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 VM_BUG_ON(add < 0); 557cf3ad20bSMike Kravetz return add; 55896822904SAndy Whitcroft } 55996822904SAndy Whitcroft 5601dd308a7SMike Kravetz /* 5611dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5621dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5631dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5641dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5651dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5661dd308a7SMike Kravetz * not change the number of huge pages represented by the 5670db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5680db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5690db9d74eSMina Almasry * file_region structure is added. 5700db9d74eSMina Almasry * 5710db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5720db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5730db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5745e911373SMike Kravetz * 5755e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5765e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5775e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5785e911373SMike Kravetz * is needed and can not be allocated. 5791dd308a7SMike Kravetz */ 5800db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5810db9d74eSMina Almasry long *out_regions_needed) 58296822904SAndy Whitcroft { 58396822904SAndy Whitcroft long chg = 0; 58496822904SAndy Whitcroft 5857b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5865e911373SMike Kravetz 587972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 588075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 589972a3da3SWei Yang out_regions_needed); 5905e911373SMike Kravetz 5910db9d74eSMina Almasry if (*out_regions_needed == 0) 5920db9d74eSMina Almasry *out_regions_needed = 1; 5935e911373SMike Kravetz 5940db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5955e911373SMike Kravetz return -ENOMEM; 5965e911373SMike Kravetz 5970db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 59896822904SAndy Whitcroft 5997b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 60096822904SAndy Whitcroft return chg; 60196822904SAndy Whitcroft } 60296822904SAndy Whitcroft 6031dd308a7SMike Kravetz /* 6045e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6055e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6065e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6075e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6080db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6090db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6100db9d74eSMina Almasry * the adds_in_progress counter. 6115e911373SMike Kravetz * 6125e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6135e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6145e911373SMike Kravetz * arguments will match the associated region_chg call. 6155e911373SMike Kravetz */ 6160db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6170db9d74eSMina Almasry long regions_needed) 6185e911373SMike Kravetz { 6195e911373SMike Kravetz spin_lock(&resv->lock); 6205e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6210db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6225e911373SMike Kravetz spin_unlock(&resv->lock); 6235e911373SMike Kravetz } 6245e911373SMike Kravetz 6255e911373SMike Kravetz /* 626feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 627feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 628feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 629feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 630feba16e2SMike Kravetz * 631feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 632feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 633feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 634feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 635feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 636feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 637feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6381dd308a7SMike Kravetz */ 639feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 64096822904SAndy Whitcroft { 6411406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 64296822904SAndy Whitcroft struct file_region *rg, *trg; 643feba16e2SMike Kravetz struct file_region *nrg = NULL; 644feba16e2SMike Kravetz long del = 0; 64596822904SAndy Whitcroft 646feba16e2SMike Kravetz retry: 6477b24d861SDavidlohr Bueso spin_lock(&resv->lock); 648feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 649dbe409e4SMike Kravetz /* 650dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 651dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 652dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 653dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 654dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 655dbe409e4SMike Kravetz */ 656dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 657feba16e2SMike Kravetz continue; 658dbe409e4SMike Kravetz 659feba16e2SMike Kravetz if (rg->from >= t) 66096822904SAndy Whitcroft break; 66196822904SAndy Whitcroft 662feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 663feba16e2SMike Kravetz /* 664feba16e2SMike Kravetz * Check for an entry in the cache before dropping 665feba16e2SMike Kravetz * lock and attempting allocation. 666feba16e2SMike Kravetz */ 667feba16e2SMike Kravetz if (!nrg && 668feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 669feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 670feba16e2SMike Kravetz struct file_region, 671feba16e2SMike Kravetz link); 672feba16e2SMike Kravetz list_del(&nrg->link); 673feba16e2SMike Kravetz resv->region_cache_count--; 67496822904SAndy Whitcroft } 67596822904SAndy Whitcroft 676feba16e2SMike Kravetz if (!nrg) { 677feba16e2SMike Kravetz spin_unlock(&resv->lock); 678feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 679feba16e2SMike Kravetz if (!nrg) 680feba16e2SMike Kravetz return -ENOMEM; 681feba16e2SMike Kravetz goto retry; 682feba16e2SMike Kravetz } 683feba16e2SMike Kravetz 684feba16e2SMike Kravetz del += t - f; 68579aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 686d85aecf2SMiaohe Lin resv, rg, t - f, false); 687feba16e2SMike Kravetz 688feba16e2SMike Kravetz /* New entry for end of split region */ 689feba16e2SMike Kravetz nrg->from = t; 690feba16e2SMike Kravetz nrg->to = rg->to; 691075a61d0SMina Almasry 692075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 693075a61d0SMina Almasry 694feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 695feba16e2SMike Kravetz 696feba16e2SMike Kravetz /* Original entry is trimmed */ 697feba16e2SMike Kravetz rg->to = f; 698feba16e2SMike Kravetz 699feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 700feba16e2SMike Kravetz nrg = NULL; 70196822904SAndy Whitcroft break; 702feba16e2SMike Kravetz } 703feba16e2SMike Kravetz 704feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 705feba16e2SMike Kravetz del += rg->to - rg->from; 706075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 707d85aecf2SMiaohe Lin rg->to - rg->from, true); 70896822904SAndy Whitcroft list_del(&rg->link); 70996822904SAndy Whitcroft kfree(rg); 710feba16e2SMike Kravetz continue; 71196822904SAndy Whitcroft } 7127b24d861SDavidlohr Bueso 713feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 714075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 715d85aecf2SMiaohe Lin t - rg->from, false); 716075a61d0SMina Almasry 71779aa925bSMike Kravetz del += t - rg->from; 71879aa925bSMike Kravetz rg->from = t; 71979aa925bSMike Kravetz } else { /* Trim end of region */ 720075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 721d85aecf2SMiaohe Lin rg->to - f, false); 72279aa925bSMike Kravetz 72379aa925bSMike Kravetz del += rg->to - f; 72479aa925bSMike Kravetz rg->to = f; 725feba16e2SMike Kravetz } 726feba16e2SMike Kravetz } 727feba16e2SMike Kravetz 7287b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 729feba16e2SMike Kravetz kfree(nrg); 730feba16e2SMike Kravetz return del; 73196822904SAndy Whitcroft } 73296822904SAndy Whitcroft 7331dd308a7SMike Kravetz /* 734b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 735b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 736b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 737b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 738b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 739b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 740b5cec28dSMike Kravetz * counts. 741b5cec28dSMike Kravetz */ 74272e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 743b5cec28dSMike Kravetz { 744b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 745b5cec28dSMike Kravetz long rsv_adjust; 746b5cec28dSMike Kravetz 747b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 74872e2936cSzhong jiang if (rsv_adjust) { 749b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 750b5cec28dSMike Kravetz 751b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 752b5cec28dSMike Kravetz } 753b5cec28dSMike Kravetz } 754b5cec28dSMike Kravetz 755b5cec28dSMike Kravetz /* 7561dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7571dd308a7SMike Kravetz * that intersect with the range [f, t). 7581dd308a7SMike Kravetz */ 7591406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 76084afd99bSAndy Whitcroft { 7611406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 76284afd99bSAndy Whitcroft struct file_region *rg; 76384afd99bSAndy Whitcroft long chg = 0; 76484afd99bSAndy Whitcroft 7657b24d861SDavidlohr Bueso spin_lock(&resv->lock); 76684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 76784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 768f2135a4aSWang Sheng-Hui long seg_from; 769f2135a4aSWang Sheng-Hui long seg_to; 77084afd99bSAndy Whitcroft 77184afd99bSAndy Whitcroft if (rg->to <= f) 77284afd99bSAndy Whitcroft continue; 77384afd99bSAndy Whitcroft if (rg->from >= t) 77484afd99bSAndy Whitcroft break; 77584afd99bSAndy Whitcroft 77684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 77784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 77884afd99bSAndy Whitcroft 77984afd99bSAndy Whitcroft chg += seg_to - seg_from; 78084afd99bSAndy Whitcroft } 7817b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 78284afd99bSAndy Whitcroft 78384afd99bSAndy Whitcroft return chg; 78484afd99bSAndy Whitcroft } 78584afd99bSAndy Whitcroft 78696822904SAndy Whitcroft /* 787e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 788e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 789e7c4b0bfSAndy Whitcroft */ 790a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 791a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 792e7c4b0bfSAndy Whitcroft { 793a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 794a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 795e7c4b0bfSAndy Whitcroft } 796e7c4b0bfSAndy Whitcroft 7970fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 7980fe6e20bSNaoya Horiguchi unsigned long address) 7990fe6e20bSNaoya Horiguchi { 8000fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8010fe6e20bSNaoya Horiguchi } 802dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8030fe6e20bSNaoya Horiguchi 80484afd99bSAndy Whitcroft /* 80508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 80608fba699SMel Gorman * cases this will be same size as used by the page table entries. 80708fba699SMel Gorman */ 80808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 80908fba699SMel Gorman { 81005ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 81105ea8860SDan Williams return vma->vm_ops->pagesize(vma); 81208fba699SMel Gorman return PAGE_SIZE; 81308fba699SMel Gorman } 814f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 81508fba699SMel Gorman 81608fba699SMel Gorman /* 8173340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8183340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 81909135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 82009135cc5SDan Williams * version of this symbol is required. 8213340289dSMel Gorman */ 82209135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8233340289dSMel Gorman { 8243340289dSMel Gorman return vma_kernel_pagesize(vma); 8253340289dSMel Gorman } 8263340289dSMel Gorman 8273340289dSMel Gorman /* 82884afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 82984afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 83084afd99bSAndy Whitcroft * alignment. 83184afd99bSAndy Whitcroft */ 83284afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 83384afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 83404f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 83584afd99bSAndy Whitcroft 836a1e78772SMel Gorman /* 837a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 838a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 839a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 840a1e78772SMel Gorman * 841a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 842a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 843a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 844a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 84584afd99bSAndy Whitcroft * 84684afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 84784afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 84884afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 84984afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 85084afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 85184afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 85284afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 85384afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 854a1e78772SMel Gorman */ 855e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 856e7c4b0bfSAndy Whitcroft { 857e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 858e7c4b0bfSAndy Whitcroft } 859e7c4b0bfSAndy Whitcroft 860e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 861e7c4b0bfSAndy Whitcroft unsigned long value) 862e7c4b0bfSAndy Whitcroft { 863e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 864e7c4b0bfSAndy Whitcroft } 865e7c4b0bfSAndy Whitcroft 866e9fe92aeSMina Almasry static void 867e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 868e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 869e9fe92aeSMina Almasry struct hstate *h) 870e9fe92aeSMina Almasry { 871e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 872e9fe92aeSMina Almasry if (!h_cg || !h) { 873e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 874e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 875e9fe92aeSMina Almasry resv_map->css = NULL; 876e9fe92aeSMina Almasry } else { 877e9fe92aeSMina Almasry resv_map->reservation_counter = 878e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 879e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 880e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 881e9fe92aeSMina Almasry } 882e9fe92aeSMina Almasry #endif 883e9fe92aeSMina Almasry } 884e9fe92aeSMina Almasry 8859119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 88684afd99bSAndy Whitcroft { 88784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8885e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8895e911373SMike Kravetz 8905e911373SMike Kravetz if (!resv_map || !rg) { 8915e911373SMike Kravetz kfree(resv_map); 8925e911373SMike Kravetz kfree(rg); 89384afd99bSAndy Whitcroft return NULL; 8945e911373SMike Kravetz } 89584afd99bSAndy Whitcroft 89684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 8977b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 89884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 89984afd99bSAndy Whitcroft 9005e911373SMike Kravetz resv_map->adds_in_progress = 0; 901e9fe92aeSMina Almasry /* 902e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 903e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 904e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 905e9fe92aeSMina Almasry * reservations are to be un-charged from here. 906e9fe92aeSMina Almasry */ 907e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9085e911373SMike Kravetz 9095e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9105e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9115e911373SMike Kravetz resv_map->region_cache_count = 1; 9125e911373SMike Kravetz 91384afd99bSAndy Whitcroft return resv_map; 91484afd99bSAndy Whitcroft } 91584afd99bSAndy Whitcroft 9169119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 91784afd99bSAndy Whitcroft { 91884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9195e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9205e911373SMike Kravetz struct file_region *rg, *trg; 92184afd99bSAndy Whitcroft 92284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 923feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9245e911373SMike Kravetz 9255e911373SMike Kravetz /* ... and any entries left in the cache */ 9265e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9275e911373SMike Kravetz list_del(&rg->link); 9285e911373SMike Kravetz kfree(rg); 9295e911373SMike Kravetz } 9305e911373SMike Kravetz 9315e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9325e911373SMike Kravetz 93384afd99bSAndy Whitcroft kfree(resv_map); 93484afd99bSAndy Whitcroft } 93584afd99bSAndy Whitcroft 9364e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9374e35f483SJoonsoo Kim { 938f27a5136SMike Kravetz /* 939f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 940f27a5136SMike Kravetz * address space within the inode. This original address space 941f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 942f27a5136SMike Kravetz * address space embedded within the inode. 943f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 944f27a5136SMike Kravetz * this may not be true for device special inodes. 945f27a5136SMike Kravetz */ 946f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9474e35f483SJoonsoo Kim } 9484e35f483SJoonsoo Kim 94984afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 950a1e78772SMel Gorman { 95181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9524e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9534e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9544e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9554e35f483SJoonsoo Kim 9564e35f483SJoonsoo Kim return inode_resv_map(inode); 9574e35f483SJoonsoo Kim 9584e35f483SJoonsoo Kim } else { 95984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 96084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9614e35f483SJoonsoo Kim } 962a1e78772SMel Gorman } 963a1e78772SMel Gorman 96484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 965a1e78772SMel Gorman { 96681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 96781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 968a1e78772SMel Gorman 96984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 97084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 97104f2cbe3SMel Gorman } 97204f2cbe3SMel Gorman 97304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 97404f2cbe3SMel Gorman { 97581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 97681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 977e7c4b0bfSAndy Whitcroft 978e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 97904f2cbe3SMel Gorman } 98004f2cbe3SMel Gorman 98104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 98204f2cbe3SMel Gorman { 98381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 984e7c4b0bfSAndy Whitcroft 985e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 986a1e78772SMel Gorman } 987a1e78772SMel Gorman 98804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 989a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 990a1e78772SMel Gorman { 99181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 992f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 993a1e78772SMel Gorman vma->vm_private_data = (void *)0; 994a1e78772SMel Gorman } 995a1e78772SMel Gorman 996a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 997559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 998a1e78772SMel Gorman { 999af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1000af0ed73eSJoonsoo Kim /* 1001af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1002af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1003af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1004af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1005af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1006af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1007af0ed73eSJoonsoo Kim * properly, so add work-around here. 1008af0ed73eSJoonsoo Kim */ 1009af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1010559ec2f8SNicholas Krause return true; 1011af0ed73eSJoonsoo Kim else 1012559ec2f8SNicholas Krause return false; 1013af0ed73eSJoonsoo Kim } 1014a63884e9SJoonsoo Kim 1015a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10161fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10171fb1b0e9SMike Kravetz /* 10181fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10191fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10201fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10217c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10221fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10231fb1b0e9SMike Kravetz */ 10241fb1b0e9SMike Kravetz if (chg) 10251fb1b0e9SMike Kravetz return false; 10261fb1b0e9SMike Kravetz else 1027559ec2f8SNicholas Krause return true; 10281fb1b0e9SMike Kravetz } 1029a63884e9SJoonsoo Kim 1030a63884e9SJoonsoo Kim /* 1031a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1032a63884e9SJoonsoo Kim * private mappings. 1033a63884e9SJoonsoo Kim */ 103467961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 103567961f9dSMike Kravetz /* 103667961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 103767961f9dSMike Kravetz * could have been performed on the private mapping. 103867961f9dSMike Kravetz * Examine the value of chg to determine if reserves 103967961f9dSMike Kravetz * actually exist or were previously consumed. 104067961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 104167961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 104267961f9dSMike Kravetz * private mappings has different (opposite) semantics 104367961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 104467961f9dSMike Kravetz * has already taken this difference in semantics into 104567961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 104667961f9dSMike Kravetz * as in the shared case above. Code could easily be 104767961f9dSMike Kravetz * combined, but keeping it separate draws attention to 104867961f9dSMike Kravetz * subtle differences. 104967961f9dSMike Kravetz */ 105067961f9dSMike Kravetz if (chg) 105167961f9dSMike Kravetz return false; 105267961f9dSMike Kravetz else 1053559ec2f8SNicholas Krause return true; 105467961f9dSMike Kravetz } 1055a63884e9SJoonsoo Kim 1056559ec2f8SNicholas Krause return false; 1057a1e78772SMel Gorman } 1058a1e78772SMel Gorman 1059a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10601da177e4SLinus Torvalds { 10611da177e4SLinus Torvalds int nid = page_to_nid(page); 10620edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1063a5516438SAndi Kleen h->free_huge_pages++; 1064a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10656c037149SMike Kravetz SetHPageFreed(page); 10661da177e4SLinus Torvalds } 10671da177e4SLinus Torvalds 106894310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1069bf50bab2SNaoya Horiguchi { 1070bf50bab2SNaoya Horiguchi struct page *page; 1071bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1072bf50bab2SNaoya Horiguchi 1073bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1074bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1075bbe88753SJoonsoo Kim continue; 1076bbe88753SJoonsoo Kim 10776664bfc8SWei Yang if (PageHWPoison(page)) 10786664bfc8SWei Yang continue; 1079bbe88753SJoonsoo Kim 10800edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1081a9869b83SNaoya Horiguchi set_page_refcounted(page); 10826c037149SMike Kravetz ClearHPageFreed(page); 1083bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1084bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1085bf50bab2SNaoya Horiguchi return page; 1086bf50bab2SNaoya Horiguchi } 1087bf50bab2SNaoya Horiguchi 10886664bfc8SWei Yang return NULL; 10896664bfc8SWei Yang } 10906664bfc8SWei Yang 10913e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10923e59fcb0SMichal Hocko nodemask_t *nmask) 109394310cbcSAnshuman Khandual { 10943e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 10953e59fcb0SMichal Hocko struct zonelist *zonelist; 10963e59fcb0SMichal Hocko struct zone *zone; 10973e59fcb0SMichal Hocko struct zoneref *z; 109898fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 10993e59fcb0SMichal Hocko 11003e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11013e59fcb0SMichal Hocko 11023e59fcb0SMichal Hocko retry_cpuset: 11033e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11043e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 110594310cbcSAnshuman Khandual struct page *page; 110694310cbcSAnshuman Khandual 11073e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11083e59fcb0SMichal Hocko continue; 11093e59fcb0SMichal Hocko /* 11103e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11113e59fcb0SMichal Hocko * zone aware 11123e59fcb0SMichal Hocko */ 11133e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11143e59fcb0SMichal Hocko continue; 11153e59fcb0SMichal Hocko node = zone_to_nid(zone); 111694310cbcSAnshuman Khandual 111794310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 111894310cbcSAnshuman Khandual if (page) 111994310cbcSAnshuman Khandual return page; 112094310cbcSAnshuman Khandual } 11213e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11223e59fcb0SMichal Hocko goto retry_cpuset; 11233e59fcb0SMichal Hocko 112494310cbcSAnshuman Khandual return NULL; 112594310cbcSAnshuman Khandual } 112694310cbcSAnshuman Khandual 1127a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1128a5516438SAndi Kleen struct vm_area_struct *vma, 1129af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1130af0ed73eSJoonsoo Kim long chg) 11311da177e4SLinus Torvalds { 11323e59fcb0SMichal Hocko struct page *page; 1133480eccf9SLee Schermerhorn struct mempolicy *mpol; 113404ec6264SVlastimil Babka gfp_t gfp_mask; 11353e59fcb0SMichal Hocko nodemask_t *nodemask; 113604ec6264SVlastimil Babka int nid; 11371da177e4SLinus Torvalds 1138a1e78772SMel Gorman /* 1139a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1140a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1141a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1142a1e78772SMel Gorman */ 1143af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1144a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1145c0ff7453SMiao Xie goto err; 1146a1e78772SMel Gorman 114704f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1148a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11496eab04a8SJustin P. Mattock goto err; 115004f2cbe3SMel Gorman 115104ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 115204ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11533e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11543e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1155d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1156a63884e9SJoonsoo Kim h->resv_huge_pages--; 1157bf50bab2SNaoya Horiguchi } 1158cc9a6c87SMel Gorman 1159cc9a6c87SMel Gorman mpol_cond_put(mpol); 1160cc9a6c87SMel Gorman return page; 1161cc9a6c87SMel Gorman 1162c0ff7453SMiao Xie err: 1163cc9a6c87SMel Gorman return NULL; 11641da177e4SLinus Torvalds } 11651da177e4SLinus Torvalds 11661cac6f2cSLuiz Capitulino /* 11671cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11681cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11691cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11701cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11711cac6f2cSLuiz Capitulino * node for alloc or free. 11721cac6f2cSLuiz Capitulino */ 11731cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11741cac6f2cSLuiz Capitulino { 11750edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11761cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11771cac6f2cSLuiz Capitulino 11781cac6f2cSLuiz Capitulino return nid; 11791cac6f2cSLuiz Capitulino } 11801cac6f2cSLuiz Capitulino 11811cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11821cac6f2cSLuiz Capitulino { 11831cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11841cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11851cac6f2cSLuiz Capitulino return nid; 11861cac6f2cSLuiz Capitulino } 11871cac6f2cSLuiz Capitulino 11881cac6f2cSLuiz Capitulino /* 11891cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11901cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11911cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11921cac6f2cSLuiz Capitulino * mask. 11931cac6f2cSLuiz Capitulino */ 11941cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 11951cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 11961cac6f2cSLuiz Capitulino { 11971cac6f2cSLuiz Capitulino int nid; 11981cac6f2cSLuiz Capitulino 11991cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12001cac6f2cSLuiz Capitulino 12011cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12021cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12031cac6f2cSLuiz Capitulino 12041cac6f2cSLuiz Capitulino return nid; 12051cac6f2cSLuiz Capitulino } 12061cac6f2cSLuiz Capitulino 12071cac6f2cSLuiz Capitulino /* 12081cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 12091cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12101cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12111cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12121cac6f2cSLuiz Capitulino */ 12131cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12141cac6f2cSLuiz Capitulino { 12151cac6f2cSLuiz Capitulino int nid; 12161cac6f2cSLuiz Capitulino 12171cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12181cac6f2cSLuiz Capitulino 12191cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12201cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12211cac6f2cSLuiz Capitulino 12221cac6f2cSLuiz Capitulino return nid; 12231cac6f2cSLuiz Capitulino } 12241cac6f2cSLuiz Capitulino 12251cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12261cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12271cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12281cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12291cac6f2cSLuiz Capitulino nr_nodes--) 12301cac6f2cSLuiz Capitulino 12311cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(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_free(hs, mask)) || 1); \ 12351cac6f2cSLuiz Capitulino nr_nodes--) 12361cac6f2cSLuiz Capitulino 1237e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1238944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1239d00181b9SKirill A. Shutemov unsigned int order) 1240944d9fecSLuiz Capitulino { 1241944d9fecSLuiz Capitulino int i; 1242944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1243944d9fecSLuiz Capitulino struct page *p = page + 1; 1244944d9fecSLuiz Capitulino 1245c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 124647e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 124747e29d32SJohn Hubbard 1248944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12491d798ca3SKirill A. Shutemov clear_compound_head(p); 1250944d9fecSLuiz Capitulino set_page_refcounted(p); 1251944d9fecSLuiz Capitulino } 1252944d9fecSLuiz Capitulino 1253944d9fecSLuiz Capitulino set_compound_order(page, 0); 1254ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1255944d9fecSLuiz Capitulino __ClearPageHead(page); 1256944d9fecSLuiz Capitulino } 1257944d9fecSLuiz Capitulino 1258d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1259944d9fecSLuiz Capitulino { 1260cf11e85fSRoman Gushchin /* 1261cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1262cf11e85fSRoman Gushchin * cma_release() returns false. 1263cf11e85fSRoman Gushchin */ 1264dbda8feaSBarry Song #ifdef CONFIG_CMA 1265dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1266cf11e85fSRoman Gushchin return; 1267dbda8feaSBarry Song #endif 1268cf11e85fSRoman Gushchin 1269944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1270944d9fecSLuiz Capitulino } 1271944d9fecSLuiz Capitulino 12724eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1273d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1274d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1275944d9fecSLuiz Capitulino { 12765e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1277953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1278953f064aSLi Xinhai nid = numa_mem_id(); 1279944d9fecSLuiz Capitulino 1280dbda8feaSBarry Song #ifdef CONFIG_CMA 1281dbda8feaSBarry Song { 1282cf11e85fSRoman Gushchin struct page *page; 1283cf11e85fSRoman Gushchin int node; 1284cf11e85fSRoman Gushchin 1285953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1286953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1287953f064aSLi Xinhai huge_page_order(h), true); 1288953f064aSLi Xinhai if (page) 1289953f064aSLi Xinhai return page; 1290953f064aSLi Xinhai } 1291953f064aSLi Xinhai 1292953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1293cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1294953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1295cf11e85fSRoman Gushchin continue; 1296cf11e85fSRoman Gushchin 1297cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1298cf11e85fSRoman Gushchin huge_page_order(h), true); 1299cf11e85fSRoman Gushchin if (page) 1300cf11e85fSRoman Gushchin return page; 1301cf11e85fSRoman Gushchin } 1302cf11e85fSRoman Gushchin } 1303953f064aSLi Xinhai } 1304dbda8feaSBarry Song #endif 1305cf11e85fSRoman Gushchin 13065e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1307944d9fecSLuiz Capitulino } 1308944d9fecSLuiz Capitulino 1309944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1310d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 13114eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13124eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13134eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13144eb0716eSAlexandre Ghiti { 13154eb0716eSAlexandre Ghiti return NULL; 13164eb0716eSAlexandre Ghiti } 13174eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1318944d9fecSLuiz Capitulino 1319e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1320d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13214eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13224eb0716eSAlexandre Ghiti { 13234eb0716eSAlexandre Ghiti return NULL; 13244eb0716eSAlexandre Ghiti } 1325d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1326944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1327d00181b9SKirill A. Shutemov unsigned int order) { } 1328944d9fecSLuiz Capitulino #endif 1329944d9fecSLuiz Capitulino 1330a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 13316af2acb6SAdam Litke { 13326af2acb6SAdam Litke int i; 1333dbfee5aeSMike Kravetz struct page *subpage = page; 1334a5516438SAndi Kleen 13354eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1336944d9fecSLuiz Capitulino return; 133718229df5SAndy Whitcroft 1338a5516438SAndi Kleen h->nr_huge_pages--; 1339a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1340dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1341dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1342dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 134332f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1344a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1345a7407a27SLuiz Capitulino 1 << PG_writeback); 13466af2acb6SAdam Litke } 1347309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13481adc4d41SMina Almasry VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1349f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13506af2acb6SAdam Litke set_page_refcounted(page); 1351944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1352cf11e85fSRoman Gushchin /* 1353cf11e85fSRoman Gushchin * Temporarily drop the hugetlb_lock, because 1354cf11e85fSRoman Gushchin * we might block in free_gigantic_page(). 1355cf11e85fSRoman Gushchin */ 1356cf11e85fSRoman Gushchin spin_unlock(&hugetlb_lock); 1357944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1358944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1359cf11e85fSRoman Gushchin spin_lock(&hugetlb_lock); 1360944d9fecSLuiz Capitulino } else { 1361a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13626af2acb6SAdam Litke } 1363944d9fecSLuiz Capitulino } 13646af2acb6SAdam Litke 1365e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1366e5ff2159SAndi Kleen { 1367e5ff2159SAndi Kleen struct hstate *h; 1368e5ff2159SAndi Kleen 1369e5ff2159SAndi Kleen for_each_hstate(h) { 1370e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1371e5ff2159SAndi Kleen return h; 1372e5ff2159SAndi Kleen } 1373e5ff2159SAndi Kleen return NULL; 1374e5ff2159SAndi Kleen } 1375e5ff2159SAndi Kleen 1376c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 137727a85ef1SDavid Gibson { 1378a5516438SAndi Kleen /* 1379a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1380a5516438SAndi Kleen * compound page destructor. 1381a5516438SAndi Kleen */ 1382e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 13837893d1d5SAdam Litke int nid = page_to_nid(page); 1384d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 138507443a85SJoonsoo Kim bool restore_reserve; 138627a85ef1SDavid Gibson 1387b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1388b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 13898ace22bcSYongkai Wu 1390d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 13918ace22bcSYongkai Wu page->mapping = NULL; 1392d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1393d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 139427a85ef1SDavid Gibson 13951c5ecae3SMike Kravetz /* 1396d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 13970919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 13980919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 13990919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14006c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 14010919e1b6SMike Kravetz * remove the reserved page from the subpool. 14020919e1b6SMike Kravetz */ 14030919e1b6SMike Kravetz if (!restore_reserve) { 14040919e1b6SMike Kravetz /* 14050919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14060919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14070919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14080919e1b6SMike Kravetz * operation. 14091c5ecae3SMike Kravetz */ 14101c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14111c5ecae3SMike Kravetz restore_reserve = true; 14120919e1b6SMike Kravetz } 14131c5ecae3SMike Kravetz 141427a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 14158f251a3dSMike Kravetz ClearHPageMigratable(page); 14166d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14176d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 141808cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 141908cf9fafSMina Almasry pages_per_huge_page(h), page); 142007443a85SJoonsoo Kim if (restore_reserve) 142107443a85SJoonsoo Kim h->resv_huge_pages++; 142207443a85SJoonsoo Kim 14239157c311SMike Kravetz if (HPageTemporary(page)) { 1424ab5ac90aSMichal Hocko list_del(&page->lru); 14259157c311SMike Kravetz ClearHPageTemporary(page); 1426ab5ac90aSMichal Hocko update_and_free_page(h, page); 1427ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14280edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14290edaecfaSAneesh Kumar K.V list_del(&page->lru); 1430a5516438SAndi Kleen update_and_free_page(h, page); 1431a5516438SAndi Kleen h->surplus_huge_pages--; 1432a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 14337893d1d5SAdam Litke } else { 14345d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1435a5516438SAndi Kleen enqueue_huge_page(h, page); 14367893d1d5SAdam Litke } 143727a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 143827a85ef1SDavid Gibson } 143927a85ef1SDavid Gibson 1440c77c0a8aSWaiman Long /* 1441c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1442c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1443c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1444c77c0a8aSWaiman Long * 1445c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1446c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1447c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1448c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1449c77c0a8aSWaiman Long */ 1450c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1451c77c0a8aSWaiman Long 1452c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1453c77c0a8aSWaiman Long { 1454c77c0a8aSWaiman Long struct llist_node *node; 1455c77c0a8aSWaiman Long struct page *page; 1456c77c0a8aSWaiman Long 1457c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1458c77c0a8aSWaiman Long 1459c77c0a8aSWaiman Long while (node) { 1460c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1461c77c0a8aSWaiman Long struct page, mapping); 1462c77c0a8aSWaiman Long node = node->next; 1463c77c0a8aSWaiman Long __free_huge_page(page); 1464c77c0a8aSWaiman Long } 1465c77c0a8aSWaiman Long } 1466c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1467c77c0a8aSWaiman Long 1468c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1469c77c0a8aSWaiman Long { 1470c77c0a8aSWaiman Long /* 1471c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1472c77c0a8aSWaiman Long */ 1473c77c0a8aSWaiman Long if (!in_task()) { 1474c77c0a8aSWaiman Long /* 1475c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1476c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1477c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1478c77c0a8aSWaiman Long */ 1479c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1480c77c0a8aSWaiman Long &hpage_freelist)) 1481c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1482c77c0a8aSWaiman Long return; 1483c77c0a8aSWaiman Long } 1484c77c0a8aSWaiman Long 1485c77c0a8aSWaiman Long __free_huge_page(page); 1486c77c0a8aSWaiman Long } 1487c77c0a8aSWaiman Long 1488a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1489b7ba30c6SAndi Kleen { 14900edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1491f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1492ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14939dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 14941adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 14952f37511cSWei Yang spin_lock(&hugetlb_lock); 1496a5516438SAndi Kleen h->nr_huge_pages++; 1497a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 14986c037149SMike Kravetz ClearHPageFreed(page); 1499b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1500b7ba30c6SAndi Kleen } 1501b7ba30c6SAndi Kleen 1502d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 150320a0307cSWu Fengguang { 150420a0307cSWu Fengguang int i; 150520a0307cSWu Fengguang int nr_pages = 1 << order; 150620a0307cSWu Fengguang struct page *p = page + 1; 150720a0307cSWu Fengguang 150820a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 150920a0307cSWu Fengguang set_compound_order(page, order); 1510ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1511de09d31dSKirill A. Shutemov __SetPageHead(page); 151220a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1513ef5a22beSAndrea Arcangeli /* 1514ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1515ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1516ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15177c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1518ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1519ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1520ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1521ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1522ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1523ef5a22beSAndrea Arcangeli * after get_user_pages(). 1524ef5a22beSAndrea Arcangeli */ 1525ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 152658a84aa9SYouquan Song set_page_count(p, 0); 15271d798ca3SKirill A. Shutemov set_compound_head(p, page); 152820a0307cSWu Fengguang } 1529b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 153047e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 153120a0307cSWu Fengguang } 153220a0307cSWu Fengguang 15337795912cSAndrew Morton /* 15347795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15357795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15367795912cSAndrew Morton * details. 15377795912cSAndrew Morton */ 153820a0307cSWu Fengguang int PageHuge(struct page *page) 153920a0307cSWu Fengguang { 154020a0307cSWu Fengguang if (!PageCompound(page)) 154120a0307cSWu Fengguang return 0; 154220a0307cSWu Fengguang 154320a0307cSWu Fengguang page = compound_head(page); 1544f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 154520a0307cSWu Fengguang } 154643131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 154743131e14SNaoya Horiguchi 154827c73ae7SAndrea Arcangeli /* 154927c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 155027c73ae7SAndrea Arcangeli * normal or transparent huge pages. 155127c73ae7SAndrea Arcangeli */ 155227c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 155327c73ae7SAndrea Arcangeli { 155427c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 155527c73ae7SAndrea Arcangeli return 0; 155627c73ae7SAndrea Arcangeli 1557d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 155827c73ae7SAndrea Arcangeli } 155927c73ae7SAndrea Arcangeli 1560c0d0381aSMike Kravetz /* 1561c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1562c0d0381aSMike Kravetz * 1563336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1564336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1565336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1566c0d0381aSMike Kravetz */ 1567c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1568c0d0381aSMike Kravetz { 1569336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1570c0d0381aSMike Kravetz 1571c0d0381aSMike Kravetz if (!mapping) 1572c0d0381aSMike Kravetz return mapping; 1573c0d0381aSMike Kravetz 1574c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1575c0d0381aSMike Kravetz return mapping; 1576c0d0381aSMike Kravetz 1577c0d0381aSMike Kravetz return NULL; 1578c0d0381aSMike Kravetz } 1579c0d0381aSMike Kravetz 158013d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 158113d60f4bSZhang Yi { 158213d60f4bSZhang Yi struct page *page_head = compound_head(page); 158313d60f4bSZhang Yi pgoff_t index = page_index(page_head); 158413d60f4bSZhang Yi unsigned long compound_idx; 158513d60f4bSZhang Yi 158613d60f4bSZhang Yi if (!PageHuge(page_head)) 158713d60f4bSZhang Yi return page_index(page); 158813d60f4bSZhang Yi 158913d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 159013d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 159113d60f4bSZhang Yi else 159213d60f4bSZhang Yi compound_idx = page - page_head; 159313d60f4bSZhang Yi 159413d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 159513d60f4bSZhang Yi } 159613d60f4bSZhang Yi 15970c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1598f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1599f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16001da177e4SLinus Torvalds { 1601af0fb9dfSMichal Hocko int order = huge_page_order(h); 16021da177e4SLinus Torvalds struct page *page; 1603f60858f9SMike Kravetz bool alloc_try_hard = true; 1604f96efd58SJoe Jin 1605f60858f9SMike Kravetz /* 1606f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1607f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1608f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1609f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1610f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1611f60858f9SMike Kravetz */ 1612f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1613f60858f9SMike Kravetz alloc_try_hard = false; 1614f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1615f60858f9SMike Kravetz if (alloc_try_hard) 1616f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1617af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1618af0fb9dfSMichal Hocko nid = numa_mem_id(); 161984172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1620af0fb9dfSMichal Hocko if (page) 1621af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1622af0fb9dfSMichal Hocko else 1623af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 162463b4613cSNishanth Aravamudan 1625f60858f9SMike Kravetz /* 1626f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1627f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1628f60858f9SMike Kravetz * normal 'try hard' allocations. 1629f60858f9SMike Kravetz */ 1630f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1631f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1632f60858f9SMike Kravetz 1633f60858f9SMike Kravetz /* 1634f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1635f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1636f60858f9SMike Kravetz * overall state change. 1637f60858f9SMike Kravetz */ 1638f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1639f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1640f60858f9SMike Kravetz 164163b4613cSNishanth Aravamudan return page; 164263b4613cSNishanth Aravamudan } 164363b4613cSNishanth Aravamudan 1644af0fb9dfSMichal Hocko /* 16450c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16460c397daeSMichal Hocko * should use this function to get new hugetlb pages 16470c397daeSMichal Hocko */ 16480c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1649f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1650f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16510c397daeSMichal Hocko { 16520c397daeSMichal Hocko struct page *page; 16530c397daeSMichal Hocko 16540c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16550c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16560c397daeSMichal Hocko else 16570c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1658f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16590c397daeSMichal Hocko if (!page) 16600c397daeSMichal Hocko return NULL; 16610c397daeSMichal Hocko 16620c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16630c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16640c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16650c397daeSMichal Hocko 16660c397daeSMichal Hocko return page; 16670c397daeSMichal Hocko } 16680c397daeSMichal Hocko 16690c397daeSMichal Hocko /* 1670af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1671af0fb9dfSMichal Hocko * manner. 1672af0fb9dfSMichal Hocko */ 1673f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1674f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1675b2261026SJoonsoo Kim { 1676b2261026SJoonsoo Kim struct page *page; 1677b2261026SJoonsoo Kim int nr_nodes, node; 1678af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1679b2261026SJoonsoo Kim 1680b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1681f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1682f60858f9SMike Kravetz node_alloc_noretry); 1683af0fb9dfSMichal Hocko if (page) 1684b2261026SJoonsoo Kim break; 1685b2261026SJoonsoo Kim } 1686b2261026SJoonsoo Kim 1687af0fb9dfSMichal Hocko if (!page) 1688af0fb9dfSMichal Hocko return 0; 1689b2261026SJoonsoo Kim 1690af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1691af0fb9dfSMichal Hocko 1692af0fb9dfSMichal Hocko return 1; 1693b2261026SJoonsoo Kim } 1694b2261026SJoonsoo Kim 1695e8c5c824SLee Schermerhorn /* 1696e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1697e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1698e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1699e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1700e8c5c824SLee Schermerhorn */ 17016ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17026ae11b27SLee Schermerhorn bool acct_surplus) 1703e8c5c824SLee Schermerhorn { 1704b2261026SJoonsoo Kim int nr_nodes, node; 1705e8c5c824SLee Schermerhorn int ret = 0; 1706e8c5c824SLee Schermerhorn 1707b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1708685f3457SLee Schermerhorn /* 1709685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1710685f3457SLee Schermerhorn * nodes with surplus pages. 1711685f3457SLee Schermerhorn */ 1712b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1713b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1714e8c5c824SLee Schermerhorn struct page *page = 1715b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1716e8c5c824SLee Schermerhorn struct page, lru); 1717e8c5c824SLee Schermerhorn list_del(&page->lru); 1718e8c5c824SLee Schermerhorn h->free_huge_pages--; 1719b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1720685f3457SLee Schermerhorn if (acct_surplus) { 1721685f3457SLee Schermerhorn h->surplus_huge_pages--; 1722b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1723685f3457SLee Schermerhorn } 1724e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1725e8c5c824SLee Schermerhorn ret = 1; 17269a76db09SLee Schermerhorn break; 1727e8c5c824SLee Schermerhorn } 1728b2261026SJoonsoo Kim } 1729e8c5c824SLee Schermerhorn 1730e8c5c824SLee Schermerhorn return ret; 1731e8c5c824SLee Schermerhorn } 1732e8c5c824SLee Schermerhorn 1733c8721bbbSNaoya Horiguchi /* 1734c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1735faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1736faf53defSNaoya Horiguchi * This function returns values like below: 1737faf53defSNaoya Horiguchi * 1738faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1739faf53defSNaoya Horiguchi * (allocated or reserved.) 1740faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1741faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1742c8721bbbSNaoya Horiguchi */ 1743c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1744c8721bbbSNaoya Horiguchi { 17456bc9b564SNaoya Horiguchi int rc = -EBUSY; 1746082d5b6bSGerald Schaefer 17477ffddd49SMuchun Song retry: 1748faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1749faf53defSNaoya Horiguchi if (!PageHuge(page)) 1750faf53defSNaoya Horiguchi return 0; 1751faf53defSNaoya Horiguchi 1752c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1753faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1754faf53defSNaoya Horiguchi rc = 0; 1755faf53defSNaoya Horiguchi goto out; 1756faf53defSNaoya Horiguchi } 1757faf53defSNaoya Horiguchi 1758faf53defSNaoya Horiguchi if (!page_count(page)) { 17592247bb33SGerald Schaefer struct page *head = compound_head(page); 17602247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17612247bb33SGerald Schaefer int nid = page_to_nid(head); 17626bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1763082d5b6bSGerald Schaefer goto out; 17647ffddd49SMuchun Song 17657ffddd49SMuchun Song /* 17667ffddd49SMuchun Song * We should make sure that the page is already on the free list 17677ffddd49SMuchun Song * when it is dissolved. 17687ffddd49SMuchun Song */ 17696c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 17707ffddd49SMuchun Song spin_unlock(&hugetlb_lock); 17717ffddd49SMuchun Song cond_resched(); 17727ffddd49SMuchun Song 17737ffddd49SMuchun Song /* 17747ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 17757ffddd49SMuchun Song * encounter this race. In fact, we have a chance 17767ffddd49SMuchun Song * to successfully dissolve the page if we do a 17777ffddd49SMuchun Song * retry. Because the race window is quite small. 17787ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 17797ffddd49SMuchun Song * for increasing the success rate of dissolving page. 17807ffddd49SMuchun Song */ 17817ffddd49SMuchun Song goto retry; 17827ffddd49SMuchun Song } 17837ffddd49SMuchun Song 1784c3114a84SAnshuman Khandual /* 1785c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1786c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1787c3114a84SAnshuman Khandual */ 1788c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1789c3114a84SAnshuman Khandual SetPageHWPoison(page); 1790c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1791c3114a84SAnshuman Khandual } 17922247bb33SGerald Schaefer list_del(&head->lru); 1793c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1794c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1795c1470b33Szhong jiang h->max_huge_pages--; 17962247bb33SGerald Schaefer update_and_free_page(h, head); 17976bc9b564SNaoya Horiguchi rc = 0; 1798c8721bbbSNaoya Horiguchi } 1799082d5b6bSGerald Schaefer out: 1800c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1801082d5b6bSGerald Schaefer return rc; 1802c8721bbbSNaoya Horiguchi } 1803c8721bbbSNaoya Horiguchi 1804c8721bbbSNaoya Horiguchi /* 1805c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1806c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18072247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18082247bb33SGerald Schaefer * part of it lies within the given range. 1809082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1810082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1811c8721bbbSNaoya Horiguchi */ 1812082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1813c8721bbbSNaoya Horiguchi { 1814c8721bbbSNaoya Horiguchi unsigned long pfn; 1815eb03aa00SGerald Schaefer struct page *page; 1816082d5b6bSGerald Schaefer int rc = 0; 1817c8721bbbSNaoya Horiguchi 1818d0177639SLi Zhong if (!hugepages_supported()) 1819082d5b6bSGerald Schaefer return rc; 1820d0177639SLi Zhong 1821eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1822eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1823eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1824eb03aa00SGerald Schaefer if (rc) 1825082d5b6bSGerald Schaefer break; 1826eb03aa00SGerald Schaefer } 1827082d5b6bSGerald Schaefer 1828082d5b6bSGerald Schaefer return rc; 1829c8721bbbSNaoya Horiguchi } 1830c8721bbbSNaoya Horiguchi 1831ab5ac90aSMichal Hocko /* 1832ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1833ab5ac90aSMichal Hocko */ 18340c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1835aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18367893d1d5SAdam Litke { 18379980d744SMichal Hocko struct page *page = NULL; 18387893d1d5SAdam Litke 1839bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1840aa888a74SAndi Kleen return NULL; 1841aa888a74SAndi Kleen 1842d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 18439980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18449980d744SMichal Hocko goto out_unlock; 1845d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1846d1c3fb1fSNishanth Aravamudan 1847f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18489980d744SMichal Hocko if (!page) 18490c397daeSMichal Hocko return NULL; 1850d1c3fb1fSNishanth Aravamudan 18517893d1d5SAdam Litke spin_lock(&hugetlb_lock); 18529980d744SMichal Hocko /* 18539980d744SMichal Hocko * We could have raced with the pool size change. 18549980d744SMichal Hocko * Double check that and simply deallocate the new page 18559980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18569980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18579980d744SMichal Hocko * codeflow 18589980d744SMichal Hocko */ 18599980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18609157c311SMike Kravetz SetHPageTemporary(page); 18612bf753e6SKai Shen spin_unlock(&hugetlb_lock); 18629980d744SMichal Hocko put_page(page); 18632bf753e6SKai Shen return NULL; 18649980d744SMichal Hocko } else { 18659980d744SMichal Hocko h->surplus_huge_pages++; 18664704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18677893d1d5SAdam Litke } 18689980d744SMichal Hocko 18699980d744SMichal Hocko out_unlock: 1870d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 18717893d1d5SAdam Litke 18727893d1d5SAdam Litke return page; 18737893d1d5SAdam Litke } 18747893d1d5SAdam Litke 1875bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1876ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1877ab5ac90aSMichal Hocko { 1878ab5ac90aSMichal Hocko struct page *page; 1879ab5ac90aSMichal Hocko 1880ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1881ab5ac90aSMichal Hocko return NULL; 1882ab5ac90aSMichal Hocko 1883f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1884ab5ac90aSMichal Hocko if (!page) 1885ab5ac90aSMichal Hocko return NULL; 1886ab5ac90aSMichal Hocko 1887ab5ac90aSMichal Hocko /* 1888ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1889ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1890ab5ac90aSMichal Hocko */ 18919157c311SMike Kravetz SetHPageTemporary(page); 1892ab5ac90aSMichal Hocko 1893ab5ac90aSMichal Hocko return page; 1894ab5ac90aSMichal Hocko } 1895ab5ac90aSMichal Hocko 1896e4e574b7SAdam Litke /* 1897099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1898099730d6SDave Hansen */ 1899e0ec90eeSDave Hansen static 19000c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1901099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1902099730d6SDave Hansen { 1903aaf14e40SMichal Hocko struct page *page; 1904aaf14e40SMichal Hocko struct mempolicy *mpol; 1905aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1906aaf14e40SMichal Hocko int nid; 1907aaf14e40SMichal Hocko nodemask_t *nodemask; 1908aaf14e40SMichal Hocko 1909aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19100c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1911aaf14e40SMichal Hocko mpol_cond_put(mpol); 1912aaf14e40SMichal Hocko 1913aaf14e40SMichal Hocko return page; 1914099730d6SDave Hansen } 1915099730d6SDave Hansen 1916ab5ac90aSMichal Hocko /* page migration callback function */ 19173e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1918d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19194db9b2efSMichal Hocko { 19204db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 19214db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19223e59fcb0SMichal Hocko struct page *page; 19233e59fcb0SMichal Hocko 19243e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19253e59fcb0SMichal Hocko if (page) { 19263e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 19273e59fcb0SMichal Hocko return page; 19284db9b2efSMichal Hocko } 19294db9b2efSMichal Hocko } 19304db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 19314db9b2efSMichal Hocko 19320c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19334db9b2efSMichal Hocko } 19344db9b2efSMichal Hocko 1935ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1936389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1937389c8178SMichal Hocko unsigned long address) 1938ebd63723SMichal Hocko { 1939ebd63723SMichal Hocko struct mempolicy *mpol; 1940ebd63723SMichal Hocko nodemask_t *nodemask; 1941ebd63723SMichal Hocko struct page *page; 1942ebd63723SMichal Hocko gfp_t gfp_mask; 1943ebd63723SMichal Hocko int node; 1944ebd63723SMichal Hocko 1945ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1946ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1947d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1948ebd63723SMichal Hocko mpol_cond_put(mpol); 1949ebd63723SMichal Hocko 1950ebd63723SMichal Hocko return page; 1951ebd63723SMichal Hocko } 1952ebd63723SMichal Hocko 1953bf50bab2SNaoya Horiguchi /* 195425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1955e4e574b7SAdam Litke * of size 'delta'. 1956e4e574b7SAdam Litke */ 19570a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19581b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1959e4e574b7SAdam Litke { 1960e4e574b7SAdam Litke struct list_head surplus_list; 1961e4e574b7SAdam Litke struct page *page, *tmp; 19620a4f3d1bSLiu Xiang int ret; 19630a4f3d1bSLiu Xiang long i; 19640a4f3d1bSLiu Xiang long needed, allocated; 196528073b02SHillf Danton bool alloc_ok = true; 1966e4e574b7SAdam Litke 1967a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1968ac09b3a1SAdam Litke if (needed <= 0) { 1969a5516438SAndi Kleen h->resv_huge_pages += delta; 1970e4e574b7SAdam Litke return 0; 1971ac09b3a1SAdam Litke } 1972e4e574b7SAdam Litke 1973e4e574b7SAdam Litke allocated = 0; 1974e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1975e4e574b7SAdam Litke 1976e4e574b7SAdam Litke ret = -ENOMEM; 1977e4e574b7SAdam Litke retry: 1978e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1979e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19800c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1981aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 198228073b02SHillf Danton if (!page) { 198328073b02SHillf Danton alloc_ok = false; 198428073b02SHillf Danton break; 198528073b02SHillf Danton } 1986e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 198769ed779aSDavid Rientjes cond_resched(); 1988e4e574b7SAdam Litke } 198928073b02SHillf Danton allocated += i; 1990e4e574b7SAdam Litke 1991e4e574b7SAdam Litke /* 1992e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1993e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1994e4e574b7SAdam Litke */ 1995e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1996a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1997a5516438SAndi Kleen (h->free_huge_pages + allocated); 199828073b02SHillf Danton if (needed > 0) { 199928073b02SHillf Danton if (alloc_ok) 2000e4e574b7SAdam Litke goto retry; 200128073b02SHillf Danton /* 200228073b02SHillf Danton * We were not able to allocate enough pages to 200328073b02SHillf Danton * satisfy the entire reservation so we free what 200428073b02SHillf Danton * we've allocated so far. 200528073b02SHillf Danton */ 200628073b02SHillf Danton goto free; 200728073b02SHillf Danton } 2008e4e574b7SAdam Litke /* 2009e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 201025985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2011e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2012ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2013ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2014ac09b3a1SAdam Litke * before they are reserved. 2015e4e574b7SAdam Litke */ 2016e4e574b7SAdam Litke needed += allocated; 2017a5516438SAndi Kleen h->resv_huge_pages += delta; 2018e4e574b7SAdam Litke ret = 0; 2019a9869b83SNaoya Horiguchi 202019fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 202119fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2022e558464bSMuchun Song int zeroed; 2023e558464bSMuchun Song 202419fc3f0aSAdam Litke if ((--needed) < 0) 202519fc3f0aSAdam Litke break; 2026a9869b83SNaoya Horiguchi /* 2027a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2028a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2029a9869b83SNaoya Horiguchi */ 2030e558464bSMuchun Song zeroed = put_page_testzero(page); 2031e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2032a5516438SAndi Kleen enqueue_huge_page(h, page); 203319fc3f0aSAdam Litke } 203428073b02SHillf Danton free: 2035b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 203619fc3f0aSAdam Litke 203719fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2038c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2039a9869b83SNaoya Horiguchi put_page(page); 204019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2041e4e574b7SAdam Litke 2042e4e574b7SAdam Litke return ret; 2043e4e574b7SAdam Litke } 2044e4e574b7SAdam Litke 2045e4e574b7SAdam Litke /* 2046e5bbc8a6SMike Kravetz * This routine has two main purposes: 2047e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2048e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2049e5bbc8a6SMike Kravetz * to the associated reservation map. 2050e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2051e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2052e5bbc8a6SMike Kravetz * 2053e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2054e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2055e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2056e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2057e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2058e4e574b7SAdam Litke */ 2059a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2060a5516438SAndi Kleen unsigned long unused_resv_pages) 2061e4e574b7SAdam Litke { 2062e4e574b7SAdam Litke unsigned long nr_pages; 2063e4e574b7SAdam Litke 2064aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2065bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2066e5bbc8a6SMike Kravetz goto out; 2067aa888a74SAndi Kleen 2068e5bbc8a6SMike Kravetz /* 2069e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2070e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2071e5bbc8a6SMike Kravetz */ 2072a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2073e4e574b7SAdam Litke 2074685f3457SLee Schermerhorn /* 2075685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20769b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20779b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20789b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 20799e7ee400SRandy Dunlap * free_pool_huge_page() will balance the freed pages across the 20809b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2081e5bbc8a6SMike Kravetz * 2082e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2083e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2084e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2085685f3457SLee Schermerhorn */ 2086685f3457SLee Schermerhorn while (nr_pages--) { 2087e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2088e5bbc8a6SMike Kravetz unused_resv_pages--; 20898cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2090e5bbc8a6SMike Kravetz goto out; 20917848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2092e4e574b7SAdam Litke } 2093e5bbc8a6SMike Kravetz 2094e5bbc8a6SMike Kravetz out: 2095e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2096e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2097e4e574b7SAdam Litke } 2098e4e574b7SAdam Litke 20995e911373SMike Kravetz 2100c37f9fb1SAndy Whitcroft /* 2101feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21025e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2103cf3ad20bSMike Kravetz * 2104cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2105cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2106cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2107cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2108cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2109feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2110feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2111feba16e2SMike Kravetz * is called in such cases. 2112cf3ad20bSMike Kravetz * 2113cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2114cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2115cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2116cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2117cf3ad20bSMike Kravetz * take appropriate action. 211896b96a96SMike Kravetz * 211996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 212096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 212196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 212296b96a96SMike Kravetz * reservation exists. 2123c37f9fb1SAndy Whitcroft */ 21245e911373SMike Kravetz enum vma_resv_mode { 21255e911373SMike Kravetz VMA_NEEDS_RESV, 21265e911373SMike Kravetz VMA_COMMIT_RESV, 2127feba16e2SMike Kravetz VMA_END_RESV, 212896b96a96SMike Kravetz VMA_ADD_RESV, 21295e911373SMike Kravetz }; 2130cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2131cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21325e911373SMike Kravetz enum vma_resv_mode mode) 2133c37f9fb1SAndy Whitcroft { 21344e35f483SJoonsoo Kim struct resv_map *resv; 21354e35f483SJoonsoo Kim pgoff_t idx; 2136cf3ad20bSMike Kravetz long ret; 21370db9d74eSMina Almasry long dummy_out_regions_needed; 2138c37f9fb1SAndy Whitcroft 21394e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21404e35f483SJoonsoo Kim if (!resv) 2141c37f9fb1SAndy Whitcroft return 1; 2142c37f9fb1SAndy Whitcroft 21434e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21445e911373SMike Kravetz switch (mode) { 21455e911373SMike Kravetz case VMA_NEEDS_RESV: 21460db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21470db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21480db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21490db9d74eSMina Almasry * ever require 1 region. 21500db9d74eSMina Almasry */ 21510db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21525e911373SMike Kravetz break; 21535e911373SMike Kravetz case VMA_COMMIT_RESV: 2154075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21550db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21560db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21575e911373SMike Kravetz break; 2158feba16e2SMike Kravetz case VMA_END_RESV: 21590db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21605e911373SMike Kravetz ret = 0; 21615e911373SMike Kravetz break; 216296b96a96SMike Kravetz case VMA_ADD_RESV: 21630db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2164075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21650db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21660db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21670db9d74eSMina Almasry } else { 21680db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 216996b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 217096b96a96SMike Kravetz } 217196b96a96SMike Kravetz break; 21725e911373SMike Kravetz default: 21735e911373SMike Kravetz BUG(); 21745e911373SMike Kravetz } 217584afd99bSAndy Whitcroft 21764e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2177cf3ad20bSMike Kravetz return ret; 217867961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 217967961f9dSMike Kravetz /* 218067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 218167961f9dSMike Kravetz * However, a file associated with mapping could have been 218267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 218367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 218467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 218567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 218667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 218767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 218867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 218967961f9dSMike Kravetz * return value of this routine is the opposite of the 219067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 219167961f9dSMike Kravetz */ 219267961f9dSMike Kravetz if (ret) 219367961f9dSMike Kravetz return 0; 219467961f9dSMike Kravetz else 219567961f9dSMike Kravetz return 1; 219667961f9dSMike Kravetz } 21974e35f483SJoonsoo Kim else 2198cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 219984afd99bSAndy Whitcroft } 2200cf3ad20bSMike Kravetz 2201cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2202a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2203c37f9fb1SAndy Whitcroft { 22045e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2205cf3ad20bSMike Kravetz } 2206c37f9fb1SAndy Whitcroft 2207cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2208cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2209cf3ad20bSMike Kravetz { 22105e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22115e911373SMike Kravetz } 22125e911373SMike Kravetz 2213feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22145e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22155e911373SMike Kravetz { 2216feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2217c37f9fb1SAndy Whitcroft } 2218c37f9fb1SAndy Whitcroft 221996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 222096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 222196b96a96SMike Kravetz { 222296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 222396b96a96SMike Kravetz } 222496b96a96SMike Kravetz 222596b96a96SMike Kravetz /* 222696b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 222796b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 222896b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 2229d6995da3SMike Kravetz * alloc_huge_page would have consumed the reservation and set 2230d6995da3SMike Kravetz * HPageRestoreReserve in the newly allocated page. When the page is freed 2231d6995da3SMike Kravetz * via free_huge_page, the global reservation count will be incremented if 2232d6995da3SMike Kravetz * HPageRestoreReserve is set. However, free_huge_page can not adjust the 2233d6995da3SMike Kravetz * reserve map. Adjust the reserve map here to be consistent with global 2234d6995da3SMike Kravetz * reserve count adjustments to be made by free_huge_page. 223596b96a96SMike Kravetz */ 223696b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 223796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 223896b96a96SMike Kravetz struct page *page) 223996b96a96SMike Kravetz { 2240d6995da3SMike Kravetz if (unlikely(HPageRestoreReserve(page))) { 224196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 224296b96a96SMike Kravetz 224396b96a96SMike Kravetz if (unlikely(rc < 0)) { 224496b96a96SMike Kravetz /* 224596b96a96SMike Kravetz * Rare out of memory condition in reserve map 2246d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 224796b96a96SMike Kravetz * global reserve count will not be incremented 224896b96a96SMike Kravetz * by free_huge_page. This will make it appear 224996b96a96SMike Kravetz * as though the reservation for this page was 225096b96a96SMike Kravetz * consumed. This may prevent the task from 225196b96a96SMike Kravetz * faulting in the page at a later time. This 225296b96a96SMike Kravetz * is better than inconsistent global huge page 225396b96a96SMike Kravetz * accounting of reserve counts. 225496b96a96SMike Kravetz */ 2255d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 225696b96a96SMike Kravetz } else if (rc) { 225796b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 225896b96a96SMike Kravetz if (unlikely(rc < 0)) 225996b96a96SMike Kravetz /* 226096b96a96SMike Kravetz * See above comment about rare out of 226196b96a96SMike Kravetz * memory condition. 226296b96a96SMike Kravetz */ 2263d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 226496b96a96SMike Kravetz } else 226596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 226696b96a96SMike Kravetz } 226796b96a96SMike Kravetz } 226896b96a96SMike Kravetz 226970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 227004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2271348ea204SAdam Litke { 227290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2273a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2274348ea204SAdam Litke struct page *page; 2275d85f69b0SMike Kravetz long map_chg, map_commit; 2276d85f69b0SMike Kravetz long gbl_chg; 22776d76dcf4SAneesh Kumar K.V int ret, idx; 22786d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 227908cf9fafSMina Almasry bool deferred_reserve; 22802fc39cecSAdam Litke 22816d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2282a1e78772SMel Gorman /* 2283d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2284d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2285d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2286a1e78772SMel Gorman */ 2287d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2288d85f69b0SMike Kravetz if (map_chg < 0) 228976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2290d85f69b0SMike Kravetz 2291d85f69b0SMike Kravetz /* 2292d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2293d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2294d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2295d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2296d85f69b0SMike Kravetz * checked against any subpool limit. 2297d85f69b0SMike Kravetz */ 2298d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2299d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2300d85f69b0SMike Kravetz if (gbl_chg < 0) { 2301feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 230276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 23035e911373SMike Kravetz } 230490d8b7e6SAdam Litke 2305d85f69b0SMike Kravetz /* 2306d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2307d85f69b0SMike Kravetz * map, there could be reservations associated with the 2308d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2309d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2310d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2311d85f69b0SMike Kravetz * the subpool reservations. 2312d85f69b0SMike Kravetz */ 2313d85f69b0SMike Kravetz if (avoid_reserve) 2314d85f69b0SMike Kravetz gbl_chg = 1; 2315d85f69b0SMike Kravetz } 2316d85f69b0SMike Kravetz 231708cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 231808cf9fafSMina Almasry */ 231908cf9fafSMina Almasry deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); 232008cf9fafSMina Almasry if (deferred_reserve) { 232108cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 232208cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 23238f34af6fSJianyu Zhan if (ret) 23248f34af6fSJianyu Zhan goto out_subpool_put; 232508cf9fafSMina Almasry } 232608cf9fafSMina Almasry 232708cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 232808cf9fafSMina Almasry if (ret) 232908cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 23308f34af6fSJianyu Zhan 2331a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2332d85f69b0SMike Kravetz /* 2333d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2334d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2335d85f69b0SMike Kravetz * a reservation exists for the allocation. 2336d85f69b0SMike Kravetz */ 2337d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 233881a6fcaeSJoonsoo Kim if (!page) { 233994ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 23400c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 23418f34af6fSJianyu Zhan if (!page) 23428f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2343a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2344d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2345a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2346a88c7695SNaoya Horiguchi } 234779dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 234815a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 234981a6fcaeSJoonsoo Kim /* Fall through */ 2350a1e78772SMel Gorman } 235181a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 235208cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 235308cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 235408cf9fafSMina Almasry */ 235508cf9fafSMina Almasry if (deferred_reserve) { 235608cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 235708cf9fafSMina Almasry h_cg, page); 235808cf9fafSMina Almasry } 235908cf9fafSMina Almasry 236081a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2361a1e78772SMel Gorman 2362d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2363a1e78772SMel Gorman 2364d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2365d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 236633039678SMike Kravetz /* 236733039678SMike Kravetz * The page was added to the reservation map between 236833039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 236933039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 237033039678SMike Kravetz * Adjust for the subpool count incremented above AND 237133039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 237233039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 237333039678SMike Kravetz * no longer applies. 237433039678SMike Kravetz */ 237533039678SMike Kravetz long rsv_adjust; 237633039678SMike Kravetz 237733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 237833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 237979aa925bSMike Kravetz if (deferred_reserve) 238079aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 238179aa925bSMike Kravetz pages_per_huge_page(h), page); 238233039678SMike Kravetz } 23837893d1d5SAdam Litke return page; 23848f34af6fSJianyu Zhan 23858f34af6fSJianyu Zhan out_uncharge_cgroup: 23868f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 238708cf9fafSMina Almasry out_uncharge_cgroup_reservation: 238808cf9fafSMina Almasry if (deferred_reserve) 238908cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 239008cf9fafSMina Almasry h_cg); 23918f34af6fSJianyu Zhan out_subpool_put: 2392d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 23938f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2394feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 23958f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2396b45b5bd6SDavid Gibson } 2397b45b5bd6SDavid Gibson 2398e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2399e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2400e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2401aa888a74SAndi Kleen { 2402aa888a74SAndi Kleen struct huge_bootmem_page *m; 2403b2261026SJoonsoo Kim int nr_nodes, node; 2404aa888a74SAndi Kleen 2405b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2406aa888a74SAndi Kleen void *addr; 2407aa888a74SAndi Kleen 2408eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 24098b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 241097ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2411aa888a74SAndi Kleen if (addr) { 2412aa888a74SAndi Kleen /* 2413aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2414aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2415aa888a74SAndi Kleen * puts them into the mem_map). 2416aa888a74SAndi Kleen */ 2417aa888a74SAndi Kleen m = addr; 2418aa888a74SAndi Kleen goto found; 2419aa888a74SAndi Kleen } 2420aa888a74SAndi Kleen } 2421aa888a74SAndi Kleen return 0; 2422aa888a74SAndi Kleen 2423aa888a74SAndi Kleen found: 2424df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2425aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2426330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2427aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2428aa888a74SAndi Kleen m->hstate = h; 2429aa888a74SAndi Kleen return 1; 2430aa888a74SAndi Kleen } 2431aa888a74SAndi Kleen 2432d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2433d00181b9SKirill A. Shutemov unsigned int order) 243418229df5SAndy Whitcroft { 243518229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 243618229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 243718229df5SAndy Whitcroft else 243818229df5SAndy Whitcroft prep_compound_page(page, order); 243918229df5SAndy Whitcroft } 244018229df5SAndy Whitcroft 2441aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2442aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2443aa888a74SAndi Kleen { 2444aa888a74SAndi Kleen struct huge_bootmem_page *m; 2445aa888a74SAndi Kleen 2446aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 244740d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2448aa888a74SAndi Kleen struct hstate *h = m->hstate; 2449ee8f248dSBecky Bruce 2450aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2451c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2452ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2453aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2454af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2455af0fb9dfSMichal Hocko 2456b0320c7bSRafael Aquini /* 2457b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2458b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2459b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2460b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2461b0320c7bSRafael Aquini */ 2462bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2463c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2464520495feSCannon Matthews cond_resched(); 2465aa888a74SAndi Kleen } 2466aa888a74SAndi Kleen } 2467aa888a74SAndi Kleen 24688faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 24691da177e4SLinus Torvalds { 24701da177e4SLinus Torvalds unsigned long i; 2471f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2472f60858f9SMike Kravetz 2473f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2474f60858f9SMike Kravetz /* 2475f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2476f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2477f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2478f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2479f60858f9SMike Kravetz */ 2480f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2481f60858f9SMike Kravetz GFP_KERNEL); 2482f60858f9SMike Kravetz } else { 2483f60858f9SMike Kravetz /* allocations done at boot time */ 2484f60858f9SMike Kravetz node_alloc_noretry = NULL; 2485f60858f9SMike Kravetz } 2486f60858f9SMike Kravetz 2487f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2488f60858f9SMike Kravetz if (node_alloc_noretry) 2489f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 24901da177e4SLinus Torvalds 2491e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2492bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2493dbda8feaSBarry Song if (hugetlb_cma_size) { 2494cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 24957ecc9565SChen Wandun goto free; 2496cf11e85fSRoman Gushchin } 2497aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2498aa888a74SAndi Kleen break; 24990c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2500f60858f9SMike Kravetz &node_states[N_MEMORY], 2501f60858f9SMike Kravetz node_alloc_noretry)) 25021da177e4SLinus Torvalds break; 250369ed779aSDavid Rientjes cond_resched(); 25041da177e4SLinus Torvalds } 2505d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2506d715cf80SLiam R. Howlett char buf[32]; 2507d715cf80SLiam R. Howlett 2508c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2509d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2510d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 25118faa8b07SAndi Kleen h->max_huge_pages = i; 2512e5ff2159SAndi Kleen } 25137ecc9565SChen Wandun free: 2514f60858f9SMike Kravetz kfree(node_alloc_noretry); 2515d715cf80SLiam R. Howlett } 2516e5ff2159SAndi Kleen 2517e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2518e5ff2159SAndi Kleen { 2519e5ff2159SAndi Kleen struct hstate *h; 2520e5ff2159SAndi Kleen 2521e5ff2159SAndi Kleen for_each_hstate(h) { 2522641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2523641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2524641844f5SNaoya Horiguchi 25258faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2526bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 25278faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2528e5ff2159SAndi Kleen } 2529641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2530e5ff2159SAndi Kleen } 2531e5ff2159SAndi Kleen 2532e5ff2159SAndi Kleen static void __init report_hugepages(void) 2533e5ff2159SAndi Kleen { 2534e5ff2159SAndi Kleen struct hstate *h; 2535e5ff2159SAndi Kleen 2536e5ff2159SAndi Kleen for_each_hstate(h) { 25374abd32dbSAndi Kleen char buf[32]; 2538c6247f72SMatthew Wilcox 2539c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2540ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2541c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2542e5ff2159SAndi Kleen } 2543e5ff2159SAndi Kleen } 2544e5ff2159SAndi Kleen 25451da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 25466ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 25476ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25481da177e4SLinus Torvalds { 25494415cc8dSChristoph Lameter int i; 25504415cc8dSChristoph Lameter 2551bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2552aa888a74SAndi Kleen return; 2553aa888a74SAndi Kleen 25546ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 25551da177e4SLinus Torvalds struct page *page, *next; 2556a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2557a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2558a5516438SAndi Kleen if (count >= h->nr_huge_pages) 25596b0c880dSAdam Litke return; 25601da177e4SLinus Torvalds if (PageHighMem(page)) 25611da177e4SLinus Torvalds continue; 25621da177e4SLinus Torvalds list_del(&page->lru); 2563e5ff2159SAndi Kleen update_and_free_page(h, page); 2564a5516438SAndi Kleen h->free_huge_pages--; 2565a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 25661da177e4SLinus Torvalds } 25671da177e4SLinus Torvalds } 25681da177e4SLinus Torvalds } 25691da177e4SLinus Torvalds #else 25706ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 25716ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25721da177e4SLinus Torvalds { 25731da177e4SLinus Torvalds } 25741da177e4SLinus Torvalds #endif 25751da177e4SLinus Torvalds 257620a0307cSWu Fengguang /* 257720a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 257820a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 257920a0307cSWu Fengguang * Returns 1 if an adjustment was made. 258020a0307cSWu Fengguang */ 25816ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 25826ae11b27SLee Schermerhorn int delta) 258320a0307cSWu Fengguang { 2584b2261026SJoonsoo Kim int nr_nodes, node; 258520a0307cSWu Fengguang 258620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 258720a0307cSWu Fengguang 2588e8c5c824SLee Schermerhorn if (delta < 0) { 2589b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2590b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2591b2261026SJoonsoo Kim goto found; 2592b2261026SJoonsoo Kim } 2593b2261026SJoonsoo Kim } else { 2594b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2595b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2596b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2597b2261026SJoonsoo Kim goto found; 2598e8c5c824SLee Schermerhorn } 25999a76db09SLee Schermerhorn } 2600b2261026SJoonsoo Kim return 0; 260120a0307cSWu Fengguang 2602b2261026SJoonsoo Kim found: 260320a0307cSWu Fengguang h->surplus_huge_pages += delta; 2604b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2605b2261026SJoonsoo Kim return 1; 260620a0307cSWu Fengguang } 260720a0307cSWu Fengguang 2608a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2609fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 26106ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26111da177e4SLinus Torvalds { 26127893d1d5SAdam Litke unsigned long min_count, ret; 2613f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2614f60858f9SMike Kravetz 2615f60858f9SMike Kravetz /* 2616f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2617f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2618f60858f9SMike Kravetz * the requested huge pages. 2619f60858f9SMike Kravetz */ 2620f60858f9SMike Kravetz if (node_alloc_noretry) 2621f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2622f60858f9SMike Kravetz else 2623f60858f9SMike Kravetz return -ENOMEM; 26241da177e4SLinus Torvalds 26254eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 26264eb0716eSAlexandre Ghiti 26274eb0716eSAlexandre Ghiti /* 2628fd875dcaSMike Kravetz * Check for a node specific request. 2629fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2630fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2631fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2632fd875dcaSMike Kravetz */ 2633fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2634fd875dcaSMike Kravetz unsigned long old_count = count; 2635fd875dcaSMike Kravetz 2636fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2637fd875dcaSMike Kravetz /* 2638fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2639fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2640fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2641fd875dcaSMike Kravetz * largest possible value to align with their intention. 2642fd875dcaSMike Kravetz */ 2643fd875dcaSMike Kravetz if (count < old_count) 2644fd875dcaSMike Kravetz count = ULONG_MAX; 2645fd875dcaSMike Kravetz } 2646fd875dcaSMike Kravetz 2647fd875dcaSMike Kravetz /* 26484eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 26494eb0716eSAlexandre Ghiti * page range allocation. 26504eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 26514eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 26524eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 26534eb0716eSAlexandre Ghiti */ 26544eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 26554eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 26564eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2657f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 26584eb0716eSAlexandre Ghiti return -EINVAL; 26594eb0716eSAlexandre Ghiti } 26604eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 26614eb0716eSAlexandre Ghiti } 2662aa888a74SAndi Kleen 26637893d1d5SAdam Litke /* 26647893d1d5SAdam Litke * Increase the pool size 26657893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 26667893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2667d1c3fb1fSNishanth Aravamudan * 26680c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2669d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2670d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2671d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2672d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 26737893d1d5SAdam Litke */ 2674a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 26756ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 26767893d1d5SAdam Litke break; 26777893d1d5SAdam Litke } 26787893d1d5SAdam Litke 2679a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 26807893d1d5SAdam Litke /* 26817893d1d5SAdam Litke * If this allocation races such that we no longer need the 26827893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 26837893d1d5SAdam Litke * and reducing the surplus. 26847893d1d5SAdam Litke */ 26857893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2686649920c6SJia He 2687649920c6SJia He /* yield cpu to avoid soft lockup */ 2688649920c6SJia He cond_resched(); 2689649920c6SJia He 2690f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2691f60858f9SMike Kravetz node_alloc_noretry); 26927893d1d5SAdam Litke spin_lock(&hugetlb_lock); 26937893d1d5SAdam Litke if (!ret) 26947893d1d5SAdam Litke goto out; 26957893d1d5SAdam Litke 2696536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2697536240f2SMel Gorman if (signal_pending(current)) 2698536240f2SMel Gorman goto out; 26997893d1d5SAdam Litke } 27007893d1d5SAdam Litke 27017893d1d5SAdam Litke /* 27027893d1d5SAdam Litke * Decrease the pool size 27037893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27047893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27057893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27067893d1d5SAdam Litke * to the desired size as pages become free. 2707d1c3fb1fSNishanth Aravamudan * 2708d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2709d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2710d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27110c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2712d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2713d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2714d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27157893d1d5SAdam Litke */ 2716a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27176b0c880dSAdam Litke min_count = max(count, min_count); 27186ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2719a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 27206ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 27211da177e4SLinus Torvalds break; 272255f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 27231da177e4SLinus Torvalds } 2724a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 27256ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 27267893d1d5SAdam Litke break; 27277893d1d5SAdam Litke } 27287893d1d5SAdam Litke out: 27294eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 27301da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 27314eb0716eSAlexandre Ghiti 2732f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2733f60858f9SMike Kravetz 27344eb0716eSAlexandre Ghiti return 0; 27351da177e4SLinus Torvalds } 27361da177e4SLinus Torvalds 2737a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2738a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2739a3437870SNishanth Aravamudan 2740a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2741a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2742a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2743a3437870SNishanth Aravamudan 2744a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2745a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2746a3437870SNishanth Aravamudan 27479a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 27489a305230SLee Schermerhorn 27499a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2750a3437870SNishanth Aravamudan { 2751a3437870SNishanth Aravamudan int i; 27529a305230SLee Schermerhorn 2753a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 27549a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 27559a305230SLee Schermerhorn if (nidp) 27569a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2757a3437870SNishanth Aravamudan return &hstates[i]; 27589a305230SLee Schermerhorn } 27599a305230SLee Schermerhorn 27609a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2761a3437870SNishanth Aravamudan } 2762a3437870SNishanth Aravamudan 276306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2764a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2765a3437870SNishanth Aravamudan { 27669a305230SLee Schermerhorn struct hstate *h; 27679a305230SLee Schermerhorn unsigned long nr_huge_pages; 27689a305230SLee Schermerhorn int nid; 27699a305230SLee Schermerhorn 27709a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27719a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27729a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 27739a305230SLee Schermerhorn else 27749a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 27759a305230SLee Schermerhorn 2776ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2777a3437870SNishanth Aravamudan } 2778adbe8726SEric B Munson 2779238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2780238d3c13SDavid Rientjes struct hstate *h, int nid, 2781238d3c13SDavid Rientjes unsigned long count, size_t len) 2782a3437870SNishanth Aravamudan { 2783a3437870SNishanth Aravamudan int err; 27842d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2785a3437870SNishanth Aravamudan 27862d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 27872d0adf7eSOscar Salvador return -EINVAL; 2788adbe8726SEric B Munson 27899a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 27909a305230SLee Schermerhorn /* 27919a305230SLee Schermerhorn * global hstate attribute 27929a305230SLee Schermerhorn */ 27939a305230SLee Schermerhorn if (!(obey_mempolicy && 27942d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 27952d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 27962d0adf7eSOscar Salvador else 27972d0adf7eSOscar Salvador n_mask = &nodes_allowed; 27982d0adf7eSOscar Salvador } else { 27999a305230SLee Schermerhorn /* 2800fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2801fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28029a305230SLee Schermerhorn */ 28032d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28042d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2805fd875dcaSMike Kravetz } 28069a305230SLee Schermerhorn 28072d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 280806808b08SLee Schermerhorn 28094eb0716eSAlexandre Ghiti return err ? err : len; 281006808b08SLee Schermerhorn } 281106808b08SLee Schermerhorn 2812238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2813238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2814238d3c13SDavid Rientjes size_t len) 2815238d3c13SDavid Rientjes { 2816238d3c13SDavid Rientjes struct hstate *h; 2817238d3c13SDavid Rientjes unsigned long count; 2818238d3c13SDavid Rientjes int nid; 2819238d3c13SDavid Rientjes int err; 2820238d3c13SDavid Rientjes 2821238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2822238d3c13SDavid Rientjes if (err) 2823238d3c13SDavid Rientjes return err; 2824238d3c13SDavid Rientjes 2825238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2826238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2827238d3c13SDavid Rientjes } 2828238d3c13SDavid Rientjes 282906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 283006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 283106808b08SLee Schermerhorn { 283206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 283306808b08SLee Schermerhorn } 283406808b08SLee Schermerhorn 283506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 283606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 283706808b08SLee Schermerhorn { 2838238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2839a3437870SNishanth Aravamudan } 2840a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2841a3437870SNishanth Aravamudan 284206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 284306808b08SLee Schermerhorn 284406808b08SLee Schermerhorn /* 284506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 284606808b08SLee Schermerhorn * huge page alloc/free. 284706808b08SLee Schermerhorn */ 284806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 2849ae7a927dSJoe Perches struct kobj_attribute *attr, 2850ae7a927dSJoe Perches char *buf) 285106808b08SLee Schermerhorn { 285206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 285306808b08SLee Schermerhorn } 285406808b08SLee Schermerhorn 285506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 285606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 285706808b08SLee Schermerhorn { 2858238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 285906808b08SLee Schermerhorn } 286006808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 286106808b08SLee Schermerhorn #endif 286206808b08SLee Schermerhorn 286306808b08SLee Schermerhorn 2864a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2865a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2866a3437870SNishanth Aravamudan { 28679a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2868ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 2869a3437870SNishanth Aravamudan } 2870adbe8726SEric B Munson 2871a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2872a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2873a3437870SNishanth Aravamudan { 2874a3437870SNishanth Aravamudan int err; 2875a3437870SNishanth Aravamudan unsigned long input; 28769a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2877a3437870SNishanth Aravamudan 2878bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2879adbe8726SEric B Munson return -EINVAL; 2880adbe8726SEric B Munson 28813dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2882a3437870SNishanth Aravamudan if (err) 288373ae31e5SEric B Munson return err; 2884a3437870SNishanth Aravamudan 2885a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2886a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2887a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2888a3437870SNishanth Aravamudan 2889a3437870SNishanth Aravamudan return count; 2890a3437870SNishanth Aravamudan } 2891a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2892a3437870SNishanth Aravamudan 2893a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2894a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2895a3437870SNishanth Aravamudan { 28969a305230SLee Schermerhorn struct hstate *h; 28979a305230SLee Schermerhorn unsigned long free_huge_pages; 28989a305230SLee Schermerhorn int nid; 28999a305230SLee Schermerhorn 29009a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29019a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29029a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29039a305230SLee Schermerhorn else 29049a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29059a305230SLee Schermerhorn 2906ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 2907a3437870SNishanth Aravamudan } 2908a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2909a3437870SNishanth Aravamudan 2910a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2911a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2912a3437870SNishanth Aravamudan { 29139a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2914ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 2915a3437870SNishanth Aravamudan } 2916a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2917a3437870SNishanth Aravamudan 2918a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2919a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2920a3437870SNishanth Aravamudan { 29219a305230SLee Schermerhorn struct hstate *h; 29229a305230SLee Schermerhorn unsigned long surplus_huge_pages; 29239a305230SLee Schermerhorn int nid; 29249a305230SLee Schermerhorn 29259a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29269a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29279a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 29289a305230SLee Schermerhorn else 29299a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 29309a305230SLee Schermerhorn 2931ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 2932a3437870SNishanth Aravamudan } 2933a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2934a3437870SNishanth Aravamudan 2935a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2936a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2937a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2938a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2939a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2940a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 294106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 294206808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 294306808b08SLee Schermerhorn #endif 2944a3437870SNishanth Aravamudan NULL, 2945a3437870SNishanth Aravamudan }; 2946a3437870SNishanth Aravamudan 294767e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2948a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2949a3437870SNishanth Aravamudan }; 2950a3437870SNishanth Aravamudan 2951094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 29529a305230SLee Schermerhorn struct kobject **hstate_kobjs, 295367e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2954a3437870SNishanth Aravamudan { 2955a3437870SNishanth Aravamudan int retval; 2956972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2957a3437870SNishanth Aravamudan 29589a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 29599a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2960a3437870SNishanth Aravamudan return -ENOMEM; 2961a3437870SNishanth Aravamudan 29629a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2963cc2205a6SMiaohe Lin if (retval) { 29649a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2965cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 2966cc2205a6SMiaohe Lin } 2967a3437870SNishanth Aravamudan 2968a3437870SNishanth Aravamudan return retval; 2969a3437870SNishanth Aravamudan } 2970a3437870SNishanth Aravamudan 2971a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2972a3437870SNishanth Aravamudan { 2973a3437870SNishanth Aravamudan struct hstate *h; 2974a3437870SNishanth Aravamudan int err; 2975a3437870SNishanth Aravamudan 2976a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2977a3437870SNishanth Aravamudan if (!hugepages_kobj) 2978a3437870SNishanth Aravamudan return; 2979a3437870SNishanth Aravamudan 2980a3437870SNishanth Aravamudan for_each_hstate(h) { 29819a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 29829a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2983a3437870SNishanth Aravamudan if (err) 2984282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 2985a3437870SNishanth Aravamudan } 2986a3437870SNishanth Aravamudan } 2987a3437870SNishanth Aravamudan 29889a305230SLee Schermerhorn #ifdef CONFIG_NUMA 29899a305230SLee Schermerhorn 29909a305230SLee Schermerhorn /* 29919a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 299210fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 299310fbcf4cSKay Sievers * index of a node device or _hstate == node id. 299410fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 29959a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 29969a305230SLee Schermerhorn */ 29979a305230SLee Schermerhorn struct node_hstate { 29989a305230SLee Schermerhorn struct kobject *hugepages_kobj; 29999a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30009a305230SLee Schermerhorn }; 3001b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30029a305230SLee Schermerhorn 30039a305230SLee Schermerhorn /* 300410fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30059a305230SLee Schermerhorn */ 30069a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30079a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30089a305230SLee Schermerhorn &free_hugepages_attr.attr, 30099a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30109a305230SLee Schermerhorn NULL, 30119a305230SLee Schermerhorn }; 30129a305230SLee Schermerhorn 301367e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30149a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30159a305230SLee Schermerhorn }; 30169a305230SLee Schermerhorn 30179a305230SLee Schermerhorn /* 301810fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30199a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30209a305230SLee Schermerhorn */ 30219a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 30229a305230SLee Schermerhorn { 30239a305230SLee Schermerhorn int nid; 30249a305230SLee Schermerhorn 30259a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 30269a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 30279a305230SLee Schermerhorn int i; 30289a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 30299a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 30309a305230SLee Schermerhorn if (nidp) 30319a305230SLee Schermerhorn *nidp = nid; 30329a305230SLee Schermerhorn return &hstates[i]; 30339a305230SLee Schermerhorn } 30349a305230SLee Schermerhorn } 30359a305230SLee Schermerhorn 30369a305230SLee Schermerhorn BUG(); 30379a305230SLee Schermerhorn return NULL; 30389a305230SLee Schermerhorn } 30399a305230SLee Schermerhorn 30409a305230SLee Schermerhorn /* 304110fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 30429a305230SLee Schermerhorn * No-op if no hstate attributes attached. 30439a305230SLee Schermerhorn */ 30443cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 30459a305230SLee Schermerhorn { 30469a305230SLee Schermerhorn struct hstate *h; 304710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30489a305230SLee Schermerhorn 30499a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30509b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 30519a305230SLee Schermerhorn 3052972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3053972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3054972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3055972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3056972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3057972dc4deSAneesh Kumar K.V } 30589a305230SLee Schermerhorn } 30599a305230SLee Schermerhorn 30609a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 30619a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 30629a305230SLee Schermerhorn } 30639a305230SLee Schermerhorn 30649a305230SLee Schermerhorn 30659a305230SLee Schermerhorn /* 306610fbcf4cSKay Sievers * Register hstate attributes for a single node device. 30679a305230SLee Schermerhorn * No-op if attributes already registered. 30689a305230SLee Schermerhorn */ 30693cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 30709a305230SLee Schermerhorn { 30719a305230SLee Schermerhorn struct hstate *h; 307210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30739a305230SLee Schermerhorn int err; 30749a305230SLee Schermerhorn 30759a305230SLee Schermerhorn if (nhs->hugepages_kobj) 30769a305230SLee Schermerhorn return; /* already allocated */ 30779a305230SLee Schermerhorn 30789a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 307910fbcf4cSKay Sievers &node->dev.kobj); 30809a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30819a305230SLee Schermerhorn return; 30829a305230SLee Schermerhorn 30839a305230SLee Schermerhorn for_each_hstate(h) { 30849a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 30859a305230SLee Schermerhorn nhs->hstate_kobjs, 30869a305230SLee Schermerhorn &per_node_hstate_attr_group); 30879a305230SLee Schermerhorn if (err) { 3088282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 308910fbcf4cSKay Sievers h->name, node->dev.id); 30909a305230SLee Schermerhorn hugetlb_unregister_node(node); 30919a305230SLee Schermerhorn break; 30929a305230SLee Schermerhorn } 30939a305230SLee Schermerhorn } 30949a305230SLee Schermerhorn } 30959a305230SLee Schermerhorn 30969a305230SLee Schermerhorn /* 30979b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 309810fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 309910fbcf4cSKay Sievers * registered their associated device by this time. 31009a305230SLee Schermerhorn */ 31017d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31029a305230SLee Schermerhorn { 31039a305230SLee Schermerhorn int nid; 31049a305230SLee Schermerhorn 31058cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31068732794bSWen Congyang struct node *node = node_devices[nid]; 310710fbcf4cSKay Sievers if (node->dev.id == nid) 31089a305230SLee Schermerhorn hugetlb_register_node(node); 31099a305230SLee Schermerhorn } 31109a305230SLee Schermerhorn 31119a305230SLee Schermerhorn /* 311210fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31139a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31149a305230SLee Schermerhorn */ 31159a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31169a305230SLee Schermerhorn hugetlb_unregister_node); 31179a305230SLee Schermerhorn } 31189a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31199a305230SLee Schermerhorn 31209a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31219a305230SLee Schermerhorn { 31229a305230SLee Schermerhorn BUG(); 31239a305230SLee Schermerhorn if (nidp) 31249a305230SLee Schermerhorn *nidp = -1; 31259a305230SLee Schermerhorn return NULL; 31269a305230SLee Schermerhorn } 31279a305230SLee Schermerhorn 31289a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 31299a305230SLee Schermerhorn 31309a305230SLee Schermerhorn #endif 31319a305230SLee Schermerhorn 3132a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3133a3437870SNishanth Aravamudan { 31348382d914SDavidlohr Bueso int i; 31358382d914SDavidlohr Bueso 3136d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3137d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3138d6995da3SMike Kravetz 3139c2833a5bSMike Kravetz if (!hugepages_supported()) { 3140c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3141c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 31420ef89d25SBenjamin Herrenschmidt return 0; 3143d715cf80SLiam R. Howlett } 3144d715cf80SLiam R. Howlett 3145282f4214SMike Kravetz /* 3146282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3147282f4214SMike Kravetz * architectures depend on setup being done here. 3148282f4214SMike Kravetz */ 3149a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3150282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3151282f4214SMike Kravetz /* 3152282f4214SMike Kravetz * If we did not parse a default huge page size, set 3153282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3154282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3155282f4214SMike Kravetz * specified, set that here as well. 3156282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3157282f4214SMike Kravetz * setting. A warning will be printed in this case. 3158282f4214SMike Kravetz */ 3159282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3160f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3161282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3162282f4214SMike Kravetz char buf[32]; 3163282f4214SMike Kravetz 3164282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3165282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3166282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3167282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3168282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3169282f4214SMike Kravetz default_hstate_max_huge_pages); 3170282f4214SMike Kravetz } 3171282f4214SMike Kravetz default_hstate.max_huge_pages = 3172282f4214SMike Kravetz default_hstate_max_huge_pages; 3173282f4214SMike Kravetz } 3174f8b74815SVaishali Thakkar } 3175a3437870SNishanth Aravamudan 3176cf11e85fSRoman Gushchin hugetlb_cma_check(); 3177a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3178aa888a74SAndi Kleen gather_bootmem_prealloc(); 3179a3437870SNishanth Aravamudan report_hugepages(); 3180a3437870SNishanth Aravamudan 3181a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 31829a305230SLee Schermerhorn hugetlb_register_all_nodes(); 31837179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 31849a305230SLee Schermerhorn 31858382d914SDavidlohr Bueso #ifdef CONFIG_SMP 31868382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 31878382d914SDavidlohr Bueso #else 31888382d914SDavidlohr Bueso num_fault_mutexes = 1; 31898382d914SDavidlohr Bueso #endif 3190c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 31916da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 31926da2ec56SKees Cook GFP_KERNEL); 3193c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 31948382d914SDavidlohr Bueso 31958382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3196c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3197a3437870SNishanth Aravamudan return 0; 3198a3437870SNishanth Aravamudan } 31993e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3200a3437870SNishanth Aravamudan 3201ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3202ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32039fee021dSVaishali Thakkar { 3204ae94da89SMike Kravetz return size == HPAGE_SIZE; 32059fee021dSVaishali Thakkar } 32069fee021dSVaishali Thakkar 3207d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3208a3437870SNishanth Aravamudan { 3209a3437870SNishanth Aravamudan struct hstate *h; 32108faa8b07SAndi Kleen unsigned long i; 32118faa8b07SAndi Kleen 3212a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3213a3437870SNishanth Aravamudan return; 3214a3437870SNishanth Aravamudan } 321547d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3216a3437870SNishanth Aravamudan BUG_ON(order == 0); 321747d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3218a3437870SNishanth Aravamudan h->order = order; 3219aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 32208faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32218faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 32220edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 322354f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 322454f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3225a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3226a3437870SNishanth Aravamudan huge_page_size(h)/1024); 32278faa8b07SAndi Kleen 3228a3437870SNishanth Aravamudan parsed_hstate = h; 3229a3437870SNishanth Aravamudan } 3230a3437870SNishanth Aravamudan 3231282f4214SMike Kravetz /* 3232282f4214SMike Kravetz * hugepages command line processing 3233282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3234282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3235282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3236282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3237282f4214SMike Kravetz */ 3238282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3239a3437870SNishanth Aravamudan { 3240a3437870SNishanth Aravamudan unsigned long *mhp; 32418faa8b07SAndi Kleen static unsigned long *last_mhp; 3242a3437870SNishanth Aravamudan 32439fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3244282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 32459fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3246282f4214SMike Kravetz return 0; 32479fee021dSVaishali Thakkar } 3248282f4214SMike Kravetz 3249a3437870SNishanth Aravamudan /* 3250282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3251282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3252282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3253282f4214SMike Kravetz * default_hugepagesz. 3254a3437870SNishanth Aravamudan */ 32559fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3256a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3257a3437870SNishanth Aravamudan else 3258a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3259a3437870SNishanth Aravamudan 32608faa8b07SAndi Kleen if (mhp == last_mhp) { 3261282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3262282f4214SMike Kravetz return 0; 32638faa8b07SAndi Kleen } 32648faa8b07SAndi Kleen 3265a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3266a3437870SNishanth Aravamudan *mhp = 0; 3267a3437870SNishanth Aravamudan 32688faa8b07SAndi Kleen /* 32698faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 32708faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 32718faa8b07SAndi Kleen * use the bootmem allocator. 32728faa8b07SAndi Kleen */ 327347d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 32748faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 32758faa8b07SAndi Kleen 32768faa8b07SAndi Kleen last_mhp = mhp; 32778faa8b07SAndi Kleen 3278a3437870SNishanth Aravamudan return 1; 3279a3437870SNishanth Aravamudan } 3280282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3281e11bfbfcSNick Piggin 3282282f4214SMike Kravetz /* 3283282f4214SMike Kravetz * hugepagesz command line processing 3284282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3285282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3286282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3287282f4214SMike Kravetz * hugepagesz argument was valid. 3288282f4214SMike Kravetz */ 3289359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3290e11bfbfcSNick Piggin { 3291359f2544SMike Kravetz unsigned long size; 3292282f4214SMike Kravetz struct hstate *h; 3293282f4214SMike Kravetz 3294282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3295359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3296359f2544SMike Kravetz 3297359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3298282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3299359f2544SMike Kravetz return 0; 3300359f2544SMike Kravetz } 3301359f2544SMike Kravetz 3302282f4214SMike Kravetz h = size_to_hstate(size); 3303282f4214SMike Kravetz if (h) { 3304282f4214SMike Kravetz /* 3305282f4214SMike Kravetz * hstate for this size already exists. This is normally 3306282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3307282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3308282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3309282f4214SMike Kravetz * previously specified. 3310282f4214SMike Kravetz */ 3311282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3312282f4214SMike Kravetz default_hstate.max_huge_pages) { 3313282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 331438237830SMike Kravetz return 0; 331538237830SMike Kravetz } 331638237830SMike Kravetz 3317282f4214SMike Kravetz /* 3318282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3319282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3320282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3321282f4214SMike Kravetz */ 3322282f4214SMike Kravetz parsed_hstate = h; 3323282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3324e11bfbfcSNick Piggin return 1; 3325e11bfbfcSNick Piggin } 3326282f4214SMike Kravetz 3327359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3328282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3329359f2544SMike Kravetz return 1; 3330359f2544SMike Kravetz } 3331359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3332359f2544SMike Kravetz 3333282f4214SMike Kravetz /* 3334282f4214SMike Kravetz * default_hugepagesz command line input 3335282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3336282f4214SMike Kravetz */ 3337ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3338e11bfbfcSNick Piggin { 3339ae94da89SMike Kravetz unsigned long size; 3340ae94da89SMike Kravetz 3341282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3342282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3343282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3344282f4214SMike Kravetz return 0; 3345282f4214SMike Kravetz } 3346282f4214SMike Kravetz 3347282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3348282f4214SMike Kravetz 3349282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3350282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3351282f4214SMike Kravetz return 0; 3352282f4214SMike Kravetz } 3353282f4214SMike Kravetz 3354282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3355282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3356282f4214SMike Kravetz parsed_default_hugepagesz = true; 3357282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3358282f4214SMike Kravetz 3359282f4214SMike Kravetz /* 3360282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3361282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3362282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3363282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3364282f4214SMike Kravetz * allocated here from bootmem allocator. 3365282f4214SMike Kravetz */ 3366282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3367282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3368282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3369282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3370282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3371282f4214SMike Kravetz } 3372282f4214SMike Kravetz 3373e11bfbfcSNick Piggin return 1; 3374e11bfbfcSNick Piggin } 3375ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3376a3437870SNishanth Aravamudan 33778ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 33788a213460SNishanth Aravamudan { 33798a213460SNishanth Aravamudan int node; 33808a213460SNishanth Aravamudan unsigned int nr = 0; 33818ca39e68SMuchun Song nodemask_t *mpol_allowed; 33828ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 33838ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 33848a213460SNishanth Aravamudan 33858ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 33868ca39e68SMuchun Song 33878ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3388c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 33898a213460SNishanth Aravamudan nr += array[node]; 33908ca39e68SMuchun Song } 33918a213460SNishanth Aravamudan 33928a213460SNishanth Aravamudan return nr; 33938a213460SNishanth Aravamudan } 33948a213460SNishanth Aravamudan 33958a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 339617743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 339717743798SMuchun Song void *buffer, size_t *length, 339817743798SMuchun Song loff_t *ppos, unsigned long *out) 339917743798SMuchun Song { 340017743798SMuchun Song struct ctl_table dup_table; 340117743798SMuchun Song 340217743798SMuchun Song /* 340317743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 340417743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 340517743798SMuchun Song */ 340617743798SMuchun Song dup_table = *table; 340717743798SMuchun Song dup_table.data = out; 340817743798SMuchun Song 340917743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 341017743798SMuchun Song } 341117743798SMuchun Song 341206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 341306808b08SLee Schermerhorn struct ctl_table *table, int write, 341432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34151da177e4SLinus Torvalds { 3416e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3417238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 341808d4a246SMichal Hocko int ret; 3419e5ff2159SAndi Kleen 3420457c1b27SNishanth Aravamudan if (!hugepages_supported()) 342186613628SJan Stancek return -EOPNOTSUPP; 3422457c1b27SNishanth Aravamudan 342317743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 342417743798SMuchun Song &tmp); 342508d4a246SMichal Hocko if (ret) 342608d4a246SMichal Hocko goto out; 3427e5ff2159SAndi Kleen 3428238d3c13SDavid Rientjes if (write) 3429238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3430238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 343108d4a246SMichal Hocko out: 343208d4a246SMichal Hocko return ret; 34331da177e4SLinus Torvalds } 3434396faf03SMel Gorman 343506808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 343632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 343706808b08SLee Schermerhorn { 343806808b08SLee Schermerhorn 343906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 344006808b08SLee Schermerhorn buffer, length, ppos); 344106808b08SLee Schermerhorn } 344206808b08SLee Schermerhorn 344306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 344406808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 344532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 344606808b08SLee Schermerhorn { 344706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 344806808b08SLee Schermerhorn buffer, length, ppos); 344906808b08SLee Schermerhorn } 345006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 345106808b08SLee Schermerhorn 3452a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 345332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3454a3d0c6aaSNishanth Aravamudan { 3455a5516438SAndi Kleen struct hstate *h = &default_hstate; 3456e5ff2159SAndi Kleen unsigned long tmp; 345708d4a246SMichal Hocko int ret; 3458e5ff2159SAndi Kleen 3459457c1b27SNishanth Aravamudan if (!hugepages_supported()) 346086613628SJan Stancek return -EOPNOTSUPP; 3461457c1b27SNishanth Aravamudan 3462e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3463e5ff2159SAndi Kleen 3464bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3465adbe8726SEric B Munson return -EINVAL; 3466adbe8726SEric B Munson 346717743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 346817743798SMuchun Song &tmp); 346908d4a246SMichal Hocko if (ret) 347008d4a246SMichal Hocko goto out; 3471e5ff2159SAndi Kleen 3472e5ff2159SAndi Kleen if (write) { 3473064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3474e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3475a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3476e5ff2159SAndi Kleen } 347708d4a246SMichal Hocko out: 347808d4a246SMichal Hocko return ret; 3479a3d0c6aaSNishanth Aravamudan } 3480a3d0c6aaSNishanth Aravamudan 34811da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 34821da177e4SLinus Torvalds 3483e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 34841da177e4SLinus Torvalds { 3485fcb2b0c5SRoman Gushchin struct hstate *h; 3486fcb2b0c5SRoman Gushchin unsigned long total = 0; 3487fcb2b0c5SRoman Gushchin 3488457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3489457c1b27SNishanth Aravamudan return; 3490fcb2b0c5SRoman Gushchin 3491fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3492fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3493fcb2b0c5SRoman Gushchin 3494aca78307SMiaohe Lin total += huge_page_size(h) * count; 3495fcb2b0c5SRoman Gushchin 3496fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3497e1759c21SAlexey Dobriyan seq_printf(m, 34981da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 34991da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3500b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35017893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35024f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3503fcb2b0c5SRoman Gushchin count, 3504a5516438SAndi Kleen h->free_huge_pages, 3505a5516438SAndi Kleen h->resv_huge_pages, 3506a5516438SAndi Kleen h->surplus_huge_pages, 3507aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3508fcb2b0c5SRoman Gushchin } 3509fcb2b0c5SRoman Gushchin 3510aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 35111da177e4SLinus Torvalds } 35121da177e4SLinus Torvalds 35137981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 35141da177e4SLinus Torvalds { 3515a5516438SAndi Kleen struct hstate *h = &default_hstate; 35167981593bSJoe Perches 3517457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3518457c1b27SNishanth Aravamudan return 0; 35197981593bSJoe Perches 35207981593bSJoe Perches return sysfs_emit_at(buf, len, 35211da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3522a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3523a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3524a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3525a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3526a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35271da177e4SLinus Torvalds } 35281da177e4SLinus Torvalds 3529949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3530949f7ec5SDavid Rientjes { 3531949f7ec5SDavid Rientjes struct hstate *h; 3532949f7ec5SDavid Rientjes int nid; 3533949f7ec5SDavid Rientjes 3534457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3535457c1b27SNishanth Aravamudan return; 3536457c1b27SNishanth Aravamudan 3537949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3538949f7ec5SDavid Rientjes for_each_hstate(h) 3539949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3540949f7ec5SDavid Rientjes nid, 3541949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3542949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3543949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3544aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3545949f7ec5SDavid Rientjes } 3546949f7ec5SDavid Rientjes 35475d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 35485d317b2bSNaoya Horiguchi { 35495d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 35505d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 35515d317b2bSNaoya Horiguchi } 35525d317b2bSNaoya Horiguchi 35531da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 35541da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 35551da177e4SLinus Torvalds { 3556d0028588SWanpeng Li struct hstate *h; 3557d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3558d0028588SWanpeng Li 3559d0028588SWanpeng Li for_each_hstate(h) 3560d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3561d0028588SWanpeng Li return nr_total_pages; 35621da177e4SLinus Torvalds } 35631da177e4SLinus Torvalds 3564a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3565fc1b8a73SMel Gorman { 3566fc1b8a73SMel Gorman int ret = -ENOMEM; 3567fc1b8a73SMel Gorman 35680aa7f354SMiaohe Lin if (!delta) 35690aa7f354SMiaohe Lin return 0; 35700aa7f354SMiaohe Lin 3571fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3572fc1b8a73SMel Gorman /* 3573fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3574fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3575fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3576fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3577fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3578fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3579fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3580fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3581fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3582fc1b8a73SMel Gorman * 3583fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3584fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3585fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3586fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3587fc1b8a73SMel Gorman * semantics that cpuset has. 35888ca39e68SMuchun Song * 35898ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 35908ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 35918ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 35928ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 35938ca39e68SMuchun Song * above. 3594fc1b8a73SMel Gorman */ 3595fc1b8a73SMel Gorman if (delta > 0) { 3596a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3597fc1b8a73SMel Gorman goto out; 3598fc1b8a73SMel Gorman 35998ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3600a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3601fc1b8a73SMel Gorman goto out; 3602fc1b8a73SMel Gorman } 3603fc1b8a73SMel Gorman } 3604fc1b8a73SMel Gorman 3605fc1b8a73SMel Gorman ret = 0; 3606fc1b8a73SMel Gorman if (delta < 0) 3607a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3608fc1b8a73SMel Gorman 3609fc1b8a73SMel Gorman out: 3610fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3611fc1b8a73SMel Gorman return ret; 3612fc1b8a73SMel Gorman } 3613fc1b8a73SMel Gorman 361484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 361584afd99bSAndy Whitcroft { 3616f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 361784afd99bSAndy Whitcroft 361884afd99bSAndy Whitcroft /* 361984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 362084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 362184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 362225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 362384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 362484afd99bSAndy Whitcroft * new reference here without additional locking. 362584afd99bSAndy Whitcroft */ 36264e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3627f522c3acSJoonsoo Kim kref_get(&resv->refs); 362884afd99bSAndy Whitcroft } 362984afd99bSAndy Whitcroft 3630a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3631a1e78772SMel Gorman { 3632a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3633f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 363490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36354e35f483SJoonsoo Kim unsigned long reserve, start, end; 36361c5ecae3SMike Kravetz long gbl_reserve; 363784afd99bSAndy Whitcroft 36384e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36394e35f483SJoonsoo Kim return; 36404e35f483SJoonsoo Kim 3641a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3642a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 364384afd99bSAndy Whitcroft 36444e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3645e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36467251ff78SAdam Litke if (reserve) { 36471c5ecae3SMike Kravetz /* 36481c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36491c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36501c5ecae3SMike Kravetz */ 36511c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36521c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36537251ff78SAdam Litke } 3654e9fe92aeSMina Almasry 3655e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3656a1e78772SMel Gorman } 3657a1e78772SMel Gorman 365831383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 365931383c68SDan Williams { 366031383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 366131383c68SDan Williams return -EINVAL; 366231383c68SDan Williams return 0; 366331383c68SDan Williams } 366431383c68SDan Williams 366505ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 366605ea8860SDan Williams { 3667aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 366805ea8860SDan Williams } 366905ea8860SDan Williams 36701da177e4SLinus Torvalds /* 36711da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 36721da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 36736c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 36741da177e4SLinus Torvalds * this far. 36751da177e4SLinus Torvalds */ 3676b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 36771da177e4SLinus Torvalds { 36781da177e4SLinus Torvalds BUG(); 3679d0217ac0SNick Piggin return 0; 36801da177e4SLinus Torvalds } 36811da177e4SLinus Torvalds 3682eec3636aSJane Chu /* 3683eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3684eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3685eec3636aSJane Chu * This is because under System V memory model, mappings created via 3686eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3687eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3688eec3636aSJane Chu */ 3689f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3690d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 369184afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3692a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3693dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 369405ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 36951da177e4SLinus Torvalds }; 36961da177e4SLinus Torvalds 36971e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 36981e8f889bSDavid Gibson int writable) 369963551ae0SDavid Gibson { 370063551ae0SDavid Gibson pte_t entry; 370163551ae0SDavid Gibson 37021e8f889bSDavid Gibson if (writable) { 3703106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3704106c992aSGerald Schaefer vma->vm_page_prot))); 370563551ae0SDavid Gibson } else { 3706106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3707106c992aSGerald Schaefer vma->vm_page_prot)); 370863551ae0SDavid Gibson } 370963551ae0SDavid Gibson entry = pte_mkyoung(entry); 371063551ae0SDavid Gibson entry = pte_mkhuge(entry); 3711d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 371263551ae0SDavid Gibson 371363551ae0SDavid Gibson return entry; 371463551ae0SDavid Gibson } 371563551ae0SDavid Gibson 37161e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37171e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37181e8f889bSDavid Gibson { 37191e8f889bSDavid Gibson pte_t entry; 37201e8f889bSDavid Gibson 3721106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 372232f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37234b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37241e8f889bSDavid Gibson } 37251e8f889bSDavid Gibson 3726d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37274a705fefSNaoya Horiguchi { 37284a705fefSNaoya Horiguchi swp_entry_t swp; 37294a705fefSNaoya Horiguchi 37304a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3731d5ed7444SAneesh Kumar K.V return false; 37324a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3733d79d176aSBaoquan He if (is_migration_entry(swp)) 3734d5ed7444SAneesh Kumar K.V return true; 37354a705fefSNaoya Horiguchi else 3736d5ed7444SAneesh Kumar K.V return false; 37374a705fefSNaoya Horiguchi } 37384a705fefSNaoya Horiguchi 37393e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 37404a705fefSNaoya Horiguchi { 37414a705fefSNaoya Horiguchi swp_entry_t swp; 37424a705fefSNaoya Horiguchi 37434a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37443e5c3600SBaoquan He return false; 37454a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3746d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 37473e5c3600SBaoquan He return true; 37484a705fefSNaoya Horiguchi else 37493e5c3600SBaoquan He return false; 37504a705fefSNaoya Horiguchi } 37511e8f889bSDavid Gibson 37524eae4efaSPeter Xu static void 37534eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 37544eae4efaSPeter Xu struct page *new_page) 37554eae4efaSPeter Xu { 37564eae4efaSPeter Xu __SetPageUptodate(new_page); 37574eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 37584eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 37594eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 37604eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 37614eae4efaSPeter Xu SetHPageMigratable(new_page); 37624eae4efaSPeter Xu } 37634eae4efaSPeter Xu 376463551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 376563551ae0SDavid Gibson struct vm_area_struct *vma) 376663551ae0SDavid Gibson { 37675e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 376863551ae0SDavid Gibson struct page *ptepage; 37691c59827dSHugh Dickins unsigned long addr; 3770ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 3771a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3772a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 37734eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 3774c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3775ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3776e8569dd2SAndreas Sandberg int ret = 0; 37771e8f889bSDavid Gibson 3778ac46d4f3SJérôme Glisse if (cow) { 37797269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 37806f4f13e8SJérôme Glisse vma->vm_start, 3781ac46d4f3SJérôme Glisse vma->vm_end); 3782ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3783c0d0381aSMike Kravetz } else { 3784c0d0381aSMike Kravetz /* 3785c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3786c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3787c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3788c0d0381aSMike Kravetz * huge_pmd_unshare. 3789c0d0381aSMike Kravetz */ 3790c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3791ac46d4f3SJérôme Glisse } 3792e8569dd2SAndreas Sandberg 3793a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3794cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 37957868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3796c74df32cSHugh Dickins if (!src_pte) 3797c74df32cSHugh Dickins continue; 3798aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 3799e8569dd2SAndreas Sandberg if (!dst_pte) { 3800e8569dd2SAndreas Sandberg ret = -ENOMEM; 3801e8569dd2SAndreas Sandberg break; 3802e8569dd2SAndreas Sandberg } 3803c5c99429SLarry Woodman 38045e41540cSMike Kravetz /* 38055e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38065e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38075e41540cSMike Kravetz * 38085e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38095e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38105e41540cSMike Kravetz * Check here before taking page table lock, and once again 38115e41540cSMike Kravetz * after taking the lock below. 38125e41540cSMike Kravetz */ 38135e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38145e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3815c5c99429SLarry Woodman continue; 3816c5c99429SLarry Woodman 3817cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3818cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3819cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38204a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38215e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38224eae4efaSPeter Xu again: 38235e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38245e41540cSMike Kravetz /* 38255e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38265e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38275e41540cSMike Kravetz * sharing with another vma. 38285e41540cSMike Kravetz */ 38294a705fefSNaoya Horiguchi ; 38304a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38314a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38324a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38334a705fefSNaoya Horiguchi 38344a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38354a705fefSNaoya Horiguchi /* 38364a705fefSNaoya Horiguchi * COW mappings require pages in both 38374a705fefSNaoya Horiguchi * parent and child to be set to read. 38384a705fefSNaoya Horiguchi */ 38394a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38404a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3841e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3842e5251fd4SPunit Agrawal entry, sz); 38434a705fefSNaoya Horiguchi } 3844e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38454a705fefSNaoya Horiguchi } else { 38464eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 38474eae4efaSPeter Xu ptepage = pte_page(entry); 38484eae4efaSPeter Xu get_page(ptepage); 38494eae4efaSPeter Xu 38504eae4efaSPeter Xu /* 38514eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 38524eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 38534eae4efaSPeter Xu * COW earlier during fork. 38544eae4efaSPeter Xu * 38554eae4efaSPeter Xu * When pre-allocating the page or copying data, we 38564eae4efaSPeter Xu * need to be without the pgtable locks since we could 38574eae4efaSPeter Xu * sleep during the process. 38584eae4efaSPeter Xu */ 38594eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 38604eae4efaSPeter Xu pte_t src_pte_old = entry; 38614eae4efaSPeter Xu struct page *new; 38624eae4efaSPeter Xu 38634eae4efaSPeter Xu spin_unlock(src_ptl); 38644eae4efaSPeter Xu spin_unlock(dst_ptl); 38654eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 38664eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 38674eae4efaSPeter Xu if (IS_ERR(new)) { 38684eae4efaSPeter Xu put_page(ptepage); 38694eae4efaSPeter Xu ret = PTR_ERR(new); 38704eae4efaSPeter Xu break; 38714eae4efaSPeter Xu } 38724eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 38734eae4efaSPeter Xu npages); 38744eae4efaSPeter Xu put_page(ptepage); 38754eae4efaSPeter Xu 38764eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 38774eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 38784eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 38794eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38804eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 38814eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 38824eae4efaSPeter Xu put_page(new); 38834eae4efaSPeter Xu /* dst_entry won't change as in child */ 38844eae4efaSPeter Xu goto again; 38854eae4efaSPeter Xu } 38864eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 38874eae4efaSPeter Xu spin_unlock(src_ptl); 38884eae4efaSPeter Xu spin_unlock(dst_ptl); 38894eae4efaSPeter Xu continue; 38904eae4efaSPeter Xu } 38914eae4efaSPeter Xu 389234ee645eSJoerg Roedel if (cow) { 38930f10851eSJérôme Glisse /* 38940f10851eSJérôme Glisse * No need to notify as we are downgrading page 38950f10851eSJérôme Glisse * table protection not changing it to point 38960f10851eSJérôme Glisse * to a new page. 38970f10851eSJérôme Glisse * 3898ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 38990f10851eSJérôme Glisse */ 39007f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 390134ee645eSJoerg Roedel } 39024eae4efaSPeter Xu 390353f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 390463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 39054eae4efaSPeter Xu hugetlb_count_add(npages, dst); 39061c59827dSHugh Dickins } 3907cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3908cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 390963551ae0SDavid Gibson } 391063551ae0SDavid Gibson 3911e8569dd2SAndreas Sandberg if (cow) 3912ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3913c0d0381aSMike Kravetz else 3914c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3915e8569dd2SAndreas Sandberg 3916e8569dd2SAndreas Sandberg return ret; 391763551ae0SDavid Gibson } 391863551ae0SDavid Gibson 391924669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 392024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 392124669e58SAneesh Kumar K.V struct page *ref_page) 392263551ae0SDavid Gibson { 392363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 392463551ae0SDavid Gibson unsigned long address; 3925c7546f8fSDavid Gibson pte_t *ptep; 392663551ae0SDavid Gibson pte_t pte; 3927cb900f41SKirill A. Shutemov spinlock_t *ptl; 392863551ae0SDavid Gibson struct page *page; 3929a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3930a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3931ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3932a5516438SAndi Kleen 393363551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3934a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3935a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 393663551ae0SDavid Gibson 393707e32661SAneesh Kumar K.V /* 393807e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 393907e32661SAneesh Kumar K.V * to huge page. 394007e32661SAneesh Kumar K.V */ 3941ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 394224669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3943dff11abeSMike Kravetz 3944dff11abeSMike Kravetz /* 3945dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3946dff11abeSMike Kravetz */ 39476f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39486f4f13e8SJérôme Glisse end); 3949ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3950ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3951569f48b8SHillf Danton address = start; 3952569f48b8SHillf Danton for (; address < end; address += sz) { 39537868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3954c7546f8fSDavid Gibson if (!ptep) 3955c7546f8fSDavid Gibson continue; 3956c7546f8fSDavid Gibson 3957cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 395834ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 395931d49da5SAneesh Kumar K.V spin_unlock(ptl); 3960dff11abeSMike Kravetz /* 3961dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3962dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3963dff11abeSMike Kravetz */ 396431d49da5SAneesh Kumar K.V continue; 396531d49da5SAneesh Kumar K.V } 396639dde65cSChen, Kenneth W 39676629326bSHillf Danton pte = huge_ptep_get(ptep); 396831d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 396931d49da5SAneesh Kumar K.V spin_unlock(ptl); 397031d49da5SAneesh Kumar K.V continue; 397131d49da5SAneesh Kumar K.V } 39726629326bSHillf Danton 39736629326bSHillf Danton /* 39749fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39759fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39766629326bSHillf Danton */ 39779fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39789386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 397931d49da5SAneesh Kumar K.V spin_unlock(ptl); 398031d49da5SAneesh Kumar K.V continue; 39818c4894c6SNaoya Horiguchi } 39826629326bSHillf Danton 39836629326bSHillf Danton page = pte_page(pte); 398404f2cbe3SMel Gorman /* 398504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 398604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 398704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 398804f2cbe3SMel Gorman */ 398904f2cbe3SMel Gorman if (ref_page) { 399031d49da5SAneesh Kumar K.V if (page != ref_page) { 399131d49da5SAneesh Kumar K.V spin_unlock(ptl); 399231d49da5SAneesh Kumar K.V continue; 399331d49da5SAneesh Kumar K.V } 399404f2cbe3SMel Gorman /* 399504f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 399604f2cbe3SMel Gorman * future faults in this VMA will fail rather than 399704f2cbe3SMel Gorman * looking like data was lost 399804f2cbe3SMel Gorman */ 399904f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 400004f2cbe3SMel Gorman } 400104f2cbe3SMel Gorman 4002c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4003b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4004106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 40056649a386SKen Chen set_page_dirty(page); 40069e81130bSHillf Danton 40075d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4008d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 400931d49da5SAneesh Kumar K.V 4010cb900f41SKirill A. Shutemov spin_unlock(ptl); 4011e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 401224669e58SAneesh Kumar K.V /* 401331d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 401424669e58SAneesh Kumar K.V */ 401531d49da5SAneesh Kumar K.V if (ref_page) 401631d49da5SAneesh Kumar K.V break; 4017fe1668aeSChen, Kenneth W } 4018ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 401924669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 40201da177e4SLinus Torvalds } 402163551ae0SDavid Gibson 4022d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4023d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4024d833352aSMel Gorman unsigned long end, struct page *ref_page) 4025d833352aSMel Gorman { 4026d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4027d833352aSMel Gorman 4028d833352aSMel Gorman /* 4029d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4030d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4031d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4032d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4033d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4034c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4035d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4036c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4037d833352aSMel Gorman */ 4038d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4039d833352aSMel Gorman } 4040d833352aSMel Gorman 4041502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 404204f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4043502717f4SChen, Kenneth W { 404424669e58SAneesh Kumar K.V struct mmu_gather tlb; 4045dff11abeSMike Kravetz 4046a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 404724669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4048ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4049502717f4SChen, Kenneth W } 4050502717f4SChen, Kenneth W 405104f2cbe3SMel Gorman /* 405204f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4053578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 405404f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 405504f2cbe3SMel Gorman * same region. 405604f2cbe3SMel Gorman */ 40572f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40582a4b3dedSHarvey Harrison struct page *page, unsigned long address) 405904f2cbe3SMel Gorman { 40607526674dSAdam Litke struct hstate *h = hstate_vma(vma); 406104f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 406204f2cbe3SMel Gorman struct address_space *mapping; 406304f2cbe3SMel Gorman pgoff_t pgoff; 406404f2cbe3SMel Gorman 406504f2cbe3SMel Gorman /* 406604f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 406704f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 406804f2cbe3SMel Gorman */ 40697526674dSAdam Litke address = address & huge_page_mask(h); 407036e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 407136e4f20aSMichal Hocko vma->vm_pgoff; 407293c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 407304f2cbe3SMel Gorman 40744eb2b1dcSMel Gorman /* 40754eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40764eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40774eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40784eb2b1dcSMel Gorman */ 407983cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40806b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 408104f2cbe3SMel Gorman /* Do not unmap the current VMA */ 408204f2cbe3SMel Gorman if (iter_vma == vma) 408304f2cbe3SMel Gorman continue; 408404f2cbe3SMel Gorman 408504f2cbe3SMel Gorman /* 40862f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40872f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 40882f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 40892f84a899SMel Gorman */ 40902f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 40912f84a899SMel Gorman continue; 40922f84a899SMel Gorman 40932f84a899SMel Gorman /* 409404f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 409504f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 409604f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 409704f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 409804f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 409904f2cbe3SMel Gorman */ 410004f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 410124669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 410224669e58SAneesh Kumar K.V address + huge_page_size(h), page); 410304f2cbe3SMel Gorman } 410483cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 410504f2cbe3SMel Gorman } 410604f2cbe3SMel Gorman 41070fe6e20bSNaoya Horiguchi /* 41080fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4109ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4110ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4111ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 41120fe6e20bSNaoya Horiguchi */ 41132b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4114974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4115cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 41161e8f889bSDavid Gibson { 41173999f52eSAneesh Kumar K.V pte_t pte; 4118a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41191e8f889bSDavid Gibson struct page *old_page, *new_page; 41202b740303SSouptick Joarder int outside_reserve = 0; 41212b740303SSouptick Joarder vm_fault_t ret = 0; 4122974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4123ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 41241e8f889bSDavid Gibson 41253999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 41261e8f889bSDavid Gibson old_page = pte_page(pte); 41271e8f889bSDavid Gibson 412804f2cbe3SMel Gorman retry_avoidcopy: 41291e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 41301e8f889bSDavid Gibson * and just make the page writable */ 413137a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 41325a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 41335b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 413483c54070SNick Piggin return 0; 41351e8f889bSDavid Gibson } 41361e8f889bSDavid Gibson 413704f2cbe3SMel Gorman /* 413804f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 413904f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 414004f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 414104f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 414204f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 414304f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 414404f2cbe3SMel Gorman * of the full address range. 414504f2cbe3SMel Gorman */ 41465944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 414704f2cbe3SMel Gorman old_page != pagecache_page) 414804f2cbe3SMel Gorman outside_reserve = 1; 414904f2cbe3SMel Gorman 415009cbfeafSKirill A. Shutemov get_page(old_page); 4151b76c8cfbSLarry Woodman 4152ad4404a2SDavidlohr Bueso /* 4153ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4154ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4155ad4404a2SDavidlohr Bueso */ 4156cb900f41SKirill A. Shutemov spin_unlock(ptl); 41575b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41581e8f889bSDavid Gibson 41592fc39cecSAdam Litke if (IS_ERR(new_page)) { 416004f2cbe3SMel Gorman /* 416104f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 416204f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 416304f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 416404f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 416504f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 416604f2cbe3SMel Gorman */ 416704f2cbe3SMel Gorman if (outside_reserve) { 4168e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4169e7dd91c4SMike Kravetz pgoff_t idx; 4170e7dd91c4SMike Kravetz u32 hash; 4171e7dd91c4SMike Kravetz 417209cbfeafSKirill A. Shutemov put_page(old_page); 417304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4174e7dd91c4SMike Kravetz /* 4175e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4176e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4177e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4178e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4179e7dd91c4SMike Kravetz * PMD sharing. 4180e7dd91c4SMike Kravetz * 4181e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4182e7dd91c4SMike Kravetz */ 4183e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4184e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4185e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4186e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4187e7dd91c4SMike Kravetz 41885b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4189e7dd91c4SMike Kravetz 4190e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4191e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4192cb900f41SKirill A. Shutemov spin_lock(ptl); 41935b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4194a9af0c5dSNaoya Horiguchi if (likely(ptep && 4195a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 419604f2cbe3SMel Gorman goto retry_avoidcopy; 4197a734bcc8SHillf Danton /* 4198cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4199cb900f41SKirill A. Shutemov * lock, and our job is done. 4200a734bcc8SHillf Danton */ 4201a734bcc8SHillf Danton return 0; 420204f2cbe3SMel Gorman } 420304f2cbe3SMel Gorman 42042b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4205ad4404a2SDavidlohr Bueso goto out_release_old; 42061e8f889bSDavid Gibson } 42071e8f889bSDavid Gibson 42080fe6e20bSNaoya Horiguchi /* 42090fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 42100fe6e20bSNaoya Horiguchi * anon_vma prepared. 42110fe6e20bSNaoya Horiguchi */ 421244e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4213ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4214ad4404a2SDavidlohr Bueso goto out_release_all; 421544e2aa93SDean Nelson } 42160fe6e20bSNaoya Horiguchi 4217974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 421847ad8475SAndrea Arcangeli pages_per_huge_page(h)); 42190ed361deSNick Piggin __SetPageUptodate(new_page); 42201e8f889bSDavid Gibson 42217269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 42226f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4223ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4224ad4404a2SDavidlohr Bueso 4225b76c8cfbSLarry Woodman /* 4226cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4227b76c8cfbSLarry Woodman * before the page tables are altered 4228b76c8cfbSLarry Woodman */ 4229cb900f41SKirill A. Shutemov spin_lock(ptl); 42305b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4231a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4232d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 423307443a85SJoonsoo Kim 42341e8f889bSDavid Gibson /* Break COW */ 42355b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4236ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 42375b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 42381e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4239d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 42405b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 42418f251a3dSMike Kravetz SetHPageMigratable(new_page); 42421e8f889bSDavid Gibson /* Make the old page be freed below */ 42431e8f889bSDavid Gibson new_page = old_page; 42441e8f889bSDavid Gibson } 4245cb900f41SKirill A. Shutemov spin_unlock(ptl); 4246ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4247ad4404a2SDavidlohr Bueso out_release_all: 42485b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 424909cbfeafSKirill A. Shutemov put_page(new_page); 4250ad4404a2SDavidlohr Bueso out_release_old: 425109cbfeafSKirill A. Shutemov put_page(old_page); 42528312034fSJoonsoo Kim 4253ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4254ad4404a2SDavidlohr Bueso return ret; 42551e8f889bSDavid Gibson } 42561e8f889bSDavid Gibson 425704f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4258a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4259a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 426004f2cbe3SMel Gorman { 426104f2cbe3SMel Gorman struct address_space *mapping; 4262e7c4b0bfSAndy Whitcroft pgoff_t idx; 426304f2cbe3SMel Gorman 426404f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4265a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 426604f2cbe3SMel Gorman 426704f2cbe3SMel Gorman return find_lock_page(mapping, idx); 426804f2cbe3SMel Gorman } 426904f2cbe3SMel Gorman 42703ae77f43SHugh Dickins /* 42713ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42723ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42733ae77f43SHugh Dickins */ 42743ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42752a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42762a15efc9SHugh Dickins { 42772a15efc9SHugh Dickins struct address_space *mapping; 42782a15efc9SHugh Dickins pgoff_t idx; 42792a15efc9SHugh Dickins struct page *page; 42802a15efc9SHugh Dickins 42812a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42822a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42832a15efc9SHugh Dickins 42842a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42852a15efc9SHugh Dickins if (page) 42862a15efc9SHugh Dickins put_page(page); 42872a15efc9SHugh Dickins return page != NULL; 42882a15efc9SHugh Dickins } 42892a15efc9SHugh Dickins 4290ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4291ab76ad54SMike Kravetz pgoff_t idx) 4292ab76ad54SMike Kravetz { 4293ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4294ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4295ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4296ab76ad54SMike Kravetz 4297ab76ad54SMike Kravetz if (err) 4298ab76ad54SMike Kravetz return err; 4299d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4300ab76ad54SMike Kravetz 430122146c3cSMike Kravetz /* 430222146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 430322146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 430422146c3cSMike Kravetz */ 430522146c3cSMike Kravetz set_page_dirty(page); 430622146c3cSMike Kravetz 4307ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4308ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4309ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4310ab76ad54SMike Kravetz return 0; 4311ab76ad54SMike Kravetz } 4312ab76ad54SMike Kravetz 43132b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 43142b740303SSouptick Joarder struct vm_area_struct *vma, 43158382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4316788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4317ac9b9c66SHugh Dickins { 4318a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43192b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4320409eb8c2SHillf Danton int anon_rmap = 0; 43214c887265SAdam Litke unsigned long size; 43224c887265SAdam Litke struct page *page; 43231e8f889bSDavid Gibson pte_t new_pte; 4324cb900f41SKirill A. Shutemov spinlock_t *ptl; 4325285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4326cb6acd01SMike Kravetz bool new_page = false; 43274c887265SAdam Litke 432804f2cbe3SMel Gorman /* 432904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 433004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 433125985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 433204f2cbe3SMel Gorman */ 433304f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4334910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 433504f2cbe3SMel Gorman current->pid); 433604f2cbe3SMel Gorman return ret; 433704f2cbe3SMel Gorman } 433804f2cbe3SMel Gorman 43394c887265SAdam Litke /* 434087bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 434187bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 434287bf91d3SMike Kravetz * once for faults beyond end of file. 43434c887265SAdam Litke */ 4344a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4345ebed4bfcSHugh Dickins if (idx >= size) 4346ebed4bfcSHugh Dickins goto out; 43471a1aad8aSMike Kravetz 434887bf91d3SMike Kravetz retry: 434987bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 435087bf91d3SMike Kravetz if (!page) { 43511a1aad8aSMike Kravetz /* 43521a1aad8aSMike Kravetz * Check for page in userfault range 43531a1aad8aSMike Kravetz */ 43541a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43551a1aad8aSMike Kravetz u32 hash; 43561a1aad8aSMike Kravetz struct vm_fault vmf = { 43571a1aad8aSMike Kravetz .vma = vma, 4358285b8dcaSHuang Ying .address = haddr, 43591a1aad8aSMike Kravetz .flags = flags, 43601a1aad8aSMike Kravetz /* 43611a1aad8aSMike Kravetz * Hard to debug if it ends up being 43621a1aad8aSMike Kravetz * used by a callee that assumes 43631a1aad8aSMike Kravetz * something about the other 43641a1aad8aSMike Kravetz * uninitialized fields... same as in 43651a1aad8aSMike Kravetz * memory.c 43661a1aad8aSMike Kravetz */ 43671a1aad8aSMike Kravetz }; 43681a1aad8aSMike Kravetz 43691a1aad8aSMike Kravetz /* 4370c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4371c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4372c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43731a1aad8aSMike Kravetz */ 4374188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43751a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4376c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43771a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4378c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 43791a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43801a1aad8aSMike Kravetz goto out; 43811a1aad8aSMike Kravetz } 43821a1aad8aSMike Kravetz 4383285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43842fc39cecSAdam Litke if (IS_ERR(page)) { 43854643d67eSMike Kravetz /* 43864643d67eSMike Kravetz * Returning error will result in faulting task being 43874643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43884643d67eSMike Kravetz * tasks from racing to fault in the same page which 43894643d67eSMike Kravetz * could result in false unable to allocate errors. 43904643d67eSMike Kravetz * Page migration does not take the fault mutex, but 43914643d67eSMike Kravetz * does a clear then write of pte's under page table 43924643d67eSMike Kravetz * lock. Page fault code could race with migration, 43934643d67eSMike Kravetz * notice the clear pte and try to allocate a page 43944643d67eSMike Kravetz * here. Before returning error, get ptl and make 43954643d67eSMike Kravetz * sure there really is no pte entry. 43964643d67eSMike Kravetz */ 43974643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 43984643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 43994643d67eSMike Kravetz ret = 0; 44004643d67eSMike Kravetz spin_unlock(ptl); 44014643d67eSMike Kravetz goto out; 44024643d67eSMike Kravetz } 44034643d67eSMike Kravetz spin_unlock(ptl); 44042b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 44056bda666aSChristoph Lameter goto out; 44066bda666aSChristoph Lameter } 440747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 44080ed361deSNick Piggin __SetPageUptodate(page); 4409cb6acd01SMike Kravetz new_page = true; 4410ac9b9c66SHugh Dickins 4411f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4412ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 44136bda666aSChristoph Lameter if (err) { 44146bda666aSChristoph Lameter put_page(page); 44156bda666aSChristoph Lameter if (err == -EEXIST) 44166bda666aSChristoph Lameter goto retry; 44176bda666aSChristoph Lameter goto out; 44186bda666aSChristoph Lameter } 441923be7468SMel Gorman } else { 44206bda666aSChristoph Lameter lock_page(page); 44210fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 44220fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 44230fe6e20bSNaoya Horiguchi goto backout_unlocked; 442423be7468SMel Gorman } 4425409eb8c2SHillf Danton anon_rmap = 1; 44260fe6e20bSNaoya Horiguchi } 44270fe6e20bSNaoya Horiguchi } else { 442857303d80SAndy Whitcroft /* 4429998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4430998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4431998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4432fd6a03edSNaoya Horiguchi */ 4433fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 44340eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4435972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4436fd6a03edSNaoya Horiguchi goto backout_unlocked; 44376bda666aSChristoph Lameter } 4438998b4382SNaoya Horiguchi } 44391e8f889bSDavid Gibson 444057303d80SAndy Whitcroft /* 444157303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 444257303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 444357303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 444457303d80SAndy Whitcroft * the spinlock. 444557303d80SAndy Whitcroft */ 44465e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4447285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44482b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44492b26736cSAndy Whitcroft goto backout_unlocked; 44502b26736cSAndy Whitcroft } 44515e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4452285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44535e911373SMike Kravetz } 445457303d80SAndy Whitcroft 44558bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 445683c54070SNick Piggin ret = 0; 44577f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44584c887265SAdam Litke goto backout; 44594c887265SAdam Litke 446007443a85SJoonsoo Kim if (anon_rmap) { 4461d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4462285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4463ac714904SChoi Gi-yong } else 446453f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44651e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44661e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4467285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44681e8f889bSDavid Gibson 44695d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4470788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44711e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4472974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44731e8f889bSDavid Gibson } 44741e8f889bSDavid Gibson 4475cb900f41SKirill A. Shutemov spin_unlock(ptl); 4476cb6acd01SMike Kravetz 4477cb6acd01SMike Kravetz /* 44788f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 44798f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 44808f251a3dSMike Kravetz * been isolated for migration. 4481cb6acd01SMike Kravetz */ 4482cb6acd01SMike Kravetz if (new_page) 44838f251a3dSMike Kravetz SetHPageMigratable(page); 4484cb6acd01SMike Kravetz 44854c887265SAdam Litke unlock_page(page); 44864c887265SAdam Litke out: 4487ac9b9c66SHugh Dickins return ret; 44884c887265SAdam Litke 44894c887265SAdam Litke backout: 4490cb900f41SKirill A. Shutemov spin_unlock(ptl); 44912b26736cSAndy Whitcroft backout_unlocked: 44924c887265SAdam Litke unlock_page(page); 4493285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 44944c887265SAdam Litke put_page(page); 44954c887265SAdam Litke goto out; 4496ac9b9c66SHugh Dickins } 4497ac9b9c66SHugh Dickins 44988382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4499188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45008382d914SDavidlohr Bueso { 45018382d914SDavidlohr Bueso unsigned long key[2]; 45028382d914SDavidlohr Bueso u32 hash; 45038382d914SDavidlohr Bueso 45048382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 45058382d914SDavidlohr Bueso key[1] = idx; 45068382d914SDavidlohr Bueso 450755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 45088382d914SDavidlohr Bueso 45098382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 45108382d914SDavidlohr Bueso } 45118382d914SDavidlohr Bueso #else 45128382d914SDavidlohr Bueso /* 45136c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 45148382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 45158382d914SDavidlohr Bueso */ 4516188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45178382d914SDavidlohr Bueso { 45188382d914SDavidlohr Bueso return 0; 45198382d914SDavidlohr Bueso } 45208382d914SDavidlohr Bueso #endif 45218382d914SDavidlohr Bueso 45222b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4523788c7df4SHugh Dickins unsigned long address, unsigned int flags) 452486e5216fSAdam Litke { 45258382d914SDavidlohr Bueso pte_t *ptep, entry; 4526cb900f41SKirill A. Shutemov spinlock_t *ptl; 45272b740303SSouptick Joarder vm_fault_t ret; 45288382d914SDavidlohr Bueso u32 hash; 45298382d914SDavidlohr Bueso pgoff_t idx; 45300fe6e20bSNaoya Horiguchi struct page *page = NULL; 453157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4532a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45338382d914SDavidlohr Bueso struct address_space *mapping; 45340f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4535285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 453686e5216fSAdam Litke 4537285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4538fd6a03edSNaoya Horiguchi if (ptep) { 4539c0d0381aSMike Kravetz /* 4540c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4541c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4542c0d0381aSMike Kravetz * not actually modifying content here. 4543c0d0381aSMike Kravetz */ 4544fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4545290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4546cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4547290408d4SNaoya Horiguchi return 0; 4548290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4549aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4550972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4551b43a9990SMike Kravetz } 45528382d914SDavidlohr Bueso 4553c0d0381aSMike Kravetz /* 4554c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 455587bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 455687bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 455787bf91d3SMike Kravetz * and making the ptep no longer valid. 455887bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4559c0d0381aSMike Kravetz * 4560c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4561c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4562c0d0381aSMike Kravetz * something has changed. 4563c0d0381aSMike Kravetz */ 4564ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4565c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4566aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 4567c0d0381aSMike Kravetz if (!ptep) { 4568c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4569c0d0381aSMike Kravetz return VM_FAULT_OOM; 4570c0d0381aSMike Kravetz } 4571ddeaab32SMike Kravetz 45723935baa9SDavid Gibson /* 45733935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45743935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45753935baa9SDavid Gibson * the same page in the page cache. 45763935baa9SDavid Gibson */ 4577c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4578188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4579c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45808382d914SDavidlohr Bueso 45817f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45827f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45838382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4584b4d1d99fSDavid Gibson goto out_mutex; 45853935baa9SDavid Gibson } 458686e5216fSAdam Litke 458783c54070SNick Piggin ret = 0; 45881e8f889bSDavid Gibson 458957303d80SAndy Whitcroft /* 45900f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45910f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45927c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45937c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 45947c8de358SEthon Paul * properly handle it. 45950f792cf9SNaoya Horiguchi */ 45960f792cf9SNaoya Horiguchi if (!pte_present(entry)) 45970f792cf9SNaoya Horiguchi goto out_mutex; 45980f792cf9SNaoya Horiguchi 45990f792cf9SNaoya Horiguchi /* 460057303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 460157303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 460257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 460357303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 460457303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 460557303d80SAndy Whitcroft * consumed. 460657303d80SAndy Whitcroft */ 4607106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4608285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 46092b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4610b4d1d99fSDavid Gibson goto out_mutex; 46112b26736cSAndy Whitcroft } 46125e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4613285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 461457303d80SAndy Whitcroft 4615f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 461657303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4617285b8dcaSHuang Ying vma, haddr); 461857303d80SAndy Whitcroft } 461957303d80SAndy Whitcroft 46200f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 46210fe6e20bSNaoya Horiguchi 46221e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4623b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4624cb900f41SKirill A. Shutemov goto out_ptl; 4625b4d1d99fSDavid Gibson 46260f792cf9SNaoya Horiguchi /* 46270f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 46280f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 46290f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 46300f792cf9SNaoya Horiguchi */ 46310f792cf9SNaoya Horiguchi page = pte_page(entry); 46320f792cf9SNaoya Horiguchi if (page != pagecache_page) 46330f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 46340f792cf9SNaoya Horiguchi need_wait_lock = 1; 46350f792cf9SNaoya Horiguchi goto out_ptl; 46360f792cf9SNaoya Horiguchi } 46370f792cf9SNaoya Horiguchi 46380f792cf9SNaoya Horiguchi get_page(page); 4639b4d1d99fSDavid Gibson 4640788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4641106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4642974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4643cb900f41SKirill A. Shutemov pagecache_page, ptl); 46440f792cf9SNaoya Horiguchi goto out_put_page; 4645b4d1d99fSDavid Gibson } 4646106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4647b4d1d99fSDavid Gibson } 4648b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4649285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4650788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4651285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46520f792cf9SNaoya Horiguchi out_put_page: 46530f792cf9SNaoya Horiguchi if (page != pagecache_page) 46540f792cf9SNaoya Horiguchi unlock_page(page); 46550f792cf9SNaoya Horiguchi put_page(page); 4656cb900f41SKirill A. Shutemov out_ptl: 4657cb900f41SKirill A. Shutemov spin_unlock(ptl); 465857303d80SAndy Whitcroft 465957303d80SAndy Whitcroft if (pagecache_page) { 466057303d80SAndy Whitcroft unlock_page(pagecache_page); 466157303d80SAndy Whitcroft put_page(pagecache_page); 466257303d80SAndy Whitcroft } 4663b4d1d99fSDavid Gibson out_mutex: 4664c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4665c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46660f792cf9SNaoya Horiguchi /* 46670f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46680f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46690f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46700f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46710f792cf9SNaoya Horiguchi * here without taking refcount. 46720f792cf9SNaoya Horiguchi */ 46730f792cf9SNaoya Horiguchi if (need_wait_lock) 46740f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46751e8f889bSDavid Gibson return ret; 467686e5216fSAdam Litke } 467786e5216fSAdam Litke 46788fb5debcSMike Kravetz /* 46798fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46808fb5debcSMike Kravetz * modifications for huge pages. 46818fb5debcSMike Kravetz */ 46828fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46838fb5debcSMike Kravetz pte_t *dst_pte, 46848fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46858fb5debcSMike Kravetz unsigned long dst_addr, 46868fb5debcSMike Kravetz unsigned long src_addr, 46878fb5debcSMike Kravetz struct page **pagep) 46888fb5debcSMike Kravetz { 46891e392147SAndrea Arcangeli struct address_space *mapping; 46901e392147SAndrea Arcangeli pgoff_t idx; 46911e392147SAndrea Arcangeli unsigned long size; 46921c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46938fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 46948fb5debcSMike Kravetz pte_t _dst_pte; 46958fb5debcSMike Kravetz spinlock_t *ptl; 46968fb5debcSMike Kravetz int ret; 46978fb5debcSMike Kravetz struct page *page; 46988fb5debcSMike Kravetz 46998fb5debcSMike Kravetz if (!*pagep) { 47008fb5debcSMike Kravetz ret = -ENOMEM; 47018fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 47028fb5debcSMike Kravetz if (IS_ERR(page)) 47038fb5debcSMike Kravetz goto out; 47048fb5debcSMike Kravetz 47058fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 47068fb5debcSMike Kravetz (const void __user *) src_addr, 4707810a56b9SMike Kravetz pages_per_huge_page(h), false); 47088fb5debcSMike Kravetz 4709c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 47108fb5debcSMike Kravetz if (unlikely(ret)) { 47119e368259SAndrea Arcangeli ret = -ENOENT; 47128fb5debcSMike Kravetz *pagep = page; 47138fb5debcSMike Kravetz /* don't free the page */ 47148fb5debcSMike Kravetz goto out; 47158fb5debcSMike Kravetz } 47168fb5debcSMike Kravetz } else { 47178fb5debcSMike Kravetz page = *pagep; 47188fb5debcSMike Kravetz *pagep = NULL; 47198fb5debcSMike Kravetz } 47208fb5debcSMike Kravetz 47218fb5debcSMike Kravetz /* 47228fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 47238fb5debcSMike Kravetz * preceding stores to the page contents become visible before 47248fb5debcSMike Kravetz * the set_pte_at() write. 47258fb5debcSMike Kravetz */ 47268fb5debcSMike Kravetz __SetPageUptodate(page); 47278fb5debcSMike Kravetz 47281e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 47291e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 47301e392147SAndrea Arcangeli 47311c9e8defSMike Kravetz /* 47321c9e8defSMike Kravetz * If shared, add to page cache 47331c9e8defSMike Kravetz */ 47341c9e8defSMike Kravetz if (vm_shared) { 47351e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47361e392147SAndrea Arcangeli ret = -EFAULT; 47371e392147SAndrea Arcangeli if (idx >= size) 47381e392147SAndrea Arcangeli goto out_release_nounlock; 47391c9e8defSMike Kravetz 47401e392147SAndrea Arcangeli /* 47411e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 47421e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 47431e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47441e392147SAndrea Arcangeli * the caller. 47451e392147SAndrea Arcangeli */ 47461c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47471c9e8defSMike Kravetz if (ret) 47481c9e8defSMike Kravetz goto out_release_nounlock; 47491c9e8defSMike Kravetz } 47501c9e8defSMike Kravetz 47518fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47528fb5debcSMike Kravetz spin_lock(ptl); 47538fb5debcSMike Kravetz 47541e392147SAndrea Arcangeli /* 47551e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47561e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47571e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47581e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47591e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47601e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47611e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47621e392147SAndrea Arcangeli */ 47631e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47641e392147SAndrea Arcangeli ret = -EFAULT; 47651e392147SAndrea Arcangeli if (idx >= size) 47661e392147SAndrea Arcangeli goto out_release_unlock; 47671e392147SAndrea Arcangeli 47688fb5debcSMike Kravetz ret = -EEXIST; 47698fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47708fb5debcSMike Kravetz goto out_release_unlock; 47718fb5debcSMike Kravetz 47721c9e8defSMike Kravetz if (vm_shared) { 47731c9e8defSMike Kravetz page_dup_rmap(page, true); 47741c9e8defSMike Kravetz } else { 4775d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 47768fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47771c9e8defSMike Kravetz } 47788fb5debcSMike Kravetz 47798fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47808fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47818fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47828fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47838fb5debcSMike Kravetz 47848fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47858fb5debcSMike Kravetz 47868fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47878fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47888fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47898fb5debcSMike Kravetz 47908fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47918fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47928fb5debcSMike Kravetz 47938fb5debcSMike Kravetz spin_unlock(ptl); 47948f251a3dSMike Kravetz SetHPageMigratable(page); 47951c9e8defSMike Kravetz if (vm_shared) 47961c9e8defSMike Kravetz unlock_page(page); 47978fb5debcSMike Kravetz ret = 0; 47988fb5debcSMike Kravetz out: 47998fb5debcSMike Kravetz return ret; 48008fb5debcSMike Kravetz out_release_unlock: 48018fb5debcSMike Kravetz spin_unlock(ptl); 48021c9e8defSMike Kravetz if (vm_shared) 48031c9e8defSMike Kravetz unlock_page(page); 48045af10dfdSAndrea Arcangeli out_release_nounlock: 48058fb5debcSMike Kravetz put_page(page); 48068fb5debcSMike Kravetz goto out; 48078fb5debcSMike Kravetz } 48088fb5debcSMike Kravetz 480982e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 481082e5d378SJoao Martins int refs, struct page **pages, 481182e5d378SJoao Martins struct vm_area_struct **vmas) 481282e5d378SJoao Martins { 481382e5d378SJoao Martins int nr; 481482e5d378SJoao Martins 481582e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 481682e5d378SJoao Martins if (likely(pages)) 481782e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 481882e5d378SJoao Martins if (vmas) 481982e5d378SJoao Martins vmas[nr] = vma; 482082e5d378SJoao Martins } 482182e5d378SJoao Martins } 482282e5d378SJoao Martins 482328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 482463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 482528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 48264f6da934SPeter Xu long i, unsigned int flags, int *locked) 482763551ae0SDavid Gibson { 4828d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4829d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 483028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4831a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48320fa5bc40SJoao Martins int err = -EFAULT, refs; 483363551ae0SDavid Gibson 483463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 483563551ae0SDavid Gibson pte_t *pte; 4836cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 48372a15efc9SHugh Dickins int absent; 483863551ae0SDavid Gibson struct page *page; 483963551ae0SDavid Gibson 48404c887265SAdam Litke /* 484102057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 484202057967SDavid Rientjes * potentially allocating memory. 484302057967SDavid Rientjes */ 4844fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 484502057967SDavid Rientjes remainder = 0; 484602057967SDavid Rientjes break; 484702057967SDavid Rientjes } 484802057967SDavid Rientjes 484902057967SDavid Rientjes /* 48504c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48512a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48524c887265SAdam Litke * first, for the page indexing below to work. 4853cb900f41SKirill A. Shutemov * 4854cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48554c887265SAdam Litke */ 48567868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48577868a208SPunit Agrawal huge_page_size(h)); 4858cb900f41SKirill A. Shutemov if (pte) 4859cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48602a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 486163551ae0SDavid Gibson 48622a15efc9SHugh Dickins /* 48632a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48643ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48653ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48663ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48673ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48682a15efc9SHugh Dickins */ 48693ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48703ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4871cb900f41SKirill A. Shutemov if (pte) 4872cb900f41SKirill A. Shutemov spin_unlock(ptl); 48732a15efc9SHugh Dickins remainder = 0; 48742a15efc9SHugh Dickins break; 48752a15efc9SHugh Dickins } 48762a15efc9SHugh Dickins 48779cc3a5bdSNaoya Horiguchi /* 48789cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48799cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48809cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 48819cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 48829cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 48839cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 48849cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 48859cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 48869cc3a5bdSNaoya Horiguchi */ 48879cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4888106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4889106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 48902b740303SSouptick Joarder vm_fault_t ret; 489187ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 48924c887265SAdam Litke 4893cb900f41SKirill A. Shutemov if (pte) 4894cb900f41SKirill A. Shutemov spin_unlock(ptl); 489587ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 489687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 48974f6da934SPeter Xu if (locked) 489871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 489971335f37SPeter Xu FAULT_FLAG_KILLABLE; 490087ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 490187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 490287ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 490387ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 49044426e945SPeter Xu /* 49054426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 49064426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 49074426e945SPeter Xu */ 490887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 490987ffc118SAndrea Arcangeli } 491087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 491187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 49122be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 49131c59827dSHugh Dickins remainder = 0; 49141c59827dSHugh Dickins break; 49151c59827dSHugh Dickins } 491687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 49174f6da934SPeter Xu if (locked && 49181ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 49194f6da934SPeter Xu *locked = 0; 492087ffc118SAndrea Arcangeli *nr_pages = 0; 492187ffc118SAndrea Arcangeli /* 492287ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 492387ffc118SAndrea Arcangeli * error, it will return zero 492487ffc118SAndrea Arcangeli * instead. 492587ffc118SAndrea Arcangeli * 492687ffc118SAndrea Arcangeli * No need to update "position" as the 492787ffc118SAndrea Arcangeli * caller will not check it after 492887ffc118SAndrea Arcangeli * *nr_pages is set to 0. 492987ffc118SAndrea Arcangeli */ 493087ffc118SAndrea Arcangeli return i; 493187ffc118SAndrea Arcangeli } 493287ffc118SAndrea Arcangeli continue; 493387ffc118SAndrea Arcangeli } 493463551ae0SDavid Gibson 4935a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 49367f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 49378fde12caSLinus Torvalds 49388fde12caSLinus Torvalds /* 4939acbfb087SZhigang Lu * If subpage information not requested, update counters 4940acbfb087SZhigang Lu * and skip the same_page loop below. 4941acbfb087SZhigang Lu */ 4942acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4943acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4944acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4945acbfb087SZhigang Lu vaddr += huge_page_size(h); 4946acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4947acbfb087SZhigang Lu i += pages_per_huge_page(h); 4948acbfb087SZhigang Lu spin_unlock(ptl); 4949acbfb087SZhigang Lu continue; 4950acbfb087SZhigang Lu } 4951acbfb087SZhigang Lu 495282e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 495382e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 49540fa5bc40SJoao Martins 495582e5d378SJoao Martins if (pages || vmas) 495682e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 495782e5d378SJoao Martins vma, refs, 495882e5d378SJoao Martins likely(pages) ? pages + i : NULL, 495982e5d378SJoao Martins vmas ? vmas + i : NULL); 496063551ae0SDavid Gibson 496182e5d378SJoao Martins if (pages) { 49620fa5bc40SJoao Martins /* 49630fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 49640fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 49650fa5bc40SJoao Martins * checked that the huge page is present in the page 49660fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 49670fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 49680fa5bc40SJoao Martins * head page and tail pages from being rearranged in 49690fa5bc40SJoao Martins * any way. So this page must be available at this 49700fa5bc40SJoao Martins * point, unless the page refcount overflowed: 49710fa5bc40SJoao Martins */ 497282e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 49730fa5bc40SJoao Martins refs, 49740fa5bc40SJoao Martins flags))) { 49750fa5bc40SJoao Martins spin_unlock(ptl); 49760fa5bc40SJoao Martins remainder = 0; 49770fa5bc40SJoao Martins err = -ENOMEM; 49780fa5bc40SJoao Martins break; 49790fa5bc40SJoao Martins } 4980d5d4b0aaSChen, Kenneth W } 498182e5d378SJoao Martins 498282e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 498382e5d378SJoao Martins remainder -= refs; 498482e5d378SJoao Martins i += refs; 498582e5d378SJoao Martins 4986cb900f41SKirill A. Shutemov spin_unlock(ptl); 498763551ae0SDavid Gibson } 498828a35716SMichel Lespinasse *nr_pages = remainder; 498987ffc118SAndrea Arcangeli /* 499087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 499187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 499287ffc118SAndrea Arcangeli * branch. 499387ffc118SAndrea Arcangeli */ 499463551ae0SDavid Gibson *position = vaddr; 499563551ae0SDavid Gibson 49962be7cfedSDaniel Jordan return i ? i : err; 499763551ae0SDavid Gibson } 49988f860591SZhang, Yanmin 49995491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 50005491ae7bSAneesh Kumar K.V /* 50015491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 50025491ae7bSAneesh Kumar K.V * implement this. 50035491ae7bSAneesh Kumar K.V */ 50045491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 50055491ae7bSAneesh Kumar K.V #endif 50065491ae7bSAneesh Kumar K.V 50077da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 50088f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 50098f860591SZhang, Yanmin { 50108f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 50118f860591SZhang, Yanmin unsigned long start = address; 50128f860591SZhang, Yanmin pte_t *ptep; 50138f860591SZhang, Yanmin pte_t pte; 5014a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 50157da4d641SPeter Zijlstra unsigned long pages = 0; 5016dff11abeSMike Kravetz bool shared_pmd = false; 5017ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5018dff11abeSMike Kravetz 5019dff11abeSMike Kravetz /* 5020dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5021ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5022dff11abeSMike Kravetz * range if PMD sharing is possible. 5023dff11abeSMike Kravetz */ 50247269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 50257269f999SJérôme Glisse 0, vma, mm, start, end); 5026ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50278f860591SZhang, Yanmin 50288f860591SZhang, Yanmin BUG_ON(address >= end); 5029ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 50308f860591SZhang, Yanmin 5031ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 503283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5033a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5034cb900f41SKirill A. Shutemov spinlock_t *ptl; 50357868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 50368f860591SZhang, Yanmin if (!ptep) 50378f860591SZhang, Yanmin continue; 5038cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 503934ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 50407da4d641SPeter Zijlstra pages++; 5041cb900f41SKirill A. Shutemov spin_unlock(ptl); 5042dff11abeSMike Kravetz shared_pmd = true; 504339dde65cSChen, Kenneth W continue; 50447da4d641SPeter Zijlstra } 5045a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5046a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5047a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5048a8bda28dSNaoya Horiguchi continue; 5049a8bda28dSNaoya Horiguchi } 5050a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5051a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5052a8bda28dSNaoya Horiguchi 5053a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5054a8bda28dSNaoya Horiguchi pte_t newpte; 5055a8bda28dSNaoya Horiguchi 5056a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5057a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5058e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5059e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5060a8bda28dSNaoya Horiguchi pages++; 5061a8bda28dSNaoya Horiguchi } 5062a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5063a8bda28dSNaoya Horiguchi continue; 5064a8bda28dSNaoya Horiguchi } 5065a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5066023bdd00SAneesh Kumar K.V pte_t old_pte; 5067023bdd00SAneesh Kumar K.V 5068023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5069023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5070be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5071023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50727da4d641SPeter Zijlstra pages++; 50738f860591SZhang, Yanmin } 5074cb900f41SKirill A. Shutemov spin_unlock(ptl); 50758f860591SZhang, Yanmin } 5076d833352aSMel Gorman /* 5077c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5078d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5079c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5080dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5081dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5082d833352aSMel Gorman */ 5083dff11abeSMike Kravetz if (shared_pmd) 5084ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5085dff11abeSMike Kravetz else 50865491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50870f10851eSJérôme Glisse /* 50880f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 50890f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 50900f10851eSJérôme Glisse * 5091ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50920f10851eSJérôme Glisse */ 509383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5094ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50957da4d641SPeter Zijlstra 50967da4d641SPeter Zijlstra return pages << h->order; 50978f860591SZhang, Yanmin } 50988f860591SZhang, Yanmin 509933b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 510033b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5101a1e78772SMel Gorman long from, long to, 51025a6fe125SMel Gorman struct vm_area_struct *vma, 5103ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5104e4e574b7SAdam Litke { 510533b8f84aSMike Kravetz long chg, add = -1; 5106a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 510790481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 51089119a41eSJoonsoo Kim struct resv_map *resv_map; 5109075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 51100db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5111e4e574b7SAdam Litke 511263489f8eSMike Kravetz /* This should never happen */ 511363489f8eSMike Kravetz if (from > to) { 511463489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 511533b8f84aSMike Kravetz return false; 511663489f8eSMike Kravetz } 511763489f8eSMike Kravetz 5118a1e78772SMel Gorman /* 511917c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 512017c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 512190481622SDavid Gibson * without using reserves 512217c9d12eSMel Gorman */ 5123ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 512433b8f84aSMike Kravetz return true; 512517c9d12eSMel Gorman 512617c9d12eSMel Gorman /* 5127a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5128a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5129a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5130a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5131a1e78772SMel Gorman */ 51329119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5133f27a5136SMike Kravetz /* 5134f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5135f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5136f27a5136SMike Kravetz * hugetlbfs_get_inode). 5137f27a5136SMike Kravetz */ 51384e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 51399119a41eSJoonsoo Kim 51400db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 51419119a41eSJoonsoo Kim 51429119a41eSJoonsoo Kim } else { 5143e9fe92aeSMina Almasry /* Private mapping. */ 51449119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51455a6fe125SMel Gorman if (!resv_map) 514633b8f84aSMike Kravetz return false; 51475a6fe125SMel Gorman 514817c9d12eSMel Gorman chg = to - from; 514917c9d12eSMel Gorman 51505a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51515a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51525a6fe125SMel Gorman } 51535a6fe125SMel Gorman 515433b8f84aSMike Kravetz if (chg < 0) 5155c50ac050SDave Hansen goto out_err; 515617c9d12eSMel Gorman 515733b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 515833b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5159075a61d0SMina Almasry goto out_err; 5160075a61d0SMina Almasry 5161075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5162075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5163075a61d0SMina Almasry * of the resv_map. 5164075a61d0SMina Almasry */ 5165075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5166075a61d0SMina Almasry } 5167075a61d0SMina Almasry 51681c5ecae3SMike Kravetz /* 51691c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51701c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51711c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51721c5ecae3SMike Kravetz */ 51731c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 517433b8f84aSMike Kravetz if (gbl_reserve < 0) 5175075a61d0SMina Almasry goto out_uncharge_cgroup; 517617c9d12eSMel Gorman 517717c9d12eSMel Gorman /* 517817c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 517990481622SDavid Gibson * Hand the pages back to the subpool if there are not 518017c9d12eSMel Gorman */ 518133b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5182075a61d0SMina Almasry goto out_put_pages; 518317c9d12eSMel Gorman 518417c9d12eSMel Gorman /* 518517c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 518617c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 518717c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 518817c9d12eSMel Gorman * the reservation was and the page cache tells how much of 518917c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 519017c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 519117c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 519217c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 519317c9d12eSMel Gorman * else has to be done for private mappings here 519417c9d12eSMel Gorman */ 519533039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5196075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 519733039678SMike Kravetz 51980db9d74eSMina Almasry if (unlikely(add < 0)) { 51990db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5200075a61d0SMina Almasry goto out_put_pages; 52010db9d74eSMina Almasry } else if (unlikely(chg > add)) { 520233039678SMike Kravetz /* 520333039678SMike Kravetz * pages in this range were added to the reserve 520433039678SMike Kravetz * map between region_chg and region_add. This 520533039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 520633039678SMike Kravetz * the subpool and reserve counts modified above 520733039678SMike Kravetz * based on the difference. 520833039678SMike Kravetz */ 520933039678SMike Kravetz long rsv_adjust; 521033039678SMike Kravetz 5211d85aecf2SMiaohe Lin /* 5212d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5213d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5214d85aecf2SMiaohe Lin */ 5215075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5216075a61d0SMina Almasry hstate_index(h), 5217075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5218075a61d0SMina Almasry 521933039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 522033039678SMike Kravetz chg - add); 522133039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5222d85aecf2SMiaohe Lin } else if (h_cg) { 5223d85aecf2SMiaohe Lin /* 5224d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5225d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5226d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5227d85aecf2SMiaohe Lin * done. 5228d85aecf2SMiaohe Lin */ 5229d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 523033039678SMike Kravetz } 523133039678SMike Kravetz } 523233b8f84aSMike Kravetz return true; 523333b8f84aSMike Kravetz 5234075a61d0SMina Almasry out_put_pages: 5235075a61d0SMina Almasry /* put back original number of pages, chg */ 5236075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5237075a61d0SMina Almasry out_uncharge_cgroup: 5238075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5239075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5240c50ac050SDave Hansen out_err: 52415e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 52420db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 52430db9d74eSMina Almasry * region_add failed or didn't run. 52440db9d74eSMina Almasry */ 52450db9d74eSMina Almasry if (chg >= 0 && add < 0) 52460db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5247f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5248f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 524933b8f84aSMike Kravetz return false; 5250a43a8c39SChen, Kenneth W } 5251a43a8c39SChen, Kenneth W 5252b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5253b5cec28dSMike Kravetz long freed) 5254a43a8c39SChen, Kenneth W { 5255a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52564e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52579119a41eSJoonsoo Kim long chg = 0; 525890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52591c5ecae3SMike Kravetz long gbl_reserve; 526045c682a6SKen Chen 5261f27a5136SMike Kravetz /* 5262f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5263f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5264f27a5136SMike Kravetz */ 5265b5cec28dSMike Kravetz if (resv_map) { 5266b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5267b5cec28dSMike Kravetz /* 5268b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5269b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5270b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5271b5cec28dSMike Kravetz */ 5272b5cec28dSMike Kravetz if (chg < 0) 5273b5cec28dSMike Kravetz return chg; 5274b5cec28dSMike Kravetz } 5275b5cec28dSMike Kravetz 527645c682a6SKen Chen spin_lock(&inode->i_lock); 5277e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 527845c682a6SKen Chen spin_unlock(&inode->i_lock); 527945c682a6SKen Chen 52801c5ecae3SMike Kravetz /* 52811c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52821c5ecae3SMike Kravetz * reservations to be released may be adjusted. 52831c5ecae3SMike Kravetz */ 52841c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52851c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5286b5cec28dSMike Kravetz 5287b5cec28dSMike Kravetz return 0; 5288a43a8c39SChen, Kenneth W } 528993f70f90SNaoya Horiguchi 52903212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 52913212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 52923212b535SSteve Capper struct vm_area_struct *vma, 52933212b535SSteve Capper unsigned long addr, pgoff_t idx) 52943212b535SSteve Capper { 52953212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52963212b535SSteve Capper svma->vm_start; 52973212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52983212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 52993212b535SSteve Capper 53003212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5301de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5302de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 53033212b535SSteve Capper 53043212b535SSteve Capper /* 53053212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 53063212b535SSteve Capper * page table page. 53073212b535SSteve Capper */ 53083212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 53093212b535SSteve Capper vm_flags != svm_flags || 531007e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 53113212b535SSteve Capper return 0; 53123212b535SSteve Capper 53133212b535SSteve Capper return saddr; 53143212b535SSteve Capper } 53153212b535SSteve Capper 531631aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 53173212b535SSteve Capper { 53183212b535SSteve Capper unsigned long base = addr & PUD_MASK; 53193212b535SSteve Capper unsigned long end = base + PUD_SIZE; 53203212b535SSteve Capper 53213212b535SSteve Capper /* 53223212b535SSteve Capper * check on proper vm_flags and page table alignment 53233212b535SSteve Capper */ 5324017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 532531aafb45SNicholas Krause return true; 532631aafb45SNicholas Krause return false; 53273212b535SSteve Capper } 53283212b535SSteve Capper 5329*c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5330*c1991e07SPeter Xu { 5331*c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5332*c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5333*c1991e07SPeter Xu return false; 5334*c1991e07SPeter Xu #endif 5335*c1991e07SPeter Xu return vma_shareable(vma, addr); 5336*c1991e07SPeter Xu } 5337*c1991e07SPeter Xu 53383212b535SSteve Capper /* 5339017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5340017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5341017b1660SMike Kravetz * shared pmd mappings. 5342017b1660SMike Kravetz */ 5343017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5344017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5345017b1660SMike Kravetz { 5346a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5347a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5348017b1660SMike Kravetz 5349a1ba9da8SLi Xinhai /* 5350a1ba9da8SLi Xinhai * vma need span at least one aligned PUD size and the start,end range 5351a1ba9da8SLi Xinhai * must at least partialy within it. 5352a1ba9da8SLi Xinhai */ 5353a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5354a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5355017b1660SMike Kravetz return; 5356017b1660SMike Kravetz 535775802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5358a1ba9da8SLi Xinhai if (*start > v_start) 5359a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5360017b1660SMike Kravetz 5361a1ba9da8SLi Xinhai if (*end < v_end) 5362a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5363017b1660SMike Kravetz } 5364017b1660SMike Kravetz 5365017b1660SMike Kravetz /* 53663212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53673212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53683212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5369c0d0381aSMike Kravetz * code much cleaner. 5370c0d0381aSMike Kravetz * 53710bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 53720bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 53730bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 53740bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 53750bf7b64eSMike Kravetz * 53760bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 53770bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 53780bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 53790bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 53800bf7b64eSMike Kravetz * only required for subsequent processing. 53813212b535SSteve Capper */ 5382aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5383aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 53843212b535SSteve Capper { 53853212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 53863212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 53873212b535SSteve Capper vma->vm_pgoff; 53883212b535SSteve Capper struct vm_area_struct *svma; 53893212b535SSteve Capper unsigned long saddr; 53903212b535SSteve Capper pte_t *spte = NULL; 53913212b535SSteve Capper pte_t *pte; 5392cb900f41SKirill A. Shutemov spinlock_t *ptl; 53933212b535SSteve Capper 53940bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 53953212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 53963212b535SSteve Capper if (svma == vma) 53973212b535SSteve Capper continue; 53983212b535SSteve Capper 53993212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 54003212b535SSteve Capper if (saddr) { 54017868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 54027868a208SPunit Agrawal vma_mmu_pagesize(svma)); 54033212b535SSteve Capper if (spte) { 54043212b535SSteve Capper get_page(virt_to_page(spte)); 54053212b535SSteve Capper break; 54063212b535SSteve Capper } 54073212b535SSteve Capper } 54083212b535SSteve Capper } 54093212b535SSteve Capper 54103212b535SSteve Capper if (!spte) 54113212b535SSteve Capper goto out; 54123212b535SSteve Capper 54138bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5414dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 54153212b535SSteve Capper pud_populate(mm, pud, 54163212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5417c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5418dc6c9a35SKirill A. Shutemov } else { 54193212b535SSteve Capper put_page(virt_to_page(spte)); 5420dc6c9a35SKirill A. Shutemov } 5421cb900f41SKirill A. Shutemov spin_unlock(ptl); 54223212b535SSteve Capper out: 54233212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54243212b535SSteve Capper return pte; 54253212b535SSteve Capper } 54263212b535SSteve Capper 54273212b535SSteve Capper /* 54283212b535SSteve Capper * unmap huge page backed by shared pte. 54293212b535SSteve Capper * 54303212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 54313212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 54323212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 54333212b535SSteve Capper * 5434c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 54353212b535SSteve Capper * 54363212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 54373212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 54383212b535SSteve Capper */ 543934ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 544034ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 54413212b535SSteve Capper { 54423212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5443c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5444c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 54453212b535SSteve Capper 544634ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 54473212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 54483212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 54493212b535SSteve Capper return 0; 54503212b535SSteve Capper 54513212b535SSteve Capper pud_clear(pud); 54523212b535SSteve Capper put_page(virt_to_page(ptep)); 5453dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54543212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54553212b535SSteve Capper return 1; 54563212b535SSteve Capper } 5457*c1991e07SPeter Xu 54589e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5459aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5460aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 54619e5fc74cSSteve Capper { 54629e5fc74cSSteve Capper return NULL; 54639e5fc74cSSteve Capper } 5464e81f2d22SZhang Zhen 546534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 546634ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5467e81f2d22SZhang Zhen { 5468e81f2d22SZhang Zhen return 0; 5469e81f2d22SZhang Zhen } 5470017b1660SMike Kravetz 5471017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5472017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5473017b1660SMike Kravetz { 5474017b1660SMike Kravetz } 5475*c1991e07SPeter Xu 5476*c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5477*c1991e07SPeter Xu { 5478*c1991e07SPeter Xu return false; 5479*c1991e07SPeter Xu } 54803212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54813212b535SSteve Capper 54829e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5483aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 54849e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54859e5fc74cSSteve Capper { 54869e5fc74cSSteve Capper pgd_t *pgd; 5487c2febafcSKirill A. Shutemov p4d_t *p4d; 54889e5fc74cSSteve Capper pud_t *pud; 54899e5fc74cSSteve Capper pte_t *pte = NULL; 54909e5fc74cSSteve Capper 54919e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5492f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5493f4f0a3d8SKirill A. Shutemov if (!p4d) 5494f4f0a3d8SKirill A. Shutemov return NULL; 5495c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 54969e5fc74cSSteve Capper if (pud) { 54979e5fc74cSSteve Capper if (sz == PUD_SIZE) { 54989e5fc74cSSteve Capper pte = (pte_t *)pud; 54999e5fc74cSSteve Capper } else { 55009e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5501*c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5502aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 55039e5fc74cSSteve Capper else 55049e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 55059e5fc74cSSteve Capper } 55069e5fc74cSSteve Capper } 55074e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 55089e5fc74cSSteve Capper 55099e5fc74cSSteve Capper return pte; 55109e5fc74cSSteve Capper } 55119e5fc74cSSteve Capper 55129b19df29SPunit Agrawal /* 55139b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 55149b19df29SPunit Agrawal * entry at address @addr 55159b19df29SPunit Agrawal * 55168ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 55178ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 55189b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 55199b19df29SPunit Agrawal * table. 55209b19df29SPunit Agrawal */ 55217868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 55227868a208SPunit Agrawal unsigned long addr, unsigned long sz) 55239e5fc74cSSteve Capper { 55249e5fc74cSSteve Capper pgd_t *pgd; 5525c2febafcSKirill A. Shutemov p4d_t *p4d; 55268ac0b81aSLi Xinhai pud_t *pud; 55278ac0b81aSLi Xinhai pmd_t *pmd; 55289e5fc74cSSteve Capper 55299e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5530c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5531c2febafcSKirill A. Shutemov return NULL; 5532c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5533c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5534c2febafcSKirill A. Shutemov return NULL; 55359b19df29SPunit Agrawal 5536c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 55378ac0b81aSLi Xinhai if (sz == PUD_SIZE) 55388ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 55399e5fc74cSSteve Capper return (pte_t *)pud; 55408ac0b81aSLi Xinhai if (!pud_present(*pud)) 55418ac0b81aSLi Xinhai return NULL; 55428ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 55439b19df29SPunit Agrawal 55449e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 55458ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 55469e5fc74cSSteve Capper return (pte_t *)pmd; 55479e5fc74cSSteve Capper } 55489e5fc74cSSteve Capper 554961f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 555061f77edaSNaoya Horiguchi 555161f77edaSNaoya Horiguchi /* 555261f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 555361f77edaSNaoya Horiguchi * behavior. 555461f77edaSNaoya Horiguchi */ 555561f77edaSNaoya Horiguchi struct page * __weak 555661f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 555761f77edaSNaoya Horiguchi int write) 555861f77edaSNaoya Horiguchi { 555961f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 556061f77edaSNaoya Horiguchi } 556161f77edaSNaoya Horiguchi 556261f77edaSNaoya Horiguchi struct page * __weak 55634dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55644dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55654dc71451SAneesh Kumar K.V { 55664dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55674dc71451SAneesh Kumar K.V return NULL; 55684dc71451SAneesh Kumar K.V } 55694dc71451SAneesh Kumar K.V 55704dc71451SAneesh Kumar K.V struct page * __weak 55719e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5572e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55739e5fc74cSSteve Capper { 5574e66f17ffSNaoya Horiguchi struct page *page = NULL; 5575e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5576c9d398faSNaoya Horiguchi pte_t pte; 55773faa52c0SJohn Hubbard 55783faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55793faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55803faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55813faa52c0SJohn Hubbard return NULL; 55823faa52c0SJohn Hubbard 5583e66f17ffSNaoya Horiguchi retry: 5584e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5585e66f17ffSNaoya Horiguchi spin_lock(ptl); 5586e66f17ffSNaoya Horiguchi /* 5587e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5588e66f17ffSNaoya Horiguchi * unmapped from other threads. 5589e66f17ffSNaoya Horiguchi */ 5590e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5591e66f17ffSNaoya Horiguchi goto out; 5592c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5593c9d398faSNaoya Horiguchi if (pte_present(pte)) { 559497534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 55953faa52c0SJohn Hubbard /* 55963faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 55973faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 55983faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 55993faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 56003faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 56013faa52c0SJohn Hubbard * unless the page refcount overflowed: 56023faa52c0SJohn Hubbard */ 56033faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 56043faa52c0SJohn Hubbard page = NULL; 56053faa52c0SJohn Hubbard goto out; 56063faa52c0SJohn Hubbard } 5607e66f17ffSNaoya Horiguchi } else { 5608c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5609e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5610e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5611e66f17ffSNaoya Horiguchi goto retry; 5612e66f17ffSNaoya Horiguchi } 5613e66f17ffSNaoya Horiguchi /* 5614e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5615e66f17ffSNaoya Horiguchi * follow_page_mask(). 5616e66f17ffSNaoya Horiguchi */ 5617e66f17ffSNaoya Horiguchi } 5618e66f17ffSNaoya Horiguchi out: 5619e66f17ffSNaoya Horiguchi spin_unlock(ptl); 56209e5fc74cSSteve Capper return page; 56219e5fc74cSSteve Capper } 56229e5fc74cSSteve Capper 562361f77edaSNaoya Horiguchi struct page * __weak 56249e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5625e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 56269e5fc74cSSteve Capper { 56273faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5628e66f17ffSNaoya Horiguchi return NULL; 56299e5fc74cSSteve Capper 5630e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 56319e5fc74cSSteve Capper } 56329e5fc74cSSteve Capper 5633faaa5b62SAnshuman Khandual struct page * __weak 5634faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5635faaa5b62SAnshuman Khandual { 56363faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5637faaa5b62SAnshuman Khandual return NULL; 5638faaa5b62SAnshuman Khandual 5639faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5640faaa5b62SAnshuman Khandual } 5641faaa5b62SAnshuman Khandual 564231caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 564331caf665SNaoya Horiguchi { 5644bcc54222SNaoya Horiguchi bool ret = true; 5645bcc54222SNaoya Horiguchi 564631caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 56478f251a3dSMike Kravetz if (!PageHeadHuge(page) || 56488f251a3dSMike Kravetz !HPageMigratable(page) || 56490eb2df2bSMuchun Song !get_page_unless_zero(page)) { 5650bcc54222SNaoya Horiguchi ret = false; 5651bcc54222SNaoya Horiguchi goto unlock; 5652bcc54222SNaoya Horiguchi } 56538f251a3dSMike Kravetz ClearHPageMigratable(page); 565431caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5655bcc54222SNaoya Horiguchi unlock: 565631caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5657bcc54222SNaoya Horiguchi return ret; 565831caf665SNaoya Horiguchi } 565931caf665SNaoya Horiguchi 566031caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 566131caf665SNaoya Horiguchi { 566231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 56638f251a3dSMike Kravetz SetHPageMigratable(page); 566431caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 566531caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 566631caf665SNaoya Horiguchi put_page(page); 566731caf665SNaoya Horiguchi } 5668ab5ac90aSMichal Hocko 5669ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5670ab5ac90aSMichal Hocko { 5671ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5672ab5ac90aSMichal Hocko 5673ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5674ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5675ab5ac90aSMichal Hocko 5676ab5ac90aSMichal Hocko /* 5677ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5678ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5679ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5680ab5ac90aSMichal Hocko * the temporary status. 5681ab5ac90aSMichal Hocko * 5682ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5683ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5684ab5ac90aSMichal Hocko * the per-node's. 5685ab5ac90aSMichal Hocko */ 56869157c311SMike Kravetz if (HPageTemporary(newpage)) { 5687ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5688ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5689ab5ac90aSMichal Hocko 56909157c311SMike Kravetz SetHPageTemporary(oldpage); 56919157c311SMike Kravetz ClearHPageTemporary(newpage); 5692ab5ac90aSMichal Hocko 5693ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5694ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5695ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5696ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5697ab5ac90aSMichal Hocko } 5698ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5699ab5ac90aSMichal Hocko } 5700ab5ac90aSMichal Hocko } 5701cf11e85fSRoman Gushchin 5702cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5703cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5704cf11e85fSRoman Gushchin 5705cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5706cf11e85fSRoman Gushchin { 5707cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5708cf11e85fSRoman Gushchin return 0; 5709cf11e85fSRoman Gushchin } 5710cf11e85fSRoman Gushchin 5711cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5712cf11e85fSRoman Gushchin 5713cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5714cf11e85fSRoman Gushchin { 5715cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5716cf11e85fSRoman Gushchin int nid; 5717cf11e85fSRoman Gushchin 5718cf11e85fSRoman Gushchin cma_reserve_called = true; 5719cf11e85fSRoman Gushchin 5720cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5721cf11e85fSRoman Gushchin return; 5722cf11e85fSRoman Gushchin 5723cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5724cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5725cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5726cf11e85fSRoman Gushchin return; 5727cf11e85fSRoman Gushchin } 5728cf11e85fSRoman Gushchin 5729cf11e85fSRoman Gushchin /* 5730cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5731cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5732cf11e85fSRoman Gushchin */ 5733cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5734cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5735cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5736cf11e85fSRoman Gushchin 5737cf11e85fSRoman Gushchin reserved = 0; 5738cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5739cf11e85fSRoman Gushchin int res; 57402281f797SBarry Song char name[CMA_MAX_NAME]; 5741cf11e85fSRoman Gushchin 5742cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5743cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5744cf11e85fSRoman Gushchin 57452281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5746cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 574729d0f41dSBarry Song 0, false, name, 5748cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5749cf11e85fSRoman Gushchin if (res) { 5750cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5751cf11e85fSRoman Gushchin res, nid); 5752cf11e85fSRoman Gushchin continue; 5753cf11e85fSRoman Gushchin } 5754cf11e85fSRoman Gushchin 5755cf11e85fSRoman Gushchin reserved += size; 5756cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5757cf11e85fSRoman Gushchin size / SZ_1M, nid); 5758cf11e85fSRoman Gushchin 5759cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5760cf11e85fSRoman Gushchin break; 5761cf11e85fSRoman Gushchin } 5762cf11e85fSRoman Gushchin } 5763cf11e85fSRoman Gushchin 5764cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5765cf11e85fSRoman Gushchin { 5766cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5767cf11e85fSRoman Gushchin return; 5768cf11e85fSRoman Gushchin 5769cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5770cf11e85fSRoman Gushchin } 5771cf11e85fSRoman Gushchin 5772cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5773