu Zhan if (ret) 23408f34af6fSJianyu Zhan goto out_subpool_put; 234108cf9fafSMina Almasry } 234208cf9fafSMina Almasry 234308cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 234408cf9fafSMina Almasry if (ret) 234508cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 23468f34af6fSJianyu Zhan 2347a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2348d85f69b0SMike Kravetz /* 2349d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2350d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2351d85f69b0SMike Kravetz * a reservation exists for the allocation. 2352d85f69b0SMike Kravetz */ 2353d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 235481a6fcaeSJoonsoo Kim if (!page) { 235594ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 23560c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 23578f34af6fSJianyu Zhan if (!page) 23588f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2359a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2360d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2361a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2362a88c7695SNaoya Horiguchi } 236379dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 236415a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 236581a6fcaeSJoonsoo Kim /* Fall through */ 2366a1e78772SMel Gorman } 236781a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 236808cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 236908cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 237008cf9fafSMina Almasry */ 237108cf9fafSMina Almasry if (deferred_reserve) { 237208cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 237308cf9fafSMina Almasry h_cg, page); 237408cf9fafSMina Almasry } 237508cf9fafSMina Almasry 237681a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2377a1e78772SMel Gorman 2378d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2379a1e78772SMel Gorman 2380d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2381d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 238233039678SMike Kravetz /* 238333039678SMike Kravetz * The page was added to the reservation map between 238433039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 238533039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 238633039678SMike Kravetz * Adjust for the subpool count incremented above AND 238733039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 238833039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 238933039678SMike Kravetz * no longer applies. 239033039678SMike Kravetz */ 239133039678SMike Kravetz long rsv_adjust; 239233039678SMike Kravetz 239333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 239433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 239579aa925bSMike Kravetz if (deferred_reserve) 239679aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 239779aa925bSMike Kravetz pages_per_huge_page(h), page); 239833039678SMike Kravetz } 23997893d1d5SAdam Litke return page; 24008f34af6fSJianyu Zhan 24018f34af6fSJianyu Zhan out_uncharge_cgroup: 24028f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 240308cf9fafSMina Almasry out_uncharge_cgroup_reservation: 240408cf9fafSMina Almasry if (deferred_reserve) 240508cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 240608cf9fafSMina Almasry h_cg); 24078f34af6fSJianyu Zhan out_subpool_put: 2408d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 24098f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2410feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 24118f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2412b45b5bd6SDavid Gibson } 2413b45b5bd6SDavid Gibson 2414e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2415e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2416e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2417aa888a74SAndi Kleen { 2418aa888a74SAndi Kleen struct huge_bootmem_page *m; 2419b2261026SJoonsoo Kim int nr_nodes, node; 2420aa888a74SAndi Kleen 2421b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2422aa888a74SAndi Kleen void *addr; 2423aa888a74SAndi Kleen 2424eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 24258b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 242697ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2427aa888a74SAndi Kleen if (addr) { 2428aa888a74SAndi Kleen /* 2429aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2430aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2431aa888a74SAndi Kleen * puts them into the mem_map). 2432aa888a74SAndi Kleen */ 2433aa888a74SAndi Kleen m = addr; 2434aa888a74SAndi Kleen goto found; 2435aa888a74SAndi Kleen } 2436aa888a74SAndi Kleen } 2437aa888a74SAndi Kleen return 0; 2438aa888a74SAndi Kleen 2439aa888a74SAndi Kleen found: 2440df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2441aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2442330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2443aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2444aa888a74SAndi Kleen m->hstate = h; 2445aa888a74SAndi Kleen return 1; 2446aa888a74SAndi Kleen } 2447aa888a74SAndi Kleen 2448d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2449d00181b9SKirill A. Shutemov unsigned int order) 245018229df5SAndy Whitcroft { 245118229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 245218229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 245318229df5SAndy Whitcroft else 245418229df5SAndy Whitcroft prep_compound_page(page, order); 245518229df5SAndy Whitcroft } 245618229df5SAndy Whitcroft 2457aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2458aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2459aa888a74SAndi Kleen { 2460aa888a74SAndi Kleen struct huge_bootmem_page *m; 2461aa888a74SAndi Kleen 2462aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 246340d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2464aa888a74SAndi Kleen struct hstate *h = m->hstate; 2465ee8f248dSBecky Bruce 2466aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2467c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2468ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2469aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2470af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2471af0fb9dfSMichal Hocko 2472b0320c7bSRafael Aquini /* 2473b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2474b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2475b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2476b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2477b0320c7bSRafael Aquini */ 2478bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2479c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2480520495feSCannon Matthews cond_resched(); 2481aa888a74SAndi Kleen } 2482aa888a74SAndi Kleen } 2483aa888a74SAndi Kleen 24848faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 24851da177e4SLinus Torvalds { 24861da177e4SLinus Torvalds unsigned long i; 2487f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2488f60858f9SMike Kravetz 2489f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2490f60858f9SMike Kravetz /* 2491f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2492f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2493f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2494f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2495f60858f9SMike Kravetz */ 2496f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2497f60858f9SMike Kravetz GFP_KERNEL); 2498f60858f9SMike Kravetz } else { 2499f60858f9SMike Kravetz /* allocations done at boot time */ 2500f60858f9SMike Kravetz node_alloc_noretry = NULL; 2501f60858f9SMike Kravetz } 2502f60858f9SMike Kravetz 2503f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2504f60858f9SMike Kravetz if (node_alloc_noretry) 2505f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 25061da177e4SLinus Torvalds 2507e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2508bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2509dbda8feaSBarry Song if (hugetlb_cma_size) { 2510cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 25117ecc9565SChen Wandun goto free; 2512cf11e85fSRoman Gushchin } 2513aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2514aa888a74SAndi Kleen break; 25150c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2516f60858f9SMike Kravetz &node_states[N_MEMORY], 2517f60858f9SMike Kravetz node_alloc_noretry)) 25181da177e4SLinus Torvalds break; 251969ed779aSDavid Rientjes cond_resched(); 25201da177e4SLinus Torvalds } 2521d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2522d715cf80SLiam R. Howlett char buf[32]; 2523d715cf80SLiam R. Howlett 2524c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2525d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2526d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 25278faa8b07SAndi Kleen h->max_huge_pages = i; 2528e5ff2159SAndi Kleen } 25297ecc9565SChen Wandun free: 2530f60858f9SMike Kravetz kfree(node_alloc_noretry); 2531d715cf80SLiam R. Howlett } 2532e5ff2159SAndi Kleen 2533e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2534e5ff2159SAndi Kleen { 2535e5ff2159SAndi Kleen struct hstate *h; 2536e5ff2159SAndi Kleen 2537e5ff2159SAndi Kleen for_each_hstate(h) { 2538641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2539641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2540641844f5SNaoya Horiguchi 25418faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2542bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 25438faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2544e5ff2159SAndi Kleen } 2545641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2546e5ff2159SAndi Kleen } 2547e5ff2159SAndi Kleen 2548e5ff2159SAndi Kleen static void __init report_hugepages(void) 2549e5ff2159SAndi Kleen { 2550e5ff2159SAndi Kleen struct hstate *h; 2551e5ff2159SAndi Kleen 2552e5ff2159SAndi Kleen for_each_hstate(h) { 25534abd32dbSAndi Kleen char buf[32]; 2554c6247f72SMatthew Wilcox 2555c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2556ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2557c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2558e5ff2159SAndi Kleen } 2559e5ff2159SAndi Kleen } 2560e5ff2159SAndi Kleen 25611da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 25626ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 25636ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25641da177e4SLinus Torvalds { 25654415cc8dSChristoph Lameter int i; 25664415cc8dSChristoph Lameter 2567bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2568aa888a74SAndi Kleen return; 2569aa888a74SAndi Kleen 25706ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 25711da177e4SLinus Torvalds struct page *page, *next; 2572a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2573a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2574a5516438SAndi Kleen if (count >= h->nr_huge_pages) 25756b0c880dSAdam Litke return; 25761da177e4SLinus Torvalds if (PageHighMem(page)) 25771da177e4SLinus Torvalds continue; 2578*6eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 2579e5ff2159SAndi Kleen update_and_free_page(h, page); 25801da177e4SLinus Torvalds } 25811da177e4SLinus Torvalds } 25821da177e4SLinus Torvalds } 25831da177e4SLinus Torvalds #else 25846ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 25856ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25861da177e4SLinus Torvalds { 25871da177e4SLinus Torvalds } 25881da177e4SLinus Torvalds #endif 25891da177e4SLinus Torvalds 259020a0307cSWu Fengguang /* 259120a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 259220a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 259320a0307cSWu Fengguang * Returns 1 if an adjustment was made. 259420a0307cSWu Fengguang */ 25956ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 25966ae11b27SLee Schermerhorn int delta) 259720a0307cSWu Fengguang { 2598b2261026SJoonsoo Kim int nr_nodes, node; 259920a0307cSWu Fengguang 260020a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 260120a0307cSWu Fengguang 2602e8c5c824SLee Schermerhorn if (delta < 0) { 2603b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2604b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2605b2261026SJoonsoo Kim goto found; 2606b2261026SJoonsoo Kim } 2607b2261026SJoonsoo Kim } else { 2608b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2609b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2610b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2611b2261026SJoonsoo Kim goto found; 2612e8c5c824SLee Schermerhorn } 26139a76db09SLee Schermerhorn } 2614b2261026SJoonsoo Kim return 0; 261520a0307cSWu Fengguang 2616b2261026SJoonsoo Kim found: 261720a0307cSWu Fengguang h->surplus_huge_pages += delta; 2618b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2619b2261026SJoonsoo Kim return 1; 262020a0307cSWu Fengguang } 262120a0307cSWu Fengguang 2622a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2623fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 26246ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26251da177e4SLinus Torvalds { 26267893d1d5SAdam Litke unsigned long min_count, ret; 2627f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2628f60858f9SMike Kravetz 2629f60858f9SMike Kravetz /* 2630f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2631f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2632f60858f9SMike Kravetz * the requested huge pages. 2633f60858f9SMike Kravetz */ 2634f60858f9SMike Kravetz if (node_alloc_noretry) 2635f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2636f60858f9SMike Kravetz else 2637f60858f9SMike Kravetz return -ENOMEM; 26381da177e4SLinus Torvalds 263929383967SMike Kravetz /* 264029383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 264129383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 264229383967SMike Kravetz */ 264329383967SMike Kravetz mutex_lock(&h->resize_lock); 26444eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 26454eb0716eSAlexandre Ghiti 26464eb0716eSAlexandre Ghiti /* 2647fd875dcaSMike Kravetz * Check for a node specific request. 2648fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2649fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2650fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2651fd875dcaSMike Kravetz */ 2652fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2653fd875dcaSMike Kravetz unsigned long old_count = count; 2654fd875dcaSMike Kravetz 2655fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2656fd875dcaSMike Kravetz /* 2657fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2658fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2659fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2660fd875dcaSMike Kravetz * largest possible value to align with their intention. 2661fd875dcaSMike Kravetz */ 2662fd875dcaSMike Kravetz if (count < old_count) 2663fd875dcaSMike Kravetz count = ULONG_MAX; 2664fd875dcaSMike Kravetz } 2665fd875dcaSMike Kravetz 2666fd875dcaSMike Kravetz /* 26674eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 26684eb0716eSAlexandre Ghiti * page range allocation. 26694eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 26704eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 26714eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 26724eb0716eSAlexandre Ghiti */ 26734eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 26744eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 26754eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 267629383967SMike Kravetz mutex_unlock(&h->resize_lock); 2677f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 26784eb0716eSAlexandre Ghiti return -EINVAL; 26794eb0716eSAlexandre Ghiti } 26804eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 26814eb0716eSAlexandre Ghiti } 2682aa888a74SAndi Kleen 26837893d1d5SAdam Litke /* 26847893d1d5SAdam Litke * Increase the pool size 26857893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 26867893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2687d1c3fb1fSNishanth Aravamudan * 26880c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2689d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2690d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2691d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2692d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 26937893d1d5SAdam Litke */ 2694a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 26956ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 26967893d1d5SAdam Litke break; 26977893d1d5SAdam Litke } 26987893d1d5SAdam Litke 2699a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 27007893d1d5SAdam Litke /* 27017893d1d5SAdam Litke * If this allocation races such that we no longer need the 27027893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 27037893d1d5SAdam Litke * and reducing the surplus. 27047893d1d5SAdam Litke */ 27057893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2706649920c6SJia He 2707649920c6SJia He /* yield cpu to avoid soft lockup */ 2708649920c6SJia He cond_resched(); 2709649920c6SJia He 2710f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2711f60858f9SMike Kravetz node_alloc_noretry); 27127893d1d5SAdam Litke spin_lock(&hugetlb_lock); 27137893d1d5SAdam Litke if (!ret) 27147893d1d5SAdam Litke goto out; 27157893d1d5SAdam Litke 2716536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2717536240f2SMel Gorman if (signal_pending(current)) 2718536240f2SMel Gorman goto out; 27197893d1d5SAdam Litke } 27207893d1d5SAdam Litke 27217893d1d5SAdam Litke /* 27227893d1d5SAdam Litke * Decrease the pool size 27237893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27247893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27257893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27267893d1d5SAdam Litke * to the desired size as pages become free. 2727d1c3fb1fSNishanth Aravamudan * 2728d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2729d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2730d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27310c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2732d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2733d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2734d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27357893d1d5SAdam Litke */ 2736a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27376b0c880dSAdam Litke min_count = max(count, min_count); 27386ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2739a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 27406ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 27411da177e4SLinus Torvalds break; 274255f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 27431da177e4SLinus Torvalds } 2744a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 27456ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 27467893d1d5SAdam Litke break; 27477893d1d5SAdam Litke } 27487893d1d5SAdam Litke out: 27494eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 27501da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 275129383967SMike Kravetz mutex_unlock(&h->resize_lock); 27524eb0716eSAlexandre Ghiti 2753f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2754f60858f9SMike Kravetz 27554eb0716eSAlexandre Ghiti return 0; 27561da177e4SLinus Torvalds } 27571da177e4SLinus Torvalds 2758a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2759a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2760a3437870SNishanth Aravamudan 2761a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2762a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2763a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2764a3437870SNishanth Aravamudan 2765a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2766a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2767a3437870SNishanth Aravamudan 27689a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 27699a305230SLee Schermerhorn 27709a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2771a3437870SNishanth Aravamudan { 2772a3437870SNishanth Aravamudan int i; 27739a305230SLee Schermerhorn 2774a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 27759a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 27769a305230SLee Schermerhorn if (nidp) 27779a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2778a3437870SNishanth Aravamudan return &hstates[i]; 27799a305230SLee Schermerhorn } 27809a305230SLee Schermerhorn 27819a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2782a3437870SNishanth Aravamudan } 2783a3437870SNishanth Aravamudan 278406808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2785a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2786a3437870SNishanth Aravamudan { 27879a305230SLee Schermerhorn struct hstate *h; 27889a305230SLee Schermerhorn unsigned long nr_huge_pages; 27899a305230SLee Schermerhorn int nid; 27909a305230SLee Schermerhorn 27919a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27929a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27939a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 27949a305230SLee Schermerhorn else 27959a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 27969a305230SLee Schermerhorn 2797ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2798a3437870SNishanth Aravamudan } 2799adbe8726SEric B Munson 2800238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2801238d3c13SDavid Rientjes struct hstate *h, int nid, 2802238d3c13SDavid Rientjes unsigned long count, size_t len) 2803a3437870SNishanth Aravamudan { 2804a3437870SNishanth Aravamudan int err; 28052d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2806a3437870SNishanth Aravamudan 28072d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 28082d0adf7eSOscar Salvador return -EINVAL; 2809adbe8726SEric B Munson 28109a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 28119a305230SLee Schermerhorn /* 28129a305230SLee Schermerhorn * global hstate attribute 28139a305230SLee Schermerhorn */ 28149a305230SLee Schermerhorn if (!(obey_mempolicy && 28152d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 28162d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 28172d0adf7eSOscar Salvador else 28182d0adf7eSOscar Salvador n_mask = &nodes_allowed; 28192d0adf7eSOscar Salvador } else { 28209a305230SLee Schermerhorn /* 2821fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2822fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28239a305230SLee Schermerhorn */ 28242d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28252d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2826fd875dcaSMike Kravetz } 28279a305230SLee Schermerhorn 28282d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 282906808b08SLee Schermerhorn 28304eb0716eSAlexandre Ghiti return err ? err : len; 283106808b08SLee Schermerhorn } 283206808b08SLee Schermerhorn 2833238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2834238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2835238d3c13SDavid Rientjes size_t len) 2836238d3c13SDavid Rientjes { 2837238d3c13SDavid Rientjes struct hstate *h; 2838238d3c13SDavid Rientjes unsigned long count; 2839238d3c13SDavid Rientjes int nid; 2840238d3c13SDavid Rientjes int err; 2841238d3c13SDavid Rientjes 2842238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2843238d3c13SDavid Rientjes if (err) 2844238d3c13SDavid Rientjes return err; 2845238d3c13SDavid Rientjes 2846238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2847238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2848238d3c13SDavid Rientjes } 2849238d3c13SDavid Rientjes 285006808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 285106808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 285206808b08SLee Schermerhorn { 285306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 285406808b08SLee Schermerhorn } 285506808b08SLee Schermerhorn 285606808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 285706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 285806808b08SLee Schermerhorn { 2859238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2860a3437870SNishanth Aravamudan } 2861a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2862a3437870SNishanth Aravamudan 286306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 286406808b08SLee Schermerhorn 286506808b08SLee Schermerhorn /* 286606808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 286706808b08SLee Schermerhorn * huge page alloc/free. 286806808b08SLee Schermerhorn */ 286906808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 2870ae7a927dSJoe Perches struct kobj_attribute *attr, 2871ae7a927dSJoe Perches char *buf) 287206808b08SLee Schermerhorn { 287306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 287406808b08SLee Schermerhorn } 287506808b08SLee Schermerhorn 287606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 287706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 287806808b08SLee Schermerhorn { 2879238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 288006808b08SLee Schermerhorn } 288106808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 288206808b08SLee Schermerhorn #endif 288306808b08SLee Schermerhorn 288406808b08SLee Schermerhorn 2885a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2886a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2887a3437870SNishanth Aravamudan { 28889a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2889ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 2890a3437870SNishanth Aravamudan } 2891adbe8726SEric B Munson 2892a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2893a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2894a3437870SNishanth Aravamudan { 2895a3437870SNishanth Aravamudan int err; 2896a3437870SNishanth Aravamudan unsigned long input; 28979a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2898a3437870SNishanth Aravamudan 2899bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2900adbe8726SEric B Munson return -EINVAL; 2901adbe8726SEric B Munson 29023dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2903a3437870SNishanth Aravamudan if (err) 290473ae31e5SEric B Munson return err; 2905a3437870SNishanth Aravamudan 2906a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2907a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2908a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2909a3437870SNishanth Aravamudan 2910a3437870SNishanth Aravamudan return count; 2911a3437870SNishanth Aravamudan } 2912a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2913a3437870SNishanth Aravamudan 2914a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2915a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2916a3437870SNishanth Aravamudan { 29179a305230SLee Schermerhorn struct hstate *h; 29189a305230SLee Schermerhorn unsigned long free_huge_pages; 29199a305230SLee Schermerhorn int nid; 29209a305230SLee Schermerhorn 29219a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29229a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29239a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29249a305230SLee Schermerhorn else 29259a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29269a305230SLee Schermerhorn 2927ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 2928a3437870SNishanth Aravamudan } 2929a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2930a3437870SNishanth Aravamudan 2931a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2932a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2933a3437870SNishanth Aravamudan { 29349a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2935ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 2936a3437870SNishanth Aravamudan } 2937a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2938a3437870SNishanth Aravamudan 2939a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2940a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2941a3437870SNishanth Aravamudan { 29429a305230SLee Schermerhorn struct hstate *h; 29439a305230SLee Schermerhorn unsigned long surplus_huge_pages; 29449a305230SLee Schermerhorn int nid; 29459a305230SLee Schermerhorn 29469a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29479a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29489a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 29499a305230SLee Schermerhorn else 29509a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 29519a305230SLee Schermerhorn 2952ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 2953a3437870SNishanth Aravamudan } 2954a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2955a3437870SNishanth Aravamudan 2956a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2957a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2958a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2959a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2960a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2961a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 296206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 296306808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 296406808b08SLee Schermerhorn #endif 2965a3437870SNishanth Aravamudan NULL, 2966a3437870SNishanth Aravamudan }; 2967a3437870SNishanth Aravamudan 296867e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2969a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2970a3437870SNishanth Aravamudan }; 2971a3437870SNishanth Aravamudan 2972094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 29739a305230SLee Schermerhorn struct kobject **hstate_kobjs, 297467e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2975a3437870SNishanth Aravamudan { 2976a3437870SNishanth Aravamudan int retval; 2977972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2978a3437870SNishanth Aravamudan 29799a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 29809a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2981a3437870SNishanth Aravamudan return -ENOMEM; 2982a3437870SNishanth Aravamudan 29839a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2984cc2205a6SMiaohe Lin if (retval) { 29859a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2986cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 2987cc2205a6SMiaohe Lin } 2988a3437870SNishanth Aravamudan 2989a3437870SNishanth Aravamudan return retval; 2990a3437870SNishanth Aravamudan } 2991a3437870SNishanth Aravamudan 2992a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2993a3437870SNishanth Aravamudan { 2994a3437870SNishanth Aravamudan struct hstate *h; 2995a3437870SNishanth Aravamudan int err; 2996a3437870SNishanth Aravamudan 2997a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2998a3437870SNishanth Aravamudan if (!hugepages_kobj) 2999a3437870SNishanth Aravamudan return; 3000a3437870SNishanth Aravamudan 3001a3437870SNishanth Aravamudan for_each_hstate(h) { 30029a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 30039a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3004a3437870SNishanth Aravamudan if (err) 3005282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3006a3437870SNishanth Aravamudan } 3007a3437870SNishanth Aravamudan } 3008a3437870SNishanth Aravamudan 30099a305230SLee Schermerhorn #ifdef CONFIG_NUMA 30109a305230SLee Schermerhorn 30119a305230SLee Schermerhorn /* 30129a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 301310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 301410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 301510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 30169a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 30179a305230SLee Schermerhorn */ 30189a305230SLee Schermerhorn struct node_hstate { 30199a305230SLee Schermerhorn struct kobject *hugepages_kobj; 30209a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30219a305230SLee Schermerhorn }; 3022b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30239a305230SLee Schermerhorn 30249a305230SLee Schermerhorn /* 302510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30269a305230SLee Schermerhorn */ 30279a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30289a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30299a305230SLee Schermerhorn &free_hugepages_attr.attr, 30309a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30319a305230SLee Schermerhorn NULL, 30329a305230SLee Schermerhorn }; 30339a305230SLee Schermerhorn 303467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30359a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30369a305230SLee Schermerhorn }; 30379a305230SLee Schermerhorn 30389a305230SLee Schermerhorn /* 303910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30409a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30419a305230SLee Schermerhorn */ 30429a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 30439a305230SLee Schermerhorn { 30449a305230SLee Schermerhorn int nid; 30459a305230SLee Schermerhorn 30469a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 30479a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 30489a305230SLee Schermerhorn int i; 30499a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 30509a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 30519a305230SLee Schermerhorn if (nidp) 30529a305230SLee Schermerhorn *nidp = nid; 30539a305230SLee Schermerhorn return &hstates[i]; 30549a305230SLee Schermerhorn } 30559a305230SLee Schermerhorn } 30569a305230SLee Schermerhorn 30579a305230SLee Schermerhorn BUG(); 30589a305230SLee Schermerhorn return NULL; 30599a305230SLee Schermerhorn } 30609a305230SLee Schermerhorn 30619a305230SLee Schermerhorn /* 306210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 30639a305230SLee Schermerhorn * No-op if no hstate attributes attached. 30649a305230SLee Schermerhorn */ 30653cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 30669a305230SLee Schermerhorn { 30679a305230SLee Schermerhorn struct hstate *h; 306810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30699a305230SLee Schermerhorn 30709a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30719b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 30729a305230SLee Schermerhorn 3073972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3074972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3075972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3076972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3077972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3078972dc4deSAneesh Kumar K.V } 30799a305230SLee Schermerhorn } 30809a305230SLee Schermerhorn 30819a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 30829a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 30839a305230SLee Schermerhorn } 30849a305230SLee Schermerhorn 30859a305230SLee Schermerhorn 30869a305230SLee Schermerhorn /* 308710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 30889a305230SLee Schermerhorn * No-op if attributes already registered. 30899a305230SLee Schermerhorn */ 30903cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 30919a305230SLee Schermerhorn { 30929a305230SLee Schermerhorn struct hstate *h; 309310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30949a305230SLee Schermerhorn int err; 30959a305230SLee Schermerhorn 30969a305230SLee Schermerhorn if (nhs->hugepages_kobj) 30979a305230SLee Schermerhorn return; /* already allocated */ 30989a305230SLee Schermerhorn 30999a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 310010fbcf4cSKay Sievers &node->dev.kobj); 31019a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31029a305230SLee Schermerhorn return; 31039a305230SLee Schermerhorn 31049a305230SLee Schermerhorn for_each_hstate(h) { 31059a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 31069a305230SLee Schermerhorn nhs->hstate_kobjs, 31079a305230SLee Schermerhorn &per_node_hstate_attr_group); 31089a305230SLee Schermerhorn if (err) { 3109282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 311010fbcf4cSKay Sievers h->name, node->dev.id); 31119a305230SLee Schermerhorn hugetlb_unregister_node(node); 31129a305230SLee Schermerhorn break; 31139a305230SLee Schermerhorn } 31149a305230SLee Schermerhorn } 31159a305230SLee Schermerhorn } 31169a305230SLee Schermerhorn 31179a305230SLee Schermerhorn /* 31189b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 311910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 312010fbcf4cSKay Sievers * registered their associated device by this time. 31219a305230SLee Schermerhorn */ 31227d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31239a305230SLee Schermerhorn { 31249a305230SLee Schermerhorn int nid; 31259a305230SLee Schermerhorn 31268cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31278732794bSWen Congyang struct node *node = node_devices[nid]; 312810fbcf4cSKay Sievers if (node->dev.id == nid) 31299a305230SLee Schermerhorn hugetlb_register_node(node); 31309a305230SLee Schermerhorn } 31319a305230SLee Schermerhorn 31329a305230SLee Schermerhorn /* 313310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31349a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31359a305230SLee Schermerhorn */ 31369a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31379a305230SLee Schermerhorn hugetlb_unregister_node); 31389a305230SLee Schermerhorn } 31399a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31409a305230SLee Schermerhorn 31419a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31429a305230SLee Schermerhorn { 31439a305230SLee Schermerhorn BUG(); 31449a305230SLee Schermerhorn if (nidp) 31459a305230SLee Schermerhorn *nidp = -1; 31469a305230SLee Schermerhorn return NULL; 31479a305230SLee Schermerhorn } 31489a305230SLee Schermerhorn 31499a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 31509a305230SLee Schermerhorn 31519a305230SLee Schermerhorn #endif 31529a305230SLee Schermerhorn 3153a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3154a3437870SNishanth Aravamudan { 31558382d914SDavidlohr Bueso int i; 31568382d914SDavidlohr Bueso 3157d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3158d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3159d6995da3SMike Kravetz 3160c2833a5bSMike Kravetz if (!hugepages_supported()) { 3161c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3162c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 31630ef89d25SBenjamin Herrenschmidt return 0; 3164d715cf80SLiam R. Howlett } 3165d715cf80SLiam R. Howlett 3166282f4214SMike Kravetz /* 3167282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3168282f4214SMike Kravetz * architectures depend on setup being done here. 3169282f4214SMike Kravetz */ 3170a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3171282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3172282f4214SMike Kravetz /* 3173282f4214SMike Kravetz * If we did not parse a default huge page size, set 3174282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3175282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3176282f4214SMike Kravetz * specified, set that here as well. 3177282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3178282f4214SMike Kravetz * setting. A warning will be printed in this case. 3179282f4214SMike Kravetz */ 3180282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3181f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3182282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3183282f4214SMike Kravetz char buf[32]; 3184282f4214SMike Kravetz 3185282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3186282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3187282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3188282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3189282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3190282f4214SMike Kravetz default_hstate_max_huge_pages); 3191282f4214SMike Kravetz } 3192282f4214SMike Kravetz default_hstate.max_huge_pages = 3193282f4214SMike Kravetz default_hstate_max_huge_pages; 3194282f4214SMike Kravetz } 3195f8b74815SVaishali Thakkar } 3196a3437870SNishanth Aravamudan 3197cf11e85fSRoman Gushchin hugetlb_cma_check(); 3198a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3199aa888a74SAndi Kleen gather_bootmem_prealloc(); 3200a3437870SNishanth Aravamudan report_hugepages(); 3201a3437870SNishanth Aravamudan 3202a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 32039a305230SLee Schermerhorn hugetlb_register_all_nodes(); 32047179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 32059a305230SLee Schermerhorn 32068382d914SDavidlohr Bueso #ifdef CONFIG_SMP 32078382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 32088382d914SDavidlohr Bueso #else 32098382d914SDavidlohr Bueso num_fault_mutexes = 1; 32108382d914SDavidlohr Bueso #endif 3211c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 32126da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 32136da2ec56SKees Cook GFP_KERNEL); 3214c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 32158382d914SDavidlohr Bueso 32168382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3217c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3218a3437870SNishanth Aravamudan return 0; 3219a3437870SNishanth Aravamudan } 32203e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3221a3437870SNishanth Aravamudan 3222ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3223ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32249fee021dSVaishali Thakkar { 3225ae94da89SMike Kravetz return size == HPAGE_SIZE; 32269fee021dSVaishali Thakkar } 32279fee021dSVaishali Thakkar 3228d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3229a3437870SNishanth Aravamudan { 3230a3437870SNishanth Aravamudan struct hstate *h; 32318faa8b07SAndi Kleen unsigned long i; 32328faa8b07SAndi Kleen 3233a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3234a3437870SNishanth Aravamudan return; 3235a3437870SNishanth Aravamudan } 323647d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3237a3437870SNishanth Aravamudan BUG_ON(order == 0); 323847d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 323929383967SMike Kravetz mutex_init(&h->resize_lock); 3240a3437870SNishanth Aravamudan h->order = order; 3241aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 32428faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32438faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 32440edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 324554f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 324654f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3247a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3248a3437870SNishanth Aravamudan huge_page_size(h)/1024); 32498faa8b07SAndi Kleen 3250a3437870SNishanth Aravamudan parsed_hstate = h; 3251a3437870SNishanth Aravamudan } 3252a3437870SNishanth Aravamudan 3253282f4214SMike Kravetz /* 3254282f4214SMike Kravetz * hugepages command line processing 3255282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3256282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3257282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3258282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3259282f4214SMike Kravetz */ 3260282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3261a3437870SNishanth Aravamudan { 3262a3437870SNishanth Aravamudan unsigned long *mhp; 32638faa8b07SAndi Kleen static unsigned long *last_mhp; 3264a3437870SNishanth Aravamudan 32659fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3266282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 32679fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3268282f4214SMike Kravetz return 0; 32699fee021dSVaishali Thakkar } 3270282f4214SMike Kravetz 3271a3437870SNishanth Aravamudan /* 3272282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3273282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3274282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3275282f4214SMike Kravetz * default_hugepagesz. 3276a3437870SNishanth Aravamudan */ 32779fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3278a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3279a3437870SNishanth Aravamudan else 3280a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3281a3437870SNishanth Aravamudan 32828faa8b07SAndi Kleen if (mhp == last_mhp) { 3283282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3284282f4214SMike Kravetz return 0; 32858faa8b07SAndi Kleen } 32868faa8b07SAndi Kleen 3287a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3288a3437870SNishanth Aravamudan *mhp = 0; 3289a3437870SNishanth Aravamudan 32908faa8b07SAndi Kleen /* 32918faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 329204adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 32938faa8b07SAndi Kleen * use the bootmem allocator. 32948faa8b07SAndi Kleen */ 329504adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 32968faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 32978faa8b07SAndi Kleen 32988faa8b07SAndi Kleen last_mhp = mhp; 32998faa8b07SAndi Kleen 3300a3437870SNishanth Aravamudan return 1; 3301a3437870SNishanth Aravamudan } 3302282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3303e11bfbfcSNick Piggin 3304282f4214SMike Kravetz /* 3305282f4214SMike Kravetz * hugepagesz command line processing 3306282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3307282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3308282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3309282f4214SMike Kravetz * hugepagesz argument was valid. 3310282f4214SMike Kravetz */ 3311359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3312e11bfbfcSNick Piggin { 3313359f2544SMike Kravetz unsigned long size; 3314282f4214SMike Kravetz struct hstate *h; 3315282f4214SMike Kravetz 3316282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3317359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3318359f2544SMike Kravetz 3319359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3320282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3321359f2544SMike Kravetz return 0; 3322359f2544SMike Kravetz } 3323359f2544SMike Kravetz 3324282f4214SMike Kravetz h = size_to_hstate(size); 3325282f4214SMike Kravetz if (h) { 3326282f4214SMike Kravetz /* 3327282f4214SMike Kravetz * hstate for this size already exists. This is normally 3328282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3329282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3330282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3331282f4214SMike Kravetz * previously specified. 3332282f4214SMike Kravetz */ 3333282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3334282f4214SMike Kravetz default_hstate.max_huge_pages) { 3335282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 333638237830SMike Kravetz return 0; 333738237830SMike Kravetz } 333838237830SMike Kravetz 3339282f4214SMike Kravetz /* 3340282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3341282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3342282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3343282f4214SMike Kravetz */ 3344282f4214SMike Kravetz parsed_hstate = h; 3345282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3346e11bfbfcSNick Piggin return 1; 3347e11bfbfcSNick Piggin } 3348282f4214SMike Kravetz 3349359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3350282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3351359f2544SMike Kravetz return 1; 3352359f2544SMike Kravetz } 3353359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3354359f2544SMike Kravetz 3355282f4214SMike Kravetz /* 3356282f4214SMike Kravetz * default_hugepagesz command line input 3357282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3358282f4214SMike Kravetz */ 3359ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3360e11bfbfcSNick Piggin { 3361ae94da89SMike Kravetz unsigned long size; 3362ae94da89SMike Kravetz 3363282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3364282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3365282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3366282f4214SMike Kravetz return 0; 3367282f4214SMike Kravetz } 3368282f4214SMike Kravetz 3369282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3370282f4214SMike Kravetz 3371282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3372282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3373282f4214SMike Kravetz return 0; 3374282f4214SMike Kravetz } 3375282f4214SMike Kravetz 3376282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3377282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3378282f4214SMike Kravetz parsed_default_hugepagesz = true; 3379282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3380282f4214SMike Kravetz 3381282f4214SMike Kravetz /* 3382282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3383282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3384282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3385282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3386282f4214SMike Kravetz * allocated here from bootmem allocator. 3387282f4214SMike Kravetz */ 3388282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3389282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3390282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3391282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3392282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3393282f4214SMike Kravetz } 3394282f4214SMike Kravetz 3395e11bfbfcSNick Piggin return 1; 3396e11bfbfcSNick Piggin } 3397ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3398a3437870SNishanth Aravamudan 33998ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 34008a213460SNishanth Aravamudan { 34018a213460SNishanth Aravamudan int node; 34028a213460SNishanth Aravamudan unsigned int nr = 0; 34038ca39e68SMuchun Song nodemask_t *mpol_allowed; 34048ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 34058ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 34068a213460SNishanth Aravamudan 34078ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 34088ca39e68SMuchun Song 34098ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3410c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 34118a213460SNishanth Aravamudan nr += array[node]; 34128ca39e68SMuchun Song } 34138a213460SNishanth Aravamudan 34148a213460SNishanth Aravamudan return nr; 34158a213460SNishanth Aravamudan } 34168a213460SNishanth Aravamudan 34178a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 341817743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 341917743798SMuchun Song void *buffer, size_t *length, 342017743798SMuchun Song loff_t *ppos, unsigned long *out) 342117743798SMuchun Song { 342217743798SMuchun Song struct ctl_table dup_table; 342317743798SMuchun Song 342417743798SMuchun Song /* 342517743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 342617743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 342717743798SMuchun Song */ 342817743798SMuchun Song dup_table = *table; 342917743798SMuchun Song dup_table.data = out; 343017743798SMuchun Song 343117743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 343217743798SMuchun Song } 343317743798SMuchun Song 343406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 343506808b08SLee Schermerhorn struct ctl_table *table, int write, 343632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34371da177e4SLinus Torvalds { 3438e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3439238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 344008d4a246SMichal Hocko int ret; 3441e5ff2159SAndi Kleen 3442457c1b27SNishanth Aravamudan if (!hugepages_supported()) 344386613628SJan Stancek return -EOPNOTSUPP; 3444457c1b27SNishanth Aravamudan 344517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 344617743798SMuchun Song &tmp); 344708d4a246SMichal Hocko if (ret) 344808d4a246SMichal Hocko goto out; 3449e5ff2159SAndi Kleen 3450238d3c13SDavid Rientjes if (write) 3451238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3452238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 345308d4a246SMichal Hocko out: 345408d4a246SMichal Hocko return ret; 34551da177e4SLinus Torvalds } 3456396faf03SMel Gorman 345706808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 345832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 345906808b08SLee Schermerhorn { 346006808b08SLee Schermerhorn 346106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 346206808b08SLee Schermerhorn buffer, length, ppos); 346306808b08SLee Schermerhorn } 346406808b08SLee Schermerhorn 346506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 346606808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 346732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 346806808b08SLee Schermerhorn { 346906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 347006808b08SLee Schermerhorn buffer, length, ppos); 347106808b08SLee Schermerhorn } 347206808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 347306808b08SLee Schermerhorn 3474a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 347532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3476a3d0c6aaSNishanth Aravamudan { 3477a5516438SAndi Kleen struct hstate *h = &default_hstate; 3478e5ff2159SAndi Kleen unsigned long tmp; 347908d4a246SMichal Hocko int ret; 3480e5ff2159SAndi Kleen 3481457c1b27SNishanth Aravamudan if (!hugepages_supported()) 348286613628SJan Stancek return -EOPNOTSUPP; 3483457c1b27SNishanth Aravamudan 3484e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3485e5ff2159SAndi Kleen 3486bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3487adbe8726SEric B Munson return -EINVAL; 3488adbe8726SEric B Munson 348917743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 349017743798SMuchun Song &tmp); 349108d4a246SMichal Hocko if (ret) 349208d4a246SMichal Hocko goto out; 3493e5ff2159SAndi Kleen 3494e5ff2159SAndi Kleen if (write) { 3495064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3496e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3497a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3498e5ff2159SAndi Kleen } 349908d4a246SMichal Hocko out: 350008d4a246SMichal Hocko return ret; 3501a3d0c6aaSNishanth Aravamudan } 3502a3d0c6aaSNishanth Aravamudan 35031da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 35041da177e4SLinus Torvalds 3505e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 35061da177e4SLinus Torvalds { 3507fcb2b0c5SRoman Gushchin struct hstate *h; 3508fcb2b0c5SRoman Gushchin unsigned long total = 0; 3509fcb2b0c5SRoman Gushchin 3510457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3511457c1b27SNishanth Aravamudan return; 3512fcb2b0c5SRoman Gushchin 3513fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3514fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3515fcb2b0c5SRoman Gushchin 3516aca78307SMiaohe Lin total += huge_page_size(h) * count; 3517fcb2b0c5SRoman Gushchin 3518fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3519e1759c21SAlexey Dobriyan seq_printf(m, 35201da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 35211da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3522b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35237893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35244f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3525fcb2b0c5SRoman Gushchin count, 3526a5516438SAndi Kleen h->free_huge_pages, 3527a5516438SAndi Kleen h->resv_huge_pages, 3528a5516438SAndi Kleen h->surplus_huge_pages, 3529aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3530fcb2b0c5SRoman Gushchin } 3531fcb2b0c5SRoman Gushchin 3532aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 35331da177e4SLinus Torvalds } 35341da177e4SLinus Torvalds 35357981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 35361da177e4SLinus Torvalds { 3537a5516438SAndi Kleen struct hstate *h = &default_hstate; 35387981593bSJoe Perches 3539457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3540457c1b27SNishanth Aravamudan return 0; 35417981593bSJoe Perches 35427981593bSJoe Perches return sysfs_emit_at(buf, len, 35431da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3544a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3545a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3546a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3547a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3548a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35491da177e4SLinus Torvalds } 35501da177e4SLinus Torvalds 3551949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3552949f7ec5SDavid Rientjes { 3553949f7ec5SDavid Rientjes struct hstate *h; 3554949f7ec5SDavid Rientjes int nid; 3555949f7ec5SDavid Rientjes 3556457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3557457c1b27SNishanth Aravamudan return; 3558457c1b27SNishanth Aravamudan 3559949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3560949f7ec5SDavid Rientjes for_each_hstate(h) 3561949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3562949f7ec5SDavid Rientjes nid, 3563949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3564949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3565949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3566aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3567949f7ec5SDavid Rientjes } 3568949f7ec5SDavid Rientjes 35695d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 35705d317b2bSNaoya Horiguchi { 35715d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 35725d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 35735d317b2bSNaoya Horiguchi } 35745d317b2bSNaoya Horiguchi 35751da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 35761da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 35771da177e4SLinus Torvalds { 3578d0028588SWanpeng Li struct hstate *h; 3579d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3580d0028588SWanpeng Li 3581d0028588SWanpeng Li for_each_hstate(h) 3582d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3583d0028588SWanpeng Li return nr_total_pages; 35841da177e4SLinus Torvalds } 35851da177e4SLinus Torvalds 3586a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3587fc1b8a73SMel Gorman { 3588fc1b8a73SMel Gorman int ret = -ENOMEM; 3589fc1b8a73SMel Gorman 35900aa7f354SMiaohe Lin if (!delta) 35910aa7f354SMiaohe Lin return 0; 35920aa7f354SMiaohe Lin 3593fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3594fc1b8a73SMel Gorman /* 3595fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3596fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3597fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3598fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3599fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3600fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3601fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3602fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3603fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3604fc1b8a73SMel Gorman * 3605fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3606fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3607fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3608fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3609fc1b8a73SMel Gorman * semantics that cpuset has. 36108ca39e68SMuchun Song * 36118ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 36128ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 36138ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 36148ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 36158ca39e68SMuchun Song * above. 3616fc1b8a73SMel Gorman */ 3617fc1b8a73SMel Gorman if (delta > 0) { 3618a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3619fc1b8a73SMel Gorman goto out; 3620fc1b8a73SMel Gorman 36218ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3622a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3623fc1b8a73SMel Gorman goto out; 3624fc1b8a73SMel Gorman } 3625fc1b8a73SMel Gorman } 3626fc1b8a73SMel Gorman 3627fc1b8a73SMel Gorman ret = 0; 3628fc1b8a73SMel Gorman if (delta < 0) 3629a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3630fc1b8a73SMel Gorman 3631fc1b8a73SMel Gorman out: 3632fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3633fc1b8a73SMel Gorman return ret; 3634fc1b8a73SMel Gorman } 3635fc1b8a73SMel Gorman 363684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 363784afd99bSAndy Whitcroft { 3638f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 363984afd99bSAndy Whitcroft 364084afd99bSAndy Whitcroft /* 364184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 364284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 364384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 364425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 364584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 364684afd99bSAndy Whitcroft * new reference here without additional locking. 364784afd99bSAndy Whitcroft */ 36484e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3649f522c3acSJoonsoo Kim kref_get(&resv->refs); 365084afd99bSAndy Whitcroft } 365184afd99bSAndy Whitcroft 3652a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3653a1e78772SMel Gorman { 3654a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3655f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 365690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36574e35f483SJoonsoo Kim unsigned long reserve, start, end; 36581c5ecae3SMike Kravetz long gbl_reserve; 365984afd99bSAndy Whitcroft 36604e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36614e35f483SJoonsoo Kim return; 36624e35f483SJoonsoo Kim 3663a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3664a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 366584afd99bSAndy Whitcroft 36664e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3667e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36687251ff78SAdam Litke if (reserve) { 36691c5ecae3SMike Kravetz /* 36701c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36711c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36721c5ecae3SMike Kravetz */ 36731c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36741c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36757251ff78SAdam Litke } 3676e9fe92aeSMina Almasry 3677e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3678a1e78772SMel Gorman } 3679a1e78772SMel Gorman 368031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 368131383c68SDan Williams { 368231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 368331383c68SDan Williams return -EINVAL; 368431383c68SDan Williams return 0; 368531383c68SDan Williams } 368631383c68SDan Williams 368705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 368805ea8860SDan Williams { 3689aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 369005ea8860SDan Williams } 369105ea8860SDan Williams 36921da177e4SLinus Torvalds /* 36931da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 36941da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 36956c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 36961da177e4SLinus Torvalds * this far. 36971da177e4SLinus Torvalds */ 3698b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 36991da177e4SLinus Torvalds { 37001da177e4SLinus Torvalds BUG(); 3701d0217ac0SNick Piggin return 0; 37021da177e4SLinus Torvalds } 37031da177e4SLinus Torvalds 3704eec3636aSJane Chu /* 3705eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3706eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3707eec3636aSJane Chu * This is because under System V memory model, mappings created via 3708eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3709eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3710eec3636aSJane Chu */ 3711f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3712d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 371384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3714a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3715dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 371605ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 37171da177e4SLinus Torvalds }; 37181da177e4SLinus Torvalds 37191e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 37201e8f889bSDavid Gibson int writable) 372163551ae0SDavid Gibson { 372263551ae0SDavid Gibson pte_t entry; 372363551ae0SDavid Gibson 37241e8f889bSDavid Gibson if (writable) { 3725106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3726106c992aSGerald Schaefer vma->vm_page_prot))); 372763551ae0SDavid Gibson } else { 3728106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3729106c992aSGerald Schaefer vma->vm_page_prot)); 373063551ae0SDavid Gibson } 373163551ae0SDavid Gibson entry = pte_mkyoung(entry); 373263551ae0SDavid Gibson entry = pte_mkhuge(entry); 3733d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 373463551ae0SDavid Gibson 373563551ae0SDavid Gibson return entry; 373663551ae0SDavid Gibson } 373763551ae0SDavid Gibson 37381e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37391e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37401e8f889bSDavid Gibson { 37411e8f889bSDavid Gibson pte_t entry; 37421e8f889bSDavid Gibson 3743106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 374432f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37454b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37461e8f889bSDavid Gibson } 37471e8f889bSDavid Gibson 3748d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37494a705fefSNaoya Horiguchi { 37504a705fefSNaoya Horiguchi swp_entry_t swp; 37514a705fefSNaoya Horiguchi 37524a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3753d5ed7444SAneesh Kumar K.V return false; 37544a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3755d79d176aSBaoquan He if (is_migration_entry(swp)) 3756d5ed7444SAneesh Kumar K.V return true; 37574a705fefSNaoya Horiguchi else 3758d5ed7444SAneesh Kumar K.V return false; 37594a705fefSNaoya Horiguchi } 37604a705fefSNaoya Horiguchi 37613e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 37624a705fefSNaoya Horiguchi { 37634a705fefSNaoya Horiguchi swp_entry_t swp; 37644a705fefSNaoya Horiguchi 37654a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37663e5c3600SBaoquan He return false; 37674a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3768d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 37693e5c3600SBaoquan He return true; 37704a705fefSNaoya Horiguchi else 37713e5c3600SBaoquan He return false; 37724a705fefSNaoya Horiguchi } 37731e8f889bSDavid Gibson 37744eae4efaSPeter Xu static void 37754eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 37764eae4efaSPeter Xu struct page *new_page) 37774eae4efaSPeter Xu { 37784eae4efaSPeter Xu __SetPageUptodate(new_page); 37794eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 37804eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 37814eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 37824eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 37834eae4efaSPeter Xu SetHPageMigratable(new_page); 37844eae4efaSPeter Xu } 37854eae4efaSPeter Xu 378663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 378763551ae0SDavid Gibson struct vm_area_struct *vma) 378863551ae0SDavid Gibson { 37895e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 379063551ae0SDavid Gibson struct page *ptepage; 37911c59827dSHugh Dickins unsigned long addr; 3792ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 3793a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3794a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 37954eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 3796c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3797ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3798e8569dd2SAndreas Sandberg int ret = 0; 37991e8f889bSDavid Gibson 3800ac46d4f3SJérôme Glisse if (cow) { 38017269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 38026f4f13e8SJérôme Glisse vma->vm_start, 3803ac46d4f3SJérôme Glisse vma->vm_end); 3804ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3805c0d0381aSMike Kravetz } else { 3806c0d0381aSMike Kravetz /* 3807c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3808c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3809c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3810c0d0381aSMike Kravetz * huge_pmd_unshare. 3811c0d0381aSMike Kravetz */ 3812c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3813ac46d4f3SJérôme Glisse } 3814e8569dd2SAndreas Sandberg 3815a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3816cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 38177868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3818c74df32cSHugh Dickins if (!src_pte) 3819c74df32cSHugh Dickins continue; 3820aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 3821e8569dd2SAndreas Sandberg if (!dst_pte) { 3822e8569dd2SAndreas Sandberg ret = -ENOMEM; 3823e8569dd2SAndreas Sandberg break; 3824e8569dd2SAndreas Sandberg } 3825c5c99429SLarry Woodman 38265e41540cSMike Kravetz /* 38275e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38285e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38295e41540cSMike Kravetz * 38305e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38315e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38325e41540cSMike Kravetz * Check here before taking page table lock, and once again 38335e41540cSMike Kravetz * after taking the lock below. 38345e41540cSMike Kravetz */ 38355e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38365e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3837c5c99429SLarry Woodman continue; 3838c5c99429SLarry Woodman 3839cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3840cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3841cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38424a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38435e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38444eae4efaSPeter Xu again: 38455e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38465e41540cSMike Kravetz /* 38475e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38485e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38495e41540cSMike Kravetz * sharing with another vma. 38505e41540cSMike Kravetz */ 38514a705fefSNaoya Horiguchi ; 38524a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38534a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38544a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38554a705fefSNaoya Horiguchi 38564a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38574a705fefSNaoya Horiguchi /* 38584a705fefSNaoya Horiguchi * COW mappings require pages in both 38594a705fefSNaoya Horiguchi * parent and child to be set to read. 38604a705fefSNaoya Horiguchi */ 38614a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38624a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3863e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3864e5251fd4SPunit Agrawal entry, sz); 38654a705fefSNaoya Horiguchi } 3866e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38674a705fefSNaoya Horiguchi } else { 38684eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 38694eae4efaSPeter Xu ptepage = pte_page(entry); 38704eae4efaSPeter Xu get_page(ptepage); 38714eae4efaSPeter Xu 38724eae4efaSPeter Xu /* 38734eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 38744eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 38754eae4efaSPeter Xu * COW earlier during fork. 38764eae4efaSPeter Xu * 38774eae4efaSPeter Xu * When pre-allocating the page or copying data, we 38784eae4efaSPeter Xu * need to be without the pgtable locks since we could 38794eae4efaSPeter Xu * sleep during the process. 38804eae4efaSPeter Xu */ 38814eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 38824eae4efaSPeter Xu pte_t src_pte_old = entry; 38834eae4efaSPeter Xu struct page *new; 38844eae4efaSPeter Xu 38854eae4efaSPeter Xu spin_unlock(src_ptl); 38864eae4efaSPeter Xu spin_unlock(dst_ptl); 38874eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 38884eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 38894eae4efaSPeter Xu if (IS_ERR(new)) { 38904eae4efaSPeter Xu put_page(ptepage); 38914eae4efaSPeter Xu ret = PTR_ERR(new); 38924eae4efaSPeter Xu break; 38934eae4efaSPeter Xu } 38944eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 38954eae4efaSPeter Xu npages); 38964eae4efaSPeter Xu put_page(ptepage); 38974eae4efaSPeter Xu 38984eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 38994eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 39004eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 39014eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 39024eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 39034eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 39044eae4efaSPeter Xu put_page(new); 39054eae4efaSPeter Xu /* dst_entry won't change as in child */ 39064eae4efaSPeter Xu goto again; 39074eae4efaSPeter Xu } 39084eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 39094eae4efaSPeter Xu spin_unlock(src_ptl); 39104eae4efaSPeter Xu spin_unlock(dst_ptl); 39114eae4efaSPeter Xu continue; 39124eae4efaSPeter Xu } 39134eae4efaSPeter Xu 391434ee645eSJoerg Roedel if (cow) { 39150f10851eSJérôme Glisse /* 39160f10851eSJérôme Glisse * No need to notify as we are downgrading page 39170f10851eSJérôme Glisse * table protection not changing it to point 39180f10851eSJérôme Glisse * to a new page. 39190f10851eSJérôme Glisse * 3920ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 39210f10851eSJérôme Glisse */ 39227f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 392334ee645eSJoerg Roedel } 39244eae4efaSPeter Xu 392553f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 392663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 39274eae4efaSPeter Xu hugetlb_count_add(npages, dst); 39281c59827dSHugh Dickins } 3929cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3930cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 393163551ae0SDavid Gibson } 393263551ae0SDavid Gibson 3933e8569dd2SAndreas Sandberg if (cow) 3934ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3935c0d0381aSMike Kravetz else 3936c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3937e8569dd2SAndreas Sandberg 3938e8569dd2SAndreas Sandberg return ret; 393963551ae0SDavid Gibson } 394063551ae0SDavid Gibson 394124669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 394224669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 394324669e58SAneesh Kumar K.V struct page *ref_page) 394463551ae0SDavid Gibson { 394563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 394663551ae0SDavid Gibson unsigned long address; 3947c7546f8fSDavid Gibson pte_t *ptep; 394863551ae0SDavid Gibson pte_t pte; 3949cb900f41SKirill A. Shutemov spinlock_t *ptl; 395063551ae0SDavid Gibson struct page *page; 3951a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3952a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3953ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3954a5516438SAndi Kleen 395563551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3956a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3957a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 395863551ae0SDavid Gibson 395907e32661SAneesh Kumar K.V /* 396007e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 396107e32661SAneesh Kumar K.V * to huge page. 396207e32661SAneesh Kumar K.V */ 3963ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 396424669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3965dff11abeSMike Kravetz 3966dff11abeSMike Kravetz /* 3967dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3968dff11abeSMike Kravetz */ 39696f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39706f4f13e8SJérôme Glisse end); 3971ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3972ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3973569f48b8SHillf Danton address = start; 3974569f48b8SHillf Danton for (; address < end; address += sz) { 39757868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3976c7546f8fSDavid Gibson if (!ptep) 3977c7546f8fSDavid Gibson continue; 3978c7546f8fSDavid Gibson 3979cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 398034ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 398131d49da5SAneesh Kumar K.V spin_unlock(ptl); 3982dff11abeSMike Kravetz /* 3983dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3984dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3985dff11abeSMike Kravetz */ 398631d49da5SAneesh Kumar K.V continue; 398731d49da5SAneesh Kumar K.V } 398839dde65cSChen, Kenneth W 39896629326bSHillf Danton pte = huge_ptep_get(ptep); 399031d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 399131d49da5SAneesh Kumar K.V spin_unlock(ptl); 399231d49da5SAneesh Kumar K.V continue; 399331d49da5SAneesh Kumar K.V } 39946629326bSHillf Danton 39956629326bSHillf Danton /* 39969fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39979fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39986629326bSHillf Danton */ 39999fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 40009386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 400131d49da5SAneesh Kumar K.V spin_unlock(ptl); 400231d49da5SAneesh Kumar K.V continue; 40038c4894c6SNaoya Horiguchi } 40046629326bSHillf Danton 40056629326bSHillf Danton page = pte_page(pte); 400604f2cbe3SMel Gorman /* 400704f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 400804f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 400904f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 401004f2cbe3SMel Gorman */ 401104f2cbe3SMel Gorman if (ref_page) { 401231d49da5SAneesh Kumar K.V if (page != ref_page) { 401331d49da5SAneesh Kumar K.V spin_unlock(ptl); 401431d49da5SAneesh Kumar K.V continue; 401531d49da5SAneesh Kumar K.V } 401604f2cbe3SMel Gorman /* 401704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 401804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 401904f2cbe3SMel Gorman * looking like data was lost 402004f2cbe3SMel Gorman */ 402104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 402204f2cbe3SMel Gorman } 402304f2cbe3SMel Gorman 4024c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4025b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4026106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 40276649a386SKen Chen set_page_dirty(page); 40289e81130bSHillf Danton 40295d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4030d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 403131d49da5SAneesh Kumar K.V 4032cb900f41SKirill A. Shutemov spin_unlock(ptl); 4033e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 403424669e58SAneesh Kumar K.V /* 403531d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 403624669e58SAneesh Kumar K.V */ 403731d49da5SAneesh Kumar K.V if (ref_page) 403831d49da5SAneesh Kumar K.V break; 4039fe1668aeSChen, Kenneth W } 4040ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 404124669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 40421da177e4SLinus Torvalds } 404363551ae0SDavid Gibson 4044d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4045d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4046d833352aSMel Gorman unsigned long end, struct page *ref_page) 4047d833352aSMel Gorman { 4048d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4049d833352aSMel Gorman 4050d833352aSMel Gorman /* 4051d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4052d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4053d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4054d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4055d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4056c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4057d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4058c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4059d833352aSMel Gorman */ 4060d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4061d833352aSMel Gorman } 4062d833352aSMel Gorman 4063502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 406404f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4065502717f4SChen, Kenneth W { 406624669e58SAneesh Kumar K.V struct mmu_gather tlb; 4067dff11abeSMike Kravetz 4068a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 406924669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4070ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4071502717f4SChen, Kenneth W } 4072502717f4SChen, Kenneth W 407304f2cbe3SMel Gorman /* 407404f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4075578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 407604f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 407704f2cbe3SMel Gorman * same region. 407804f2cbe3SMel Gorman */ 40792f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40802a4b3dedSHarvey Harrison struct page *page, unsigned long address) 408104f2cbe3SMel Gorman { 40827526674dSAdam Litke struct hstate *h = hstate_vma(vma); 408304f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 408404f2cbe3SMel Gorman struct address_space *mapping; 408504f2cbe3SMel Gorman pgoff_t pgoff; 408604f2cbe3SMel Gorman 408704f2cbe3SMel Gorman /* 408804f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 408904f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 409004f2cbe3SMel Gorman */ 40917526674dSAdam Litke address = address & huge_page_mask(h); 409236e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 409336e4f20aSMichal Hocko vma->vm_pgoff; 409493c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 409504f2cbe3SMel Gorman 40964eb2b1dcSMel Gorman /* 40974eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40984eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40994eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 41004eb2b1dcSMel Gorman */ 410183cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 41026b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 410304f2cbe3SMel Gorman /* Do not unmap the current VMA */ 410404f2cbe3SMel Gorman if (iter_vma == vma) 410504f2cbe3SMel Gorman continue; 410604f2cbe3SMel Gorman 410704f2cbe3SMel Gorman /* 41082f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 41092f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 41102f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 41112f84a899SMel Gorman */ 41122f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 41132f84a899SMel Gorman continue; 41142f84a899SMel Gorman 41152f84a899SMel Gorman /* 411604f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 411704f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 411804f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 411904f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 412004f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 412104f2cbe3SMel Gorman */ 412204f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 412324669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 412424669e58SAneesh Kumar K.V address + huge_page_size(h), page); 412504f2cbe3SMel Gorman } 412683cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 412704f2cbe3SMel Gorman } 412804f2cbe3SMel Gorman 41290fe6e20bSNaoya Horiguchi /* 41300fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4131ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4132ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4133ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 41340fe6e20bSNaoya Horiguchi */ 41352b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4136974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4137cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 41381e8f889bSDavid Gibson { 41393999f52eSAneesh Kumar K.V pte_t pte; 4140a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41411e8f889bSDavid Gibson struct page *old_page, *new_page; 41422b740303SSouptick Joarder int outside_reserve = 0; 41432b740303SSouptick Joarder vm_fault_t ret = 0; 4144974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4145ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 41461e8f889bSDavid Gibson 41473999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 41481e8f889bSDavid Gibson old_page = pte_page(pte); 41491e8f889bSDavid Gibson 415004f2cbe3SMel Gorman retry_avoidcopy: 41511e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 41521e8f889bSDavid Gibson * and just make the page writable */ 415337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 41545a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 41555b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 415683c54070SNick Piggin return 0; 41571e8f889bSDavid Gibson } 41581e8f889bSDavid Gibson 415904f2cbe3SMel Gorman /* 416004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 416104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 416204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 416304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 416404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 416504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 416604f2cbe3SMel Gorman * of the full address range. 416704f2cbe3SMel Gorman */ 41685944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 416904f2cbe3SMel Gorman old_page != pagecache_page) 417004f2cbe3SMel Gorman outside_reserve = 1; 417104f2cbe3SMel Gorman 417209cbfeafSKirill A. Shutemov get_page(old_page); 4173b76c8cfbSLarry Woodman 4174ad4404a2SDavidlohr Bueso /* 4175ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4176ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4177ad4404a2SDavidlohr Bueso */ 4178cb900f41SKirill A. Shutemov spin_unlock(ptl); 41795b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41801e8f889bSDavid Gibson 41812fc39cecSAdam Litke if (IS_ERR(new_page)) { 418204f2cbe3SMel Gorman /* 418304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 418404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 418504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 418604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 418704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 418804f2cbe3SMel Gorman */ 418904f2cbe3SMel Gorman if (outside_reserve) { 4190e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4191e7dd91c4SMike Kravetz pgoff_t idx; 4192e7dd91c4SMike Kravetz u32 hash; 4193e7dd91c4SMike Kravetz 419409cbfeafSKirill A. Shutemov put_page(old_page); 419504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4196e7dd91c4SMike Kravetz /* 4197e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4198e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4199e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4200e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4201e7dd91c4SMike Kravetz * PMD sharing. 4202e7dd91c4SMike Kravetz * 4203e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4204e7dd91c4SMike Kravetz */ 4205e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4206e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4207e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4208e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4209e7dd91c4SMike Kravetz 42105b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4211e7dd91c4SMike Kravetz 4212e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4213e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4214cb900f41SKirill A. Shutemov spin_lock(ptl); 42155b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4216a9af0c5dSNaoya Horiguchi if (likely(ptep && 4217a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 421804f2cbe3SMel Gorman goto retry_avoidcopy; 4219a734bcc8SHillf Danton /* 4220cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4221cb900f41SKirill A. Shutemov * lock, and our job is done. 4222a734bcc8SHillf Danton */ 4223a734bcc8SHillf Danton return 0; 422404f2cbe3SMel Gorman } 422504f2cbe3SMel Gorman 42262b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4227ad4404a2SDavidlohr Bueso goto out_release_old; 42281e8f889bSDavid Gibson } 42291e8f889bSDavid Gibson 42300fe6e20bSNaoya Horiguchi /* 42310fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 42320fe6e20bSNaoya Horiguchi * anon_vma prepared. 42330fe6e20bSNaoya Horiguchi */ 423444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4235ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4236ad4404a2SDavidlohr Bueso goto out_release_all; 423744e2aa93SDean Nelson } 42380fe6e20bSNaoya Horiguchi 4239974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 424047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 42410ed361deSNick Piggin __SetPageUptodate(new_page); 42421e8f889bSDavid Gibson 42437269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 42446f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4245ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4246ad4404a2SDavidlohr Bueso 4247b76c8cfbSLarry Woodman /* 4248cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4249b76c8cfbSLarry Woodman * before the page tables are altered 4250b76c8cfbSLarry Woodman */ 4251cb900f41SKirill A. Shutemov spin_lock(ptl); 42525b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4253a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4254d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 425507443a85SJoonsoo Kim 42561e8f889bSDavid Gibson /* Break COW */ 42575b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4258ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 42595b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 42601e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4261d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 42625b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 42638f251a3dSMike Kravetz SetHPageMigratable(new_page); 42641e8f889bSDavid Gibson /* Make the old page be freed below */ 42651e8f889bSDavid Gibson new_page = old_page; 42661e8f889bSDavid Gibson } 4267cb900f41SKirill A. Shutemov spin_unlock(ptl); 4268ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4269ad4404a2SDavidlohr Bueso out_release_all: 42705b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 427109cbfeafSKirill A. Shutemov put_page(new_page); 4272ad4404a2SDavidlohr Bueso out_release_old: 427309cbfeafSKirill A. Shutemov put_page(old_page); 42748312034fSJoonsoo Kim 4275ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4276ad4404a2SDavidlohr Bueso return ret; 42771e8f889bSDavid Gibson } 42781e8f889bSDavid Gibson 427904f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4280a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4281a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 428204f2cbe3SMel Gorman { 428304f2cbe3SMel Gorman struct address_space *mapping; 4284e7c4b0bfSAndy Whitcroft pgoff_t idx; 428504f2cbe3SMel Gorman 428604f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4287a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 428804f2cbe3SMel Gorman 428904f2cbe3SMel Gorman return find_lock_page(mapping, idx); 429004f2cbe3SMel Gorman } 429104f2cbe3SMel Gorman 42923ae77f43SHugh Dickins /* 42933ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42943ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42953ae77f43SHugh Dickins */ 42963ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42972a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42982a15efc9SHugh Dickins { 42992a15efc9SHugh Dickins struct address_space *mapping; 43002a15efc9SHugh Dickins pgoff_t idx; 43012a15efc9SHugh Dickins struct page *page; 43022a15efc9SHugh Dickins 43032a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 43042a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 43052a15efc9SHugh Dickins 43062a15efc9SHugh Dickins page = find_get_page(mapping, idx); 43072a15efc9SHugh Dickins if (page) 43082a15efc9SHugh Dickins put_page(page); 43092a15efc9SHugh Dickins return page != NULL; 43102a15efc9SHugh Dickins } 43112a15efc9SHugh Dickins 4312ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4313ab76ad54SMike Kravetz pgoff_t idx) 4314ab76ad54SMike Kravetz { 4315ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4316ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4317ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4318ab76ad54SMike Kravetz 4319ab76ad54SMike Kravetz if (err) 4320ab76ad54SMike Kravetz return err; 4321d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4322ab76ad54SMike Kravetz 432322146c3cSMike Kravetz /* 432422146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 432522146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 432622146c3cSMike Kravetz */ 432722146c3cSMike Kravetz set_page_dirty(page); 432822146c3cSMike Kravetz 4329ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4330ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4331ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4332ab76ad54SMike Kravetz return 0; 4333ab76ad54SMike Kravetz } 4334ab76ad54SMike Kravetz 43352b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 43362b740303SSouptick Joarder struct vm_area_struct *vma, 43378382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4338788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4339ac9b9c66SHugh Dickins { 4340a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43412b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4342409eb8c2SHillf Danton int anon_rmap = 0; 43434c887265SAdam Litke unsigned long size; 43444c887265SAdam Litke struct page *page; 43451e8f889bSDavid Gibson pte_t new_pte; 4346cb900f41SKirill A. Shutemov spinlock_t *ptl; 4347285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4348cb6acd01SMike Kravetz bool new_page = false; 43494c887265SAdam Litke 435004f2cbe3SMel Gorman /* 435104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 435204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 435325985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 435404f2cbe3SMel Gorman */ 435504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4356910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 435704f2cbe3SMel Gorman current->pid); 435804f2cbe3SMel Gorman return ret; 435904f2cbe3SMel Gorman } 436004f2cbe3SMel Gorman 43614c887265SAdam Litke /* 436287bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 436387bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 436487bf91d3SMike Kravetz * once for faults beyond end of file. 43654c887265SAdam Litke */ 4366a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4367ebed4bfcSHugh Dickins if (idx >= size) 4368ebed4bfcSHugh Dickins goto out; 43691a1aad8aSMike Kravetz 437087bf91d3SMike Kravetz retry: 437187bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 437287bf91d3SMike Kravetz if (!page) { 43731a1aad8aSMike Kravetz /* 43741a1aad8aSMike Kravetz * Check for page in userfault range 43751a1aad8aSMike Kravetz */ 43761a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43771a1aad8aSMike Kravetz u32 hash; 43781a1aad8aSMike Kravetz struct vm_fault vmf = { 43791a1aad8aSMike Kravetz .vma = vma, 4380285b8dcaSHuang Ying .address = haddr, 43811a1aad8aSMike Kravetz .flags = flags, 43821a1aad8aSMike Kravetz /* 43831a1aad8aSMike Kravetz * Hard to debug if it ends up being 43841a1aad8aSMike Kravetz * used by a callee that assumes 43851a1aad8aSMike Kravetz * something about the other 43861a1aad8aSMike Kravetz * uninitialized fields... same as in 43871a1aad8aSMike Kravetz * memory.c 43881a1aad8aSMike Kravetz */ 43891a1aad8aSMike Kravetz }; 43901a1aad8aSMike Kravetz 43911a1aad8aSMike Kravetz /* 4392c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4393c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4394c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43951a1aad8aSMike Kravetz */ 4396188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43971a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4398c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43991a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4400c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 44011a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 44021a1aad8aSMike Kravetz goto out; 44031a1aad8aSMike Kravetz } 44041a1aad8aSMike Kravetz 4405285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 44062fc39cecSAdam Litke if (IS_ERR(page)) { 44074643d67eSMike Kravetz /* 44084643d67eSMike Kravetz * Returning error will result in faulting task being 44094643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 44104643d67eSMike Kravetz * tasks from racing to fault in the same page which 44114643d67eSMike Kravetz * could result in false unable to allocate errors. 44124643d67eSMike Kravetz * Page migration does not take the fault mutex, but 44134643d67eSMike Kravetz * does a clear then write of pte's under page table 44144643d67eSMike Kravetz * lock. Page fault code could race with migration, 44154643d67eSMike Kravetz * notice the clear pte and try to allocate a page 44164643d67eSMike Kravetz * here. Before returning error, get ptl and make 44174643d67eSMike Kravetz * sure there really is no pte entry. 44184643d67eSMike Kravetz */ 44194643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 44204643d67eSMike Kravetz ret = 0; 4421d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 44222b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 4423d83e6c8aSMiaohe Lin spin_unlock(ptl); 44246bda666aSChristoph Lameter goto out; 44256bda666aSChristoph Lameter } 442647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 44270ed361deSNick Piggin __SetPageUptodate(page); 4428cb6acd01SMike Kravetz new_page = true; 4429ac9b9c66SHugh Dickins 4430f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4431ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 44326bda666aSChristoph Lameter if (err) { 44336bda666aSChristoph Lameter put_page(page); 44346bda666aSChristoph Lameter if (err == -EEXIST) 44356bda666aSChristoph Lameter goto retry; 44366bda666aSChristoph Lameter goto out; 44376bda666aSChristoph Lameter } 443823be7468SMel Gorman } else { 44396bda666aSChristoph Lameter lock_page(page); 44400fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 44410fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 44420fe6e20bSNaoya Horiguchi goto backout_unlocked; 444323be7468SMel Gorman } 4444409eb8c2SHillf Danton anon_rmap = 1; 44450fe6e20bSNaoya Horiguchi } 44460fe6e20bSNaoya Horiguchi } else { 444757303d80SAndy Whitcroft /* 4448998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4449998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4450998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4451fd6a03edSNaoya Horiguchi */ 4452fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 44530eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4454972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4455fd6a03edSNaoya Horiguchi goto backout_unlocked; 44566bda666aSChristoph Lameter } 4457998b4382SNaoya Horiguchi } 44581e8f889bSDavid Gibson 445957303d80SAndy Whitcroft /* 446057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 446157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 446257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 446357303d80SAndy Whitcroft * the spinlock. 446457303d80SAndy Whitcroft */ 44655e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4466285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44672b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44682b26736cSAndy Whitcroft goto backout_unlocked; 44692b26736cSAndy Whitcroft } 44705e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4471285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44725e911373SMike Kravetz } 447357303d80SAndy Whitcroft 44748bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 447583c54070SNick Piggin ret = 0; 44767f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44774c887265SAdam Litke goto backout; 44784c887265SAdam Litke 447907443a85SJoonsoo Kim if (anon_rmap) { 4480d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4481285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4482ac714904SChoi Gi-yong } else 448353f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44841e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44851e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4486285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44871e8f889bSDavid Gibson 44885d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4489788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44901e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4491974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44921e8f889bSDavid Gibson } 44931e8f889bSDavid Gibson 4494cb900f41SKirill A. Shutemov spin_unlock(ptl); 4495cb6acd01SMike Kravetz 4496cb6acd01SMike Kravetz /* 44978f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 44988f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 44998f251a3dSMike Kravetz * been isolated for migration. 4500cb6acd01SMike Kravetz */ 4501cb6acd01SMike Kravetz if (new_page) 45028f251a3dSMike Kravetz SetHPageMigratable(page); 4503cb6acd01SMike Kravetz 45044c887265SAdam Litke unlock_page(page); 45054c887265SAdam Litke out: 4506ac9b9c66SHugh Dickins return ret; 45074c887265SAdam Litke 45084c887265SAdam Litke backout: 4509cb900f41SKirill A. Shutemov spin_unlock(ptl); 45102b26736cSAndy Whitcroft backout_unlocked: 45114c887265SAdam Litke unlock_page(page); 4512285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 45134c887265SAdam Litke put_page(page); 45144c887265SAdam Litke goto out; 4515ac9b9c66SHugh Dickins } 4516ac9b9c66SHugh Dickins 45178382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4518188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45198382d914SDavidlohr Bueso { 45208382d914SDavidlohr Bueso unsigned long key[2]; 45218382d914SDavidlohr Bueso u32 hash; 45228382d914SDavidlohr Bueso 45238382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 45248382d914SDavidlohr Bueso key[1] = idx; 45258382d914SDavidlohr Bueso 452655254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 45278382d914SDavidlohr Bueso 45288382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 45298382d914SDavidlohr Bueso } 45308382d914SDavidlohr Bueso #else 45318382d914SDavidlohr Bueso /* 45326c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 45338382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 45348382d914SDavidlohr Bueso */ 4535188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45368382d914SDavidlohr Bueso { 45378382d914SDavidlohr Bueso return 0; 45388382d914SDavidlohr Bueso } 45398382d914SDavidlohr Bueso #endif 45408382d914SDavidlohr Bueso 45412b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4542788c7df4SHugh Dickins unsigned long address, unsigned int flags) 454386e5216fSAdam Litke { 45448382d914SDavidlohr Bueso pte_t *ptep, entry; 4545cb900f41SKirill A. Shutemov spinlock_t *ptl; 45462b740303SSouptick Joarder vm_fault_t ret; 45478382d914SDavidlohr Bueso u32 hash; 45488382d914SDavidlohr Bueso pgoff_t idx; 45490fe6e20bSNaoya Horiguchi struct page *page = NULL; 455057303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4551a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45528382d914SDavidlohr Bueso struct address_space *mapping; 45530f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4554285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 455586e5216fSAdam Litke 4556285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4557fd6a03edSNaoya Horiguchi if (ptep) { 4558c0d0381aSMike Kravetz /* 4559c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4560c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4561c0d0381aSMike Kravetz * not actually modifying content here. 4562c0d0381aSMike Kravetz */ 4563fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4564290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4565cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4566290408d4SNaoya Horiguchi return 0; 4567290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4568aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4569972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4570b43a9990SMike Kravetz } 45718382d914SDavidlohr Bueso 4572c0d0381aSMike Kravetz /* 4573c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 457487bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 457587bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 457687bf91d3SMike Kravetz * and making the ptep no longer valid. 457787bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4578c0d0381aSMike Kravetz * 4579c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4580c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4581c0d0381aSMike Kravetz * something has changed. 4582c0d0381aSMike Kravetz */ 4583ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4584c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4585aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 4586c0d0381aSMike Kravetz if (!ptep) { 4587c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4588c0d0381aSMike Kravetz return VM_FAULT_OOM; 4589c0d0381aSMike Kravetz } 4590ddeaab32SMike Kravetz 45913935baa9SDavid Gibson /* 45923935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45933935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45943935baa9SDavid Gibson * the same page in the page cache. 45953935baa9SDavid Gibson */ 4596c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4597188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4598c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45998382d914SDavidlohr Bueso 46007f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 46017f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 46028382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4603b4d1d99fSDavid Gibson goto out_mutex; 46043935baa9SDavid Gibson } 460586e5216fSAdam Litke 460683c54070SNick Piggin ret = 0; 46071e8f889bSDavid Gibson 460857303d80SAndy Whitcroft /* 46090f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 46100f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 46117c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 46127c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 46137c8de358SEthon Paul * properly handle it. 46140f792cf9SNaoya Horiguchi */ 46150f792cf9SNaoya Horiguchi if (!pte_present(entry)) 46160f792cf9SNaoya Horiguchi goto out_mutex; 46170f792cf9SNaoya Horiguchi 46180f792cf9SNaoya Horiguchi /* 461957303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 462057303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 462157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 462257303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 462357303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 462457303d80SAndy Whitcroft * consumed. 462557303d80SAndy Whitcroft */ 4626106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4627285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 46282b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4629b4d1d99fSDavid Gibson goto out_mutex; 46302b26736cSAndy Whitcroft } 46315e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4632285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 463357303d80SAndy Whitcroft 4634f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 463557303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4636285b8dcaSHuang Ying vma, haddr); 463757303d80SAndy Whitcroft } 463857303d80SAndy Whitcroft 46390f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 46400fe6e20bSNaoya Horiguchi 46411e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4642b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4643cb900f41SKirill A. Shutemov goto out_ptl; 4644b4d1d99fSDavid Gibson 46450f792cf9SNaoya Horiguchi /* 46460f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 46470f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 46480f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 46490f792cf9SNaoya Horiguchi */ 46500f792cf9SNaoya Horiguchi page = pte_page(entry); 46510f792cf9SNaoya Horiguchi if (page != pagecache_page) 46520f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 46530f792cf9SNaoya Horiguchi need_wait_lock = 1; 46540f792cf9SNaoya Horiguchi goto out_ptl; 46550f792cf9SNaoya Horiguchi } 46560f792cf9SNaoya Horiguchi 46570f792cf9SNaoya Horiguchi get_page(page); 4658b4d1d99fSDavid Gibson 4659788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4660106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4661974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4662cb900f41SKirill A. Shutemov pagecache_page, ptl); 46630f792cf9SNaoya Horiguchi goto out_put_page; 4664b4d1d99fSDavid Gibson } 4665106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4666b4d1d99fSDavid Gibson } 4667b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4668285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4669788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4670285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46710f792cf9SNaoya Horiguchi out_put_page: 46720f792cf9SNaoya Horiguchi if (page != pagecache_page) 46730f792cf9SNaoya Horiguchi unlock_page(page); 46740f792cf9SNaoya Horiguchi put_page(page); 4675cb900f41SKirill A. Shutemov out_ptl: 4676cb900f41SKirill A. Shutemov spin_unlock(ptl); 467757303d80SAndy Whitcroft 467857303d80SAndy Whitcroft if (pagecache_page) { 467957303d80SAndy Whitcroft unlock_page(pagecache_page); 468057303d80SAndy Whitcroft put_page(pagecache_page); 468157303d80SAndy Whitcroft } 4682b4d1d99fSDavid Gibson out_mutex: 4683c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4684c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46850f792cf9SNaoya Horiguchi /* 46860f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46870f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46880f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46890f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46900f792cf9SNaoya Horiguchi * here without taking refcount. 46910f792cf9SNaoya Horiguchi */ 46920f792cf9SNaoya Horiguchi if (need_wait_lock) 46930f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46941e8f889bSDavid Gibson return ret; 469586e5216fSAdam Litke } 469686e5216fSAdam Litke 46978fb5debcSMike Kravetz /* 46988fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46998fb5debcSMike Kravetz * modifications for huge pages. 47008fb5debcSMike Kravetz */ 47018fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 47028fb5debcSMike Kravetz pte_t *dst_pte, 47038fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 47048fb5debcSMike Kravetz unsigned long dst_addr, 47058fb5debcSMike Kravetz unsigned long src_addr, 47068fb5debcSMike Kravetz struct page **pagep) 47078fb5debcSMike Kravetz { 47081e392147SAndrea Arcangeli struct address_space *mapping; 47091e392147SAndrea Arcangeli pgoff_t idx; 47101e392147SAndrea Arcangeli unsigned long size; 47111c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 47128fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 47138fb5debcSMike Kravetz pte_t _dst_pte; 47148fb5debcSMike Kravetz spinlock_t *ptl; 47158fb5debcSMike Kravetz int ret; 47168fb5debcSMike Kravetz struct page *page; 47178fb5debcSMike Kravetz 47188fb5debcSMike Kravetz if (!*pagep) { 47198fb5debcSMike Kravetz ret = -ENOMEM; 47208fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 47218fb5debcSMike Kravetz if (IS_ERR(page)) 47228fb5debcSMike Kravetz goto out; 47238fb5debcSMike Kravetz 47248fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 47258fb5debcSMike Kravetz (const void __user *) src_addr, 4726810a56b9SMike Kravetz pages_per_huge_page(h), false); 47278fb5debcSMike Kravetz 4728c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 47298fb5debcSMike Kravetz if (unlikely(ret)) { 47309e368259SAndrea Arcangeli ret = -ENOENT; 47318fb5debcSMike Kravetz *pagep = page; 47328fb5debcSMike Kravetz /* don't free the page */ 47338fb5debcSMike Kravetz goto out; 47348fb5debcSMike Kravetz } 47358fb5debcSMike Kravetz } else { 47368fb5debcSMike Kravetz page = *pagep; 47378fb5debcSMike Kravetz *pagep = NULL; 47388fb5debcSMike Kravetz } 47398fb5debcSMike Kravetz 47408fb5debcSMike Kravetz /* 47418fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 47428fb5debcSMike Kravetz * preceding stores to the page contents become visible before 47438fb5debcSMike Kravetz * the set_pte_at() write. 47448fb5debcSMike Kravetz */ 47458fb5debcSMike Kravetz __SetPageUptodate(page); 47468fb5debcSMike Kravetz 47471e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 47481e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 47491e392147SAndrea Arcangeli 47501c9e8defSMike Kravetz /* 47511c9e8defSMike Kravetz * If shared, add to page cache 47521c9e8defSMike Kravetz */ 47531c9e8defSMike Kravetz if (vm_shared) { 47541e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47551e392147SAndrea Arcangeli ret = -EFAULT; 47561e392147SAndrea Arcangeli if (idx >= size) 47571e392147SAndrea Arcangeli goto out_release_nounlock; 47581c9e8defSMike Kravetz 47591e392147SAndrea Arcangeli /* 47601e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 47611e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 47621e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47631e392147SAndrea Arcangeli * the caller. 47641e392147SAndrea Arcangeli */ 47651c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47661c9e8defSMike Kravetz if (ret) 47671c9e8defSMike Kravetz goto out_release_nounlock; 47681c9e8defSMike Kravetz } 47691c9e8defSMike Kravetz 47708fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47718fb5debcSMike Kravetz spin_lock(ptl); 47728fb5debcSMike Kravetz 47731e392147SAndrea Arcangeli /* 47741e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47751e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47761e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47771e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47781e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47791e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47801e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47811e392147SAndrea Arcangeli */ 47821e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47831e392147SAndrea Arcangeli ret = -EFAULT; 47841e392147SAndrea Arcangeli if (idx >= size) 47851e392147SAndrea Arcangeli goto out_release_unlock; 47861e392147SAndrea Arcangeli 47878fb5debcSMike Kravetz ret = -EEXIST; 47888fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47898fb5debcSMike Kravetz goto out_release_unlock; 47908fb5debcSMike Kravetz 47911c9e8defSMike Kravetz if (vm_shared) { 47921c9e8defSMike Kravetz page_dup_rmap(page, true); 47931c9e8defSMike Kravetz } else { 4794d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 47958fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47961c9e8defSMike Kravetz } 47978fb5debcSMike Kravetz 47988fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47998fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 48008fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 48018fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 48028fb5debcSMike Kravetz 48038fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 48048fb5debcSMike Kravetz 48058fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 48068fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 48078fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 48088fb5debcSMike Kravetz 48098fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 48108fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 48118fb5debcSMike Kravetz 48128fb5debcSMike Kravetz spin_unlock(ptl); 48138f251a3dSMike Kravetz SetHPageMigratable(page); 48141c9e8defSMike Kravetz if (vm_shared) 48151c9e8defSMike Kravetz unlock_page(page); 48168fb5debcSMike Kravetz ret = 0; 48178fb5debcSMike Kravetz out: 48188fb5debcSMike Kravetz return ret; 48198fb5debcSMike Kravetz out_release_unlock: 48208fb5debcSMike Kravetz spin_unlock(ptl); 48211c9e8defSMike Kravetz if (vm_shared) 48221c9e8defSMike Kravetz unlock_page(page); 48235af10dfdSAndrea Arcangeli out_release_nounlock: 48248fb5debcSMike Kravetz put_page(page); 48258fb5debcSMike Kravetz goto out; 48268fb5debcSMike Kravetz } 48278fb5debcSMike Kravetz 482882e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 482982e5d378SJoao Martins int refs, struct page **pages, 483082e5d378SJoao Martins struct vm_area_struct **vmas) 483182e5d378SJoao Martins { 483282e5d378SJoao Martins int nr; 483382e5d378SJoao Martins 483482e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 483582e5d378SJoao Martins if (likely(pages)) 483682e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 483782e5d378SJoao Martins if (vmas) 483882e5d378SJoao Martins vmas[nr] = vma; 483982e5d378SJoao Martins } 484082e5d378SJoao Martins } 484182e5d378SJoao Martins 484228a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 484363551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 484428a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 48454f6da934SPeter Xu long i, unsigned int flags, int *locked) 484663551ae0SDavid Gibson { 4847d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4848d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 484928a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4850a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48510fa5bc40SJoao Martins int err = -EFAULT, refs; 485263551ae0SDavid Gibson 485363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 485463551ae0SDavid Gibson pte_t *pte; 4855cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 48562a15efc9SHugh Dickins int absent; 485763551ae0SDavid Gibson struct page *page; 485863551ae0SDavid Gibson 48594c887265SAdam Litke /* 486002057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 486102057967SDavid Rientjes * potentially allocating memory. 486202057967SDavid Rientjes */ 4863fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 486402057967SDavid Rientjes remainder = 0; 486502057967SDavid Rientjes break; 486602057967SDavid Rientjes } 486702057967SDavid Rientjes 486802057967SDavid Rientjes /* 48694c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48702a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48714c887265SAdam Litke * first, for the page indexing below to work. 4872cb900f41SKirill A. Shutemov * 4873cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48744c887265SAdam Litke */ 48757868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48767868a208SPunit Agrawal huge_page_size(h)); 4877cb900f41SKirill A. Shutemov if (pte) 4878cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48792a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 488063551ae0SDavid Gibson 48812a15efc9SHugh Dickins /* 48822a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48833ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48843ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48853ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48863ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48872a15efc9SHugh Dickins */ 48883ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48893ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4890cb900f41SKirill A. Shutemov if (pte) 4891cb900f41SKirill A. Shutemov spin_unlock(ptl); 48922a15efc9SHugh Dickins remainder = 0; 48932a15efc9SHugh Dickins break; 48942a15efc9SHugh Dickins } 48952a15efc9SHugh Dickins 48969cc3a5bdSNaoya Horiguchi /* 48979cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48989cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48999cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 49009cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 49019cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 49029cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 49039cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 49049cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 49059cc3a5bdSNaoya Horiguchi */ 49069cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4907106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4908106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 49092b740303SSouptick Joarder vm_fault_t ret; 491087ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 49114c887265SAdam Litke 4912cb900f41SKirill A. Shutemov if (pte) 4913cb900f41SKirill A. Shutemov spin_unlock(ptl); 491487ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 491587ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 49164f6da934SPeter Xu if (locked) 491771335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 491871335f37SPeter Xu FAULT_FLAG_KILLABLE; 491987ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 492087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 492187ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 492287ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 49234426e945SPeter Xu /* 49244426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 49254426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 49264426e945SPeter Xu */ 492787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 492887ffc118SAndrea Arcangeli } 492987ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 493087ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 49312be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 49321c59827dSHugh Dickins remainder = 0; 49331c59827dSHugh Dickins break; 49341c59827dSHugh Dickins } 493587ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 49364f6da934SPeter Xu if (locked && 49371ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 49384f6da934SPeter Xu *locked = 0; 493987ffc118SAndrea Arcangeli *nr_pages = 0; 494087ffc118SAndrea Arcangeli /* 494187ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 494287ffc118SAndrea Arcangeli * error, it will return zero 494387ffc118SAndrea Arcangeli * instead. 494487ffc118SAndrea Arcangeli * 494587ffc118SAndrea Arcangeli * No need to update "position" as the 494687ffc118SAndrea Arcangeli * caller will not check it after 494787ffc118SAndrea Arcangeli * *nr_pages is set to 0. 494887ffc118SAndrea Arcangeli */ 494987ffc118SAndrea Arcangeli return i; 495087ffc118SAndrea Arcangeli } 495187ffc118SAndrea Arcangeli continue; 495287ffc118SAndrea Arcangeli } 495363551ae0SDavid Gibson 4954a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 49557f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 49568fde12caSLinus Torvalds 49578fde12caSLinus Torvalds /* 4958acbfb087SZhigang Lu * If subpage information not requested, update counters 4959acbfb087SZhigang Lu * and skip the same_page loop below. 4960acbfb087SZhigang Lu */ 4961acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4962acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4963acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4964acbfb087SZhigang Lu vaddr += huge_page_size(h); 4965acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4966acbfb087SZhigang Lu i += pages_per_huge_page(h); 4967acbfb087SZhigang Lu spin_unlock(ptl); 4968acbfb087SZhigang Lu continue; 4969acbfb087SZhigang Lu } 4970acbfb087SZhigang Lu 497182e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 497282e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 49730fa5bc40SJoao Martins 497482e5d378SJoao Martins if (pages || vmas) 497582e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 497682e5d378SJoao Martins vma, refs, 497782e5d378SJoao Martins likely(pages) ? pages + i : NULL, 497882e5d378SJoao Martins vmas ? vmas + i : NULL); 497963551ae0SDavid Gibson 498082e5d378SJoao Martins if (pages) { 49810fa5bc40SJoao Martins /* 49820fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 49830fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 49840fa5bc40SJoao Martins * checked that the huge page is present in the page 49850fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 49860fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 49870fa5bc40SJoao Martins * head page and tail pages from being rearranged in 49880fa5bc40SJoao Martins * any way. So this page must be available at this 49890fa5bc40SJoao Martins * point, unless the page refcount overflowed: 49900fa5bc40SJoao Martins */ 499182e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 49920fa5bc40SJoao Martins refs, 49930fa5bc40SJoao Martins flags))) { 49940fa5bc40SJoao Martins spin_unlock(ptl); 49950fa5bc40SJoao Martins remainder = 0; 49960fa5bc40SJoao Martins err = -ENOMEM; 49970fa5bc40SJoao Martins break; 49980fa5bc40SJoao Martins } 4999d5d4b0aaSChen, Kenneth W } 500082e5d378SJoao Martins 500182e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 500282e5d378SJoao Martins remainder -= refs; 500382e5d378SJoao Martins i += refs; 500482e5d378SJoao Martins 5005cb900f41SKirill A. Shutemov spin_unlock(ptl); 500663551ae0SDavid Gibson } 500728a35716SMichel Lespinasse *nr_pages = remainder; 500887ffc118SAndrea Arcangeli /* 500987ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 501087ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 501187ffc118SAndrea Arcangeli * branch. 501287ffc118SAndrea Arcangeli */ 501363551ae0SDavid Gibson *position = vaddr; 501463551ae0SDavid Gibson 50152be7cfedSDaniel Jordan return i ? i : err; 501663551ae0SDavid Gibson } 50178f860591SZhang, Yanmin 50187da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 50198f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 50208f860591SZhang, Yanmin { 50218f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 50228f860591SZhang, Yanmin unsigned long start = address; 50238f860591SZhang, Yanmin pte_t *ptep; 50248f860591SZhang, Yanmin pte_t pte; 5025a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 50267da4d641SPeter Zijlstra unsigned long pages = 0; 5027dff11abeSMike Kravetz bool shared_pmd = false; 5028ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5029dff11abeSMike Kravetz 5030dff11abeSMike Kravetz /* 5031dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5032ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5033dff11abeSMike Kravetz * range if PMD sharing is possible. 5034dff11abeSMike Kravetz */ 50357269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 50367269f999SJérôme Glisse 0, vma, mm, start, end); 5037ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50388f860591SZhang, Yanmin 50398f860591SZhang, Yanmin BUG_ON(address >= end); 5040ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 50418f860591SZhang, Yanmin 5042ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 504383cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5044a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5045cb900f41SKirill A. Shutemov spinlock_t *ptl; 50467868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 50478f860591SZhang, Yanmin if (!ptep) 50488f860591SZhang, Yanmin continue; 5049cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 505034ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 50517da4d641SPeter Zijlstra pages++; 5052cb900f41SKirill A. Shutemov spin_unlock(ptl); 5053dff11abeSMike Kravetz shared_pmd = true; 505439dde65cSChen, Kenneth W continue; 50557da4d641SPeter Zijlstra } 5056a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5057a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5058a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5059a8bda28dSNaoya Horiguchi continue; 5060a8bda28dSNaoya Horiguchi } 5061a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5062a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5063a8bda28dSNaoya Horiguchi 5064a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5065a8bda28dSNaoya Horiguchi pte_t newpte; 5066a8bda28dSNaoya Horiguchi 5067a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5068a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5069e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5070e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5071a8bda28dSNaoya Horiguchi pages++; 5072a8bda28dSNaoya Horiguchi } 5073a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5074a8bda28dSNaoya Horiguchi continue; 5075a8bda28dSNaoya Horiguchi } 5076a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5077023bdd00SAneesh Kumar K.V pte_t old_pte; 5078023bdd00SAneesh Kumar K.V 5079023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5080023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5081be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5082023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50837da4d641SPeter Zijlstra pages++; 50848f860591SZhang, Yanmin } 5085cb900f41SKirill A. Shutemov spin_unlock(ptl); 50868f860591SZhang, Yanmin } 5087d833352aSMel Gorman /* 5088c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5089d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5090c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5091dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5092dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5093d833352aSMel Gorman */ 5094dff11abeSMike Kravetz if (shared_pmd) 5095ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5096dff11abeSMike Kravetz else 50975491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50980f10851eSJérôme Glisse /* 50990f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 51000f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 51010f10851eSJérôme Glisse * 5102ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 51030f10851eSJérôme Glisse */ 510483cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5105ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 51067da4d641SPeter Zijlstra 51077da4d641SPeter Zijlstra return pages << h->order; 51088f860591SZhang, Yanmin } 51098f860591SZhang, Yanmin 511033b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 511133b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5112a1e78772SMel Gorman long from, long to, 51135a6fe125SMel Gorman struct vm_area_struct *vma, 5114ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5115e4e574b7SAdam Litke { 511633b8f84aSMike Kravetz long chg, add = -1; 5117a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 511890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 51199119a41eSJoonsoo Kim struct resv_map *resv_map; 5120075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 51210db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5122e4e574b7SAdam Litke 512363489f8eSMike Kravetz /* This should never happen */ 512463489f8eSMike Kravetz if (from > to) { 512563489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 512633b8f84aSMike Kravetz return false; 512763489f8eSMike Kravetz } 512863489f8eSMike Kravetz 5129a1e78772SMel Gorman /* 513017c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 513117c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 513290481622SDavid Gibson * without using reserves 513317c9d12eSMel Gorman */ 5134ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 513533b8f84aSMike Kravetz return true; 513617c9d12eSMel Gorman 513717c9d12eSMel Gorman /* 5138a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5139a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5140a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5141a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5142a1e78772SMel Gorman */ 51439119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5144f27a5136SMike Kravetz /* 5145f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5146f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5147f27a5136SMike Kravetz * hugetlbfs_get_inode). 5148f27a5136SMike Kravetz */ 51494e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 51509119a41eSJoonsoo Kim 51510db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 51529119a41eSJoonsoo Kim 51539119a41eSJoonsoo Kim } else { 5154e9fe92aeSMina Almasry /* Private mapping. */ 51559119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51565a6fe125SMel Gorman if (!resv_map) 515733b8f84aSMike Kravetz return false; 51585a6fe125SMel Gorman 515917c9d12eSMel Gorman chg = to - from; 516017c9d12eSMel Gorman 51615a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51625a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51635a6fe125SMel Gorman } 51645a6fe125SMel Gorman 516533b8f84aSMike Kravetz if (chg < 0) 5166c50ac050SDave Hansen goto out_err; 516717c9d12eSMel Gorman 516833b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 516933b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5170075a61d0SMina Almasry goto out_err; 5171075a61d0SMina Almasry 5172075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5173075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5174075a61d0SMina Almasry * of the resv_map. 5175075a61d0SMina Almasry */ 5176075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5177075a61d0SMina Almasry } 5178075a61d0SMina Almasry 51791c5ecae3SMike Kravetz /* 51801c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51811c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51821c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51831c5ecae3SMike Kravetz */ 51841c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 518533b8f84aSMike Kravetz if (gbl_reserve < 0) 5186075a61d0SMina Almasry goto out_uncharge_cgroup; 518717c9d12eSMel Gorman 518817c9d12eSMel Gorman /* 518917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 519090481622SDavid Gibson * Hand the pages back to the subpool if there are not 519117c9d12eSMel Gorman */ 519233b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5193075a61d0SMina Almasry goto out_put_pages; 519417c9d12eSMel Gorman 519517c9d12eSMel Gorman /* 519617c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 519717c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 519817c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 519917c9d12eSMel Gorman * the reservation was and the page cache tells how much of 520017c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 520117c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 520217c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 520317c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 520417c9d12eSMel Gorman * else has to be done for private mappings here 520517c9d12eSMel Gorman */ 520633039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5207075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 520833039678SMike Kravetz 52090db9d74eSMina Almasry if (unlikely(add < 0)) { 52100db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5211075a61d0SMina Almasry goto out_put_pages; 52120db9d74eSMina Almasry } else if (unlikely(chg > add)) { 521333039678SMike Kravetz /* 521433039678SMike Kravetz * pages in this range were added to the reserve 521533039678SMike Kravetz * map between region_chg and region_add. This 521633039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 521733039678SMike Kravetz * the subpool and reserve counts modified above 521833039678SMike Kravetz * based on the difference. 521933039678SMike Kravetz */ 522033039678SMike Kravetz long rsv_adjust; 522133039678SMike Kravetz 5222d85aecf2SMiaohe Lin /* 5223d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5224d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5225d85aecf2SMiaohe Lin */ 5226075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5227075a61d0SMina Almasry hstate_index(h), 5228075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5229075a61d0SMina Almasry 523033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 523133039678SMike Kravetz chg - add); 523233039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5233d85aecf2SMiaohe Lin } else if (h_cg) { 5234d85aecf2SMiaohe Lin /* 5235d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5236d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5237d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5238d85aecf2SMiaohe Lin * done. 5239d85aecf2SMiaohe Lin */ 5240d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 524133039678SMike Kravetz } 524233039678SMike Kravetz } 524333b8f84aSMike Kravetz return true; 524433b8f84aSMike Kravetz 5245075a61d0SMina Almasry out_put_pages: 5246075a61d0SMina Almasry /* put back original number of pages, chg */ 5247075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5248075a61d0SMina Almasry out_uncharge_cgroup: 5249075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5250075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5251c50ac050SDave Hansen out_err: 52525e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 52530db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 52540db9d74eSMina Almasry * region_add failed or didn't run. 52550db9d74eSMina Almasry */ 52560db9d74eSMina Almasry if (chg >= 0 && add < 0) 52570db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5258f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5259f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 526033b8f84aSMike Kravetz return false; 5261a43a8c39SChen, Kenneth W } 5262a43a8c39SChen, Kenneth W 5263b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5264b5cec28dSMike Kravetz long freed) 5265a43a8c39SChen, Kenneth W { 5266a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52674e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52689119a41eSJoonsoo Kim long chg = 0; 526990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52701c5ecae3SMike Kravetz long gbl_reserve; 527145c682a6SKen Chen 5272f27a5136SMike Kravetz /* 5273f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5274f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5275f27a5136SMike Kravetz */ 5276b5cec28dSMike Kravetz if (resv_map) { 5277b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5278b5cec28dSMike Kravetz /* 5279b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5280b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5281b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5282b5cec28dSMike Kravetz */ 5283b5cec28dSMike Kravetz if (chg < 0) 5284b5cec28dSMike Kravetz return chg; 5285b5cec28dSMike Kravetz } 5286b5cec28dSMike Kravetz 528745c682a6SKen Chen spin_lock(&inode->i_lock); 5288e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 528945c682a6SKen Chen spin_unlock(&inode->i_lock); 529045c682a6SKen Chen 52911c5ecae3SMike Kravetz /* 52921c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52931c5ecae3SMike Kravetz * reservations to be released may be adjusted. 5294dddf31a4SMiaohe Lin * 5295dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 5296dddf31a4SMiaohe Lin * won't go negative. 52971c5ecae3SMike Kravetz */ 52981c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52991c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5300b5cec28dSMike Kravetz 5301b5cec28dSMike Kravetz return 0; 5302a43a8c39SChen, Kenneth W } 530393f70f90SNaoya Horiguchi 53043212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 53053212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 53063212b535SSteve Capper struct vm_area_struct *vma, 53073212b535SSteve Capper unsigned long addr, pgoff_t idx) 53083212b535SSteve Capper { 53093212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 53103212b535SSteve Capper svma->vm_start; 53113212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 53123212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 53133212b535SSteve Capper 53143212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5315de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5316de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 53173212b535SSteve Capper 53183212b535SSteve Capper /* 53193212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 53203212b535SSteve Capper * page table page. 53213212b535SSteve Capper */ 53223212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 53233212b535SSteve Capper vm_flags != svm_flags || 532407e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 53253212b535SSteve Capper return 0; 53263212b535SSteve Capper 53273212b535SSteve Capper return saddr; 53283212b535SSteve Capper } 53293212b535SSteve Capper 533031aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 53313212b535SSteve Capper { 53323212b535SSteve Capper unsigned long base = addr & PUD_MASK; 53333212b535SSteve Capper unsigned long end = base + PUD_SIZE; 53343212b535SSteve Capper 53353212b535SSteve Capper /* 53363212b535SSteve Capper * check on proper vm_flags and page table alignment 53373212b535SSteve Capper */ 5338017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 533931aafb45SNicholas Krause return true; 534031aafb45SNicholas Krause return false; 53413212b535SSteve Capper } 53423212b535SSteve Capper 5343c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5344c1991e07SPeter Xu { 5345c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5346c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5347c1991e07SPeter Xu return false; 5348c1991e07SPeter Xu #endif 5349c1991e07SPeter Xu return vma_shareable(vma, addr); 5350c1991e07SPeter Xu } 5351c1991e07SPeter Xu 53523212b535SSteve Capper /* 5353017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5354017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5355017b1660SMike Kravetz * shared pmd mappings. 5356017b1660SMike Kravetz */ 5357017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5358017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5359017b1660SMike Kravetz { 5360a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5361a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5362017b1660SMike Kravetz 5363a1ba9da8SLi Xinhai /* 5364a1ba9da8SLi Xinhai * vma need span at least one aligned PUD size and the start,end range 5365a1ba9da8SLi Xinhai * must at least partialy within it. 5366a1ba9da8SLi Xinhai */ 5367a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5368a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5369017b1660SMike Kravetz return; 5370017b1660SMike Kravetz 537175802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5372a1ba9da8SLi Xinhai if (*start > v_start) 5373a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5374017b1660SMike Kravetz 5375a1ba9da8SLi Xinhai if (*end < v_end) 5376a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5377017b1660SMike Kravetz } 5378017b1660SMike Kravetz 5379017b1660SMike Kravetz /* 53803212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53813212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53823212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5383c0d0381aSMike Kravetz * code much cleaner. 5384c0d0381aSMike Kravetz * 53850bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 53860bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 53870bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 53880bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 53890bf7b64eSMike Kravetz * 53900bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 53910bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 53920bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 53930bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 53940bf7b64eSMike Kravetz * only required for subsequent processing. 53953212b535SSteve Capper */ 5396aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5397aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 53983212b535SSteve Capper { 53993212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 54003212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 54013212b535SSteve Capper vma->vm_pgoff; 54023212b535SSteve Capper struct vm_area_struct *svma; 54033212b535SSteve Capper unsigned long saddr; 54043212b535SSteve Capper pte_t *spte = NULL; 54053212b535SSteve Capper pte_t *pte; 5406cb900f41SKirill A. Shutemov spinlock_t *ptl; 54073212b535SSteve Capper 54080bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 54093212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 54103212b535SSteve Capper if (svma == vma) 54113212b535SSteve Capper continue; 54123212b535SSteve Capper 54133212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 54143212b535SSteve Capper if (saddr) { 54157868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 54167868a208SPunit Agrawal vma_mmu_pagesize(svma)); 54173212b535SSteve Capper if (spte) { 54183212b535SSteve Capper get_page(virt_to_page(spte)); 54193212b535SSteve Capper break; 54203212b535SSteve Capper } 54213212b535SSteve Capper } 54223212b535SSteve Capper } 54233212b535SSteve Capper 54243212b535SSteve Capper if (!spte) 54253212b535SSteve Capper goto out; 54263212b535SSteve Capper 54278bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5428dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 54293212b535SSteve Capper pud_populate(mm, pud, 54303212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5431c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5432dc6c9a35SKirill A. Shutemov } else { 54333212b535SSteve Capper put_page(virt_to_page(spte)); 5434dc6c9a35SKirill A. Shutemov } 5435cb900f41SKirill A. Shutemov spin_unlock(ptl); 54363212b535SSteve Capper out: 54373212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54383212b535SSteve Capper return pte; 54393212b535SSteve Capper } 54403212b535SSteve Capper 54413212b535SSteve Capper /* 54423212b535SSteve Capper * unmap huge page backed by shared pte. 54433212b535SSteve Capper * 54443212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 54453212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 54463212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 54473212b535SSteve Capper * 5448c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 54493212b535SSteve Capper * 54503212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 54513212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 54523212b535SSteve Capper */ 545334ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 545434ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 54553212b535SSteve Capper { 54563212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5457c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5458c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 54593212b535SSteve Capper 546034ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 54613212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 54623212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 54633212b535SSteve Capper return 0; 54643212b535SSteve Capper 54653212b535SSteve Capper pud_clear(pud); 54663212b535SSteve Capper put_page(virt_to_page(ptep)); 5467dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54683212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54693212b535SSteve Capper return 1; 54703212b535SSteve Capper } 5471c1991e07SPeter Xu 54729e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5473aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5474aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 54759e5fc74cSSteve Capper { 54769e5fc74cSSteve Capper return NULL; 54779e5fc74cSSteve Capper } 5478e81f2d22SZhang Zhen 547934ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 548034ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5481e81f2d22SZhang Zhen { 5482e81f2d22SZhang Zhen return 0; 5483e81f2d22SZhang Zhen } 5484017b1660SMike Kravetz 5485017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5486017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5487017b1660SMike Kravetz { 5488017b1660SMike Kravetz } 5489c1991e07SPeter Xu 5490c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5491c1991e07SPeter Xu { 5492c1991e07SPeter Xu return false; 5493c1991e07SPeter Xu } 54943212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54953212b535SSteve Capper 54969e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5497aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 54989e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54999e5fc74cSSteve Capper { 55009e5fc74cSSteve Capper pgd_t *pgd; 5501c2febafcSKirill A. Shutemov p4d_t *p4d; 55029e5fc74cSSteve Capper pud_t *pud; 55039e5fc74cSSteve Capper pte_t *pte = NULL; 55049e5fc74cSSteve Capper 55059e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5506f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5507f4f0a3d8SKirill A. Shutemov if (!p4d) 5508f4f0a3d8SKirill A. Shutemov return NULL; 5509c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 55109e5fc74cSSteve Capper if (pud) { 55119e5fc74cSSteve Capper if (sz == PUD_SIZE) { 55129e5fc74cSSteve Capper pte = (pte_t *)pud; 55139e5fc74cSSteve Capper } else { 55149e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5515c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5516aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 55179e5fc74cSSteve Capper else 55189e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 55199e5fc74cSSteve Capper } 55209e5fc74cSSteve Capper } 55214e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 55229e5fc74cSSteve Capper 55239e5fc74cSSteve Capper return pte; 55249e5fc74cSSteve Capper } 55259e5fc74cSSteve Capper 55269b19df29SPunit Agrawal /* 55279b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 55289b19df29SPunit Agrawal * entry at address @addr 55299b19df29SPunit Agrawal * 55308ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 55318ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 55329b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 55339b19df29SPunit Agrawal * table. 55349b19df29SPunit Agrawal */ 55357868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 55367868a208SPunit Agrawal unsigned long addr, unsigned long sz) 55379e5fc74cSSteve Capper { 55389e5fc74cSSteve Capper pgd_t *pgd; 5539c2febafcSKirill A. Shutemov p4d_t *p4d; 55408ac0b81aSLi Xinhai pud_t *pud; 55418ac0b81aSLi Xinhai pmd_t *pmd; 55429e5fc74cSSteve Capper 55439e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5544c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5545c2febafcSKirill A. Shutemov return NULL; 5546c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5547c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5548c2febafcSKirill A. Shutemov return NULL; 55499b19df29SPunit Agrawal 5550c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 55518ac0b81aSLi Xinhai if (sz == PUD_SIZE) 55528ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 55539e5fc74cSSteve Capper return (pte_t *)pud; 55548ac0b81aSLi Xinhai if (!pud_present(*pud)) 55558ac0b81aSLi Xinhai return NULL; 55568ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 55579b19df29SPunit Agrawal 55589e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 55598ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 55609e5fc74cSSteve Capper return (pte_t *)pmd; 55619e5fc74cSSteve Capper } 55629e5fc74cSSteve Capper 556361f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 556461f77edaSNaoya Horiguchi 556561f77edaSNaoya Horiguchi /* 556661f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 556761f77edaSNaoya Horiguchi * behavior. 556861f77edaSNaoya Horiguchi */ 556961f77edaSNaoya Horiguchi struct page * __weak 557061f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 557161f77edaSNaoya Horiguchi int write) 557261f77edaSNaoya Horiguchi { 557361f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 557461f77edaSNaoya Horiguchi } 557561f77edaSNaoya Horiguchi 557661f77edaSNaoya Horiguchi struct page * __weak 55774dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55784dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55794dc71451SAneesh Kumar K.V { 55804dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55814dc71451SAneesh Kumar K.V return NULL; 55824dc71451SAneesh Kumar K.V } 55834dc71451SAneesh Kumar K.V 55844dc71451SAneesh Kumar K.V struct page * __weak 55859e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5586e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55879e5fc74cSSteve Capper { 5588e66f17ffSNaoya Horiguchi struct page *page = NULL; 5589e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5590c9d398faSNaoya Horiguchi pte_t pte; 55913faa52c0SJohn Hubbard 55923faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55933faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55943faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55953faa52c0SJohn Hubbard return NULL; 55963faa52c0SJohn Hubbard 5597e66f17ffSNaoya Horiguchi retry: 5598e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5599e66f17ffSNaoya Horiguchi spin_lock(ptl); 5600e66f17ffSNaoya Horiguchi /* 5601e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5602e66f17ffSNaoya Horiguchi * unmapped from other threads. 5603e66f17ffSNaoya Horiguchi */ 5604e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5605e66f17ffSNaoya Horiguchi goto out; 5606c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5607c9d398faSNaoya Horiguchi if (pte_present(pte)) { 560897534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 56093faa52c0SJohn Hubbard /* 56103faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 56113faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 56123faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 56133faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 56143faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 56153faa52c0SJohn Hubbard * unless the page refcount overflowed: 56163faa52c0SJohn Hubbard */ 56173faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 56183faa52c0SJohn Hubbard page = NULL; 56193faa52c0SJohn Hubbard goto out; 56203faa52c0SJohn Hubbard } 5621e66f17ffSNaoya Horiguchi } else { 5622c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5623e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5624e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5625e66f17ffSNaoya Horiguchi goto retry; 5626e66f17ffSNaoya Horiguchi } 5627e66f17ffSNaoya Horiguchi /* 5628e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5629e66f17ffSNaoya Horiguchi * follow_page_mask(). 5630e66f17ffSNaoya Horiguchi */ 5631e66f17ffSNaoya Horiguchi } 5632e66f17ffSNaoya Horiguchi out: 5633e66f17ffSNaoya Horiguchi spin_unlock(ptl); 56349e5fc74cSSteve Capper return page; 56359e5fc74cSSteve Capper } 56369e5fc74cSSteve Capper 563761f77edaSNaoya Horiguchi struct page * __weak 56389e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5639e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 56409e5fc74cSSteve Capper { 56413faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5642e66f17ffSNaoya Horiguchi return NULL; 56439e5fc74cSSteve Capper 5644e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 56459e5fc74cSSteve Capper } 56469e5fc74cSSteve Capper 5647faaa5b62SAnshuman Khandual struct page * __weak 5648faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5649faaa5b62SAnshuman Khandual { 56503faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5651faaa5b62SAnshuman Khandual return NULL; 5652faaa5b62SAnshuman Khandual 5653faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5654faaa5b62SAnshuman Khandual } 5655faaa5b62SAnshuman Khandual 565631caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 565731caf665SNaoya Horiguchi { 5658bcc54222SNaoya Horiguchi bool ret = true; 5659bcc54222SNaoya Horiguchi 566031caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 56618f251a3dSMike Kravetz if (!PageHeadHuge(page) || 56628f251a3dSMike Kravetz !HPageMigratable(page) || 56630eb2df2bSMuchun Song !get_page_unless_zero(page)) { 5664bcc54222SNaoya Horiguchi ret = false; 5665bcc54222SNaoya Horiguchi goto unlock; 5666bcc54222SNaoya Horiguchi } 56678f251a3dSMike Kravetz ClearHPageMigratable(page); 566831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5669bcc54222SNaoya Horiguchi unlock: 567031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5671bcc54222SNaoya Horiguchi return ret; 567231caf665SNaoya Horiguchi } 567331caf665SNaoya Horiguchi 567431caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 567531caf665SNaoya Horiguchi { 567631caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 56778f251a3dSMike Kravetz SetHPageMigratable(page); 567831caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 567931caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 568031caf665SNaoya Horiguchi put_page(page); 568131caf665SNaoya Horiguchi } 5682ab5ac90aSMichal Hocko 5683ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5684ab5ac90aSMichal Hocko { 5685ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5686ab5ac90aSMichal Hocko 5687ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5688ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5689ab5ac90aSMichal Hocko 5690ab5ac90aSMichal Hocko /* 5691ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5692ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5693ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5694ab5ac90aSMichal Hocko * the temporary status. 5695ab5ac90aSMichal Hocko * 5696ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5697ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5698ab5ac90aSMichal Hocko * the per-node's. 5699ab5ac90aSMichal Hocko */ 57009157c311SMike Kravetz if (HPageTemporary(newpage)) { 5701ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5702ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5703ab5ac90aSMichal Hocko 57049157c311SMike Kravetz SetHPageTemporary(oldpage); 57059157c311SMike Kravetz ClearHPageTemporary(newpage); 5706ab5ac90aSMichal Hocko 57075af1ab1dSMiaohe Lin /* 57085af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 57095af1ab1dSMiaohe Lin * when we do not cross the node. 57105af1ab1dSMiaohe Lin */ 57115af1ab1dSMiaohe Lin if (new_nid == old_nid) 57125af1ab1dSMiaohe Lin return; 5713ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5714ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5715ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5716ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5717ab5ac90aSMichal Hocko } 5718ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5719ab5ac90aSMichal Hocko } 5720ab5ac90aSMichal Hocko } 5721cf11e85fSRoman Gushchin 57226dfeaff9SPeter Xu /* 57236dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 57246dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 57256dfeaff9SPeter Xu */ 57266dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 57276dfeaff9SPeter Xu { 57286dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 57296dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 57306dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 57316dfeaff9SPeter Xu struct mmu_notifier_range range; 57326dfeaff9SPeter Xu unsigned long address, start, end; 57336dfeaff9SPeter Xu spinlock_t *ptl; 57346dfeaff9SPeter Xu pte_t *ptep; 57356dfeaff9SPeter Xu 57366dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 57376dfeaff9SPeter Xu return; 57386dfeaff9SPeter Xu 57396dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 57406dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 57416dfeaff9SPeter Xu 57426dfeaff9SPeter Xu if (start >= end) 57436dfeaff9SPeter Xu return; 57446dfeaff9SPeter Xu 57456dfeaff9SPeter Xu /* 57466dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 57476dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 57486dfeaff9SPeter Xu */ 57496dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 57506dfeaff9SPeter Xu start, end); 57516dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 57526dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 57536dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 57546dfeaff9SPeter Xu unsigned long tmp = address; 57556dfeaff9SPeter Xu 57566dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 57576dfeaff9SPeter Xu if (!ptep) 57586dfeaff9SPeter Xu continue; 57596dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 57606dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 57616dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 57626dfeaff9SPeter Xu spin_unlock(ptl); 57636dfeaff9SPeter Xu } 57646dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 57656dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 57666dfeaff9SPeter Xu /* 57676dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 57686dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 57696dfeaff9SPeter Xu */ 57706dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 57716dfeaff9SPeter Xu } 57726dfeaff9SPeter Xu 5773cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5774cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5775cf11e85fSRoman Gushchin 5776cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5777cf11e85fSRoman Gushchin { 5778cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5779cf11e85fSRoman Gushchin return 0; 5780cf11e85fSRoman Gushchin } 5781cf11e85fSRoman Gushchin 5782cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5783cf11e85fSRoman Gushchin 5784cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5785cf11e85fSRoman Gushchin { 5786cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5787cf11e85fSRoman Gushchin int nid; 5788cf11e85fSRoman Gushchin 5789cf11e85fSRoman Gushchin cma_reserve_called = true; 5790cf11e85fSRoman Gushchin 5791cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5792cf11e85fSRoman Gushchin return; 5793cf11e85fSRoman Gushchin 5794cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5795cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5796cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5797cf11e85fSRoman Gushchin return; 5798cf11e85fSRoman Gushchin } 5799cf11e85fSRoman Gushchin 5800cf11e85fSRoman Gushchin /* 5801cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5802cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5803cf11e85fSRoman Gushchin */ 5804cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5805cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5806cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5807cf11e85fSRoman Gushchin 5808cf11e85fSRoman Gushchin reserved = 0; 5809cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5810cf11e85fSRoman Gushchin int res; 58112281f797SBarry Song char name[CMA_MAX_NAME]; 5812cf11e85fSRoman Gushchin 5813cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5814cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5815cf11e85fSRoman Gushchin 58162281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5817cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 581829d0f41dSBarry Song 0, false, name, 5819cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5820cf11e85fSRoman Gushchin if (res) { 5821cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5822cf11e85fSRoman Gushchin res, nid); 5823cf11e85fSRoman Gushchin continue; 5824cf11e85fSRoman Gushchin } 5825cf11e85fSRoman Gushchin 5826cf11e85fSRoman Gushchin reserved += size; 5827cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5828cf11e85fSRoman Gushchin size / SZ_1M, nid); 5829cf11e85fSRoman Gushchin 5830cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5831cf11e85fSRoman Gushchin break; 5832cf11e85fSRoman Gushchin } 5833cf11e85fSRoman Gushchin } 5834cf11e85fSRoman Gushchin 5835cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5836cf11e85fSRoman Gushchin { 5837cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5838cf11e85fSRoman Gushchin return; 5839cf11e85fSRoman Gushchin 5840cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5841cf11e85fSRoman Gushchin } 5842cf11e85fSRoman Gushchin 5843cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5844