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 97db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 98db71ef79SMike Kravetz unsigned long irq_flags) 9990481622SDavid Gibson { 100db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 10190481622SDavid Gibson 10290481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1037c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1047ca02d0aSMike Kravetz * free the subpool */ 1051d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1067ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1077ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1087ca02d0aSMike Kravetz -spool->min_hpages); 10990481622SDavid Gibson kfree(spool); 11090481622SDavid Gibson } 1117ca02d0aSMike Kravetz } 11290481622SDavid Gibson 1137ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1147ca02d0aSMike Kravetz long min_hpages) 11590481622SDavid Gibson { 11690481622SDavid Gibson struct hugepage_subpool *spool; 11790481622SDavid Gibson 118c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 11990481622SDavid Gibson if (!spool) 12090481622SDavid Gibson return NULL; 12190481622SDavid Gibson 12290481622SDavid Gibson spin_lock_init(&spool->lock); 12390481622SDavid Gibson spool->count = 1; 1247ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1257ca02d0aSMike Kravetz spool->hstate = h; 1267ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1277ca02d0aSMike Kravetz 1287ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1297ca02d0aSMike Kravetz kfree(spool); 1307ca02d0aSMike Kravetz return NULL; 1317ca02d0aSMike Kravetz } 1327ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 13390481622SDavid Gibson 13490481622SDavid Gibson return spool; 13590481622SDavid Gibson } 13690481622SDavid Gibson 13790481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 13890481622SDavid Gibson { 139db71ef79SMike Kravetz unsigned long flags; 140db71ef79SMike Kravetz 141db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 14290481622SDavid Gibson BUG_ON(!spool->count); 14390481622SDavid Gibson spool->count--; 144db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 14590481622SDavid Gibson } 14690481622SDavid Gibson 1471c5ecae3SMike Kravetz /* 1481c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1491c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1509e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1511c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1521c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1537c8de358SEthon Paul * a subpool minimum size must be maintained. 1541c5ecae3SMike Kravetz */ 1551c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 15690481622SDavid Gibson long delta) 15790481622SDavid Gibson { 1581c5ecae3SMike Kravetz long ret = delta; 15990481622SDavid Gibson 16090481622SDavid Gibson if (!spool) 1611c5ecae3SMike Kravetz return ret; 16290481622SDavid Gibson 163db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 16490481622SDavid Gibson 1651c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1661c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1671c5ecae3SMike Kravetz spool->used_hpages += delta; 1681c5ecae3SMike Kravetz else { 1691c5ecae3SMike Kravetz ret = -ENOMEM; 1701c5ecae3SMike Kravetz goto unlock_ret; 1711c5ecae3SMike Kravetz } 1721c5ecae3SMike Kravetz } 1731c5ecae3SMike Kravetz 17409a95e29SMike Kravetz /* minimum size accounting */ 17509a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1761c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1771c5ecae3SMike Kravetz /* 1781c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1791c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1801c5ecae3SMike Kravetz */ 1811c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1821c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1831c5ecae3SMike Kravetz } else { 1841c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1851c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1861c5ecae3SMike Kravetz } 1871c5ecae3SMike Kravetz } 1881c5ecae3SMike Kravetz 1891c5ecae3SMike Kravetz unlock_ret: 190db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 19190481622SDavid Gibson return ret; 19290481622SDavid Gibson } 19390481622SDavid Gibson 1941c5ecae3SMike Kravetz /* 1951c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1961c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1971c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1981c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1991c5ecae3SMike Kravetz */ 2001c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 20190481622SDavid Gibson long delta) 20290481622SDavid Gibson { 2031c5ecae3SMike Kravetz long ret = delta; 204db71ef79SMike Kravetz unsigned long flags; 2051c5ecae3SMike Kravetz 20690481622SDavid Gibson if (!spool) 2071c5ecae3SMike Kravetz return delta; 20890481622SDavid Gibson 209db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2101c5ecae3SMike Kravetz 2111c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 21290481622SDavid Gibson spool->used_hpages -= delta; 2131c5ecae3SMike Kravetz 21409a95e29SMike Kravetz /* minimum size accounting */ 21509a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2161c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2171c5ecae3SMike Kravetz ret = 0; 2181c5ecae3SMike Kravetz else 2191c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2201c5ecae3SMike Kravetz 2211c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2221c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2231c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2241c5ecae3SMike Kravetz } 2251c5ecae3SMike Kravetz 2261c5ecae3SMike Kravetz /* 2271c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2281c5ecae3SMike Kravetz * quota reference, free it now. 2291c5ecae3SMike Kravetz */ 230db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2311c5ecae3SMike Kravetz 2321c5ecae3SMike Kravetz return ret; 23390481622SDavid Gibson } 23490481622SDavid Gibson 23590481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 23690481622SDavid Gibson { 23790481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 23890481622SDavid Gibson } 23990481622SDavid Gibson 24090481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 24190481622SDavid Gibson { 242496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 24390481622SDavid Gibson } 24490481622SDavid Gibson 2450db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2460db9d74eSMina Almasry * it for use. 2470db9d74eSMina Almasry */ 2480db9d74eSMina Almasry static struct file_region * 2490db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2500db9d74eSMina Almasry { 2510db9d74eSMina Almasry struct file_region *nrg = NULL; 2520db9d74eSMina Almasry 2530db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2540db9d74eSMina Almasry 2550db9d74eSMina Almasry resv->region_cache_count--; 2560db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2570db9d74eSMina Almasry list_del(&nrg->link); 2580db9d74eSMina Almasry 2590db9d74eSMina Almasry nrg->from = from; 2600db9d74eSMina Almasry nrg->to = to; 2610db9d74eSMina Almasry 2620db9d74eSMina Almasry return nrg; 2630db9d74eSMina Almasry } 2640db9d74eSMina Almasry 265075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 266075a61d0SMina Almasry struct file_region *rg) 267075a61d0SMina Almasry { 268075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 269075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 270075a61d0SMina Almasry nrg->css = rg->css; 271075a61d0SMina Almasry if (rg->css) 272075a61d0SMina Almasry css_get(rg->css); 273075a61d0SMina Almasry #endif 274075a61d0SMina Almasry } 275075a61d0SMina Almasry 276075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 277075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 278075a61d0SMina Almasry struct hstate *h, 279075a61d0SMina Almasry struct resv_map *resv, 280075a61d0SMina Almasry struct file_region *nrg) 281075a61d0SMina Almasry { 282075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 283075a61d0SMina Almasry if (h_cg) { 284075a61d0SMina Almasry nrg->reservation_counter = 285075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 286075a61d0SMina Almasry nrg->css = &h_cg->css; 287d85aecf2SMiaohe Lin /* 288d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 289d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 290d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 291d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 292d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 293d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 294d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 295d85aecf2SMiaohe Lin * untouched. 296d85aecf2SMiaohe Lin */ 297d85aecf2SMiaohe Lin css_get(&h_cg->css); 298075a61d0SMina Almasry if (!resv->pages_per_hpage) 299075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 300075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 301075a61d0SMina Almasry * a resv_map. 302075a61d0SMina Almasry */ 303075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 304075a61d0SMina Almasry } else { 305075a61d0SMina Almasry nrg->reservation_counter = NULL; 306075a61d0SMina Almasry nrg->css = NULL; 307075a61d0SMina Almasry } 308075a61d0SMina Almasry #endif 309075a61d0SMina Almasry } 310075a61d0SMina Almasry 311d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 312d85aecf2SMiaohe Lin { 313d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 314d85aecf2SMiaohe Lin if (rg->css) 315d85aecf2SMiaohe Lin css_put(rg->css); 316d85aecf2SMiaohe Lin #endif 317d85aecf2SMiaohe Lin } 318d85aecf2SMiaohe Lin 319a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 320a9b3f867SMina Almasry struct file_region *org) 321a9b3f867SMina Almasry { 322a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 323a9b3f867SMina Almasry return rg && org && 324a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 325a9b3f867SMina Almasry rg->css == org->css; 326a9b3f867SMina Almasry 327a9b3f867SMina Almasry #else 328a9b3f867SMina Almasry return true; 329a9b3f867SMina Almasry #endif 330a9b3f867SMina Almasry } 331a9b3f867SMina Almasry 332a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 333a9b3f867SMina Almasry { 334a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 335a9b3f867SMina Almasry 336a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 337a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 338a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 339a9b3f867SMina Almasry prg->to = rg->to; 340a9b3f867SMina Almasry 341a9b3f867SMina Almasry list_del(&rg->link); 342d85aecf2SMiaohe Lin put_uncharge_info(rg); 343a9b3f867SMina Almasry kfree(rg); 344a9b3f867SMina Almasry 3457db5e7b6SWei Yang rg = prg; 346a9b3f867SMina Almasry } 347a9b3f867SMina Almasry 348a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 349a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 350a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 351a9b3f867SMina Almasry nrg->from = rg->from; 352a9b3f867SMina Almasry 353a9b3f867SMina Almasry list_del(&rg->link); 354d85aecf2SMiaohe Lin put_uncharge_info(rg); 355a9b3f867SMina Almasry kfree(rg); 356a9b3f867SMina Almasry } 357a9b3f867SMina Almasry } 358a9b3f867SMina Almasry 3592103cf9cSPeter Xu static inline long 3602103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3612103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3622103cf9cSPeter Xu long *regions_needed) 3632103cf9cSPeter Xu { 3642103cf9cSPeter Xu struct file_region *nrg; 3652103cf9cSPeter Xu 3662103cf9cSPeter Xu if (!regions_needed) { 3672103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3682103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3692103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3702103cf9cSPeter Xu coalesce_file_region(map, nrg); 3712103cf9cSPeter Xu } else 3722103cf9cSPeter Xu *regions_needed += 1; 3732103cf9cSPeter Xu 3742103cf9cSPeter Xu return to - from; 3752103cf9cSPeter Xu } 3762103cf9cSPeter Xu 377972a3da3SWei Yang /* 378972a3da3SWei Yang * Must be called with resv->lock held. 379972a3da3SWei Yang * 380972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 381972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 382972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 383972a3da3SWei Yang * the regions for this range. 384d75c6af9SMina Almasry */ 385d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 386075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 387972a3da3SWei Yang struct hstate *h, long *regions_needed) 388d75c6af9SMina Almasry { 3890db9d74eSMina Almasry long add = 0; 390d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3910db9d74eSMina Almasry long last_accounted_offset = f; 3922103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 393d75c6af9SMina Almasry 3940db9d74eSMina Almasry if (regions_needed) 3950db9d74eSMina Almasry *regions_needed = 0; 396d75c6af9SMina Almasry 3970db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3980db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3990db9d74eSMina Almasry * bounds checking. 4000db9d74eSMina Almasry */ 4010db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4020db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4030db9d74eSMina Almasry if (rg->from < f) { 4040db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4050db9d74eSMina Almasry * then we need to update last_accounted_offset. 4060db9d74eSMina Almasry */ 4070db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4080db9d74eSMina Almasry last_accounted_offset = rg->to; 4090db9d74eSMina Almasry continue; 4100db9d74eSMina Almasry } 411d75c6af9SMina Almasry 4120db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4130db9d74eSMina Almasry * finished. 4140db9d74eSMina Almasry */ 415ca7e0457SPeter Xu if (rg->from >= t) 416d75c6af9SMina Almasry break; 417d75c6af9SMina Almasry 4180db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4190db9d74eSMina Almasry * update last_accounted_offset. 420d75c6af9SMina Almasry */ 4212103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4222103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4232103cf9cSPeter Xu last_accounted_offset, 4242103cf9cSPeter Xu rg->from, h, h_cg, 4252103cf9cSPeter Xu regions_needed); 426d75c6af9SMina Almasry 4270db9d74eSMina Almasry last_accounted_offset = rg->to; 4280db9d74eSMina Almasry } 4290db9d74eSMina Almasry 4300db9d74eSMina Almasry /* Handle the case where our range extends beyond 4310db9d74eSMina Almasry * last_accounted_offset. 4320db9d74eSMina Almasry */ 4332103cf9cSPeter Xu if (last_accounted_offset < t) 4342103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4352103cf9cSPeter Xu t, h, h_cg, regions_needed); 4360db9d74eSMina Almasry 4370db9d74eSMina Almasry VM_BUG_ON(add < 0); 4380db9d74eSMina Almasry return add; 4390db9d74eSMina Almasry } 4400db9d74eSMina Almasry 4410db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4420db9d74eSMina Almasry */ 4430db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4440db9d74eSMina Almasry int regions_needed) 4450db9d74eSMina Almasry __must_hold(&resv->lock) 4460db9d74eSMina Almasry { 4470db9d74eSMina Almasry struct list_head allocated_regions; 4480db9d74eSMina Almasry int to_allocate = 0, i = 0; 4490db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4500db9d74eSMina Almasry 4510db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4520db9d74eSMina Almasry 4530db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4540db9d74eSMina Almasry 4550db9d74eSMina Almasry /* 4560db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4570db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4580db9d74eSMina Almasry * 4590db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4600db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4610db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4620db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4630db9d74eSMina Almasry */ 4640db9d74eSMina Almasry while (resv->region_cache_count < 4650db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4660db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4670db9d74eSMina Almasry resv->region_cache_count; 4680db9d74eSMina Almasry 4690db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4700db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4710db9d74eSMina Almasry * needing to allocate for regions_needed. 4720db9d74eSMina Almasry */ 4730db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4740db9d74eSMina Almasry 4750db9d74eSMina Almasry spin_unlock(&resv->lock); 4760db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4770db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4780db9d74eSMina Almasry if (!trg) 4790db9d74eSMina Almasry goto out_of_memory; 4800db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4810db9d74eSMina Almasry } 4820db9d74eSMina Almasry 4830db9d74eSMina Almasry spin_lock(&resv->lock); 4840db9d74eSMina Almasry 485d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 486d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4870db9d74eSMina Almasry } 4880db9d74eSMina Almasry 4890db9d74eSMina Almasry return 0; 4900db9d74eSMina Almasry 4910db9d74eSMina Almasry out_of_memory: 4920db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 493d75c6af9SMina Almasry list_del(&rg->link); 494d75c6af9SMina Almasry kfree(rg); 495d75c6af9SMina Almasry } 4960db9d74eSMina Almasry return -ENOMEM; 497d75c6af9SMina Almasry } 498d75c6af9SMina Almasry 4991dd308a7SMike Kravetz /* 5001dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5010db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5020db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5030db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5040db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5050db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 506cf3ad20bSMike Kravetz * 5070db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5080db9d74eSMina Almasry * 5090db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5100db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5117c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5120db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5130db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5140db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5151dd308a7SMike Kravetz */ 5160db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 517075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 518075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 51996822904SAndy Whitcroft { 5200db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 52196822904SAndy Whitcroft 5227b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5230db9d74eSMina Almasry retry: 5240db9d74eSMina Almasry 5250db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 526972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 527972a3da3SWei Yang &actual_regions_needed); 52896822904SAndy Whitcroft 5295e911373SMike Kravetz /* 5300db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5310db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5320db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5330db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5340db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5350db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5360db9d74eSMina Almasry * operation. 5375e911373SMike Kravetz */ 5380db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5390db9d74eSMina Almasry resv->region_cache_count < 5400db9d74eSMina Almasry resv->adds_in_progress + 5410db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5420db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5430db9d74eSMina Almasry * allocate file_region entries. 5440db9d74eSMina Almasry */ 5450db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5465e911373SMike Kravetz 5470db9d74eSMina Almasry if (allocate_file_region_entries( 5480db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5490db9d74eSMina Almasry return -ENOMEM; 5505e911373SMike Kravetz } 5515e911373SMike Kravetz 5520db9d74eSMina Almasry goto retry; 5530db9d74eSMina Almasry } 554cf3ad20bSMike Kravetz 555972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5560db9d74eSMina Almasry 5570db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5580db9d74eSMina Almasry 5597b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 560cf3ad20bSMike Kravetz return add; 56196822904SAndy Whitcroft } 56296822904SAndy Whitcroft 5631dd308a7SMike Kravetz /* 5641dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5651dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5661dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5671dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5681dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5691dd308a7SMike Kravetz * not change the number of huge pages represented by the 5700db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5710db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5720db9d74eSMina Almasry * file_region structure is added. 5730db9d74eSMina Almasry * 5740db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5750db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5760db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5775e911373SMike Kravetz * 5785e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5795e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5805e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5815e911373SMike Kravetz * is needed and can not be allocated. 5821dd308a7SMike Kravetz */ 5830db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5840db9d74eSMina Almasry long *out_regions_needed) 58596822904SAndy Whitcroft { 58696822904SAndy Whitcroft long chg = 0; 58796822904SAndy Whitcroft 5887b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5895e911373SMike Kravetz 590972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 591075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 592972a3da3SWei Yang out_regions_needed); 5935e911373SMike Kravetz 5940db9d74eSMina Almasry if (*out_regions_needed == 0) 5950db9d74eSMina Almasry *out_regions_needed = 1; 5965e911373SMike Kravetz 5970db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5985e911373SMike Kravetz return -ENOMEM; 5995e911373SMike Kravetz 6000db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 60196822904SAndy Whitcroft 6027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 60396822904SAndy Whitcroft return chg; 60496822904SAndy Whitcroft } 60596822904SAndy Whitcroft 6061dd308a7SMike Kravetz /* 6075e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6085e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6095e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6105e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6110db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6120db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6130db9d74eSMina Almasry * the adds_in_progress counter. 6145e911373SMike Kravetz * 6155e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6165e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6175e911373SMike Kravetz * arguments will match the associated region_chg call. 6185e911373SMike Kravetz */ 6190db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6200db9d74eSMina Almasry long regions_needed) 6215e911373SMike Kravetz { 6225e911373SMike Kravetz spin_lock(&resv->lock); 6235e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6240db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6255e911373SMike Kravetz spin_unlock(&resv->lock); 6265e911373SMike Kravetz } 6275e911373SMike Kravetz 6285e911373SMike Kravetz /* 629feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 630feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 631feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 632feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 633feba16e2SMike Kravetz * 634feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 635feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 636feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 637feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 638feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 639feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 640feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6411dd308a7SMike Kravetz */ 642feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 64396822904SAndy Whitcroft { 6441406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 64596822904SAndy Whitcroft struct file_region *rg, *trg; 646feba16e2SMike Kravetz struct file_region *nrg = NULL; 647feba16e2SMike Kravetz long del = 0; 64896822904SAndy Whitcroft 649feba16e2SMike Kravetz retry: 6507b24d861SDavidlohr Bueso spin_lock(&resv->lock); 651feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 652dbe409e4SMike Kravetz /* 653dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 654dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 655dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 656dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 657dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 658dbe409e4SMike Kravetz */ 659dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 660feba16e2SMike Kravetz continue; 661dbe409e4SMike Kravetz 662feba16e2SMike Kravetz if (rg->from >= t) 66396822904SAndy Whitcroft break; 66496822904SAndy Whitcroft 665feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 666feba16e2SMike Kravetz /* 667feba16e2SMike Kravetz * Check for an entry in the cache before dropping 668feba16e2SMike Kravetz * lock and attempting allocation. 669feba16e2SMike Kravetz */ 670feba16e2SMike Kravetz if (!nrg && 671feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 672feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 673feba16e2SMike Kravetz struct file_region, 674feba16e2SMike Kravetz link); 675feba16e2SMike Kravetz list_del(&nrg->link); 676feba16e2SMike Kravetz resv->region_cache_count--; 67796822904SAndy Whitcroft } 67896822904SAndy Whitcroft 679feba16e2SMike Kravetz if (!nrg) { 680feba16e2SMike Kravetz spin_unlock(&resv->lock); 681feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 682feba16e2SMike Kravetz if (!nrg) 683feba16e2SMike Kravetz return -ENOMEM; 684feba16e2SMike Kravetz goto retry; 685feba16e2SMike Kravetz } 686feba16e2SMike Kravetz 687feba16e2SMike Kravetz del += t - f; 68879aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 689d85aecf2SMiaohe Lin resv, rg, t - f, false); 690feba16e2SMike Kravetz 691feba16e2SMike Kravetz /* New entry for end of split region */ 692feba16e2SMike Kravetz nrg->from = t; 693feba16e2SMike Kravetz nrg->to = rg->to; 694075a61d0SMina Almasry 695075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 696075a61d0SMina Almasry 697feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 698feba16e2SMike Kravetz 699feba16e2SMike Kravetz /* Original entry is trimmed */ 700feba16e2SMike Kravetz rg->to = f; 701feba16e2SMike Kravetz 702feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 703feba16e2SMike Kravetz nrg = NULL; 70496822904SAndy Whitcroft break; 705feba16e2SMike Kravetz } 706feba16e2SMike Kravetz 707feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 708feba16e2SMike Kravetz del += rg->to - rg->from; 709075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 710d85aecf2SMiaohe Lin rg->to - rg->from, true); 71196822904SAndy Whitcroft list_del(&rg->link); 71296822904SAndy Whitcroft kfree(rg); 713feba16e2SMike Kravetz continue; 71496822904SAndy Whitcroft } 7157b24d861SDavidlohr Bueso 716feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 717075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 718d85aecf2SMiaohe Lin t - rg->from, false); 719075a61d0SMina Almasry 72079aa925bSMike Kravetz del += t - rg->from; 72179aa925bSMike Kravetz rg->from = t; 72279aa925bSMike Kravetz } else { /* Trim end of region */ 723075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 724d85aecf2SMiaohe Lin rg->to - f, false); 72579aa925bSMike Kravetz 72679aa925bSMike Kravetz del += rg->to - f; 72779aa925bSMike Kravetz rg->to = f; 728feba16e2SMike Kravetz } 729feba16e2SMike Kravetz } 730feba16e2SMike Kravetz 7317b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 732feba16e2SMike Kravetz kfree(nrg); 733feba16e2SMike Kravetz return del; 73496822904SAndy Whitcroft } 73596822904SAndy Whitcroft 7361dd308a7SMike Kravetz /* 737b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 738b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 739b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 740b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 741b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 742b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 743b5cec28dSMike Kravetz * counts. 744b5cec28dSMike Kravetz */ 74572e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 746b5cec28dSMike Kravetz { 747b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 748b5cec28dSMike Kravetz long rsv_adjust; 749da56388cSMiaohe Lin bool reserved = false; 750b5cec28dSMike Kravetz 751b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 752da56388cSMiaohe Lin if (rsv_adjust > 0) { 753b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 754b5cec28dSMike Kravetz 755da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 756da56388cSMiaohe Lin reserved = true; 757da56388cSMiaohe Lin } else if (!rsv_adjust) { 758da56388cSMiaohe Lin reserved = true; 759b5cec28dSMike Kravetz } 760da56388cSMiaohe Lin 761da56388cSMiaohe Lin if (!reserved) 762da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 763b5cec28dSMike Kravetz } 764b5cec28dSMike Kravetz 765b5cec28dSMike Kravetz /* 7661dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7671dd308a7SMike Kravetz * that intersect with the range [f, t). 7681dd308a7SMike Kravetz */ 7691406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 77084afd99bSAndy Whitcroft { 7711406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 77284afd99bSAndy Whitcroft struct file_region *rg; 77384afd99bSAndy Whitcroft long chg = 0; 77484afd99bSAndy Whitcroft 7757b24d861SDavidlohr Bueso spin_lock(&resv->lock); 77684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 77784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 778f2135a4aSWang Sheng-Hui long seg_from; 779f2135a4aSWang Sheng-Hui long seg_to; 78084afd99bSAndy Whitcroft 78184afd99bSAndy Whitcroft if (rg->to <= f) 78284afd99bSAndy Whitcroft continue; 78384afd99bSAndy Whitcroft if (rg->from >= t) 78484afd99bSAndy Whitcroft break; 78584afd99bSAndy Whitcroft 78684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 78784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 78884afd99bSAndy Whitcroft 78984afd99bSAndy Whitcroft chg += seg_to - seg_from; 79084afd99bSAndy Whitcroft } 7917b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 79284afd99bSAndy Whitcroft 79384afd99bSAndy Whitcroft return chg; 79484afd99bSAndy Whitcroft } 79584afd99bSAndy Whitcroft 79696822904SAndy Whitcroft /* 797e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 798e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 799e7c4b0bfSAndy Whitcroft */ 800a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 801a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 802e7c4b0bfSAndy Whitcroft { 803a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 804a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 805e7c4b0bfSAndy Whitcroft } 806e7c4b0bfSAndy Whitcroft 8070fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8080fe6e20bSNaoya Horiguchi unsigned long address) 8090fe6e20bSNaoya Horiguchi { 8100fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8110fe6e20bSNaoya Horiguchi } 812dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8130fe6e20bSNaoya Horiguchi 81484afd99bSAndy Whitcroft /* 81508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 81608fba699SMel Gorman * cases this will be same size as used by the page table entries. 81708fba699SMel Gorman */ 81808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 81908fba699SMel Gorman { 82005ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 82105ea8860SDan Williams return vma->vm_ops->pagesize(vma); 82208fba699SMel Gorman return PAGE_SIZE; 82308fba699SMel Gorman } 824f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 82508fba699SMel Gorman 82608fba699SMel Gorman /* 8273340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8283340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 82909135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 83009135cc5SDan Williams * version of this symbol is required. 8313340289dSMel Gorman */ 83209135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8333340289dSMel Gorman { 8343340289dSMel Gorman return vma_kernel_pagesize(vma); 8353340289dSMel Gorman } 8363340289dSMel Gorman 8373340289dSMel Gorman /* 83884afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 83984afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 84084afd99bSAndy Whitcroft * alignment. 84184afd99bSAndy Whitcroft */ 84284afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 84384afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 84404f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 84584afd99bSAndy Whitcroft 846a1e78772SMel Gorman /* 847a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 848a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 849a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 850a1e78772SMel Gorman * 851a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 852a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 853a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 854a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 85584afd99bSAndy Whitcroft * 85684afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 85784afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 85884afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 85984afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 86084afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 86184afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 86284afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 86384afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 864a1e78772SMel Gorman */ 865e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 866e7c4b0bfSAndy Whitcroft { 867e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 868e7c4b0bfSAndy Whitcroft } 869e7c4b0bfSAndy Whitcroft 870e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 871e7c4b0bfSAndy Whitcroft unsigned long value) 872e7c4b0bfSAndy Whitcroft { 873e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 874e7c4b0bfSAndy Whitcroft } 875e7c4b0bfSAndy Whitcroft 876e9fe92aeSMina Almasry static void 877e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 878e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 879e9fe92aeSMina Almasry struct hstate *h) 880e9fe92aeSMina Almasry { 881e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 882e9fe92aeSMina Almasry if (!h_cg || !h) { 883e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 884e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 885e9fe92aeSMina Almasry resv_map->css = NULL; 886e9fe92aeSMina Almasry } else { 887e9fe92aeSMina Almasry resv_map->reservation_counter = 888e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 889e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 890e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 891e9fe92aeSMina Almasry } 892e9fe92aeSMina Almasry #endif 893e9fe92aeSMina Almasry } 894e9fe92aeSMina Almasry 8959119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 89684afd99bSAndy Whitcroft { 89784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8985e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8995e911373SMike Kravetz 9005e911373SMike Kravetz if (!resv_map || !rg) { 9015e911373SMike Kravetz kfree(resv_map); 9025e911373SMike Kravetz kfree(rg); 90384afd99bSAndy Whitcroft return NULL; 9045e911373SMike Kravetz } 90584afd99bSAndy Whitcroft 90684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9077b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 90884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 90984afd99bSAndy Whitcroft 9105e911373SMike Kravetz resv_map->adds_in_progress = 0; 911e9fe92aeSMina Almasry /* 912e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 913e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 914e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 915e9fe92aeSMina Almasry * reservations are to be un-charged from here. 916e9fe92aeSMina Almasry */ 917e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9185e911373SMike Kravetz 9195e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9205e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9215e911373SMike Kravetz resv_map->region_cache_count = 1; 9225e911373SMike Kravetz 92384afd99bSAndy Whitcroft return resv_map; 92484afd99bSAndy Whitcroft } 92584afd99bSAndy Whitcroft 9269119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 92784afd99bSAndy Whitcroft { 92884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9295e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9305e911373SMike Kravetz struct file_region *rg, *trg; 93184afd99bSAndy Whitcroft 93284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 933feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9345e911373SMike Kravetz 9355e911373SMike Kravetz /* ... and any entries left in the cache */ 9365e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9375e911373SMike Kravetz list_del(&rg->link); 9385e911373SMike Kravetz kfree(rg); 9395e911373SMike Kravetz } 9405e911373SMike Kravetz 9415e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9425e911373SMike Kravetz 94384afd99bSAndy Whitcroft kfree(resv_map); 94484afd99bSAndy Whitcroft } 94584afd99bSAndy Whitcroft 9464e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9474e35f483SJoonsoo Kim { 948f27a5136SMike Kravetz /* 949f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 950f27a5136SMike Kravetz * address space within the inode. This original address space 951f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 952f27a5136SMike Kravetz * address space embedded within the inode. 953f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 954f27a5136SMike Kravetz * this may not be true for device special inodes. 955f27a5136SMike Kravetz */ 956f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9574e35f483SJoonsoo Kim } 9584e35f483SJoonsoo Kim 95984afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 960a1e78772SMel Gorman { 96181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9624e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9634e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9644e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9654e35f483SJoonsoo Kim 9664e35f483SJoonsoo Kim return inode_resv_map(inode); 9674e35f483SJoonsoo Kim 9684e35f483SJoonsoo Kim } else { 96984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 97084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9714e35f483SJoonsoo Kim } 972a1e78772SMel Gorman } 973a1e78772SMel Gorman 97484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 975a1e78772SMel Gorman { 97681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 97781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 978a1e78772SMel Gorman 97984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 98084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 98104f2cbe3SMel Gorman } 98204f2cbe3SMel Gorman 98304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 98404f2cbe3SMel Gorman { 98581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 987e7c4b0bfSAndy Whitcroft 988e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 98904f2cbe3SMel Gorman } 99004f2cbe3SMel Gorman 99104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 99204f2cbe3SMel Gorman { 99381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 994e7c4b0bfSAndy Whitcroft 995e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 996a1e78772SMel Gorman } 997a1e78772SMel Gorman 99804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 999a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1000a1e78772SMel Gorman { 100181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1002f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1003a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1004a1e78772SMel Gorman } 1005a1e78772SMel Gorman 1006a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1007559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1008a1e78772SMel Gorman { 1009af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1010af0ed73eSJoonsoo Kim /* 1011af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1012af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1013af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1014af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1015af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1016af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1017af0ed73eSJoonsoo Kim * properly, so add work-around here. 1018af0ed73eSJoonsoo Kim */ 1019af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1020559ec2f8SNicholas Krause return true; 1021af0ed73eSJoonsoo Kim else 1022559ec2f8SNicholas Krause return false; 1023af0ed73eSJoonsoo Kim } 1024a63884e9SJoonsoo Kim 1025a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10261fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10271fb1b0e9SMike Kravetz /* 10281fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10291fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10301fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10317c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10321fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10331fb1b0e9SMike Kravetz */ 10341fb1b0e9SMike Kravetz if (chg) 10351fb1b0e9SMike Kravetz return false; 10361fb1b0e9SMike Kravetz else 1037559ec2f8SNicholas Krause return true; 10381fb1b0e9SMike Kravetz } 1039a63884e9SJoonsoo Kim 1040a63884e9SJoonsoo Kim /* 1041a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1042a63884e9SJoonsoo Kim * private mappings. 1043a63884e9SJoonsoo Kim */ 104467961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 104567961f9dSMike Kravetz /* 104667961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 104767961f9dSMike Kravetz * could have been performed on the private mapping. 104867961f9dSMike Kravetz * Examine the value of chg to determine if reserves 104967961f9dSMike Kravetz * actually exist or were previously consumed. 105067961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 105167961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 105267961f9dSMike Kravetz * private mappings has different (opposite) semantics 105367961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 105467961f9dSMike Kravetz * has already taken this difference in semantics into 105567961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 105667961f9dSMike Kravetz * as in the shared case above. Code could easily be 105767961f9dSMike Kravetz * combined, but keeping it separate draws attention to 105867961f9dSMike Kravetz * subtle differences. 105967961f9dSMike Kravetz */ 106067961f9dSMike Kravetz if (chg) 106167961f9dSMike Kravetz return false; 106267961f9dSMike Kravetz else 1063559ec2f8SNicholas Krause return true; 106467961f9dSMike Kravetz } 1065a63884e9SJoonsoo Kim 1066559ec2f8SNicholas Krause return false; 1067a1e78772SMel Gorman } 1068a1e78772SMel Gorman 1069a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10701da177e4SLinus Torvalds { 10711da177e4SLinus Torvalds int nid = page_to_nid(page); 10729487ca60SMike Kravetz 10739487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 10740edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1075a5516438SAndi Kleen h->free_huge_pages++; 1076a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10776c037149SMike Kravetz SetHPageFreed(page); 10781da177e4SLinus Torvalds } 10791da177e4SLinus Torvalds 108094310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1081bf50bab2SNaoya Horiguchi { 1082bf50bab2SNaoya Horiguchi struct page *page; 1083bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1084bf50bab2SNaoya Horiguchi 10859487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1086bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1087bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1088bbe88753SJoonsoo Kim continue; 1089bbe88753SJoonsoo Kim 10906664bfc8SWei Yang if (PageHWPoison(page)) 10916664bfc8SWei Yang continue; 1092bbe88753SJoonsoo Kim 10930edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1094a9869b83SNaoya Horiguchi set_page_refcounted(page); 10956c037149SMike Kravetz ClearHPageFreed(page); 1096bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1097bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1098bf50bab2SNaoya Horiguchi return page; 1099bf50bab2SNaoya Horiguchi } 1100bf50bab2SNaoya Horiguchi 11016664bfc8SWei Yang return NULL; 11026664bfc8SWei Yang } 11036664bfc8SWei Yang 11043e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11053e59fcb0SMichal Hocko nodemask_t *nmask) 110694310cbcSAnshuman Khandual { 11073e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11083e59fcb0SMichal Hocko struct zonelist *zonelist; 11093e59fcb0SMichal Hocko struct zone *zone; 11103e59fcb0SMichal Hocko struct zoneref *z; 111198fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11123e59fcb0SMichal Hocko 11133e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11143e59fcb0SMichal Hocko 11153e59fcb0SMichal Hocko retry_cpuset: 11163e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11173e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 111894310cbcSAnshuman Khandual struct page *page; 111994310cbcSAnshuman Khandual 11203e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11213e59fcb0SMichal Hocko continue; 11223e59fcb0SMichal Hocko /* 11233e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11243e59fcb0SMichal Hocko * zone aware 11253e59fcb0SMichal Hocko */ 11263e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11273e59fcb0SMichal Hocko continue; 11283e59fcb0SMichal Hocko node = zone_to_nid(zone); 112994310cbcSAnshuman Khandual 113094310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 113194310cbcSAnshuman Khandual if (page) 113294310cbcSAnshuman Khandual return page; 113394310cbcSAnshuman Khandual } 11343e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11353e59fcb0SMichal Hocko goto retry_cpuset; 11363e59fcb0SMichal Hocko 113794310cbcSAnshuman Khandual return NULL; 113894310cbcSAnshuman Khandual } 113994310cbcSAnshuman Khandual 1140a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1141a5516438SAndi Kleen struct vm_area_struct *vma, 1142af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1143af0ed73eSJoonsoo Kim long chg) 11441da177e4SLinus Torvalds { 11453e59fcb0SMichal Hocko struct page *page; 1146480eccf9SLee Schermerhorn struct mempolicy *mpol; 114704ec6264SVlastimil Babka gfp_t gfp_mask; 11483e59fcb0SMichal Hocko nodemask_t *nodemask; 114904ec6264SVlastimil Babka int nid; 11501da177e4SLinus Torvalds 1151a1e78772SMel Gorman /* 1152a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1153a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1154a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1155a1e78772SMel Gorman */ 1156af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1157a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1158c0ff7453SMiao Xie goto err; 1159a1e78772SMel Gorman 116004f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1161a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11626eab04a8SJustin P. Mattock goto err; 116304f2cbe3SMel Gorman 116404ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 116504ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11663e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11673e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1168d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1169a63884e9SJoonsoo Kim h->resv_huge_pages--; 1170bf50bab2SNaoya Horiguchi } 1171cc9a6c87SMel Gorman 1172cc9a6c87SMel Gorman mpol_cond_put(mpol); 1173cc9a6c87SMel Gorman return page; 1174cc9a6c87SMel Gorman 1175c0ff7453SMiao Xie err: 1176cc9a6c87SMel Gorman return NULL; 11771da177e4SLinus Torvalds } 11781da177e4SLinus Torvalds 11791cac6f2cSLuiz Capitulino /* 11801cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11811cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11821cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11831cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11841cac6f2cSLuiz Capitulino * node for alloc or free. 11851cac6f2cSLuiz Capitulino */ 11861cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11871cac6f2cSLuiz Capitulino { 11880edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11891cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11901cac6f2cSLuiz Capitulino 11911cac6f2cSLuiz Capitulino return nid; 11921cac6f2cSLuiz Capitulino } 11931cac6f2cSLuiz Capitulino 11941cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11951cac6f2cSLuiz Capitulino { 11961cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11971cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11981cac6f2cSLuiz Capitulino return nid; 11991cac6f2cSLuiz Capitulino } 12001cac6f2cSLuiz Capitulino 12011cac6f2cSLuiz Capitulino /* 12021cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12031cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12041cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12051cac6f2cSLuiz Capitulino * mask. 12061cac6f2cSLuiz Capitulino */ 12071cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12081cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12091cac6f2cSLuiz Capitulino { 12101cac6f2cSLuiz Capitulino int nid; 12111cac6f2cSLuiz Capitulino 12121cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12131cac6f2cSLuiz Capitulino 12141cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12151cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12161cac6f2cSLuiz Capitulino 12171cac6f2cSLuiz Capitulino return nid; 12181cac6f2cSLuiz Capitulino } 12191cac6f2cSLuiz Capitulino 12201cac6f2cSLuiz Capitulino /* 122110c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12221cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12231cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12241cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12251cac6f2cSLuiz Capitulino */ 12261cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12271cac6f2cSLuiz Capitulino { 12281cac6f2cSLuiz Capitulino int nid; 12291cac6f2cSLuiz Capitulino 12301cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12311cac6f2cSLuiz Capitulino 12321cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12331cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12341cac6f2cSLuiz Capitulino 12351cac6f2cSLuiz Capitulino return nid; 12361cac6f2cSLuiz Capitulino } 12371cac6f2cSLuiz Capitulino 12381cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12391cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12401cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12411cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12421cac6f2cSLuiz Capitulino nr_nodes--) 12431cac6f2cSLuiz Capitulino 12441cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12451cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12461cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12471cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12481cac6f2cSLuiz Capitulino nr_nodes--) 12491cac6f2cSLuiz Capitulino 1250e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1251944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1252d00181b9SKirill A. Shutemov unsigned int order) 1253944d9fecSLuiz Capitulino { 1254944d9fecSLuiz Capitulino int i; 1255944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1256944d9fecSLuiz Capitulino struct page *p = page + 1; 1257944d9fecSLuiz Capitulino 1258c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 125947e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 126047e29d32SJohn Hubbard 1261944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12621d798ca3SKirill A. Shutemov clear_compound_head(p); 1263944d9fecSLuiz Capitulino set_page_refcounted(p); 1264944d9fecSLuiz Capitulino } 1265944d9fecSLuiz Capitulino 1266944d9fecSLuiz Capitulino set_compound_order(page, 0); 1267ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1268944d9fecSLuiz Capitulino __ClearPageHead(page); 1269944d9fecSLuiz Capitulino } 1270944d9fecSLuiz Capitulino 1271d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1272944d9fecSLuiz Capitulino { 1273cf11e85fSRoman Gushchin /* 1274cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1275cf11e85fSRoman Gushchin * cma_release() returns false. 1276cf11e85fSRoman Gushchin */ 1277dbda8feaSBarry Song #ifdef CONFIG_CMA 1278dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1279cf11e85fSRoman Gushchin return; 1280dbda8feaSBarry Song #endif 1281cf11e85fSRoman Gushchin 1282944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1283944d9fecSLuiz Capitulino } 1284944d9fecSLuiz Capitulino 12854eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1286d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1287d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1288944d9fecSLuiz Capitulino { 128904adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1290953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1291953f064aSLi Xinhai nid = numa_mem_id(); 1292944d9fecSLuiz Capitulino 1293dbda8feaSBarry Song #ifdef CONFIG_CMA 1294dbda8feaSBarry Song { 1295cf11e85fSRoman Gushchin struct page *page; 1296cf11e85fSRoman Gushchin int node; 1297cf11e85fSRoman Gushchin 1298953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1299953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1300953f064aSLi Xinhai huge_page_order(h), true); 1301953f064aSLi Xinhai if (page) 1302953f064aSLi Xinhai return page; 1303953f064aSLi Xinhai } 1304953f064aSLi Xinhai 1305953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1306cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1307953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1308cf11e85fSRoman Gushchin continue; 1309cf11e85fSRoman Gushchin 1310cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1311cf11e85fSRoman Gushchin huge_page_order(h), true); 1312cf11e85fSRoman Gushchin if (page) 1313cf11e85fSRoman Gushchin return page; 1314cf11e85fSRoman Gushchin } 1315cf11e85fSRoman Gushchin } 1316953f064aSLi Xinhai } 1317dbda8feaSBarry Song #endif 1318cf11e85fSRoman Gushchin 13195e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1320944d9fecSLuiz Capitulino } 1321944d9fecSLuiz Capitulino 1322944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1323d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 13244eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13254eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13264eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13274eb0716eSAlexandre Ghiti { 13284eb0716eSAlexandre Ghiti return NULL; 13294eb0716eSAlexandre Ghiti } 13304eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1331944d9fecSLuiz Capitulino 1332e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1333d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13344eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13354eb0716eSAlexandre Ghiti { 13364eb0716eSAlexandre Ghiti return NULL; 13374eb0716eSAlexandre Ghiti } 1338d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1339944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1340d00181b9SKirill A. Shutemov unsigned int order) { } 1341944d9fecSLuiz Capitulino #endif 1342944d9fecSLuiz Capitulino 13436eb4e88aSMike Kravetz /* 13446eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 13456eb4e88aSMike Kravetz * as just a compound page. A reference is held on the page. 13466eb4e88aSMike Kravetz * 13476eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 13486eb4e88aSMike Kravetz */ 13496eb4e88aSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 13506eb4e88aSMike Kravetz bool adjust_surplus) 13516eb4e88aSMike Kravetz { 13526eb4e88aSMike Kravetz int nid = page_to_nid(page); 13536eb4e88aSMike Kravetz 13546eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13556eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 13566eb4e88aSMike Kravetz 13579487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 13586eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 13596eb4e88aSMike Kravetz return; 13606eb4e88aSMike Kravetz 13616eb4e88aSMike Kravetz list_del(&page->lru); 13626eb4e88aSMike Kravetz 13636eb4e88aSMike Kravetz if (HPageFreed(page)) { 13646eb4e88aSMike Kravetz h->free_huge_pages--; 13656eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 13666eb4e88aSMike Kravetz } 13676eb4e88aSMike Kravetz if (adjust_surplus) { 13686eb4e88aSMike Kravetz h->surplus_huge_pages--; 13696eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 13706eb4e88aSMike Kravetz } 13716eb4e88aSMike Kravetz 13726eb4e88aSMike Kravetz set_page_refcounted(page); 13736eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13746eb4e88aSMike Kravetz 13756eb4e88aSMike Kravetz h->nr_huge_pages--; 13766eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 13776eb4e88aSMike Kravetz } 13786eb4e88aSMike Kravetz 1379a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 13806af2acb6SAdam Litke { 13816af2acb6SAdam Litke int i; 1382dbfee5aeSMike Kravetz struct page *subpage = page; 1383a5516438SAndi Kleen 13844eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1385944d9fecSLuiz Capitulino return; 138618229df5SAndy Whitcroft 1387dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1388dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1389dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 139032f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1391a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1392a7407a27SLuiz Capitulino 1 << PG_writeback); 13936af2acb6SAdam Litke } 1394944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1395944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1396944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1397944d9fecSLuiz Capitulino } else { 1398a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13996af2acb6SAdam Litke } 1400944d9fecSLuiz Capitulino } 14016af2acb6SAdam Litke 140210c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 140310c6ec49SMike Kravetz { 140410c6ec49SMike Kravetz struct page *page, *t_page; 140510c6ec49SMike Kravetz 140610c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 140710c6ec49SMike Kravetz update_and_free_page(h, page); 140810c6ec49SMike Kravetz cond_resched(); 140910c6ec49SMike Kravetz } 141010c6ec49SMike Kravetz } 141110c6ec49SMike Kravetz 1412e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1413e5ff2159SAndi Kleen { 1414e5ff2159SAndi Kleen struct hstate *h; 1415e5ff2159SAndi Kleen 1416e5ff2159SAndi Kleen for_each_hstate(h) { 1417e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1418e5ff2159SAndi Kleen return h; 1419e5ff2159SAndi Kleen } 1420e5ff2159SAndi Kleen return NULL; 1421e5ff2159SAndi Kleen } 1422e5ff2159SAndi Kleen 1423db71ef79SMike Kravetz void free_huge_page(struct page *page) 142427a85ef1SDavid Gibson { 1425a5516438SAndi Kleen /* 1426a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1427a5516438SAndi Kleen * compound page destructor. 1428a5516438SAndi Kleen */ 1429e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 14307893d1d5SAdam Litke int nid = page_to_nid(page); 1431d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 143207443a85SJoonsoo Kim bool restore_reserve; 1433db71ef79SMike Kravetz unsigned long flags; 143427a85ef1SDavid Gibson 1435b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1436b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14378ace22bcSYongkai Wu 1438d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14398ace22bcSYongkai Wu page->mapping = NULL; 1440d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1441d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 144227a85ef1SDavid Gibson 14431c5ecae3SMike Kravetz /* 1444d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 14450919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14460919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14470919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14486c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 14490919e1b6SMike Kravetz * remove the reserved page from the subpool. 14500919e1b6SMike Kravetz */ 14510919e1b6SMike Kravetz if (!restore_reserve) { 14520919e1b6SMike Kravetz /* 14530919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14540919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14550919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14560919e1b6SMike Kravetz * operation. 14571c5ecae3SMike Kravetz */ 14581c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14591c5ecae3SMike Kravetz restore_reserve = true; 14600919e1b6SMike Kravetz } 14611c5ecae3SMike Kravetz 1462db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 14638f251a3dSMike Kravetz ClearHPageMigratable(page); 14646d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14656d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 146608cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 146708cf9fafSMina Almasry pages_per_huge_page(h), page); 146807443a85SJoonsoo Kim if (restore_reserve) 146907443a85SJoonsoo Kim h->resv_huge_pages++; 147007443a85SJoonsoo Kim 14719157c311SMike Kravetz if (HPageTemporary(page)) { 14726eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1473db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1474ab5ac90aSMichal Hocko update_and_free_page(h, page); 1475ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14760edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14776eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1478db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1479a5516438SAndi Kleen update_and_free_page(h, page); 14807893d1d5SAdam Litke } else { 14815d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1482a5516438SAndi Kleen enqueue_huge_page(h, page); 1483db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 148427a85ef1SDavid Gibson } 14851121828aSMike Kravetz } 148627a85ef1SDavid Gibson 1487d3d99fccSOscar Salvador /* 1488d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1489d3d99fccSOscar Salvador */ 1490d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1491d3d99fccSOscar Salvador { 1492d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1493d3d99fccSOscar Salvador h->nr_huge_pages++; 1494d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1495d3d99fccSOscar Salvador } 1496d3d99fccSOscar Salvador 1497d3d99fccSOscar Salvador static void __prep_new_huge_page(struct page *page) 1498b7ba30c6SAndi Kleen { 14990edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1500f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1501ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 15029dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15031adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1504d3d99fccSOscar Salvador } 1505d3d99fccSOscar Salvador 1506d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1507d3d99fccSOscar Salvador { 1508d3d99fccSOscar Salvador __prep_new_huge_page(page); 1509db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1510d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1511db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1512b7ba30c6SAndi Kleen } 1513b7ba30c6SAndi Kleen 1514d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 151520a0307cSWu Fengguang { 151620a0307cSWu Fengguang int i; 151720a0307cSWu Fengguang int nr_pages = 1 << order; 151820a0307cSWu Fengguang struct page *p = page + 1; 151920a0307cSWu Fengguang 152020a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 152120a0307cSWu Fengguang set_compound_order(page, order); 1522ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1523de09d31dSKirill A. Shutemov __SetPageHead(page); 152420a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1525ef5a22beSAndrea Arcangeli /* 1526ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1527ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1528ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15297c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1530ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1531ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1532ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1533ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1534ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1535ef5a22beSAndrea Arcangeli * after get_user_pages(). 1536ef5a22beSAndrea Arcangeli */ 1537ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 153858a84aa9SYouquan Song set_page_count(p, 0); 15391d798ca3SKirill A. Shutemov set_compound_head(p, page); 154020a0307cSWu Fengguang } 1541b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 154247e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 154320a0307cSWu Fengguang } 154420a0307cSWu Fengguang 15457795912cSAndrew Morton /* 15467795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15477795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15487795912cSAndrew Morton * details. 15497795912cSAndrew Morton */ 155020a0307cSWu Fengguang int PageHuge(struct page *page) 155120a0307cSWu Fengguang { 155220a0307cSWu Fengguang if (!PageCompound(page)) 155320a0307cSWu Fengguang return 0; 155420a0307cSWu Fengguang 155520a0307cSWu Fengguang page = compound_head(page); 1556f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 155720a0307cSWu Fengguang } 155843131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 155943131e14SNaoya Horiguchi 156027c73ae7SAndrea Arcangeli /* 156127c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 156227c73ae7SAndrea Arcangeli * normal or transparent huge pages. 156327c73ae7SAndrea Arcangeli */ 156427c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 156527c73ae7SAndrea Arcangeli { 156627c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 156727c73ae7SAndrea Arcangeli return 0; 156827c73ae7SAndrea Arcangeli 1569d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 157027c73ae7SAndrea Arcangeli } 157127c73ae7SAndrea Arcangeli 1572c0d0381aSMike Kravetz /* 1573c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1574c0d0381aSMike Kravetz * 1575336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1576336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1577336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1578c0d0381aSMike Kravetz */ 1579c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1580c0d0381aSMike Kravetz { 1581336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1582c0d0381aSMike Kravetz 1583c0d0381aSMike Kravetz if (!mapping) 1584c0d0381aSMike Kravetz return mapping; 1585c0d0381aSMike Kravetz 1586c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1587c0d0381aSMike Kravetz return mapping; 1588c0d0381aSMike Kravetz 1589c0d0381aSMike Kravetz return NULL; 1590c0d0381aSMike Kravetz } 1591c0d0381aSMike Kravetz 159213d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 159313d60f4bSZhang Yi { 159413d60f4bSZhang Yi struct page *page_head = compound_head(page); 159513d60f4bSZhang Yi pgoff_t index = page_index(page_head); 159613d60f4bSZhang Yi unsigned long compound_idx; 159713d60f4bSZhang Yi 159813d60f4bSZhang Yi if (!PageHuge(page_head)) 159913d60f4bSZhang Yi return page_index(page); 160013d60f4bSZhang Yi 160113d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 160213d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 160313d60f4bSZhang Yi else 160413d60f4bSZhang Yi compound_idx = page - page_head; 160513d60f4bSZhang Yi 160613d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 160713d60f4bSZhang Yi } 160813d60f4bSZhang Yi 16090c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1610f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1611f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16121da177e4SLinus Torvalds { 1613af0fb9dfSMichal Hocko int order = huge_page_order(h); 16141da177e4SLinus Torvalds struct page *page; 1615f60858f9SMike Kravetz bool alloc_try_hard = true; 1616f96efd58SJoe Jin 1617f60858f9SMike Kravetz /* 1618f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1619f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1620f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1621f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1622f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1623f60858f9SMike Kravetz */ 1624f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1625f60858f9SMike Kravetz alloc_try_hard = false; 1626f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1627f60858f9SMike Kravetz if (alloc_try_hard) 1628f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1629af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1630af0fb9dfSMichal Hocko nid = numa_mem_id(); 163184172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1632af0fb9dfSMichal Hocko if (page) 1633af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1634af0fb9dfSMichal Hocko else 1635af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 163663b4613cSNishanth Aravamudan 1637f60858f9SMike Kravetz /* 1638f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1639f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1640f60858f9SMike Kravetz * normal 'try hard' allocations. 1641f60858f9SMike Kravetz */ 1642f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1643f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1644f60858f9SMike Kravetz 1645f60858f9SMike Kravetz /* 1646f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1647f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1648f60858f9SMike Kravetz * overall state change. 1649f60858f9SMike Kravetz */ 1650f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1651f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1652f60858f9SMike Kravetz 165363b4613cSNishanth Aravamudan return page; 165463b4613cSNishanth Aravamudan } 165563b4613cSNishanth Aravamudan 1656af0fb9dfSMichal Hocko /* 16570c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16580c397daeSMichal Hocko * should use this function to get new hugetlb pages 16590c397daeSMichal Hocko */ 16600c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1661f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1662f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16630c397daeSMichal Hocko { 16640c397daeSMichal Hocko struct page *page; 16650c397daeSMichal Hocko 16660c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16670c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16680c397daeSMichal Hocko else 16690c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1670f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16710c397daeSMichal Hocko if (!page) 16720c397daeSMichal Hocko return NULL; 16730c397daeSMichal Hocko 16740c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16750c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16760c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16770c397daeSMichal Hocko 16780c397daeSMichal Hocko return page; 16790c397daeSMichal Hocko } 16800c397daeSMichal Hocko 16810c397daeSMichal Hocko /* 1682af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1683af0fb9dfSMichal Hocko * manner. 1684af0fb9dfSMichal Hocko */ 1685f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1686f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1687b2261026SJoonsoo Kim { 1688b2261026SJoonsoo Kim struct page *page; 1689b2261026SJoonsoo Kim int nr_nodes, node; 1690af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1691b2261026SJoonsoo Kim 1692b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1693f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1694f60858f9SMike Kravetz node_alloc_noretry); 1695af0fb9dfSMichal Hocko if (page) 1696b2261026SJoonsoo Kim break; 1697b2261026SJoonsoo Kim } 1698b2261026SJoonsoo Kim 1699af0fb9dfSMichal Hocko if (!page) 1700af0fb9dfSMichal Hocko return 0; 1701b2261026SJoonsoo Kim 1702af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1703af0fb9dfSMichal Hocko 1704af0fb9dfSMichal Hocko return 1; 1705b2261026SJoonsoo Kim } 1706b2261026SJoonsoo Kim 1707e8c5c824SLee Schermerhorn /* 170810c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 170910c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 171010c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 171110c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 1712e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1713e8c5c824SLee Schermerhorn */ 171410c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 171510c6ec49SMike Kravetz nodemask_t *nodes_allowed, 17166ae11b27SLee Schermerhorn bool acct_surplus) 1717e8c5c824SLee Schermerhorn { 1718b2261026SJoonsoo Kim int nr_nodes, node; 171910c6ec49SMike Kravetz struct page *page = NULL; 1720e8c5c824SLee Schermerhorn 17219487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1722b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1723685f3457SLee Schermerhorn /* 1724685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1725685f3457SLee Schermerhorn * nodes with surplus pages. 1726685f3457SLee Schermerhorn */ 1727b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1728b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 172910c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 1730e8c5c824SLee Schermerhorn struct page, lru); 17316eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 17329a76db09SLee Schermerhorn break; 1733e8c5c824SLee Schermerhorn } 1734b2261026SJoonsoo Kim } 1735e8c5c824SLee Schermerhorn 173610c6ec49SMike Kravetz return page; 1737e8c5c824SLee Schermerhorn } 1738e8c5c824SLee Schermerhorn 1739c8721bbbSNaoya Horiguchi /* 1740c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1741faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1742faf53defSNaoya Horiguchi * This function returns values like below: 1743faf53defSNaoya Horiguchi * 1744faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1745faf53defSNaoya Horiguchi * (allocated or reserved.) 1746faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1747faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1748c8721bbbSNaoya Horiguchi */ 1749c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1750c8721bbbSNaoya Horiguchi { 17516bc9b564SNaoya Horiguchi int rc = -EBUSY; 1752082d5b6bSGerald Schaefer 17537ffddd49SMuchun Song retry: 1754faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1755faf53defSNaoya Horiguchi if (!PageHuge(page)) 1756faf53defSNaoya Horiguchi return 0; 1757faf53defSNaoya Horiguchi 1758db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1759faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1760faf53defSNaoya Horiguchi rc = 0; 1761faf53defSNaoya Horiguchi goto out; 1762faf53defSNaoya Horiguchi } 1763faf53defSNaoya Horiguchi 1764faf53defSNaoya Horiguchi if (!page_count(page)) { 17652247bb33SGerald Schaefer struct page *head = compound_head(page); 17662247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17676bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1768082d5b6bSGerald Schaefer goto out; 17697ffddd49SMuchun Song 17707ffddd49SMuchun Song /* 17717ffddd49SMuchun Song * We should make sure that the page is already on the free list 17727ffddd49SMuchun Song * when it is dissolved. 17737ffddd49SMuchun Song */ 17746c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 1775db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 17767ffddd49SMuchun Song cond_resched(); 17777ffddd49SMuchun Song 17787ffddd49SMuchun Song /* 17797ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 17807ffddd49SMuchun Song * encounter this race. In fact, we have a chance 17817ffddd49SMuchun Song * to successfully dissolve the page if we do a 17827ffddd49SMuchun Song * retry. Because the race window is quite small. 17837ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 17847ffddd49SMuchun Song * for increasing the success rate of dissolving page. 17857ffddd49SMuchun Song */ 17867ffddd49SMuchun Song goto retry; 17877ffddd49SMuchun Song } 17887ffddd49SMuchun Song 1789c3114a84SAnshuman Khandual /* 1790c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1791c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1792c3114a84SAnshuman Khandual */ 1793c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1794c3114a84SAnshuman Khandual SetPageHWPoison(page); 1795c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1796c3114a84SAnshuman Khandual } 17976eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1798c1470b33Szhong jiang h->max_huge_pages--; 1799db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18002247bb33SGerald Schaefer update_and_free_page(h, head); 18011121828aSMike Kravetz return 0; 1802c8721bbbSNaoya Horiguchi } 1803082d5b6bSGerald Schaefer out: 1804db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1805082d5b6bSGerald Schaefer return rc; 1806c8721bbbSNaoya Horiguchi } 1807c8721bbbSNaoya Horiguchi 1808c8721bbbSNaoya Horiguchi /* 1809c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1810c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18112247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18122247bb33SGerald Schaefer * part of it lies within the given range. 1813082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1814082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1815c8721bbbSNaoya Horiguchi */ 1816082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1817c8721bbbSNaoya Horiguchi { 1818c8721bbbSNaoya Horiguchi unsigned long pfn; 1819eb03aa00SGerald Schaefer struct page *page; 1820082d5b6bSGerald Schaefer int rc = 0; 1821c8721bbbSNaoya Horiguchi 1822d0177639SLi Zhong if (!hugepages_supported()) 1823082d5b6bSGerald Schaefer return rc; 1824d0177639SLi Zhong 1825eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1826eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1827eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1828eb03aa00SGerald Schaefer if (rc) 1829082d5b6bSGerald Schaefer break; 1830eb03aa00SGerald Schaefer } 1831082d5b6bSGerald Schaefer 1832082d5b6bSGerald Schaefer return rc; 1833c8721bbbSNaoya Horiguchi } 1834c8721bbbSNaoya Horiguchi 1835ab5ac90aSMichal Hocko /* 1836ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1837ab5ac90aSMichal Hocko */ 18380c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1839aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18407893d1d5SAdam Litke { 18419980d744SMichal Hocko struct page *page = NULL; 18427893d1d5SAdam Litke 1843bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1844aa888a74SAndi Kleen return NULL; 1845aa888a74SAndi Kleen 1846db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 18479980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18489980d744SMichal Hocko goto out_unlock; 1849db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1850d1c3fb1fSNishanth Aravamudan 1851f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18529980d744SMichal Hocko if (!page) 18530c397daeSMichal Hocko return NULL; 1854d1c3fb1fSNishanth Aravamudan 1855db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 18569980d744SMichal Hocko /* 18579980d744SMichal Hocko * We could have raced with the pool size change. 18589980d744SMichal Hocko * Double check that and simply deallocate the new page 18599980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18609980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18619980d744SMichal Hocko * codeflow 18629980d744SMichal Hocko */ 18639980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18649157c311SMike Kravetz SetHPageTemporary(page); 1865db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18669980d744SMichal Hocko put_page(page); 18672bf753e6SKai Shen return NULL; 18689980d744SMichal Hocko } else { 18699980d744SMichal Hocko h->surplus_huge_pages++; 18704704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18717893d1d5SAdam Litke } 18729980d744SMichal Hocko 18739980d744SMichal Hocko out_unlock: 1874db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18757893d1d5SAdam Litke 18767893d1d5SAdam Litke return page; 18777893d1d5SAdam Litke } 18787893d1d5SAdam Litke 1879bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1880ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1881ab5ac90aSMichal Hocko { 1882ab5ac90aSMichal Hocko struct page *page; 1883ab5ac90aSMichal Hocko 1884ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1885ab5ac90aSMichal Hocko return NULL; 1886ab5ac90aSMichal Hocko 1887f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1888ab5ac90aSMichal Hocko if (!page) 1889ab5ac90aSMichal Hocko return NULL; 1890ab5ac90aSMichal Hocko 1891ab5ac90aSMichal Hocko /* 1892ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1893ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1894ab5ac90aSMichal Hocko */ 18959157c311SMike Kravetz SetHPageTemporary(page); 1896ab5ac90aSMichal Hocko 1897ab5ac90aSMichal Hocko return page; 1898ab5ac90aSMichal Hocko } 1899ab5ac90aSMichal Hocko 1900e4e574b7SAdam Litke /* 1901099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1902099730d6SDave Hansen */ 1903e0ec90eeSDave Hansen static 19040c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1905099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1906099730d6SDave Hansen { 1907aaf14e40SMichal Hocko struct page *page; 1908aaf14e40SMichal Hocko struct mempolicy *mpol; 1909aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1910aaf14e40SMichal Hocko int nid; 1911aaf14e40SMichal Hocko nodemask_t *nodemask; 1912aaf14e40SMichal Hocko 1913aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19140c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1915aaf14e40SMichal Hocko mpol_cond_put(mpol); 1916aaf14e40SMichal Hocko 1917aaf14e40SMichal Hocko return page; 1918099730d6SDave Hansen } 1919099730d6SDave Hansen 1920ab5ac90aSMichal Hocko /* page migration callback function */ 19213e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1922d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19234db9b2efSMichal Hocko { 1924db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 19254db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19263e59fcb0SMichal Hocko struct page *page; 19273e59fcb0SMichal Hocko 19283e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19293e59fcb0SMichal Hocko if (page) { 1930db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19313e59fcb0SMichal Hocko return page; 19324db9b2efSMichal Hocko } 19334db9b2efSMichal Hocko } 1934db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19354db9b2efSMichal Hocko 19360c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19374db9b2efSMichal Hocko } 19384db9b2efSMichal Hocko 1939ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1940389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1941389c8178SMichal Hocko unsigned long address) 1942ebd63723SMichal Hocko { 1943ebd63723SMichal Hocko struct mempolicy *mpol; 1944ebd63723SMichal Hocko nodemask_t *nodemask; 1945ebd63723SMichal Hocko struct page *page; 1946ebd63723SMichal Hocko gfp_t gfp_mask; 1947ebd63723SMichal Hocko int node; 1948ebd63723SMichal Hocko 1949ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1950ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1951d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1952ebd63723SMichal Hocko mpol_cond_put(mpol); 1953ebd63723SMichal Hocko 1954ebd63723SMichal Hocko return page; 1955ebd63723SMichal Hocko } 1956ebd63723SMichal Hocko 1957bf50bab2SNaoya Horiguchi /* 195825985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1959e4e574b7SAdam Litke * of size 'delta'. 1960e4e574b7SAdam Litke */ 19610a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19621b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1963e4e574b7SAdam Litke { 1964e4e574b7SAdam Litke struct list_head surplus_list; 1965e4e574b7SAdam Litke struct page *page, *tmp; 19660a4f3d1bSLiu Xiang int ret; 19670a4f3d1bSLiu Xiang long i; 19680a4f3d1bSLiu Xiang long needed, allocated; 196928073b02SHillf Danton bool alloc_ok = true; 1970e4e574b7SAdam Litke 19719487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1972a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1973ac09b3a1SAdam Litke if (needed <= 0) { 1974a5516438SAndi Kleen h->resv_huge_pages += delta; 1975e4e574b7SAdam Litke return 0; 1976ac09b3a1SAdam Litke } 1977e4e574b7SAdam Litke 1978e4e574b7SAdam Litke allocated = 0; 1979e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1980e4e574b7SAdam Litke 1981e4e574b7SAdam Litke ret = -ENOMEM; 1982e4e574b7SAdam Litke retry: 1983db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1984e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19850c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1986aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 198728073b02SHillf Danton if (!page) { 198828073b02SHillf Danton alloc_ok = false; 198928073b02SHillf Danton break; 199028073b02SHillf Danton } 1991e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 199269ed779aSDavid Rientjes cond_resched(); 1993e4e574b7SAdam Litke } 199428073b02SHillf Danton allocated += i; 1995e4e574b7SAdam Litke 1996e4e574b7SAdam Litke /* 1997e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1998e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1999e4e574b7SAdam Litke */ 2000db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2001a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2002a5516438SAndi Kleen (h->free_huge_pages + allocated); 200328073b02SHillf Danton if (needed > 0) { 200428073b02SHillf Danton if (alloc_ok) 2005e4e574b7SAdam Litke goto retry; 200628073b02SHillf Danton /* 200728073b02SHillf Danton * We were not able to allocate enough pages to 200828073b02SHillf Danton * satisfy the entire reservation so we free what 200928073b02SHillf Danton * we've allocated so far. 201028073b02SHillf Danton */ 201128073b02SHillf Danton goto free; 201228073b02SHillf Danton } 2013e4e574b7SAdam Litke /* 2014e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 201525985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2016e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2017ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2018ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2019ac09b3a1SAdam Litke * before they are reserved. 2020e4e574b7SAdam Litke */ 2021e4e574b7SAdam Litke needed += allocated; 2022a5516438SAndi Kleen h->resv_huge_pages += delta; 2023e4e574b7SAdam Litke ret = 0; 2024a9869b83SNaoya Horiguchi 202519fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 202619fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2027e558464bSMuchun Song int zeroed; 2028e558464bSMuchun Song 202919fc3f0aSAdam Litke if ((--needed) < 0) 203019fc3f0aSAdam Litke break; 2031a9869b83SNaoya Horiguchi /* 2032a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2033a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2034a9869b83SNaoya Horiguchi */ 2035e558464bSMuchun Song zeroed = put_page_testzero(page); 2036e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2037a5516438SAndi Kleen enqueue_huge_page(h, page); 203819fc3f0aSAdam Litke } 203928073b02SHillf Danton free: 2040db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 204119fc3f0aSAdam Litke 204219fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2043c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2044a9869b83SNaoya Horiguchi put_page(page); 2045db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2046e4e574b7SAdam Litke 2047e4e574b7SAdam Litke return ret; 2048e4e574b7SAdam Litke } 2049e4e574b7SAdam Litke 2050e4e574b7SAdam Litke /* 2051e5bbc8a6SMike Kravetz * This routine has two main purposes: 2052e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2053e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2054e5bbc8a6SMike Kravetz * to the associated reservation map. 2055e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2056e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2057e4e574b7SAdam Litke */ 2058a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2059a5516438SAndi Kleen unsigned long unused_resv_pages) 2060e4e574b7SAdam Litke { 2061e4e574b7SAdam Litke unsigned long nr_pages; 206210c6ec49SMike Kravetz struct page *page; 206310c6ec49SMike Kravetz LIST_HEAD(page_list); 206410c6ec49SMike Kravetz 20659487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 206610c6ec49SMike Kravetz /* Uncommit the reservation */ 206710c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2068e4e574b7SAdam Litke 2069aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2070bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2071e5bbc8a6SMike Kravetz goto out; 2072aa888a74SAndi Kleen 2073e5bbc8a6SMike Kravetz /* 2074e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2075e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2076e5bbc8a6SMike Kravetz */ 2077a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2078e4e574b7SAdam Litke 2079685f3457SLee Schermerhorn /* 2080685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20819b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20829b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20839b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 208410c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 20859b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2086685f3457SLee Schermerhorn */ 2087685f3457SLee Schermerhorn while (nr_pages--) { 208810c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 208910c6ec49SMike Kravetz if (!page) 2090e5bbc8a6SMike Kravetz goto out; 209110c6ec49SMike Kravetz 209210c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2093e4e574b7SAdam Litke } 2094e5bbc8a6SMike Kravetz 2095e5bbc8a6SMike Kravetz out: 2096db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 209710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2098db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2099e4e574b7SAdam Litke } 2100e4e574b7SAdam Litke 21015e911373SMike Kravetz 2102c37f9fb1SAndy Whitcroft /* 2103feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21045e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2105cf3ad20bSMike Kravetz * 2106cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2107cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2108cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2109cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2110cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2111feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2112feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2113feba16e2SMike Kravetz * is called in such cases. 2114cf3ad20bSMike Kravetz * 2115cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2116cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2117cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2118cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2119cf3ad20bSMike Kravetz * take appropriate action. 212096b96a96SMike Kravetz * 212196b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 212296b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 212396b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 212496b96a96SMike Kravetz * reservation exists. 2125c37f9fb1SAndy Whitcroft */ 21265e911373SMike Kravetz enum vma_resv_mode { 21275e911373SMike Kravetz VMA_NEEDS_RESV, 21285e911373SMike Kravetz VMA_COMMIT_RESV, 2129feba16e2SMike Kravetz VMA_END_RESV, 213096b96a96SMike Kravetz VMA_ADD_RESV, 21315e911373SMike Kravetz }; 2132cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2133cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21345e911373SMike Kravetz enum vma_resv_mode mode) 2135c37f9fb1SAndy Whitcroft { 21364e35f483SJoonsoo Kim struct resv_map *resv; 21374e35f483SJoonsoo Kim pgoff_t idx; 2138cf3ad20bSMike Kravetz long ret; 21390db9d74eSMina Almasry long dummy_out_regions_needed; 2140c37f9fb1SAndy Whitcroft 21414e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21424e35f483SJoonsoo Kim if (!resv) 2143c37f9fb1SAndy Whitcroft return 1; 2144c37f9fb1SAndy Whitcroft 21454e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21465e911373SMike Kravetz switch (mode) { 21475e911373SMike Kravetz case VMA_NEEDS_RESV: 21480db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21490db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21500db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21510db9d74eSMina Almasry * ever require 1 region. 21520db9d74eSMina Almasry */ 21530db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21545e911373SMike Kravetz break; 21555e911373SMike Kravetz case VMA_COMMIT_RESV: 2156075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21570db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21580db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21595e911373SMike Kravetz break; 2160feba16e2SMike Kravetz case VMA_END_RESV: 21610db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21625e911373SMike Kravetz ret = 0; 21635e911373SMike Kravetz break; 216496b96a96SMike Kravetz case VMA_ADD_RESV: 21650db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2166075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21670db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21680db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21690db9d74eSMina Almasry } else { 21700db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 217196b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 217296b96a96SMike Kravetz } 217396b96a96SMike Kravetz break; 21745e911373SMike Kravetz default: 21755e911373SMike Kravetz BUG(); 21765e911373SMike Kravetz } 217784afd99bSAndy Whitcroft 21784e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2179cf3ad20bSMike Kravetz return ret; 218067961f9dSMike Kravetz /* 2181bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2182bf3d12b9SMiaohe Lin * 218367961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 218467961f9dSMike Kravetz * However, a file associated with mapping could have been 218567961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 218667961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 218767961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 218867961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 218967961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 219067961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 219167961f9dSMike Kravetz * reservation has already been consumed. As a result, the 219267961f9dSMike Kravetz * return value of this routine is the opposite of the 219367961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 219467961f9dSMike Kravetz */ 2195bf3d12b9SMiaohe Lin if (ret > 0) 219667961f9dSMike Kravetz return 0; 2197bf3d12b9SMiaohe Lin if (ret == 0) 219867961f9dSMike Kravetz return 1; 2199bf3d12b9SMiaohe Lin return ret; 220084afd99bSAndy Whitcroft } 2201cf3ad20bSMike Kravetz 2202cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2203a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2204c37f9fb1SAndy Whitcroft { 22055e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2206cf3ad20bSMike Kravetz } 2207c37f9fb1SAndy Whitcroft 2208cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2209cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2210cf3ad20bSMike Kravetz { 22115e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22125e911373SMike Kravetz } 22135e911373SMike Kravetz 2214feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22155e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22165e911373SMike Kravetz { 2217feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2218c37f9fb1SAndy Whitcroft } 2219c37f9fb1SAndy Whitcroft 222096b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 222196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 222296b96a96SMike Kravetz { 222396b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 222496b96a96SMike Kravetz } 222596b96a96SMike Kravetz 222696b96a96SMike Kravetz /* 222796b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 222896b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 222996b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 2230d6995da3SMike Kravetz * alloc_huge_page would have consumed the reservation and set 2231d6995da3SMike Kravetz * HPageRestoreReserve in the newly allocated page. When the page is freed 2232d6995da3SMike Kravetz * via free_huge_page, the global reservation count will be incremented if 2233d6995da3SMike Kravetz * HPageRestoreReserve is set. However, free_huge_page can not adjust the 2234d6995da3SMike Kravetz * reserve map. Adjust the reserve map here to be consistent with global 2235d6995da3SMike Kravetz * reserve count adjustments to be made by free_huge_page. 223696b96a96SMike Kravetz */ 223796b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 223896b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 223996b96a96SMike Kravetz struct page *page) 224096b96a96SMike Kravetz { 2241d6995da3SMike Kravetz if (unlikely(HPageRestoreReserve(page))) { 224296b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 224396b96a96SMike Kravetz 224496b96a96SMike Kravetz if (unlikely(rc < 0)) { 224596b96a96SMike Kravetz /* 224696b96a96SMike Kravetz * Rare out of memory condition in reserve map 2247d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 224896b96a96SMike Kravetz * global reserve count will not be incremented 224996b96a96SMike Kravetz * by free_huge_page. This will make it appear 225096b96a96SMike Kravetz * as though the reservation for this page was 225196b96a96SMike Kravetz * consumed. This may prevent the task from 225296b96a96SMike Kravetz * faulting in the page at a later time. This 225396b96a96SMike Kravetz * is better than inconsistent global huge page 225496b96a96SMike Kravetz * accounting of reserve counts. 225596b96a96SMike Kravetz */ 2256d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 225796b96a96SMike Kravetz } else if (rc) { 225896b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 225996b96a96SMike Kravetz if (unlikely(rc < 0)) 226096b96a96SMike Kravetz /* 226196b96a96SMike Kravetz * See above comment about rare out of 226296b96a96SMike Kravetz * memory condition. 226396b96a96SMike Kravetz */ 2264d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 226596b96a96SMike Kravetz } else 226696b96a96SMike Kravetz vma_end_reservation(h, vma, address); 226796b96a96SMike Kravetz } 226896b96a96SMike Kravetz } 226996b96a96SMike Kravetz 2270369fa227SOscar Salvador /* 2271369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2272369fa227SOscar Salvador * @h: struct hstate old page belongs to 2273369fa227SOscar Salvador * @old_page: Old page to dissolve 2274*ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2275369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2276369fa227SOscar Salvador */ 2277*ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2278*ae37c7ffSOscar Salvador struct list_head *list) 2279369fa227SOscar Salvador { 2280369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2281369fa227SOscar Salvador int nid = page_to_nid(old_page); 2282369fa227SOscar Salvador struct page *new_page; 2283369fa227SOscar Salvador int ret = 0; 2284369fa227SOscar Salvador 2285369fa227SOscar Salvador /* 2286369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2287369fa227SOscar Salvador * pool to remain stable. Using alloc_buddy_huge_page() allows us to 2288369fa227SOscar Salvador * not having to deal with prep_new_huge_page() and avoids dealing of any 2289369fa227SOscar Salvador * counters. This simplifies and let us do the whole thing under the 2290369fa227SOscar Salvador * lock. 2291369fa227SOscar Salvador */ 2292369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2293369fa227SOscar Salvador if (!new_page) 2294369fa227SOscar Salvador return -ENOMEM; 2295369fa227SOscar Salvador 2296369fa227SOscar Salvador retry: 2297369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2298369fa227SOscar Salvador if (!PageHuge(old_page)) { 2299369fa227SOscar Salvador /* 2300369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2301369fa227SOscar Salvador */ 2302369fa227SOscar Salvador goto free_new; 2303369fa227SOscar Salvador } else if (page_count(old_page)) { 2304369fa227SOscar Salvador /* 2305*ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2306*ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2307369fa227SOscar Salvador */ 2308*ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2309*ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2310369fa227SOscar Salvador ret = -EBUSY; 2311*ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2312369fa227SOscar Salvador goto free_new; 2313369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2314369fa227SOscar Salvador /* 2315369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2316369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2317369fa227SOscar Salvador * we retry. 2318369fa227SOscar Salvador */ 2319369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2320369fa227SOscar Salvador cond_resched(); 2321369fa227SOscar Salvador goto retry; 2322369fa227SOscar Salvador } else { 2323369fa227SOscar Salvador /* 2324369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2325369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2326369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2327369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2328369fa227SOscar Salvador * stable since this happens under the lock. 2329369fa227SOscar Salvador */ 2330369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2331369fa227SOscar Salvador 2332369fa227SOscar Salvador /* 2333369fa227SOscar Salvador * new_page needs to be initialized with the standard hugetlb 2334369fa227SOscar Salvador * state. This is normally done by prep_new_huge_page() but 2335369fa227SOscar Salvador * that takes hugetlb_lock which is already held so we need to 2336369fa227SOscar Salvador * open code it here. 2337369fa227SOscar Salvador * Reference count trick is needed because allocator gives us 2338369fa227SOscar Salvador * referenced page but the pool requires pages with 0 refcount. 2339369fa227SOscar Salvador */ 2340369fa227SOscar Salvador __prep_new_huge_page(new_page); 2341369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2342369fa227SOscar Salvador page_ref_dec(new_page); 2343369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2344369fa227SOscar Salvador 2345369fa227SOscar Salvador /* 2346369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2347369fa227SOscar Salvador */ 2348369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2349369fa227SOscar Salvador update_and_free_page(h, old_page); 2350369fa227SOscar Salvador } 2351369fa227SOscar Salvador 2352369fa227SOscar Salvador return ret; 2353369fa227SOscar Salvador 2354369fa227SOscar Salvador free_new: 2355369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2356369fa227SOscar Salvador __free_pages(new_page, huge_page_order(h)); 2357369fa227SOscar Salvador 2358369fa227SOscar Salvador return ret; 2359369fa227SOscar Salvador } 2360369fa227SOscar Salvador 2361*ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2362369fa227SOscar Salvador { 2363369fa227SOscar Salvador struct hstate *h; 2364369fa227SOscar Salvador struct page *head; 2365*ae37c7ffSOscar Salvador int ret = -EBUSY; 2366369fa227SOscar Salvador 2367369fa227SOscar Salvador /* 2368369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2369369fa227SOscar Salvador * to carefully check the state under the lock. 2370369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2371369fa227SOscar Salvador */ 2372369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2373369fa227SOscar Salvador if (PageHuge(page)) { 2374369fa227SOscar Salvador head = compound_head(page); 2375369fa227SOscar Salvador h = page_hstate(head); 2376369fa227SOscar Salvador } else { 2377369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2378369fa227SOscar Salvador return 0; 2379369fa227SOscar Salvador } 2380369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2381369fa227SOscar Salvador 2382369fa227SOscar Salvador /* 2383369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2384369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2385369fa227SOscar Salvador * of bailing out right away without further retrying. 2386369fa227SOscar Salvador */ 2387369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2388369fa227SOscar Salvador return -ENOMEM; 2389369fa227SOscar Salvador 2390*ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2391*ae37c7ffSOscar Salvador ret = 0; 2392*ae37c7ffSOscar Salvador else if (!page_count(head)) 2393*ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2394*ae37c7ffSOscar Salvador 2395*ae37c7ffSOscar Salvador return ret; 2396369fa227SOscar Salvador } 2397369fa227SOscar Salvador 239870c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 239904f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2400348ea204SAdam Litke { 240190481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2402a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2403348ea204SAdam Litke struct page *page; 2404d85f69b0SMike Kravetz long map_chg, map_commit; 2405d85f69b0SMike Kravetz long gbl_chg; 24066d76dcf4SAneesh Kumar K.V int ret, idx; 24076d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 240808cf9fafSMina Almasry bool deferred_reserve; 24092fc39cecSAdam Litke 24106d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2411a1e78772SMel Gorman /* 2412d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2413d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2414d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2415a1e78772SMel Gorman */ 2416d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2417d85f69b0SMike Kravetz if (map_chg < 0) 241876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2419d85f69b0SMike Kravetz 2420d85f69b0SMike Kravetz /* 2421d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2422d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2423d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2424d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2425d85f69b0SMike Kravetz * checked against any subpool limit. 2426d85f69b0SMike Kravetz */ 2427d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2428d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2429d85f69b0SMike Kravetz if (gbl_chg < 0) { 2430feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 243176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 24325e911373SMike Kravetz } 243390d8b7e6SAdam Litke 2434d85f69b0SMike Kravetz /* 2435d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2436d85f69b0SMike Kravetz * map, there could be reservations associated with the 2437d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2438d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2439d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2440d85f69b0SMike Kravetz * the subpool reservations. 2441d85f69b0SMike Kravetz */ 2442d85f69b0SMike Kravetz if (avoid_reserve) 2443d85f69b0SMike Kravetz gbl_chg = 1; 2444d85f69b0SMike Kravetz } 2445d85f69b0SMike Kravetz 244608cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 244708cf9fafSMina Almasry */ 24486501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 244908cf9fafSMina Almasry if (deferred_reserve) { 245008cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 245108cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 24528f34af6fSJianyu Zhan if (ret) 24538f34af6fSJianyu Zhan goto out_subpool_put; 245408cf9fafSMina Almasry } 245508cf9fafSMina Almasry 245608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 245708cf9fafSMina Almasry if (ret) 245808cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 24598f34af6fSJianyu Zhan 2460db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2461d85f69b0SMike Kravetz /* 2462d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2463d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2464d85f69b0SMike Kravetz * a reservation exists for the allocation. 2465d85f69b0SMike Kravetz */ 2466d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 246781a6fcaeSJoonsoo Kim if (!page) { 2468db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24690c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 24708f34af6fSJianyu Zhan if (!page) 24718f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2472a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2473d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2474a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2475a88c7695SNaoya Horiguchi } 2476db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 247715a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 247881a6fcaeSJoonsoo Kim /* Fall through */ 2479a1e78772SMel Gorman } 248081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 248108cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 248208cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 248308cf9fafSMina Almasry */ 248408cf9fafSMina Almasry if (deferred_reserve) { 248508cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 248608cf9fafSMina Almasry h_cg, page); 248708cf9fafSMina Almasry } 248808cf9fafSMina Almasry 2489db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2490a1e78772SMel Gorman 2491d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2492a1e78772SMel Gorman 2493d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2494d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 249533039678SMike Kravetz /* 249633039678SMike Kravetz * The page was added to the reservation map between 249733039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 249833039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 249933039678SMike Kravetz * Adjust for the subpool count incremented above AND 250033039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 250133039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 250233039678SMike Kravetz * no longer applies. 250333039678SMike Kravetz */ 250433039678SMike Kravetz long rsv_adjust; 250533039678SMike Kravetz 250633039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 250733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 250879aa925bSMike Kravetz if (deferred_reserve) 250979aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 251079aa925bSMike Kravetz pages_per_huge_page(h), page); 251133039678SMike Kravetz } 25127893d1d5SAdam Litke return page; 25138f34af6fSJianyu Zhan 25148f34af6fSJianyu Zhan out_uncharge_cgroup: 25158f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 251608cf9fafSMina Almasry out_uncharge_cgroup_reservation: 251708cf9fafSMina Almasry if (deferred_reserve) 251808cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 251908cf9fafSMina Almasry h_cg); 25208f34af6fSJianyu Zhan out_subpool_put: 2521d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 25228f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2523feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 25248f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2525b45b5bd6SDavid Gibson } 2526b45b5bd6SDavid Gibson 2527e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2528e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2529e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2530aa888a74SAndi Kleen { 2531aa888a74SAndi Kleen struct huge_bootmem_page *m; 2532b2261026SJoonsoo Kim int nr_nodes, node; 2533aa888a74SAndi Kleen 2534b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2535aa888a74SAndi Kleen void *addr; 2536aa888a74SAndi Kleen 2537eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 25388b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 253997ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2540aa888a74SAndi Kleen if (addr) { 2541aa888a74SAndi Kleen /* 2542aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2543aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2544aa888a74SAndi Kleen * puts them into the mem_map). 2545aa888a74SAndi Kleen */ 2546aa888a74SAndi Kleen m = addr; 2547aa888a74SAndi Kleen goto found; 2548aa888a74SAndi Kleen } 2549aa888a74SAndi Kleen } 2550aa888a74SAndi Kleen return 0; 2551aa888a74SAndi Kleen 2552aa888a74SAndi Kleen found: 2553df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2554aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2555330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2556aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2557aa888a74SAndi Kleen m->hstate = h; 2558aa888a74SAndi Kleen return 1; 2559aa888a74SAndi Kleen } 2560aa888a74SAndi Kleen 2561d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2562d00181b9SKirill A. Shutemov unsigned int order) 256318229df5SAndy Whitcroft { 256418229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 256518229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 256618229df5SAndy Whitcroft else 256718229df5SAndy Whitcroft prep_compound_page(page, order); 256818229df5SAndy Whitcroft } 256918229df5SAndy Whitcroft 2570aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2571aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2572aa888a74SAndi Kleen { 2573aa888a74SAndi Kleen struct huge_bootmem_page *m; 2574aa888a74SAndi Kleen 2575aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 257640d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2577aa888a74SAndi Kleen struct hstate *h = m->hstate; 2578ee8f248dSBecky Bruce 2579aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2580c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2581ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2582aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2583af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2584af0fb9dfSMichal Hocko 2585b0320c7bSRafael Aquini /* 2586b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2587b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2588b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2589b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2590b0320c7bSRafael Aquini */ 2591bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2592c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2593520495feSCannon Matthews cond_resched(); 2594aa888a74SAndi Kleen } 2595aa888a74SAndi Kleen } 2596aa888a74SAndi Kleen 25978faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 25981da177e4SLinus Torvalds { 25991da177e4SLinus Torvalds unsigned long i; 2600f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2601f60858f9SMike Kravetz 2602f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2603f60858f9SMike Kravetz /* 2604f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2605f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2606f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2607f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2608f60858f9SMike Kravetz */ 2609f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2610f60858f9SMike Kravetz GFP_KERNEL); 2611f60858f9SMike Kravetz } else { 2612f60858f9SMike Kravetz /* allocations done at boot time */ 2613f60858f9SMike Kravetz node_alloc_noretry = NULL; 2614f60858f9SMike Kravetz } 2615f60858f9SMike Kravetz 2616f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2617f60858f9SMike Kravetz if (node_alloc_noretry) 2618f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 26191da177e4SLinus Torvalds 2620e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2621bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2622dbda8feaSBarry Song if (hugetlb_cma_size) { 2623cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 26247ecc9565SChen Wandun goto free; 2625cf11e85fSRoman Gushchin } 2626aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2627aa888a74SAndi Kleen break; 26280c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2629f60858f9SMike Kravetz &node_states[N_MEMORY], 2630f60858f9SMike Kravetz node_alloc_noretry)) 26311da177e4SLinus Torvalds break; 263269ed779aSDavid Rientjes cond_resched(); 26331da177e4SLinus Torvalds } 2634d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2635d715cf80SLiam R. Howlett char buf[32]; 2636d715cf80SLiam R. Howlett 2637c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2638d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2639d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 26408faa8b07SAndi Kleen h->max_huge_pages = i; 2641e5ff2159SAndi Kleen } 26427ecc9565SChen Wandun free: 2643f60858f9SMike Kravetz kfree(node_alloc_noretry); 2644d715cf80SLiam R. Howlett } 2645e5ff2159SAndi Kleen 2646e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2647e5ff2159SAndi Kleen { 2648e5ff2159SAndi Kleen struct hstate *h; 2649e5ff2159SAndi Kleen 2650e5ff2159SAndi Kleen for_each_hstate(h) { 2651641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2652641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2653641844f5SNaoya Horiguchi 26548faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2655bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 26568faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2657e5ff2159SAndi Kleen } 2658641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2659e5ff2159SAndi Kleen } 2660e5ff2159SAndi Kleen 2661e5ff2159SAndi Kleen static void __init report_hugepages(void) 2662e5ff2159SAndi Kleen { 2663e5ff2159SAndi Kleen struct hstate *h; 2664e5ff2159SAndi Kleen 2665e5ff2159SAndi Kleen for_each_hstate(h) { 26664abd32dbSAndi Kleen char buf[32]; 2667c6247f72SMatthew Wilcox 2668c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2669ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2670c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2671e5ff2159SAndi Kleen } 2672e5ff2159SAndi Kleen } 2673e5ff2159SAndi Kleen 26741da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 26756ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 26766ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26771da177e4SLinus Torvalds { 26784415cc8dSChristoph Lameter int i; 26791121828aSMike Kravetz LIST_HEAD(page_list); 26804415cc8dSChristoph Lameter 26819487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2682bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2683aa888a74SAndi Kleen return; 2684aa888a74SAndi Kleen 26851121828aSMike Kravetz /* 26861121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 26871121828aSMike Kravetz */ 26886ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 268910c6ec49SMike Kravetz struct page *page, *next; 2690a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2691a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2692a5516438SAndi Kleen if (count >= h->nr_huge_pages) 26931121828aSMike Kravetz goto out; 26941da177e4SLinus Torvalds if (PageHighMem(page)) 26951da177e4SLinus Torvalds continue; 26966eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 26971121828aSMike Kravetz list_add(&page->lru, &page_list); 26981121828aSMike Kravetz } 26991121828aSMike Kravetz } 27001121828aSMike Kravetz 27011121828aSMike Kravetz out: 2702db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 270310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2704db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 27051da177e4SLinus Torvalds } 27061da177e4SLinus Torvalds #else 27076ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 27086ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 27091da177e4SLinus Torvalds { 27101da177e4SLinus Torvalds } 27111da177e4SLinus Torvalds #endif 27121da177e4SLinus Torvalds 271320a0307cSWu Fengguang /* 271420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 271520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 271620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 271720a0307cSWu Fengguang */ 27186ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 27196ae11b27SLee Schermerhorn int delta) 272020a0307cSWu Fengguang { 2721b2261026SJoonsoo Kim int nr_nodes, node; 272220a0307cSWu Fengguang 27239487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 272420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 272520a0307cSWu Fengguang 2726e8c5c824SLee Schermerhorn if (delta < 0) { 2727b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2728b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2729b2261026SJoonsoo Kim goto found; 2730b2261026SJoonsoo Kim } 2731b2261026SJoonsoo Kim } else { 2732b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2733b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2734b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2735b2261026SJoonsoo Kim goto found; 2736e8c5c824SLee Schermerhorn } 27379a76db09SLee Schermerhorn } 2738b2261026SJoonsoo Kim return 0; 273920a0307cSWu Fengguang 2740b2261026SJoonsoo Kim found: 274120a0307cSWu Fengguang h->surplus_huge_pages += delta; 2742b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2743b2261026SJoonsoo Kim return 1; 274420a0307cSWu Fengguang } 274520a0307cSWu Fengguang 2746a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2747fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 27486ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 27491da177e4SLinus Torvalds { 27507893d1d5SAdam Litke unsigned long min_count, ret; 275110c6ec49SMike Kravetz struct page *page; 275210c6ec49SMike Kravetz LIST_HEAD(page_list); 2753f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2754f60858f9SMike Kravetz 2755f60858f9SMike Kravetz /* 2756f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2757f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2758f60858f9SMike Kravetz * the requested huge pages. 2759f60858f9SMike Kravetz */ 2760f60858f9SMike Kravetz if (node_alloc_noretry) 2761f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2762f60858f9SMike Kravetz else 2763f60858f9SMike Kravetz return -ENOMEM; 27641da177e4SLinus Torvalds 276529383967SMike Kravetz /* 276629383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 276729383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 276829383967SMike Kravetz */ 276929383967SMike Kravetz mutex_lock(&h->resize_lock); 2770db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 27714eb0716eSAlexandre Ghiti 27724eb0716eSAlexandre Ghiti /* 2773fd875dcaSMike Kravetz * Check for a node specific request. 2774fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2775fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2776fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2777fd875dcaSMike Kravetz */ 2778fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2779fd875dcaSMike Kravetz unsigned long old_count = count; 2780fd875dcaSMike Kravetz 2781fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2782fd875dcaSMike Kravetz /* 2783fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2784fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2785fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2786fd875dcaSMike Kravetz * largest possible value to align with their intention. 2787fd875dcaSMike Kravetz */ 2788fd875dcaSMike Kravetz if (count < old_count) 2789fd875dcaSMike Kravetz count = ULONG_MAX; 2790fd875dcaSMike Kravetz } 2791fd875dcaSMike Kravetz 2792fd875dcaSMike Kravetz /* 27934eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 27944eb0716eSAlexandre Ghiti * page range allocation. 27954eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 27964eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 27974eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 27984eb0716eSAlexandre Ghiti */ 27994eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 28004eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 2801db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 280229383967SMike Kravetz mutex_unlock(&h->resize_lock); 2803f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 28044eb0716eSAlexandre Ghiti return -EINVAL; 28054eb0716eSAlexandre Ghiti } 28064eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 28074eb0716eSAlexandre Ghiti } 2808aa888a74SAndi Kleen 28097893d1d5SAdam Litke /* 28107893d1d5SAdam Litke * Increase the pool size 28117893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 28127893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2813d1c3fb1fSNishanth Aravamudan * 28140c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2815d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2816d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2817d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2818d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 28197893d1d5SAdam Litke */ 2820a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 28216ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 28227893d1d5SAdam Litke break; 28237893d1d5SAdam Litke } 28247893d1d5SAdam Litke 2825a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 28267893d1d5SAdam Litke /* 28277893d1d5SAdam Litke * If this allocation races such that we no longer need the 28287893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 28297893d1d5SAdam Litke * and reducing the surplus. 28307893d1d5SAdam Litke */ 2831db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2832649920c6SJia He 2833649920c6SJia He /* yield cpu to avoid soft lockup */ 2834649920c6SJia He cond_resched(); 2835649920c6SJia He 2836f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2837f60858f9SMike Kravetz node_alloc_noretry); 2838db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 28397893d1d5SAdam Litke if (!ret) 28407893d1d5SAdam Litke goto out; 28417893d1d5SAdam Litke 2842536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2843536240f2SMel Gorman if (signal_pending(current)) 2844536240f2SMel Gorman goto out; 28457893d1d5SAdam Litke } 28467893d1d5SAdam Litke 28477893d1d5SAdam Litke /* 28487893d1d5SAdam Litke * Decrease the pool size 28497893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 28507893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 28517893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 28527893d1d5SAdam Litke * to the desired size as pages become free. 2853d1c3fb1fSNishanth Aravamudan * 2854d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2855d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2856d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 28570c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2858d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2859d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2860d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 28617893d1d5SAdam Litke */ 2862a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 28636b0c880dSAdam Litke min_count = max(count, min_count); 28646ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 286510c6ec49SMike Kravetz 286610c6ec49SMike Kravetz /* 286710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 286810c6ec49SMike Kravetz */ 2869a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 287010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 287110c6ec49SMike Kravetz if (!page) 28721da177e4SLinus Torvalds break; 287310c6ec49SMike Kravetz 287410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 28751da177e4SLinus Torvalds } 287610c6ec49SMike Kravetz /* free the pages after dropping lock */ 2877db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 287810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2879db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 288010c6ec49SMike Kravetz 2881a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 28826ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 28837893d1d5SAdam Litke break; 28847893d1d5SAdam Litke } 28857893d1d5SAdam Litke out: 28864eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 2887db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 288829383967SMike Kravetz mutex_unlock(&h->resize_lock); 28894eb0716eSAlexandre Ghiti 2890f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2891f60858f9SMike Kravetz 28924eb0716eSAlexandre Ghiti return 0; 28931da177e4SLinus Torvalds } 28941da177e4SLinus Torvalds 2895a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2896a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2897a3437870SNishanth Aravamudan 2898a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2899a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2900a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2901a3437870SNishanth Aravamudan 2902a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2903a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2904a3437870SNishanth Aravamudan 29059a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 29069a305230SLee Schermerhorn 29079a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2908a3437870SNishanth Aravamudan { 2909a3437870SNishanth Aravamudan int i; 29109a305230SLee Schermerhorn 2911a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 29129a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 29139a305230SLee Schermerhorn if (nidp) 29149a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2915a3437870SNishanth Aravamudan return &hstates[i]; 29169a305230SLee Schermerhorn } 29179a305230SLee Schermerhorn 29189a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2919a3437870SNishanth Aravamudan } 2920a3437870SNishanth Aravamudan 292106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2922a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2923a3437870SNishanth Aravamudan { 29249a305230SLee Schermerhorn struct hstate *h; 29259a305230SLee Schermerhorn unsigned long nr_huge_pages; 29269a305230SLee Schermerhorn int nid; 29279a305230SLee Schermerhorn 29289a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29299a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29309a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 29319a305230SLee Schermerhorn else 29329a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 29339a305230SLee Schermerhorn 2934ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2935a3437870SNishanth Aravamudan } 2936adbe8726SEric B Munson 2937238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2938238d3c13SDavid Rientjes struct hstate *h, int nid, 2939238d3c13SDavid Rientjes unsigned long count, size_t len) 2940a3437870SNishanth Aravamudan { 2941a3437870SNishanth Aravamudan int err; 29422d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2943a3437870SNishanth Aravamudan 29442d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 29452d0adf7eSOscar Salvador return -EINVAL; 2946adbe8726SEric B Munson 29479a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 29489a305230SLee Schermerhorn /* 29499a305230SLee Schermerhorn * global hstate attribute 29509a305230SLee Schermerhorn */ 29519a305230SLee Schermerhorn if (!(obey_mempolicy && 29522d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 29532d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 29542d0adf7eSOscar Salvador else 29552d0adf7eSOscar Salvador n_mask = &nodes_allowed; 29562d0adf7eSOscar Salvador } else { 29579a305230SLee Schermerhorn /* 2958fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2959fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 29609a305230SLee Schermerhorn */ 29612d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 29622d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2963fd875dcaSMike Kravetz } 29649a305230SLee Schermerhorn 29652d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 296606808b08SLee Schermerhorn 29674eb0716eSAlexandre Ghiti return err ? err : len; 296806808b08SLee Schermerhorn } 296906808b08SLee Schermerhorn 2970238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2971238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2972238d3c13SDavid Rientjes size_t len) 2973238d3c13SDavid Rientjes { 2974238d3c13SDavid Rientjes struct hstate *h; 2975238d3c13SDavid Rientjes unsigned long count; 2976238d3c13SDavid Rientjes int nid; 2977238d3c13SDavid Rientjes int err; 2978238d3c13SDavid Rientjes 2979238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2980238d3c13SDavid Rientjes if (err) 2981238d3c13SDavid Rientjes return err; 2982238d3c13SDavid Rientjes 2983238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2984238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2985238d3c13SDavid Rientjes } 2986238d3c13SDavid Rientjes 298706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 298806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 298906808b08SLee Schermerhorn { 299006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 299106808b08SLee Schermerhorn } 299206808b08SLee Schermerhorn 299306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 299406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 299506808b08SLee Schermerhorn { 2996238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2997a3437870SNishanth Aravamudan } 2998a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2999a3437870SNishanth Aravamudan 300006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 300106808b08SLee Schermerhorn 300206808b08SLee Schermerhorn /* 300306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 300406808b08SLee Schermerhorn * huge page alloc/free. 300506808b08SLee Schermerhorn */ 300606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3007ae7a927dSJoe Perches struct kobj_attribute *attr, 3008ae7a927dSJoe Perches char *buf) 300906808b08SLee Schermerhorn { 301006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 301106808b08SLee Schermerhorn } 301206808b08SLee Schermerhorn 301306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 301406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 301506808b08SLee Schermerhorn { 3016238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 301706808b08SLee Schermerhorn } 301806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 301906808b08SLee Schermerhorn #endif 302006808b08SLee Schermerhorn 302106808b08SLee Schermerhorn 3022a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3023a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3024a3437870SNishanth Aravamudan { 30259a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3026ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3027a3437870SNishanth Aravamudan } 3028adbe8726SEric B Munson 3029a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3030a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3031a3437870SNishanth Aravamudan { 3032a3437870SNishanth Aravamudan int err; 3033a3437870SNishanth Aravamudan unsigned long input; 30349a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3035a3437870SNishanth Aravamudan 3036bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3037adbe8726SEric B Munson return -EINVAL; 3038adbe8726SEric B Munson 30393dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3040a3437870SNishanth Aravamudan if (err) 304173ae31e5SEric B Munson return err; 3042a3437870SNishanth Aravamudan 3043db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3044a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3045db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3046a3437870SNishanth Aravamudan 3047a3437870SNishanth Aravamudan return count; 3048a3437870SNishanth Aravamudan } 3049a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3050a3437870SNishanth Aravamudan 3051a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3052a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3053a3437870SNishanth Aravamudan { 30549a305230SLee Schermerhorn struct hstate *h; 30559a305230SLee Schermerhorn unsigned long free_huge_pages; 30569a305230SLee Schermerhorn int nid; 30579a305230SLee Schermerhorn 30589a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 30599a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 30609a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 30619a305230SLee Schermerhorn else 30629a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 30639a305230SLee Schermerhorn 3064ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3065a3437870SNishanth Aravamudan } 3066a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3067a3437870SNishanth Aravamudan 3068a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3069a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3070a3437870SNishanth Aravamudan { 30719a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3072ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3073a3437870SNishanth Aravamudan } 3074a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3075a3437870SNishanth Aravamudan 3076a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3077a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3078a3437870SNishanth Aravamudan { 30799a305230SLee Schermerhorn struct hstate *h; 30809a305230SLee Schermerhorn unsigned long surplus_huge_pages; 30819a305230SLee Schermerhorn int nid; 30829a305230SLee Schermerhorn 30839a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 30849a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 30859a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 30869a305230SLee Schermerhorn else 30879a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 30889a305230SLee Schermerhorn 3089ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3090a3437870SNishanth Aravamudan } 3091a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3092a3437870SNishanth Aravamudan 3093a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3094a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3095a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3096a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3097a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3098a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 309906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 310006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 310106808b08SLee Schermerhorn #endif 3102a3437870SNishanth Aravamudan NULL, 3103a3437870SNishanth Aravamudan }; 3104a3437870SNishanth Aravamudan 310567e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3106a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3107a3437870SNishanth Aravamudan }; 3108a3437870SNishanth Aravamudan 3109094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 31109a305230SLee Schermerhorn struct kobject **hstate_kobjs, 311167e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3112a3437870SNishanth Aravamudan { 3113a3437870SNishanth Aravamudan int retval; 3114972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3115a3437870SNishanth Aravamudan 31169a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 31179a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3118a3437870SNishanth Aravamudan return -ENOMEM; 3119a3437870SNishanth Aravamudan 31209a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3121cc2205a6SMiaohe Lin if (retval) { 31229a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3123cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3124cc2205a6SMiaohe Lin } 3125a3437870SNishanth Aravamudan 3126a3437870SNishanth Aravamudan return retval; 3127a3437870SNishanth Aravamudan } 3128a3437870SNishanth Aravamudan 3129a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3130a3437870SNishanth Aravamudan { 3131a3437870SNishanth Aravamudan struct hstate *h; 3132a3437870SNishanth Aravamudan int err; 3133a3437870SNishanth Aravamudan 3134a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3135a3437870SNishanth Aravamudan if (!hugepages_kobj) 3136a3437870SNishanth Aravamudan return; 3137a3437870SNishanth Aravamudan 3138a3437870SNishanth Aravamudan for_each_hstate(h) { 31399a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 31409a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3141a3437870SNishanth Aravamudan if (err) 3142282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3143a3437870SNishanth Aravamudan } 3144a3437870SNishanth Aravamudan } 3145a3437870SNishanth Aravamudan 31469a305230SLee Schermerhorn #ifdef CONFIG_NUMA 31479a305230SLee Schermerhorn 31489a305230SLee Schermerhorn /* 31499a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 315010fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 315110fbcf4cSKay Sievers * index of a node device or _hstate == node id. 315210fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 31539a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 31549a305230SLee Schermerhorn */ 31559a305230SLee Schermerhorn struct node_hstate { 31569a305230SLee Schermerhorn struct kobject *hugepages_kobj; 31579a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 31589a305230SLee Schermerhorn }; 3159b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 31609a305230SLee Schermerhorn 31619a305230SLee Schermerhorn /* 316210fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 31639a305230SLee Schermerhorn */ 31649a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 31659a305230SLee Schermerhorn &nr_hugepages_attr.attr, 31669a305230SLee Schermerhorn &free_hugepages_attr.attr, 31679a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 31689a305230SLee Schermerhorn NULL, 31699a305230SLee Schermerhorn }; 31709a305230SLee Schermerhorn 317167e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 31729a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 31739a305230SLee Schermerhorn }; 31749a305230SLee Schermerhorn 31759a305230SLee Schermerhorn /* 317610fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 31779a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 31789a305230SLee Schermerhorn */ 31799a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31809a305230SLee Schermerhorn { 31819a305230SLee Schermerhorn int nid; 31829a305230SLee Schermerhorn 31839a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 31849a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 31859a305230SLee Schermerhorn int i; 31869a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 31879a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 31889a305230SLee Schermerhorn if (nidp) 31899a305230SLee Schermerhorn *nidp = nid; 31909a305230SLee Schermerhorn return &hstates[i]; 31919a305230SLee Schermerhorn } 31929a305230SLee Schermerhorn } 31939a305230SLee Schermerhorn 31949a305230SLee Schermerhorn BUG(); 31959a305230SLee Schermerhorn return NULL; 31969a305230SLee Schermerhorn } 31979a305230SLee Schermerhorn 31989a305230SLee Schermerhorn /* 319910fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 32009a305230SLee Schermerhorn * No-op if no hstate attributes attached. 32019a305230SLee Schermerhorn */ 32023cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 32039a305230SLee Schermerhorn { 32049a305230SLee Schermerhorn struct hstate *h; 320510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 32069a305230SLee Schermerhorn 32079a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 32089b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 32099a305230SLee Schermerhorn 3210972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3211972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3212972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3213972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3214972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3215972dc4deSAneesh Kumar K.V } 32169a305230SLee Schermerhorn } 32179a305230SLee Schermerhorn 32189a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 32199a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 32209a305230SLee Schermerhorn } 32219a305230SLee Schermerhorn 32229a305230SLee Schermerhorn 32239a305230SLee Schermerhorn /* 322410fbcf4cSKay Sievers * Register hstate attributes for a single node device. 32259a305230SLee Schermerhorn * No-op if attributes already registered. 32269a305230SLee Schermerhorn */ 32273cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 32289a305230SLee Schermerhorn { 32299a305230SLee Schermerhorn struct hstate *h; 323010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 32319a305230SLee Schermerhorn int err; 32329a305230SLee Schermerhorn 32339a305230SLee Schermerhorn if (nhs->hugepages_kobj) 32349a305230SLee Schermerhorn return; /* already allocated */ 32359a305230SLee Schermerhorn 32369a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 323710fbcf4cSKay Sievers &node->dev.kobj); 32389a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 32399a305230SLee Schermerhorn return; 32409a305230SLee Schermerhorn 32419a305230SLee Schermerhorn for_each_hstate(h) { 32429a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 32439a305230SLee Schermerhorn nhs->hstate_kobjs, 32449a305230SLee Schermerhorn &per_node_hstate_attr_group); 32459a305230SLee Schermerhorn if (err) { 3246282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 324710fbcf4cSKay Sievers h->name, node->dev.id); 32489a305230SLee Schermerhorn hugetlb_unregister_node(node); 32499a305230SLee Schermerhorn break; 32509a305230SLee Schermerhorn } 32519a305230SLee Schermerhorn } 32529a305230SLee Schermerhorn } 32539a305230SLee Schermerhorn 32549a305230SLee Schermerhorn /* 32559b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 325610fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 325710fbcf4cSKay Sievers * registered their associated device by this time. 32589a305230SLee Schermerhorn */ 32597d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 32609a305230SLee Schermerhorn { 32619a305230SLee Schermerhorn int nid; 32629a305230SLee Schermerhorn 32638cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 32648732794bSWen Congyang struct node *node = node_devices[nid]; 326510fbcf4cSKay Sievers if (node->dev.id == nid) 32669a305230SLee Schermerhorn hugetlb_register_node(node); 32679a305230SLee Schermerhorn } 32689a305230SLee Schermerhorn 32699a305230SLee Schermerhorn /* 327010fbcf4cSKay Sievers * Let the node device driver know we're here so it can 32719a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 32729a305230SLee Schermerhorn */ 32739a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 32749a305230SLee Schermerhorn hugetlb_unregister_node); 32759a305230SLee Schermerhorn } 32769a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 32779a305230SLee Schermerhorn 32789a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 32799a305230SLee Schermerhorn { 32809a305230SLee Schermerhorn BUG(); 32819a305230SLee Schermerhorn if (nidp) 32829a305230SLee Schermerhorn *nidp = -1; 32839a305230SLee Schermerhorn return NULL; 32849a305230SLee Schermerhorn } 32859a305230SLee Schermerhorn 32869a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 32879a305230SLee Schermerhorn 32889a305230SLee Schermerhorn #endif 32899a305230SLee Schermerhorn 3290a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3291a3437870SNishanth Aravamudan { 32928382d914SDavidlohr Bueso int i; 32938382d914SDavidlohr Bueso 3294d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3295d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3296d6995da3SMike Kravetz 3297c2833a5bSMike Kravetz if (!hugepages_supported()) { 3298c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3299c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 33000ef89d25SBenjamin Herrenschmidt return 0; 3301d715cf80SLiam R. Howlett } 3302d715cf80SLiam R. Howlett 3303282f4214SMike Kravetz /* 3304282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3305282f4214SMike Kravetz * architectures depend on setup being done here. 3306282f4214SMike Kravetz */ 3307a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3308282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3309282f4214SMike Kravetz /* 3310282f4214SMike Kravetz * If we did not parse a default huge page size, set 3311282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3312282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3313282f4214SMike Kravetz * specified, set that here as well. 3314282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3315282f4214SMike Kravetz * setting. A warning will be printed in this case. 3316282f4214SMike Kravetz */ 3317282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3318f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3319282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3320282f4214SMike Kravetz char buf[32]; 3321282f4214SMike Kravetz 3322282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3323282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3324282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3325282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3326282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3327282f4214SMike Kravetz default_hstate_max_huge_pages); 3328282f4214SMike Kravetz } 3329282f4214SMike Kravetz default_hstate.max_huge_pages = 3330282f4214SMike Kravetz default_hstate_max_huge_pages; 3331282f4214SMike Kravetz } 3332f8b74815SVaishali Thakkar } 3333a3437870SNishanth Aravamudan 3334cf11e85fSRoman Gushchin hugetlb_cma_check(); 3335a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3336aa888a74SAndi Kleen gather_bootmem_prealloc(); 3337a3437870SNishanth Aravamudan report_hugepages(); 3338a3437870SNishanth Aravamudan 3339a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 33409a305230SLee Schermerhorn hugetlb_register_all_nodes(); 33417179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 33429a305230SLee Schermerhorn 33438382d914SDavidlohr Bueso #ifdef CONFIG_SMP 33448382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 33458382d914SDavidlohr Bueso #else 33468382d914SDavidlohr Bueso num_fault_mutexes = 1; 33478382d914SDavidlohr Bueso #endif 3348c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 33496da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 33506da2ec56SKees Cook GFP_KERNEL); 3351c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 33528382d914SDavidlohr Bueso 33538382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3354c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3355a3437870SNishanth Aravamudan return 0; 3356a3437870SNishanth Aravamudan } 33573e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3358a3437870SNishanth Aravamudan 3359ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3360ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 33619fee021dSVaishali Thakkar { 3362ae94da89SMike Kravetz return size == HPAGE_SIZE; 33639fee021dSVaishali Thakkar } 33649fee021dSVaishali Thakkar 3365d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3366a3437870SNishanth Aravamudan { 3367a3437870SNishanth Aravamudan struct hstate *h; 33688faa8b07SAndi Kleen unsigned long i; 33698faa8b07SAndi Kleen 3370a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3371a3437870SNishanth Aravamudan return; 3372a3437870SNishanth Aravamudan } 337347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3374a3437870SNishanth Aravamudan BUG_ON(order == 0); 337547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 337629383967SMike Kravetz mutex_init(&h->resize_lock); 3377a3437870SNishanth Aravamudan h->order = order; 3378aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 33798faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 33808faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 33810edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 338254f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 338354f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3384a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3385a3437870SNishanth Aravamudan huge_page_size(h)/1024); 33868faa8b07SAndi Kleen 3387a3437870SNishanth Aravamudan parsed_hstate = h; 3388a3437870SNishanth Aravamudan } 3389a3437870SNishanth Aravamudan 3390282f4214SMike Kravetz /* 3391282f4214SMike Kravetz * hugepages command line processing 3392282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3393282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3394282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3395282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3396282f4214SMike Kravetz */ 3397282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3398a3437870SNishanth Aravamudan { 3399a3437870SNishanth Aravamudan unsigned long *mhp; 34008faa8b07SAndi Kleen static unsigned long *last_mhp; 3401a3437870SNishanth Aravamudan 34029fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3403282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 34049fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3405282f4214SMike Kravetz return 0; 34069fee021dSVaishali Thakkar } 3407282f4214SMike Kravetz 3408a3437870SNishanth Aravamudan /* 3409282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3410282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3411282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3412282f4214SMike Kravetz * default_hugepagesz. 3413a3437870SNishanth Aravamudan */ 34149fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3415a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3416a3437870SNishanth Aravamudan else 3417a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3418a3437870SNishanth Aravamudan 34198faa8b07SAndi Kleen if (mhp == last_mhp) { 3420282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3421282f4214SMike Kravetz return 0; 34228faa8b07SAndi Kleen } 34238faa8b07SAndi Kleen 3424a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3425a3437870SNishanth Aravamudan *mhp = 0; 3426a3437870SNishanth Aravamudan 34278faa8b07SAndi Kleen /* 34288faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 342904adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 34308faa8b07SAndi Kleen * use the bootmem allocator. 34318faa8b07SAndi Kleen */ 343204adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 34338faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 34348faa8b07SAndi Kleen 34358faa8b07SAndi Kleen last_mhp = mhp; 34368faa8b07SAndi Kleen 3437a3437870SNishanth Aravamudan return 1; 3438a3437870SNishanth Aravamudan } 3439282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3440e11bfbfcSNick Piggin 3441282f4214SMike Kravetz /* 3442282f4214SMike Kravetz * hugepagesz command line processing 3443282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3444282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3445282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3446282f4214SMike Kravetz * hugepagesz argument was valid. 3447282f4214SMike Kravetz */ 3448359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3449e11bfbfcSNick Piggin { 3450359f2544SMike Kravetz unsigned long size; 3451282f4214SMike Kravetz struct hstate *h; 3452282f4214SMike Kravetz 3453282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3454359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3455359f2544SMike Kravetz 3456359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3457282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3458359f2544SMike Kravetz return 0; 3459359f2544SMike Kravetz } 3460359f2544SMike Kravetz 3461282f4214SMike Kravetz h = size_to_hstate(size); 3462282f4214SMike Kravetz if (h) { 3463282f4214SMike Kravetz /* 3464282f4214SMike Kravetz * hstate for this size already exists. This is normally 3465282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3466282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3467282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3468282f4214SMike Kravetz * previously specified. 3469282f4214SMike Kravetz */ 3470282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3471282f4214SMike Kravetz default_hstate.max_huge_pages) { 3472282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 347338237830SMike Kravetz return 0; 347438237830SMike Kravetz } 347538237830SMike Kravetz 3476282f4214SMike Kravetz /* 3477282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3478282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3479282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3480282f4214SMike Kravetz */ 3481282f4214SMike Kravetz parsed_hstate = h; 3482282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3483e11bfbfcSNick Piggin return 1; 3484e11bfbfcSNick Piggin } 3485282f4214SMike Kravetz 3486359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3487282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3488359f2544SMike Kravetz return 1; 3489359f2544SMike Kravetz } 3490359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3491359f2544SMike Kravetz 3492282f4214SMike Kravetz /* 3493282f4214SMike Kravetz * default_hugepagesz command line input 3494282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3495282f4214SMike Kravetz */ 3496ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3497e11bfbfcSNick Piggin { 3498ae94da89SMike Kravetz unsigned long size; 3499ae94da89SMike Kravetz 3500282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3501282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3502282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3503282f4214SMike Kravetz return 0; 3504282f4214SMike Kravetz } 3505282f4214SMike Kravetz 3506282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3507282f4214SMike Kravetz 3508282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3509282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3510282f4214SMike Kravetz return 0; 3511282f4214SMike Kravetz } 3512282f4214SMike Kravetz 3513282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3514282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3515282f4214SMike Kravetz parsed_default_hugepagesz = true; 3516282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3517282f4214SMike Kravetz 3518282f4214SMike Kravetz /* 3519282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3520282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3521282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3522282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3523282f4214SMike Kravetz * allocated here from bootmem allocator. 3524282f4214SMike Kravetz */ 3525282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3526282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3527282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3528282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3529282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3530282f4214SMike Kravetz } 3531282f4214SMike Kravetz 3532e11bfbfcSNick Piggin return 1; 3533e11bfbfcSNick Piggin } 3534ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3535a3437870SNishanth Aravamudan 35368ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 35378a213460SNishanth Aravamudan { 35388a213460SNishanth Aravamudan int node; 35398a213460SNishanth Aravamudan unsigned int nr = 0; 35408ca39e68SMuchun Song nodemask_t *mpol_allowed; 35418ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 35428ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 35438a213460SNishanth Aravamudan 35448ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 35458ca39e68SMuchun Song 35468ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3547c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 35488a213460SNishanth Aravamudan nr += array[node]; 35498ca39e68SMuchun Song } 35508a213460SNishanth Aravamudan 35518a213460SNishanth Aravamudan return nr; 35528a213460SNishanth Aravamudan } 35538a213460SNishanth Aravamudan 35548a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 355517743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 355617743798SMuchun Song void *buffer, size_t *length, 355717743798SMuchun Song loff_t *ppos, unsigned long *out) 355817743798SMuchun Song { 355917743798SMuchun Song struct ctl_table dup_table; 356017743798SMuchun Song 356117743798SMuchun Song /* 356217743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 356317743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 356417743798SMuchun Song */ 356517743798SMuchun Song dup_table = *table; 356617743798SMuchun Song dup_table.data = out; 356717743798SMuchun Song 356817743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 356917743798SMuchun Song } 357017743798SMuchun Song 357106808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 357206808b08SLee Schermerhorn struct ctl_table *table, int write, 357332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 35741da177e4SLinus Torvalds { 3575e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3576238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 357708d4a246SMichal Hocko int ret; 3578e5ff2159SAndi Kleen 3579457c1b27SNishanth Aravamudan if (!hugepages_supported()) 358086613628SJan Stancek return -EOPNOTSUPP; 3581457c1b27SNishanth Aravamudan 358217743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 358317743798SMuchun Song &tmp); 358408d4a246SMichal Hocko if (ret) 358508d4a246SMichal Hocko goto out; 3586e5ff2159SAndi Kleen 3587238d3c13SDavid Rientjes if (write) 3588238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3589238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 359008d4a246SMichal Hocko out: 359108d4a246SMichal Hocko return ret; 35921da177e4SLinus Torvalds } 3593396faf03SMel Gorman 359406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 359532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 359606808b08SLee Schermerhorn { 359706808b08SLee Schermerhorn 359806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 359906808b08SLee Schermerhorn buffer, length, ppos); 360006808b08SLee Schermerhorn } 360106808b08SLee Schermerhorn 360206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 360306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 360432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 360506808b08SLee Schermerhorn { 360606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 360706808b08SLee Schermerhorn buffer, length, ppos); 360806808b08SLee Schermerhorn } 360906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 361006808b08SLee Schermerhorn 3611a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 361232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3613a3d0c6aaSNishanth Aravamudan { 3614a5516438SAndi Kleen struct hstate *h = &default_hstate; 3615e5ff2159SAndi Kleen unsigned long tmp; 361608d4a246SMichal Hocko int ret; 3617e5ff2159SAndi Kleen 3618457c1b27SNishanth Aravamudan if (!hugepages_supported()) 361986613628SJan Stancek return -EOPNOTSUPP; 3620457c1b27SNishanth Aravamudan 3621e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3622e5ff2159SAndi Kleen 3623bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3624adbe8726SEric B Munson return -EINVAL; 3625adbe8726SEric B Munson 362617743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 362717743798SMuchun Song &tmp); 362808d4a246SMichal Hocko if (ret) 362908d4a246SMichal Hocko goto out; 3630e5ff2159SAndi Kleen 3631e5ff2159SAndi Kleen if (write) { 3632db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3633e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3634db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3635e5ff2159SAndi Kleen } 363608d4a246SMichal Hocko out: 363708d4a246SMichal Hocko return ret; 3638a3d0c6aaSNishanth Aravamudan } 3639a3d0c6aaSNishanth Aravamudan 36401da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 36411da177e4SLinus Torvalds 3642e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 36431da177e4SLinus Torvalds { 3644fcb2b0c5SRoman Gushchin struct hstate *h; 3645fcb2b0c5SRoman Gushchin unsigned long total = 0; 3646fcb2b0c5SRoman Gushchin 3647457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3648457c1b27SNishanth Aravamudan return; 3649fcb2b0c5SRoman Gushchin 3650fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3651fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3652fcb2b0c5SRoman Gushchin 3653aca78307SMiaohe Lin total += huge_page_size(h) * count; 3654fcb2b0c5SRoman Gushchin 3655fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3656e1759c21SAlexey Dobriyan seq_printf(m, 36571da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 36581da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3659b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 36607893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 36614f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3662fcb2b0c5SRoman Gushchin count, 3663a5516438SAndi Kleen h->free_huge_pages, 3664a5516438SAndi Kleen h->resv_huge_pages, 3665a5516438SAndi Kleen h->surplus_huge_pages, 3666aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3667fcb2b0c5SRoman Gushchin } 3668fcb2b0c5SRoman Gushchin 3669aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 36701da177e4SLinus Torvalds } 36711da177e4SLinus Torvalds 36727981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 36731da177e4SLinus Torvalds { 3674a5516438SAndi Kleen struct hstate *h = &default_hstate; 36757981593bSJoe Perches 3676457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3677457c1b27SNishanth Aravamudan return 0; 36787981593bSJoe Perches 36797981593bSJoe Perches return sysfs_emit_at(buf, len, 36801da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3681a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3682a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3683a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3684a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3685a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 36861da177e4SLinus Torvalds } 36871da177e4SLinus Torvalds 3688949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3689949f7ec5SDavid Rientjes { 3690949f7ec5SDavid Rientjes struct hstate *h; 3691949f7ec5SDavid Rientjes int nid; 3692949f7ec5SDavid Rientjes 3693457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3694457c1b27SNishanth Aravamudan return; 3695457c1b27SNishanth Aravamudan 3696949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3697949f7ec5SDavid Rientjes for_each_hstate(h) 3698949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3699949f7ec5SDavid Rientjes nid, 3700949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3701949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3702949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3703aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3704949f7ec5SDavid Rientjes } 3705949f7ec5SDavid Rientjes 37065d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 37075d317b2bSNaoya Horiguchi { 37085d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 37095d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 37105d317b2bSNaoya Horiguchi } 37115d317b2bSNaoya Horiguchi 37121da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 37131da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 37141da177e4SLinus Torvalds { 3715d0028588SWanpeng Li struct hstate *h; 3716d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3717d0028588SWanpeng Li 3718d0028588SWanpeng Li for_each_hstate(h) 3719d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3720d0028588SWanpeng Li return nr_total_pages; 37211da177e4SLinus Torvalds } 37221da177e4SLinus Torvalds 3723a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3724fc1b8a73SMel Gorman { 3725fc1b8a73SMel Gorman int ret = -ENOMEM; 3726fc1b8a73SMel Gorman 37270aa7f354SMiaohe Lin if (!delta) 37280aa7f354SMiaohe Lin return 0; 37290aa7f354SMiaohe Lin 3730db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3731fc1b8a73SMel Gorman /* 3732fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3733fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3734fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3735fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3736fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3737fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3738fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3739fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3740fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3741fc1b8a73SMel Gorman * 3742fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3743fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3744fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3745fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3746fc1b8a73SMel Gorman * semantics that cpuset has. 37478ca39e68SMuchun Song * 37488ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 37498ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 37508ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 37518ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 37528ca39e68SMuchun Song * above. 3753fc1b8a73SMel Gorman */ 3754fc1b8a73SMel Gorman if (delta > 0) { 3755a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3756fc1b8a73SMel Gorman goto out; 3757fc1b8a73SMel Gorman 37588ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3759a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3760fc1b8a73SMel Gorman goto out; 3761fc1b8a73SMel Gorman } 3762fc1b8a73SMel Gorman } 3763fc1b8a73SMel Gorman 3764fc1b8a73SMel Gorman ret = 0; 3765fc1b8a73SMel Gorman if (delta < 0) 3766a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3767fc1b8a73SMel Gorman 3768fc1b8a73SMel Gorman out: 3769db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3770fc1b8a73SMel Gorman return ret; 3771fc1b8a73SMel Gorman } 3772fc1b8a73SMel Gorman 377384afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 377484afd99bSAndy Whitcroft { 3775f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 377684afd99bSAndy Whitcroft 377784afd99bSAndy Whitcroft /* 377884afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 377984afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 378084afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 378125985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 378284afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 378384afd99bSAndy Whitcroft * new reference here without additional locking. 378484afd99bSAndy Whitcroft */ 37854e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3786f522c3acSJoonsoo Kim kref_get(&resv->refs); 378784afd99bSAndy Whitcroft } 378884afd99bSAndy Whitcroft 3789a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3790a1e78772SMel Gorman { 3791a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3792f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 379390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 37944e35f483SJoonsoo Kim unsigned long reserve, start, end; 37951c5ecae3SMike Kravetz long gbl_reserve; 379684afd99bSAndy Whitcroft 37974e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 37984e35f483SJoonsoo Kim return; 37994e35f483SJoonsoo Kim 3800a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3801a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 380284afd99bSAndy Whitcroft 38034e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3804e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 38057251ff78SAdam Litke if (reserve) { 38061c5ecae3SMike Kravetz /* 38071c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 38081c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 38091c5ecae3SMike Kravetz */ 38101c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 38111c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 38127251ff78SAdam Litke } 3813e9fe92aeSMina Almasry 3814e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3815a1e78772SMel Gorman } 3816a1e78772SMel Gorman 381731383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 381831383c68SDan Williams { 381931383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 382031383c68SDan Williams return -EINVAL; 382131383c68SDan Williams return 0; 382231383c68SDan Williams } 382331383c68SDan Williams 382405ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 382505ea8860SDan Williams { 3826aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 382705ea8860SDan Williams } 382805ea8860SDan Williams 38291da177e4SLinus Torvalds /* 38301da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 38311da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 38326c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 38331da177e4SLinus Torvalds * this far. 38341da177e4SLinus Torvalds */ 3835b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 38361da177e4SLinus Torvalds { 38371da177e4SLinus Torvalds BUG(); 3838d0217ac0SNick Piggin return 0; 38391da177e4SLinus Torvalds } 38401da177e4SLinus Torvalds 3841eec3636aSJane Chu /* 3842eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3843eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3844eec3636aSJane Chu * This is because under System V memory model, mappings created via 3845eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3846eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3847eec3636aSJane Chu */ 3848f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3849d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 385084afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3851a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3852dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 385305ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 38541da177e4SLinus Torvalds }; 38551da177e4SLinus Torvalds 38561e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 38571e8f889bSDavid Gibson int writable) 385863551ae0SDavid Gibson { 385963551ae0SDavid Gibson pte_t entry; 386063551ae0SDavid Gibson 38611e8f889bSDavid Gibson if (writable) { 3862106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3863106c992aSGerald Schaefer vma->vm_page_prot))); 386463551ae0SDavid Gibson } else { 3865106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3866106c992aSGerald Schaefer vma->vm_page_prot)); 386763551ae0SDavid Gibson } 386863551ae0SDavid Gibson entry = pte_mkyoung(entry); 386963551ae0SDavid Gibson entry = pte_mkhuge(entry); 3870d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 387163551ae0SDavid Gibson 387263551ae0SDavid Gibson return entry; 387363551ae0SDavid Gibson } 387463551ae0SDavid Gibson 38751e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 38761e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 38771e8f889bSDavid Gibson { 38781e8f889bSDavid Gibson pte_t entry; 38791e8f889bSDavid Gibson 3880106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 388132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 38824b3073e1SRussell King update_mmu_cache(vma, address, ptep); 38831e8f889bSDavid Gibson } 38841e8f889bSDavid Gibson 3885d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 38864a705fefSNaoya Horiguchi { 38874a705fefSNaoya Horiguchi swp_entry_t swp; 38884a705fefSNaoya Horiguchi 38894a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3890d5ed7444SAneesh Kumar K.V return false; 38914a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3892d79d176aSBaoquan He if (is_migration_entry(swp)) 3893d5ed7444SAneesh Kumar K.V return true; 38944a705fefSNaoya Horiguchi else 3895d5ed7444SAneesh Kumar K.V return false; 38964a705fefSNaoya Horiguchi } 38974a705fefSNaoya Horiguchi 38983e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 38994a705fefSNaoya Horiguchi { 39004a705fefSNaoya Horiguchi swp_entry_t swp; 39014a705fefSNaoya Horiguchi 39024a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 39033e5c3600SBaoquan He return false; 39044a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3905d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 39063e5c3600SBaoquan He return true; 39074a705fefSNaoya Horiguchi else 39083e5c3600SBaoquan He return false; 39094a705fefSNaoya Horiguchi } 39101e8f889bSDavid Gibson 39114eae4efaSPeter Xu static void 39124eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 39134eae4efaSPeter Xu struct page *new_page) 39144eae4efaSPeter Xu { 39154eae4efaSPeter Xu __SetPageUptodate(new_page); 39164eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 39174eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 39184eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 39194eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 39204eae4efaSPeter Xu SetHPageMigratable(new_page); 39214eae4efaSPeter Xu } 39224eae4efaSPeter Xu 392363551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 392463551ae0SDavid Gibson struct vm_area_struct *vma) 392563551ae0SDavid Gibson { 39265e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 392763551ae0SDavid Gibson struct page *ptepage; 39281c59827dSHugh Dickins unsigned long addr; 3929ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 3930a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3931a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 39324eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 3933c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3934ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3935e8569dd2SAndreas Sandberg int ret = 0; 39361e8f889bSDavid Gibson 3937ac46d4f3SJérôme Glisse if (cow) { 39387269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 39396f4f13e8SJérôme Glisse vma->vm_start, 3940ac46d4f3SJérôme Glisse vma->vm_end); 3941ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3942c0d0381aSMike Kravetz } else { 3943c0d0381aSMike Kravetz /* 3944c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3945c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3946c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3947c0d0381aSMike Kravetz * huge_pmd_unshare. 3948c0d0381aSMike Kravetz */ 3949c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3950ac46d4f3SJérôme Glisse } 3951e8569dd2SAndreas Sandberg 3952a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3953cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 39547868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3955c74df32cSHugh Dickins if (!src_pte) 3956c74df32cSHugh Dickins continue; 3957aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 3958e8569dd2SAndreas Sandberg if (!dst_pte) { 3959e8569dd2SAndreas Sandberg ret = -ENOMEM; 3960e8569dd2SAndreas Sandberg break; 3961e8569dd2SAndreas Sandberg } 3962c5c99429SLarry Woodman 39635e41540cSMike Kravetz /* 39645e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 39655e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 39665e41540cSMike Kravetz * 39675e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 39685e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 39695e41540cSMike Kravetz * Check here before taking page table lock, and once again 39705e41540cSMike Kravetz * after taking the lock below. 39715e41540cSMike Kravetz */ 39725e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 39735e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3974c5c99429SLarry Woodman continue; 3975c5c99429SLarry Woodman 3976cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3977cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3978cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 39794a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 39805e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 39814eae4efaSPeter Xu again: 39825e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 39835e41540cSMike Kravetz /* 39845e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 39855e41540cSMike Kravetz * unlikely case dst entry !none as this implies 39865e41540cSMike Kravetz * sharing with another vma. 39875e41540cSMike Kravetz */ 39884a705fefSNaoya Horiguchi ; 39894a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 39904a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 39914a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 39924a705fefSNaoya Horiguchi 39934a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 39944a705fefSNaoya Horiguchi /* 39954a705fefSNaoya Horiguchi * COW mappings require pages in both 39964a705fefSNaoya Horiguchi * parent and child to be set to read. 39974a705fefSNaoya Horiguchi */ 39984a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 39994a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4000e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4001e5251fd4SPunit Agrawal entry, sz); 40024a705fefSNaoya Horiguchi } 4003e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 40044a705fefSNaoya Horiguchi } else { 40054eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 40064eae4efaSPeter Xu ptepage = pte_page(entry); 40074eae4efaSPeter Xu get_page(ptepage); 40084eae4efaSPeter Xu 40094eae4efaSPeter Xu /* 40104eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 40114eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 40124eae4efaSPeter Xu * COW earlier during fork. 40134eae4efaSPeter Xu * 40144eae4efaSPeter Xu * When pre-allocating the page or copying data, we 40154eae4efaSPeter Xu * need to be without the pgtable locks since we could 40164eae4efaSPeter Xu * sleep during the process. 40174eae4efaSPeter Xu */ 40184eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 40194eae4efaSPeter Xu pte_t src_pte_old = entry; 40204eae4efaSPeter Xu struct page *new; 40214eae4efaSPeter Xu 40224eae4efaSPeter Xu spin_unlock(src_ptl); 40234eae4efaSPeter Xu spin_unlock(dst_ptl); 40244eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 40254eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 40264eae4efaSPeter Xu if (IS_ERR(new)) { 40274eae4efaSPeter Xu put_page(ptepage); 40284eae4efaSPeter Xu ret = PTR_ERR(new); 40294eae4efaSPeter Xu break; 40304eae4efaSPeter Xu } 40314eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 40324eae4efaSPeter Xu npages); 40334eae4efaSPeter Xu put_page(ptepage); 40344eae4efaSPeter Xu 40354eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 40364eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 40374eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 40384eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 40394eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 40404eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 40414eae4efaSPeter Xu put_page(new); 40424eae4efaSPeter Xu /* dst_entry won't change as in child */ 40434eae4efaSPeter Xu goto again; 40444eae4efaSPeter Xu } 40454eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 40464eae4efaSPeter Xu spin_unlock(src_ptl); 40474eae4efaSPeter Xu spin_unlock(dst_ptl); 40484eae4efaSPeter Xu continue; 40494eae4efaSPeter Xu } 40504eae4efaSPeter Xu 405134ee645eSJoerg Roedel if (cow) { 40520f10851eSJérôme Glisse /* 40530f10851eSJérôme Glisse * No need to notify as we are downgrading page 40540f10851eSJérôme Glisse * table protection not changing it to point 40550f10851eSJérôme Glisse * to a new page. 40560f10851eSJérôme Glisse * 4057ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 40580f10851eSJérôme Glisse */ 40597f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 406034ee645eSJoerg Roedel } 40614eae4efaSPeter Xu 406253f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 406363551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 40644eae4efaSPeter Xu hugetlb_count_add(npages, dst); 40651c59827dSHugh Dickins } 4066cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4067cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 406863551ae0SDavid Gibson } 406963551ae0SDavid Gibson 4070e8569dd2SAndreas Sandberg if (cow) 4071ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4072c0d0381aSMike Kravetz else 4073c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4074e8569dd2SAndreas Sandberg 4075e8569dd2SAndreas Sandberg return ret; 407663551ae0SDavid Gibson } 407763551ae0SDavid Gibson 407824669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 407924669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 408024669e58SAneesh Kumar K.V struct page *ref_page) 408163551ae0SDavid Gibson { 408263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 408363551ae0SDavid Gibson unsigned long address; 4084c7546f8fSDavid Gibson pte_t *ptep; 408563551ae0SDavid Gibson pte_t pte; 4086cb900f41SKirill A. Shutemov spinlock_t *ptl; 408763551ae0SDavid Gibson struct page *page; 4088a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4089a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4090ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4091a5516438SAndi Kleen 409263551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4093a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4094a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 409563551ae0SDavid Gibson 409607e32661SAneesh Kumar K.V /* 409707e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 409807e32661SAneesh Kumar K.V * to huge page. 409907e32661SAneesh Kumar K.V */ 4100ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 410124669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4102dff11abeSMike Kravetz 4103dff11abeSMike Kravetz /* 4104dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4105dff11abeSMike Kravetz */ 41066f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 41076f4f13e8SJérôme Glisse end); 4108ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4109ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4110569f48b8SHillf Danton address = start; 4111569f48b8SHillf Danton for (; address < end; address += sz) { 41127868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4113c7546f8fSDavid Gibson if (!ptep) 4114c7546f8fSDavid Gibson continue; 4115c7546f8fSDavid Gibson 4116cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 411734ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 411831d49da5SAneesh Kumar K.V spin_unlock(ptl); 4119dff11abeSMike Kravetz /* 4120dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 4121dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 4122dff11abeSMike Kravetz */ 412331d49da5SAneesh Kumar K.V continue; 412431d49da5SAneesh Kumar K.V } 412539dde65cSChen, Kenneth W 41266629326bSHillf Danton pte = huge_ptep_get(ptep); 412731d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 412831d49da5SAneesh Kumar K.V spin_unlock(ptl); 412931d49da5SAneesh Kumar K.V continue; 413031d49da5SAneesh Kumar K.V } 41316629326bSHillf Danton 41326629326bSHillf Danton /* 41339fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 41349fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 41356629326bSHillf Danton */ 41369fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 41379386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 413831d49da5SAneesh Kumar K.V spin_unlock(ptl); 413931d49da5SAneesh Kumar K.V continue; 41408c4894c6SNaoya Horiguchi } 41416629326bSHillf Danton 41426629326bSHillf Danton page = pte_page(pte); 414304f2cbe3SMel Gorman /* 414404f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 414504f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 414604f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 414704f2cbe3SMel Gorman */ 414804f2cbe3SMel Gorman if (ref_page) { 414931d49da5SAneesh Kumar K.V if (page != ref_page) { 415031d49da5SAneesh Kumar K.V spin_unlock(ptl); 415131d49da5SAneesh Kumar K.V continue; 415231d49da5SAneesh Kumar K.V } 415304f2cbe3SMel Gorman /* 415404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 415504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 415604f2cbe3SMel Gorman * looking like data was lost 415704f2cbe3SMel Gorman */ 415804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 415904f2cbe3SMel Gorman } 416004f2cbe3SMel Gorman 4161c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4162b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4163106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 41646649a386SKen Chen set_page_dirty(page); 41659e81130bSHillf Danton 41665d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4167d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 416831d49da5SAneesh Kumar K.V 4169cb900f41SKirill A. Shutemov spin_unlock(ptl); 4170e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 417124669e58SAneesh Kumar K.V /* 417231d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 417324669e58SAneesh Kumar K.V */ 417431d49da5SAneesh Kumar K.V if (ref_page) 417531d49da5SAneesh Kumar K.V break; 4176fe1668aeSChen, Kenneth W } 4177ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 417824669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 41791da177e4SLinus Torvalds } 418063551ae0SDavid Gibson 4181d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4182d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4183d833352aSMel Gorman unsigned long end, struct page *ref_page) 4184d833352aSMel Gorman { 4185d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4186d833352aSMel Gorman 4187d833352aSMel Gorman /* 4188d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4189d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4190d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4191d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4192d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4193c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4194d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4195c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4196d833352aSMel Gorman */ 4197d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4198d833352aSMel Gorman } 4199d833352aSMel Gorman 4200502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 420104f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4202502717f4SChen, Kenneth W { 420324669e58SAneesh Kumar K.V struct mmu_gather tlb; 4204dff11abeSMike Kravetz 4205a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 420624669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4207ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4208502717f4SChen, Kenneth W } 4209502717f4SChen, Kenneth W 421004f2cbe3SMel Gorman /* 421104f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4212578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 421304f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 421404f2cbe3SMel Gorman * same region. 421504f2cbe3SMel Gorman */ 42162f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 42172a4b3dedSHarvey Harrison struct page *page, unsigned long address) 421804f2cbe3SMel Gorman { 42197526674dSAdam Litke struct hstate *h = hstate_vma(vma); 422004f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 422104f2cbe3SMel Gorman struct address_space *mapping; 422204f2cbe3SMel Gorman pgoff_t pgoff; 422304f2cbe3SMel Gorman 422404f2cbe3SMel Gorman /* 422504f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 422604f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 422704f2cbe3SMel Gorman */ 42287526674dSAdam Litke address = address & huge_page_mask(h); 422936e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 423036e4f20aSMichal Hocko vma->vm_pgoff; 423193c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 423204f2cbe3SMel Gorman 42334eb2b1dcSMel Gorman /* 42344eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 42354eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 42364eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 42374eb2b1dcSMel Gorman */ 423883cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 42396b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 424004f2cbe3SMel Gorman /* Do not unmap the current VMA */ 424104f2cbe3SMel Gorman if (iter_vma == vma) 424204f2cbe3SMel Gorman continue; 424304f2cbe3SMel Gorman 424404f2cbe3SMel Gorman /* 42452f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 42462f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 42472f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 42482f84a899SMel Gorman */ 42492f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 42502f84a899SMel Gorman continue; 42512f84a899SMel Gorman 42522f84a899SMel Gorman /* 425304f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 425404f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 425504f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 425604f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 425704f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 425804f2cbe3SMel Gorman */ 425904f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 426024669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 426124669e58SAneesh Kumar K.V address + huge_page_size(h), page); 426204f2cbe3SMel Gorman } 426383cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 426404f2cbe3SMel Gorman } 426504f2cbe3SMel Gorman 42660fe6e20bSNaoya Horiguchi /* 42670fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4268ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4269ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4270ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 42710fe6e20bSNaoya Horiguchi */ 42722b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4273974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4274cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 42751e8f889bSDavid Gibson { 42763999f52eSAneesh Kumar K.V pte_t pte; 4277a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42781e8f889bSDavid Gibson struct page *old_page, *new_page; 42792b740303SSouptick Joarder int outside_reserve = 0; 42802b740303SSouptick Joarder vm_fault_t ret = 0; 4281974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4282ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 42831e8f889bSDavid Gibson 42843999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 42851e8f889bSDavid Gibson old_page = pte_page(pte); 42861e8f889bSDavid Gibson 428704f2cbe3SMel Gorman retry_avoidcopy: 42881e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 42891e8f889bSDavid Gibson * and just make the page writable */ 429037a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 42915a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 42925b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 429383c54070SNick Piggin return 0; 42941e8f889bSDavid Gibson } 42951e8f889bSDavid Gibson 429604f2cbe3SMel Gorman /* 429704f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 429804f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 429904f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 430004f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 430104f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 430204f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 430304f2cbe3SMel Gorman * of the full address range. 430404f2cbe3SMel Gorman */ 43055944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 430604f2cbe3SMel Gorman old_page != pagecache_page) 430704f2cbe3SMel Gorman outside_reserve = 1; 430804f2cbe3SMel Gorman 430909cbfeafSKirill A. Shutemov get_page(old_page); 4310b76c8cfbSLarry Woodman 4311ad4404a2SDavidlohr Bueso /* 4312ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4313ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4314ad4404a2SDavidlohr Bueso */ 4315cb900f41SKirill A. Shutemov spin_unlock(ptl); 43165b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 43171e8f889bSDavid Gibson 43182fc39cecSAdam Litke if (IS_ERR(new_page)) { 431904f2cbe3SMel Gorman /* 432004f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 432104f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 432204f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 432304f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 432404f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 432504f2cbe3SMel Gorman */ 432604f2cbe3SMel Gorman if (outside_reserve) { 4327e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4328e7dd91c4SMike Kravetz pgoff_t idx; 4329e7dd91c4SMike Kravetz u32 hash; 4330e7dd91c4SMike Kravetz 433109cbfeafSKirill A. Shutemov put_page(old_page); 433204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4333e7dd91c4SMike Kravetz /* 4334e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4335e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4336e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4337e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4338e7dd91c4SMike Kravetz * PMD sharing. 4339e7dd91c4SMike Kravetz * 4340e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4341e7dd91c4SMike Kravetz */ 4342e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4343e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4344e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4345e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4346e7dd91c4SMike Kravetz 43475b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4348e7dd91c4SMike Kravetz 4349e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4350e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4351cb900f41SKirill A. Shutemov spin_lock(ptl); 43525b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4353a9af0c5dSNaoya Horiguchi if (likely(ptep && 4354a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 435504f2cbe3SMel Gorman goto retry_avoidcopy; 4356a734bcc8SHillf Danton /* 4357cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4358cb900f41SKirill A. Shutemov * lock, and our job is done. 4359a734bcc8SHillf Danton */ 4360a734bcc8SHillf Danton return 0; 436104f2cbe3SMel Gorman } 436204f2cbe3SMel Gorman 43632b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4364ad4404a2SDavidlohr Bueso goto out_release_old; 43651e8f889bSDavid Gibson } 43661e8f889bSDavid Gibson 43670fe6e20bSNaoya Horiguchi /* 43680fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 43690fe6e20bSNaoya Horiguchi * anon_vma prepared. 43700fe6e20bSNaoya Horiguchi */ 437144e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4372ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4373ad4404a2SDavidlohr Bueso goto out_release_all; 437444e2aa93SDean Nelson } 43750fe6e20bSNaoya Horiguchi 4376974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 437747ad8475SAndrea Arcangeli pages_per_huge_page(h)); 43780ed361deSNick Piggin __SetPageUptodate(new_page); 43791e8f889bSDavid Gibson 43807269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 43816f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4382ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4383ad4404a2SDavidlohr Bueso 4384b76c8cfbSLarry Woodman /* 4385cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4386b76c8cfbSLarry Woodman * before the page tables are altered 4387b76c8cfbSLarry Woodman */ 4388cb900f41SKirill A. Shutemov spin_lock(ptl); 43895b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4390a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4391d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 439207443a85SJoonsoo Kim 43931e8f889bSDavid Gibson /* Break COW */ 43945b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4395ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 43965b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 43971e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4398d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 43995b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 44008f251a3dSMike Kravetz SetHPageMigratable(new_page); 44011e8f889bSDavid Gibson /* Make the old page be freed below */ 44021e8f889bSDavid Gibson new_page = old_page; 44031e8f889bSDavid Gibson } 4404cb900f41SKirill A. Shutemov spin_unlock(ptl); 4405ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4406ad4404a2SDavidlohr Bueso out_release_all: 44075b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 440809cbfeafSKirill A. Shutemov put_page(new_page); 4409ad4404a2SDavidlohr Bueso out_release_old: 441009cbfeafSKirill A. Shutemov put_page(old_page); 44118312034fSJoonsoo Kim 4412ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4413ad4404a2SDavidlohr Bueso return ret; 44141e8f889bSDavid Gibson } 44151e8f889bSDavid Gibson 441604f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4417a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4418a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 441904f2cbe3SMel Gorman { 442004f2cbe3SMel Gorman struct address_space *mapping; 4421e7c4b0bfSAndy Whitcroft pgoff_t idx; 442204f2cbe3SMel Gorman 442304f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4424a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 442504f2cbe3SMel Gorman 442604f2cbe3SMel Gorman return find_lock_page(mapping, idx); 442704f2cbe3SMel Gorman } 442804f2cbe3SMel Gorman 44293ae77f43SHugh Dickins /* 44303ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 44313ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 44323ae77f43SHugh Dickins */ 44333ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 44342a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 44352a15efc9SHugh Dickins { 44362a15efc9SHugh Dickins struct address_space *mapping; 44372a15efc9SHugh Dickins pgoff_t idx; 44382a15efc9SHugh Dickins struct page *page; 44392a15efc9SHugh Dickins 44402a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 44412a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 44422a15efc9SHugh Dickins 44432a15efc9SHugh Dickins page = find_get_page(mapping, idx); 44442a15efc9SHugh Dickins if (page) 44452a15efc9SHugh Dickins put_page(page); 44462a15efc9SHugh Dickins return page != NULL; 44472a15efc9SHugh Dickins } 44482a15efc9SHugh Dickins 4449ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4450ab76ad54SMike Kravetz pgoff_t idx) 4451ab76ad54SMike Kravetz { 4452ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4453ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4454ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4455ab76ad54SMike Kravetz 4456ab76ad54SMike Kravetz if (err) 4457ab76ad54SMike Kravetz return err; 4458d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4459ab76ad54SMike Kravetz 446022146c3cSMike Kravetz /* 446122146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 446222146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 446322146c3cSMike Kravetz */ 446422146c3cSMike Kravetz set_page_dirty(page); 446522146c3cSMike Kravetz 4466ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4467ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4468ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4469ab76ad54SMike Kravetz return 0; 4470ab76ad54SMike Kravetz } 4471ab76ad54SMike Kravetz 44722b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 44732b740303SSouptick Joarder struct vm_area_struct *vma, 44748382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4475788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4476ac9b9c66SHugh Dickins { 4477a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44782b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4479409eb8c2SHillf Danton int anon_rmap = 0; 44804c887265SAdam Litke unsigned long size; 44814c887265SAdam Litke struct page *page; 44821e8f889bSDavid Gibson pte_t new_pte; 4483cb900f41SKirill A. Shutemov spinlock_t *ptl; 4484285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4485cb6acd01SMike Kravetz bool new_page = false; 44864c887265SAdam Litke 448704f2cbe3SMel Gorman /* 448804f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 448904f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 449025985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 449104f2cbe3SMel Gorman */ 449204f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4493910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 449404f2cbe3SMel Gorman current->pid); 449504f2cbe3SMel Gorman return ret; 449604f2cbe3SMel Gorman } 449704f2cbe3SMel Gorman 44984c887265SAdam Litke /* 449987bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 450087bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 450187bf91d3SMike Kravetz * once for faults beyond end of file. 45024c887265SAdam Litke */ 4503a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4504ebed4bfcSHugh Dickins if (idx >= size) 4505ebed4bfcSHugh Dickins goto out; 45061a1aad8aSMike Kravetz 450787bf91d3SMike Kravetz retry: 450887bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 450987bf91d3SMike Kravetz if (!page) { 45101a1aad8aSMike Kravetz /* 45111a1aad8aSMike Kravetz * Check for page in userfault range 45121a1aad8aSMike Kravetz */ 45131a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 45141a1aad8aSMike Kravetz u32 hash; 45151a1aad8aSMike Kravetz struct vm_fault vmf = { 45161a1aad8aSMike Kravetz .vma = vma, 4517285b8dcaSHuang Ying .address = haddr, 45181a1aad8aSMike Kravetz .flags = flags, 45191a1aad8aSMike Kravetz /* 45201a1aad8aSMike Kravetz * Hard to debug if it ends up being 45211a1aad8aSMike Kravetz * used by a callee that assumes 45221a1aad8aSMike Kravetz * something about the other 45231a1aad8aSMike Kravetz * uninitialized fields... same as in 45241a1aad8aSMike Kravetz * memory.c 45251a1aad8aSMike Kravetz */ 45261a1aad8aSMike Kravetz }; 45271a1aad8aSMike Kravetz 45281a1aad8aSMike Kravetz /* 4529c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4530c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4531c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 45321a1aad8aSMike Kravetz */ 4533188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 45341a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4535c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 45361a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4537c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 45381a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45391a1aad8aSMike Kravetz goto out; 45401a1aad8aSMike Kravetz } 45411a1aad8aSMike Kravetz 4542285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 45432fc39cecSAdam Litke if (IS_ERR(page)) { 45444643d67eSMike Kravetz /* 45454643d67eSMike Kravetz * Returning error will result in faulting task being 45464643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 45474643d67eSMike Kravetz * tasks from racing to fault in the same page which 45484643d67eSMike Kravetz * could result in false unable to allocate errors. 45494643d67eSMike Kravetz * Page migration does not take the fault mutex, but 45504643d67eSMike Kravetz * does a clear then write of pte's under page table 45514643d67eSMike Kravetz * lock. Page fault code could race with migration, 45524643d67eSMike Kravetz * notice the clear pte and try to allocate a page 45534643d67eSMike Kravetz * here. Before returning error, get ptl and make 45544643d67eSMike Kravetz * sure there really is no pte entry. 45554643d67eSMike Kravetz */ 45564643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 45574643d67eSMike Kravetz ret = 0; 4558d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 45592b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 4560d83e6c8aSMiaohe Lin spin_unlock(ptl); 45616bda666aSChristoph Lameter goto out; 45626bda666aSChristoph Lameter } 456347ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 45640ed361deSNick Piggin __SetPageUptodate(page); 4565cb6acd01SMike Kravetz new_page = true; 4566ac9b9c66SHugh Dickins 4567f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4568ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 45696bda666aSChristoph Lameter if (err) { 45706bda666aSChristoph Lameter put_page(page); 45716bda666aSChristoph Lameter if (err == -EEXIST) 45726bda666aSChristoph Lameter goto retry; 45736bda666aSChristoph Lameter goto out; 45746bda666aSChristoph Lameter } 457523be7468SMel Gorman } else { 45766bda666aSChristoph Lameter lock_page(page); 45770fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 45780fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 45790fe6e20bSNaoya Horiguchi goto backout_unlocked; 458023be7468SMel Gorman } 4581409eb8c2SHillf Danton anon_rmap = 1; 45820fe6e20bSNaoya Horiguchi } 45830fe6e20bSNaoya Horiguchi } else { 458457303d80SAndy Whitcroft /* 4585998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4586998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4587998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4588fd6a03edSNaoya Horiguchi */ 4589fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 45900eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4591972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4592fd6a03edSNaoya Horiguchi goto backout_unlocked; 45936bda666aSChristoph Lameter } 4594998b4382SNaoya Horiguchi } 45951e8f889bSDavid Gibson 459657303d80SAndy Whitcroft /* 459757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 459857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 459957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 460057303d80SAndy Whitcroft * the spinlock. 460157303d80SAndy Whitcroft */ 46025e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4603285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 46042b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 46052b26736cSAndy Whitcroft goto backout_unlocked; 46062b26736cSAndy Whitcroft } 46075e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4608285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 46095e911373SMike Kravetz } 461057303d80SAndy Whitcroft 46118bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 461283c54070SNick Piggin ret = 0; 46137f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 46144c887265SAdam Litke goto backout; 46154c887265SAdam Litke 461607443a85SJoonsoo Kim if (anon_rmap) { 4617d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4618285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4619ac714904SChoi Gi-yong } else 462053f9263bSKirill A. Shutemov page_dup_rmap(page, true); 46211e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 46221e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4623285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 46241e8f889bSDavid Gibson 46255d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4626788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 46271e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4628974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 46291e8f889bSDavid Gibson } 46301e8f889bSDavid Gibson 4631cb900f41SKirill A. Shutemov spin_unlock(ptl); 4632cb6acd01SMike Kravetz 4633cb6acd01SMike Kravetz /* 46348f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 46358f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 46368f251a3dSMike Kravetz * been isolated for migration. 4637cb6acd01SMike Kravetz */ 4638cb6acd01SMike Kravetz if (new_page) 46398f251a3dSMike Kravetz SetHPageMigratable(page); 4640cb6acd01SMike Kravetz 46414c887265SAdam Litke unlock_page(page); 46424c887265SAdam Litke out: 4643ac9b9c66SHugh Dickins return ret; 46444c887265SAdam Litke 46454c887265SAdam Litke backout: 4646cb900f41SKirill A. Shutemov spin_unlock(ptl); 46472b26736cSAndy Whitcroft backout_unlocked: 46484c887265SAdam Litke unlock_page(page); 4649285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 46504c887265SAdam Litke put_page(page); 46514c887265SAdam Litke goto out; 4652ac9b9c66SHugh Dickins } 4653ac9b9c66SHugh Dickins 46548382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4655188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 46568382d914SDavidlohr Bueso { 46578382d914SDavidlohr Bueso unsigned long key[2]; 46588382d914SDavidlohr Bueso u32 hash; 46598382d914SDavidlohr Bueso 46608382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 46618382d914SDavidlohr Bueso key[1] = idx; 46628382d914SDavidlohr Bueso 466355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 46648382d914SDavidlohr Bueso 46658382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 46668382d914SDavidlohr Bueso } 46678382d914SDavidlohr Bueso #else 46688382d914SDavidlohr Bueso /* 46696c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 46708382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 46718382d914SDavidlohr Bueso */ 4672188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 46738382d914SDavidlohr Bueso { 46748382d914SDavidlohr Bueso return 0; 46758382d914SDavidlohr Bueso } 46768382d914SDavidlohr Bueso #endif 46778382d914SDavidlohr Bueso 46782b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4679788c7df4SHugh Dickins unsigned long address, unsigned int flags) 468086e5216fSAdam Litke { 46818382d914SDavidlohr Bueso pte_t *ptep, entry; 4682cb900f41SKirill A. Shutemov spinlock_t *ptl; 46832b740303SSouptick Joarder vm_fault_t ret; 46848382d914SDavidlohr Bueso u32 hash; 46858382d914SDavidlohr Bueso pgoff_t idx; 46860fe6e20bSNaoya Horiguchi struct page *page = NULL; 468757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4688a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 46898382d914SDavidlohr Bueso struct address_space *mapping; 46900f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4691285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 469286e5216fSAdam Litke 4693285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4694fd6a03edSNaoya Horiguchi if (ptep) { 4695c0d0381aSMike Kravetz /* 4696c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4697c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4698c0d0381aSMike Kravetz * not actually modifying content here. 4699c0d0381aSMike Kravetz */ 4700fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4701290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4702cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4703290408d4SNaoya Horiguchi return 0; 4704290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4705aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4706972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4707b43a9990SMike Kravetz } 47088382d914SDavidlohr Bueso 4709c0d0381aSMike Kravetz /* 4710c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 471187bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 471287bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 471387bf91d3SMike Kravetz * and making the ptep no longer valid. 471487bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4715c0d0381aSMike Kravetz * 4716c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4717c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4718c0d0381aSMike Kravetz * something has changed. 4719c0d0381aSMike Kravetz */ 4720ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4721c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4722aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 4723c0d0381aSMike Kravetz if (!ptep) { 4724c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4725c0d0381aSMike Kravetz return VM_FAULT_OOM; 4726c0d0381aSMike Kravetz } 4727ddeaab32SMike Kravetz 47283935baa9SDavid Gibson /* 47293935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 47303935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 47313935baa9SDavid Gibson * the same page in the page cache. 47323935baa9SDavid Gibson */ 4733c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4734188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4735c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 47368382d914SDavidlohr Bueso 47377f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 47387f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 47398382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4740b4d1d99fSDavid Gibson goto out_mutex; 47413935baa9SDavid Gibson } 474286e5216fSAdam Litke 474383c54070SNick Piggin ret = 0; 47441e8f889bSDavid Gibson 474557303d80SAndy Whitcroft /* 47460f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 47470f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 47487c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 47497c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 47507c8de358SEthon Paul * properly handle it. 47510f792cf9SNaoya Horiguchi */ 47520f792cf9SNaoya Horiguchi if (!pte_present(entry)) 47530f792cf9SNaoya Horiguchi goto out_mutex; 47540f792cf9SNaoya Horiguchi 47550f792cf9SNaoya Horiguchi /* 475657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 475757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 475857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 475957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 476057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 476157303d80SAndy Whitcroft * consumed. 476257303d80SAndy Whitcroft */ 4763106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4764285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 47652b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4766b4d1d99fSDavid Gibson goto out_mutex; 47672b26736cSAndy Whitcroft } 47685e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4769285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 477057303d80SAndy Whitcroft 4771f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 477257303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4773285b8dcaSHuang Ying vma, haddr); 477457303d80SAndy Whitcroft } 477557303d80SAndy Whitcroft 47760f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 47770fe6e20bSNaoya Horiguchi 47781e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4779b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4780cb900f41SKirill A. Shutemov goto out_ptl; 4781b4d1d99fSDavid Gibson 47820f792cf9SNaoya Horiguchi /* 47830f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 47840f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 47850f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 47860f792cf9SNaoya Horiguchi */ 47870f792cf9SNaoya Horiguchi page = pte_page(entry); 47880f792cf9SNaoya Horiguchi if (page != pagecache_page) 47890f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 47900f792cf9SNaoya Horiguchi need_wait_lock = 1; 47910f792cf9SNaoya Horiguchi goto out_ptl; 47920f792cf9SNaoya Horiguchi } 47930f792cf9SNaoya Horiguchi 47940f792cf9SNaoya Horiguchi get_page(page); 4795b4d1d99fSDavid Gibson 4796788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4797106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4798974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4799cb900f41SKirill A. Shutemov pagecache_page, ptl); 48000f792cf9SNaoya Horiguchi goto out_put_page; 4801b4d1d99fSDavid Gibson } 4802106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4803b4d1d99fSDavid Gibson } 4804b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4805285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4806788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4807285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 48080f792cf9SNaoya Horiguchi out_put_page: 48090f792cf9SNaoya Horiguchi if (page != pagecache_page) 48100f792cf9SNaoya Horiguchi unlock_page(page); 48110f792cf9SNaoya Horiguchi put_page(page); 4812cb900f41SKirill A. Shutemov out_ptl: 4813cb900f41SKirill A. Shutemov spin_unlock(ptl); 481457303d80SAndy Whitcroft 481557303d80SAndy Whitcroft if (pagecache_page) { 481657303d80SAndy Whitcroft unlock_page(pagecache_page); 481757303d80SAndy Whitcroft put_page(pagecache_page); 481857303d80SAndy Whitcroft } 4819b4d1d99fSDavid Gibson out_mutex: 4820c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4821c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 48220f792cf9SNaoya Horiguchi /* 48230f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 48240f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 48250f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 48260f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 48270f792cf9SNaoya Horiguchi * here without taking refcount. 48280f792cf9SNaoya Horiguchi */ 48290f792cf9SNaoya Horiguchi if (need_wait_lock) 48300f792cf9SNaoya Horiguchi wait_on_page_locked(page); 48311e8f889bSDavid Gibson return ret; 483286e5216fSAdam Litke } 483386e5216fSAdam Litke 48348fb5debcSMike Kravetz /* 48358fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 48368fb5debcSMike Kravetz * modifications for huge pages. 48378fb5debcSMike Kravetz */ 48388fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 48398fb5debcSMike Kravetz pte_t *dst_pte, 48408fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 48418fb5debcSMike Kravetz unsigned long dst_addr, 48428fb5debcSMike Kravetz unsigned long src_addr, 48438fb5debcSMike Kravetz struct page **pagep) 48448fb5debcSMike Kravetz { 48451e392147SAndrea Arcangeli struct address_space *mapping; 48461e392147SAndrea Arcangeli pgoff_t idx; 48471e392147SAndrea Arcangeli unsigned long size; 48481c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 48498fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 48508fb5debcSMike Kravetz pte_t _dst_pte; 48518fb5debcSMike Kravetz spinlock_t *ptl; 48528fb5debcSMike Kravetz int ret; 48538fb5debcSMike Kravetz struct page *page; 48548fb5debcSMike Kravetz 48558fb5debcSMike Kravetz if (!*pagep) { 48568fb5debcSMike Kravetz ret = -ENOMEM; 48578fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 48588fb5debcSMike Kravetz if (IS_ERR(page)) 48598fb5debcSMike Kravetz goto out; 48608fb5debcSMike Kravetz 48618fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 48628fb5debcSMike Kravetz (const void __user *) src_addr, 4863810a56b9SMike Kravetz pages_per_huge_page(h), false); 48648fb5debcSMike Kravetz 4865c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 48668fb5debcSMike Kravetz if (unlikely(ret)) { 48679e368259SAndrea Arcangeli ret = -ENOENT; 48688fb5debcSMike Kravetz *pagep = page; 48698fb5debcSMike Kravetz /* don't free the page */ 48708fb5debcSMike Kravetz goto out; 48718fb5debcSMike Kravetz } 48728fb5debcSMike Kravetz } else { 48738fb5debcSMike Kravetz page = *pagep; 48748fb5debcSMike Kravetz *pagep = NULL; 48758fb5debcSMike Kravetz } 48768fb5debcSMike Kravetz 48778fb5debcSMike Kravetz /* 48788fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 48798fb5debcSMike Kravetz * preceding stores to the page contents become visible before 48808fb5debcSMike Kravetz * the set_pte_at() write. 48818fb5debcSMike Kravetz */ 48828fb5debcSMike Kravetz __SetPageUptodate(page); 48838fb5debcSMike Kravetz 48841e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 48851e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 48861e392147SAndrea Arcangeli 48871c9e8defSMike Kravetz /* 48881c9e8defSMike Kravetz * If shared, add to page cache 48891c9e8defSMike Kravetz */ 48901c9e8defSMike Kravetz if (vm_shared) { 48911e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 48921e392147SAndrea Arcangeli ret = -EFAULT; 48931e392147SAndrea Arcangeli if (idx >= size) 48941e392147SAndrea Arcangeli goto out_release_nounlock; 48951c9e8defSMike Kravetz 48961e392147SAndrea Arcangeli /* 48971e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 48981e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 48991e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 49001e392147SAndrea Arcangeli * the caller. 49011e392147SAndrea Arcangeli */ 49021c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 49031c9e8defSMike Kravetz if (ret) 49041c9e8defSMike Kravetz goto out_release_nounlock; 49051c9e8defSMike Kravetz } 49061c9e8defSMike Kravetz 49078fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 49088fb5debcSMike Kravetz spin_lock(ptl); 49098fb5debcSMike Kravetz 49101e392147SAndrea Arcangeli /* 49111e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 49121e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 49131e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 49141e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 49151e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 49161e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 49171e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 49181e392147SAndrea Arcangeli */ 49191e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 49201e392147SAndrea Arcangeli ret = -EFAULT; 49211e392147SAndrea Arcangeli if (idx >= size) 49221e392147SAndrea Arcangeli goto out_release_unlock; 49231e392147SAndrea Arcangeli 49248fb5debcSMike Kravetz ret = -EEXIST; 49258fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 49268fb5debcSMike Kravetz goto out_release_unlock; 49278fb5debcSMike Kravetz 49281c9e8defSMike Kravetz if (vm_shared) { 49291c9e8defSMike Kravetz page_dup_rmap(page, true); 49301c9e8defSMike Kravetz } else { 4931d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 49328fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 49331c9e8defSMike Kravetz } 49348fb5debcSMike Kravetz 49358fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 49368fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 49378fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 49388fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 49398fb5debcSMike Kravetz 49408fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 49418fb5debcSMike Kravetz 49428fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 49438fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 49448fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 49458fb5debcSMike Kravetz 49468fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 49478fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 49488fb5debcSMike Kravetz 49498fb5debcSMike Kravetz spin_unlock(ptl); 49508f251a3dSMike Kravetz SetHPageMigratable(page); 49511c9e8defSMike Kravetz if (vm_shared) 49521c9e8defSMike Kravetz unlock_page(page); 49538fb5debcSMike Kravetz ret = 0; 49548fb5debcSMike Kravetz out: 49558fb5debcSMike Kravetz return ret; 49568fb5debcSMike Kravetz out_release_unlock: 49578fb5debcSMike Kravetz spin_unlock(ptl); 49581c9e8defSMike Kravetz if (vm_shared) 49591c9e8defSMike Kravetz unlock_page(page); 49605af10dfdSAndrea Arcangeli out_release_nounlock: 49618fb5debcSMike Kravetz put_page(page); 49628fb5debcSMike Kravetz goto out; 49638fb5debcSMike Kravetz } 49648fb5debcSMike Kravetz 496582e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 496682e5d378SJoao Martins int refs, struct page **pages, 496782e5d378SJoao Martins struct vm_area_struct **vmas) 496882e5d378SJoao Martins { 496982e5d378SJoao Martins int nr; 497082e5d378SJoao Martins 497182e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 497282e5d378SJoao Martins if (likely(pages)) 497382e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 497482e5d378SJoao Martins if (vmas) 497582e5d378SJoao Martins vmas[nr] = vma; 497682e5d378SJoao Martins } 497782e5d378SJoao Martins } 497882e5d378SJoao Martins 497928a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 498063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 498128a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 49824f6da934SPeter Xu long i, unsigned int flags, int *locked) 498363551ae0SDavid Gibson { 4984d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4985d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 498628a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4987a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49880fa5bc40SJoao Martins int err = -EFAULT, refs; 498963551ae0SDavid Gibson 499063551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 499163551ae0SDavid Gibson pte_t *pte; 4992cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 49932a15efc9SHugh Dickins int absent; 499463551ae0SDavid Gibson struct page *page; 499563551ae0SDavid Gibson 49964c887265SAdam Litke /* 499702057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 499802057967SDavid Rientjes * potentially allocating memory. 499902057967SDavid Rientjes */ 5000fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 500102057967SDavid Rientjes remainder = 0; 500202057967SDavid Rientjes break; 500302057967SDavid Rientjes } 500402057967SDavid Rientjes 500502057967SDavid Rientjes /* 50064c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 50072a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 50084c887265SAdam Litke * first, for the page indexing below to work. 5009cb900f41SKirill A. Shutemov * 5010cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 50114c887265SAdam Litke */ 50127868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 50137868a208SPunit Agrawal huge_page_size(h)); 5014cb900f41SKirill A. Shutemov if (pte) 5015cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 50162a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 501763551ae0SDavid Gibson 50182a15efc9SHugh Dickins /* 50192a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 50203ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 50213ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 50223ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 50233ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 50242a15efc9SHugh Dickins */ 50253ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 50263ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5027cb900f41SKirill A. Shutemov if (pte) 5028cb900f41SKirill A. Shutemov spin_unlock(ptl); 50292a15efc9SHugh Dickins remainder = 0; 50302a15efc9SHugh Dickins break; 50312a15efc9SHugh Dickins } 50322a15efc9SHugh Dickins 50339cc3a5bdSNaoya Horiguchi /* 50349cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 50359cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 50369cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 50379cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 50389cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 50399cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 50409cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 50419cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 50429cc3a5bdSNaoya Horiguchi */ 50439cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5044106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5045106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 50462b740303SSouptick Joarder vm_fault_t ret; 504787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 50484c887265SAdam Litke 5049cb900f41SKirill A. Shutemov if (pte) 5050cb900f41SKirill A. Shutemov spin_unlock(ptl); 505187ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 505287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 50534f6da934SPeter Xu if (locked) 505471335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 505571335f37SPeter Xu FAULT_FLAG_KILLABLE; 505687ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 505787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 505887ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 505987ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 50604426e945SPeter Xu /* 50614426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 50624426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 50634426e945SPeter Xu */ 506487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 506587ffc118SAndrea Arcangeli } 506687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 506787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 50682be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 50691c59827dSHugh Dickins remainder = 0; 50701c59827dSHugh Dickins break; 50711c59827dSHugh Dickins } 507287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 50734f6da934SPeter Xu if (locked && 50741ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 50754f6da934SPeter Xu *locked = 0; 507687ffc118SAndrea Arcangeli *nr_pages = 0; 507787ffc118SAndrea Arcangeli /* 507887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 507987ffc118SAndrea Arcangeli * error, it will return zero 508087ffc118SAndrea Arcangeli * instead. 508187ffc118SAndrea Arcangeli * 508287ffc118SAndrea Arcangeli * No need to update "position" as the 508387ffc118SAndrea Arcangeli * caller will not check it after 508487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 508587ffc118SAndrea Arcangeli */ 508687ffc118SAndrea Arcangeli return i; 508787ffc118SAndrea Arcangeli } 508887ffc118SAndrea Arcangeli continue; 508987ffc118SAndrea Arcangeli } 509063551ae0SDavid Gibson 5091a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 50927f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 50938fde12caSLinus Torvalds 50948fde12caSLinus Torvalds /* 5095acbfb087SZhigang Lu * If subpage information not requested, update counters 5096acbfb087SZhigang Lu * and skip the same_page loop below. 5097acbfb087SZhigang Lu */ 5098acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 5099acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 5100acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 5101acbfb087SZhigang Lu vaddr += huge_page_size(h); 5102acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 5103acbfb087SZhigang Lu i += pages_per_huge_page(h); 5104acbfb087SZhigang Lu spin_unlock(ptl); 5105acbfb087SZhigang Lu continue; 5106acbfb087SZhigang Lu } 5107acbfb087SZhigang Lu 510882e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 510982e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 51100fa5bc40SJoao Martins 511182e5d378SJoao Martins if (pages || vmas) 511282e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 511382e5d378SJoao Martins vma, refs, 511482e5d378SJoao Martins likely(pages) ? pages + i : NULL, 511582e5d378SJoao Martins vmas ? vmas + i : NULL); 511663551ae0SDavid Gibson 511782e5d378SJoao Martins if (pages) { 51180fa5bc40SJoao Martins /* 51190fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 51200fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 51210fa5bc40SJoao Martins * checked that the huge page is present in the page 51220fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 51230fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 51240fa5bc40SJoao Martins * head page and tail pages from being rearranged in 51250fa5bc40SJoao Martins * any way. So this page must be available at this 51260fa5bc40SJoao Martins * point, unless the page refcount overflowed: 51270fa5bc40SJoao Martins */ 512882e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 51290fa5bc40SJoao Martins refs, 51300fa5bc40SJoao Martins flags))) { 51310fa5bc40SJoao Martins spin_unlock(ptl); 51320fa5bc40SJoao Martins remainder = 0; 51330fa5bc40SJoao Martins err = -ENOMEM; 51340fa5bc40SJoao Martins break; 51350fa5bc40SJoao Martins } 5136d5d4b0aaSChen, Kenneth W } 513782e5d378SJoao Martins 513882e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 513982e5d378SJoao Martins remainder -= refs; 514082e5d378SJoao Martins i += refs; 514182e5d378SJoao Martins 5142cb900f41SKirill A. Shutemov spin_unlock(ptl); 514363551ae0SDavid Gibson } 514428a35716SMichel Lespinasse *nr_pages = remainder; 514587ffc118SAndrea Arcangeli /* 514687ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 514787ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 514887ffc118SAndrea Arcangeli * branch. 514987ffc118SAndrea Arcangeli */ 515063551ae0SDavid Gibson *position = vaddr; 515163551ae0SDavid Gibson 51522be7cfedSDaniel Jordan return i ? i : err; 515363551ae0SDavid Gibson } 51548f860591SZhang, Yanmin 51557da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 51568f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 51578f860591SZhang, Yanmin { 51588f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 51598f860591SZhang, Yanmin unsigned long start = address; 51608f860591SZhang, Yanmin pte_t *ptep; 51618f860591SZhang, Yanmin pte_t pte; 5162a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 51637da4d641SPeter Zijlstra unsigned long pages = 0; 5164dff11abeSMike Kravetz bool shared_pmd = false; 5165ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5166dff11abeSMike Kravetz 5167dff11abeSMike Kravetz /* 5168dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5169ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5170dff11abeSMike Kravetz * range if PMD sharing is possible. 5171dff11abeSMike Kravetz */ 51727269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 51737269f999SJérôme Glisse 0, vma, mm, start, end); 5174ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 51758f860591SZhang, Yanmin 51768f860591SZhang, Yanmin BUG_ON(address >= end); 5177ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 51788f860591SZhang, Yanmin 5179ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 518083cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5181a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5182cb900f41SKirill A. Shutemov spinlock_t *ptl; 51837868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 51848f860591SZhang, Yanmin if (!ptep) 51858f860591SZhang, Yanmin continue; 5186cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 518734ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 51887da4d641SPeter Zijlstra pages++; 5189cb900f41SKirill A. Shutemov spin_unlock(ptl); 5190dff11abeSMike Kravetz shared_pmd = true; 519139dde65cSChen, Kenneth W continue; 51927da4d641SPeter Zijlstra } 5193a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5194a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5195a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5196a8bda28dSNaoya Horiguchi continue; 5197a8bda28dSNaoya Horiguchi } 5198a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5199a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5200a8bda28dSNaoya Horiguchi 5201a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5202a8bda28dSNaoya Horiguchi pte_t newpte; 5203a8bda28dSNaoya Horiguchi 5204a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5205a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5206e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5207e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5208a8bda28dSNaoya Horiguchi pages++; 5209a8bda28dSNaoya Horiguchi } 5210a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5211a8bda28dSNaoya Horiguchi continue; 5212a8bda28dSNaoya Horiguchi } 5213a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5214023bdd00SAneesh Kumar K.V pte_t old_pte; 5215023bdd00SAneesh Kumar K.V 5216023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5217023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5218be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5219023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 52207da4d641SPeter Zijlstra pages++; 52218f860591SZhang, Yanmin } 5222cb900f41SKirill A. Shutemov spin_unlock(ptl); 52238f860591SZhang, Yanmin } 5224d833352aSMel Gorman /* 5225c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5226d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5227c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5228dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5229dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5230d833352aSMel Gorman */ 5231dff11abeSMike Kravetz if (shared_pmd) 5232ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5233dff11abeSMike Kravetz else 52345491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 52350f10851eSJérôme Glisse /* 52360f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 52370f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 52380f10851eSJérôme Glisse * 5239ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 52400f10851eSJérôme Glisse */ 524183cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5242ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 52437da4d641SPeter Zijlstra 52447da4d641SPeter Zijlstra return pages << h->order; 52458f860591SZhang, Yanmin } 52468f860591SZhang, Yanmin 524733b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 524833b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5249a1e78772SMel Gorman long from, long to, 52505a6fe125SMel Gorman struct vm_area_struct *vma, 5251ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5252e4e574b7SAdam Litke { 525333b8f84aSMike Kravetz long chg, add = -1; 5254a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 525590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52569119a41eSJoonsoo Kim struct resv_map *resv_map; 5257075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 52580db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5259e4e574b7SAdam Litke 526063489f8eSMike Kravetz /* This should never happen */ 526163489f8eSMike Kravetz if (from > to) { 526263489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 526333b8f84aSMike Kravetz return false; 526463489f8eSMike Kravetz } 526563489f8eSMike Kravetz 5266a1e78772SMel Gorman /* 526717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 526817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 526990481622SDavid Gibson * without using reserves 527017c9d12eSMel Gorman */ 5271ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 527233b8f84aSMike Kravetz return true; 527317c9d12eSMel Gorman 527417c9d12eSMel Gorman /* 5275a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5276a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5277a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5278a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5279a1e78772SMel Gorman */ 52809119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5281f27a5136SMike Kravetz /* 5282f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5283f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5284f27a5136SMike Kravetz * hugetlbfs_get_inode). 5285f27a5136SMike Kravetz */ 52864e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 52879119a41eSJoonsoo Kim 52880db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 52899119a41eSJoonsoo Kim 52909119a41eSJoonsoo Kim } else { 5291e9fe92aeSMina Almasry /* Private mapping. */ 52929119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 52935a6fe125SMel Gorman if (!resv_map) 529433b8f84aSMike Kravetz return false; 52955a6fe125SMel Gorman 529617c9d12eSMel Gorman chg = to - from; 529717c9d12eSMel Gorman 52985a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 52995a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 53005a6fe125SMel Gorman } 53015a6fe125SMel Gorman 530233b8f84aSMike Kravetz if (chg < 0) 5303c50ac050SDave Hansen goto out_err; 530417c9d12eSMel Gorman 530533b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 530633b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5307075a61d0SMina Almasry goto out_err; 5308075a61d0SMina Almasry 5309075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5310075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5311075a61d0SMina Almasry * of the resv_map. 5312075a61d0SMina Almasry */ 5313075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5314075a61d0SMina Almasry } 5315075a61d0SMina Almasry 53161c5ecae3SMike Kravetz /* 53171c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 53181c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 53191c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 53201c5ecae3SMike Kravetz */ 53211c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 532233b8f84aSMike Kravetz if (gbl_reserve < 0) 5323075a61d0SMina Almasry goto out_uncharge_cgroup; 532417c9d12eSMel Gorman 532517c9d12eSMel Gorman /* 532617c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 532790481622SDavid Gibson * Hand the pages back to the subpool if there are not 532817c9d12eSMel Gorman */ 532933b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5330075a61d0SMina Almasry goto out_put_pages; 533117c9d12eSMel Gorman 533217c9d12eSMel Gorman /* 533317c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 533417c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 533517c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 533617c9d12eSMel Gorman * the reservation was and the page cache tells how much of 533717c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 533817c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 533917c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 534017c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 534117c9d12eSMel Gorman * else has to be done for private mappings here 534217c9d12eSMel Gorman */ 534333039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5344075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 534533039678SMike Kravetz 53460db9d74eSMina Almasry if (unlikely(add < 0)) { 53470db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5348075a61d0SMina Almasry goto out_put_pages; 53490db9d74eSMina Almasry } else if (unlikely(chg > add)) { 535033039678SMike Kravetz /* 535133039678SMike Kravetz * pages in this range were added to the reserve 535233039678SMike Kravetz * map between region_chg and region_add. This 535333039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 535433039678SMike Kravetz * the subpool and reserve counts modified above 535533039678SMike Kravetz * based on the difference. 535633039678SMike Kravetz */ 535733039678SMike Kravetz long rsv_adjust; 535833039678SMike Kravetz 5359d85aecf2SMiaohe Lin /* 5360d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5361d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5362d85aecf2SMiaohe Lin */ 5363075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5364075a61d0SMina Almasry hstate_index(h), 5365075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5366075a61d0SMina Almasry 536733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 536833039678SMike Kravetz chg - add); 536933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5370d85aecf2SMiaohe Lin } else if (h_cg) { 5371d85aecf2SMiaohe Lin /* 5372d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5373d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5374d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5375d85aecf2SMiaohe Lin * done. 5376d85aecf2SMiaohe Lin */ 5377d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 537833039678SMike Kravetz } 537933039678SMike Kravetz } 538033b8f84aSMike Kravetz return true; 538133b8f84aSMike Kravetz 5382075a61d0SMina Almasry out_put_pages: 5383075a61d0SMina Almasry /* put back original number of pages, chg */ 5384075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5385075a61d0SMina Almasry out_uncharge_cgroup: 5386075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5387075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5388c50ac050SDave Hansen out_err: 53895e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 53900db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 53910db9d74eSMina Almasry * region_add failed or didn't run. 53920db9d74eSMina Almasry */ 53930db9d74eSMina Almasry if (chg >= 0 && add < 0) 53940db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5395f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5396f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 539733b8f84aSMike Kravetz return false; 5398a43a8c39SChen, Kenneth W } 5399a43a8c39SChen, Kenneth W 5400b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5401b5cec28dSMike Kravetz long freed) 5402a43a8c39SChen, Kenneth W { 5403a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 54044e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 54059119a41eSJoonsoo Kim long chg = 0; 540690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 54071c5ecae3SMike Kravetz long gbl_reserve; 540845c682a6SKen Chen 5409f27a5136SMike Kravetz /* 5410f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5411f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5412f27a5136SMike Kravetz */ 5413b5cec28dSMike Kravetz if (resv_map) { 5414b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5415b5cec28dSMike Kravetz /* 5416b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5417b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5418b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5419b5cec28dSMike Kravetz */ 5420b5cec28dSMike Kravetz if (chg < 0) 5421b5cec28dSMike Kravetz return chg; 5422b5cec28dSMike Kravetz } 5423b5cec28dSMike Kravetz 542445c682a6SKen Chen spin_lock(&inode->i_lock); 5425e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 542645c682a6SKen Chen spin_unlock(&inode->i_lock); 542745c682a6SKen Chen 54281c5ecae3SMike Kravetz /* 54291c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 54301c5ecae3SMike Kravetz * reservations to be released may be adjusted. 5431dddf31a4SMiaohe Lin * 5432dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 5433dddf31a4SMiaohe Lin * won't go negative. 54341c5ecae3SMike Kravetz */ 54351c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 54361c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5437b5cec28dSMike Kravetz 5438b5cec28dSMike Kravetz return 0; 5439a43a8c39SChen, Kenneth W } 544093f70f90SNaoya Horiguchi 54413212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 54423212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 54433212b535SSteve Capper struct vm_area_struct *vma, 54443212b535SSteve Capper unsigned long addr, pgoff_t idx) 54453212b535SSteve Capper { 54463212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 54473212b535SSteve Capper svma->vm_start; 54483212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 54493212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 54503212b535SSteve Capper 54513212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5452de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5453de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 54543212b535SSteve Capper 54553212b535SSteve Capper /* 54563212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 54573212b535SSteve Capper * page table page. 54583212b535SSteve Capper */ 54593212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 54603212b535SSteve Capper vm_flags != svm_flags || 546107e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 54623212b535SSteve Capper return 0; 54633212b535SSteve Capper 54643212b535SSteve Capper return saddr; 54653212b535SSteve Capper } 54663212b535SSteve Capper 546731aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 54683212b535SSteve Capper { 54693212b535SSteve Capper unsigned long base = addr & PUD_MASK; 54703212b535SSteve Capper unsigned long end = base + PUD_SIZE; 54713212b535SSteve Capper 54723212b535SSteve Capper /* 54733212b535SSteve Capper * check on proper vm_flags and page table alignment 54743212b535SSteve Capper */ 5475017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 547631aafb45SNicholas Krause return true; 547731aafb45SNicholas Krause return false; 54783212b535SSteve Capper } 54793212b535SSteve Capper 5480c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5481c1991e07SPeter Xu { 5482c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5483c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5484c1991e07SPeter Xu return false; 5485c1991e07SPeter Xu #endif 5486c1991e07SPeter Xu return vma_shareable(vma, addr); 5487c1991e07SPeter Xu } 5488c1991e07SPeter Xu 54893212b535SSteve Capper /* 5490017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5491017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5492017b1660SMike Kravetz * shared pmd mappings. 5493017b1660SMike Kravetz */ 5494017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5495017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5496017b1660SMike Kravetz { 5497a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5498a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5499017b1660SMike Kravetz 5500a1ba9da8SLi Xinhai /* 5501a1ba9da8SLi Xinhai * vma need span at least one aligned PUD size and the start,end range 5502a1ba9da8SLi Xinhai * must at least partialy within it. 5503a1ba9da8SLi Xinhai */ 5504a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5505a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5506017b1660SMike Kravetz return; 5507017b1660SMike Kravetz 550875802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5509a1ba9da8SLi Xinhai if (*start > v_start) 5510a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5511017b1660SMike Kravetz 5512a1ba9da8SLi Xinhai if (*end < v_end) 5513a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5514017b1660SMike Kravetz } 5515017b1660SMike Kravetz 5516017b1660SMike Kravetz /* 55173212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 55183212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 55193212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5520c0d0381aSMike Kravetz * code much cleaner. 5521c0d0381aSMike Kravetz * 55220bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 55230bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 55240bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 55250bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 55260bf7b64eSMike Kravetz * 55270bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 55280bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 55290bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 55300bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 55310bf7b64eSMike Kravetz * only required for subsequent processing. 55323212b535SSteve Capper */ 5533aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5534aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 55353212b535SSteve Capper { 55363212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 55373212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 55383212b535SSteve Capper vma->vm_pgoff; 55393212b535SSteve Capper struct vm_area_struct *svma; 55403212b535SSteve Capper unsigned long saddr; 55413212b535SSteve Capper pte_t *spte = NULL; 55423212b535SSteve Capper pte_t *pte; 5543cb900f41SKirill A. Shutemov spinlock_t *ptl; 55443212b535SSteve Capper 55450bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 55463212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 55473212b535SSteve Capper if (svma == vma) 55483212b535SSteve Capper continue; 55493212b535SSteve Capper 55503212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 55513212b535SSteve Capper if (saddr) { 55527868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 55537868a208SPunit Agrawal vma_mmu_pagesize(svma)); 55543212b535SSteve Capper if (spte) { 55553212b535SSteve Capper get_page(virt_to_page(spte)); 55563212b535SSteve Capper break; 55573212b535SSteve Capper } 55583212b535SSteve Capper } 55593212b535SSteve Capper } 55603212b535SSteve Capper 55613212b535SSteve Capper if (!spte) 55623212b535SSteve Capper goto out; 55633212b535SSteve Capper 55648bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5565dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 55663212b535SSteve Capper pud_populate(mm, pud, 55673212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5568c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5569dc6c9a35SKirill A. Shutemov } else { 55703212b535SSteve Capper put_page(virt_to_page(spte)); 5571dc6c9a35SKirill A. Shutemov } 5572cb900f41SKirill A. Shutemov spin_unlock(ptl); 55733212b535SSteve Capper out: 55743212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 55753212b535SSteve Capper return pte; 55763212b535SSteve Capper } 55773212b535SSteve Capper 55783212b535SSteve Capper /* 55793212b535SSteve Capper * unmap huge page backed by shared pte. 55803212b535SSteve Capper * 55813212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 55823212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 55833212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 55843212b535SSteve Capper * 5585c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 55863212b535SSteve Capper * 55873212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 55883212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 55893212b535SSteve Capper */ 559034ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 559134ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 55923212b535SSteve Capper { 55933212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5594c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5595c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 55963212b535SSteve Capper 559734ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 55983212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 55993212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 56003212b535SSteve Capper return 0; 56013212b535SSteve Capper 56023212b535SSteve Capper pud_clear(pud); 56033212b535SSteve Capper put_page(virt_to_page(ptep)); 5604dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 56053212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 56063212b535SSteve Capper return 1; 56073212b535SSteve Capper } 5608c1991e07SPeter Xu 56099e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5610aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5611aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 56129e5fc74cSSteve Capper { 56139e5fc74cSSteve Capper return NULL; 56149e5fc74cSSteve Capper } 5615e81f2d22SZhang Zhen 561634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 561734ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5618e81f2d22SZhang Zhen { 5619e81f2d22SZhang Zhen return 0; 5620e81f2d22SZhang Zhen } 5621017b1660SMike Kravetz 5622017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5623017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5624017b1660SMike Kravetz { 5625017b1660SMike Kravetz } 5626c1991e07SPeter Xu 5627c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5628c1991e07SPeter Xu { 5629c1991e07SPeter Xu return false; 5630c1991e07SPeter Xu } 56313212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 56323212b535SSteve Capper 56339e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5634aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 56359e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 56369e5fc74cSSteve Capper { 56379e5fc74cSSteve Capper pgd_t *pgd; 5638c2febafcSKirill A. Shutemov p4d_t *p4d; 56399e5fc74cSSteve Capper pud_t *pud; 56409e5fc74cSSteve Capper pte_t *pte = NULL; 56419e5fc74cSSteve Capper 56429e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5643f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5644f4f0a3d8SKirill A. Shutemov if (!p4d) 5645f4f0a3d8SKirill A. Shutemov return NULL; 5646c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 56479e5fc74cSSteve Capper if (pud) { 56489e5fc74cSSteve Capper if (sz == PUD_SIZE) { 56499e5fc74cSSteve Capper pte = (pte_t *)pud; 56509e5fc74cSSteve Capper } else { 56519e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5652c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5653aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 56549e5fc74cSSteve Capper else 56559e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 56569e5fc74cSSteve Capper } 56579e5fc74cSSteve Capper } 56584e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 56599e5fc74cSSteve Capper 56609e5fc74cSSteve Capper return pte; 56619e5fc74cSSteve Capper } 56629e5fc74cSSteve Capper 56639b19df29SPunit Agrawal /* 56649b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 56659b19df29SPunit Agrawal * entry at address @addr 56669b19df29SPunit Agrawal * 56678ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 56688ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 56699b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 56709b19df29SPunit Agrawal * table. 56719b19df29SPunit Agrawal */ 56727868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 56737868a208SPunit Agrawal unsigned long addr, unsigned long sz) 56749e5fc74cSSteve Capper { 56759e5fc74cSSteve Capper pgd_t *pgd; 5676c2febafcSKirill A. Shutemov p4d_t *p4d; 56778ac0b81aSLi Xinhai pud_t *pud; 56788ac0b81aSLi Xinhai pmd_t *pmd; 56799e5fc74cSSteve Capper 56809e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5681c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5682c2febafcSKirill A. Shutemov return NULL; 5683c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5684c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5685c2febafcSKirill A. Shutemov return NULL; 56869b19df29SPunit Agrawal 5687c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 56888ac0b81aSLi Xinhai if (sz == PUD_SIZE) 56898ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 56909e5fc74cSSteve Capper return (pte_t *)pud; 56918ac0b81aSLi Xinhai if (!pud_present(*pud)) 56928ac0b81aSLi Xinhai return NULL; 56938ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 56949b19df29SPunit Agrawal 56959e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 56968ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 56979e5fc74cSSteve Capper return (pte_t *)pmd; 56989e5fc74cSSteve Capper } 56999e5fc74cSSteve Capper 570061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 570161f77edaSNaoya Horiguchi 570261f77edaSNaoya Horiguchi /* 570361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 570461f77edaSNaoya Horiguchi * behavior. 570561f77edaSNaoya Horiguchi */ 570661f77edaSNaoya Horiguchi struct page * __weak 570761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 570861f77edaSNaoya Horiguchi int write) 570961f77edaSNaoya Horiguchi { 571061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 571161f77edaSNaoya Horiguchi } 571261f77edaSNaoya Horiguchi 571361f77edaSNaoya Horiguchi struct page * __weak 57144dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 57154dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 57164dc71451SAneesh Kumar K.V { 57174dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 57184dc71451SAneesh Kumar K.V return NULL; 57194dc71451SAneesh Kumar K.V } 57204dc71451SAneesh Kumar K.V 57214dc71451SAneesh Kumar K.V struct page * __weak 57229e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5723e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 57249e5fc74cSSteve Capper { 5725e66f17ffSNaoya Horiguchi struct page *page = NULL; 5726e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5727c9d398faSNaoya Horiguchi pte_t pte; 57283faa52c0SJohn Hubbard 57293faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 57303faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 57313faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 57323faa52c0SJohn Hubbard return NULL; 57333faa52c0SJohn Hubbard 5734e66f17ffSNaoya Horiguchi retry: 5735e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5736e66f17ffSNaoya Horiguchi spin_lock(ptl); 5737e66f17ffSNaoya Horiguchi /* 5738e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5739e66f17ffSNaoya Horiguchi * unmapped from other threads. 5740e66f17ffSNaoya Horiguchi */ 5741e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5742e66f17ffSNaoya Horiguchi goto out; 5743c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5744c9d398faSNaoya Horiguchi if (pte_present(pte)) { 574597534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 57463faa52c0SJohn Hubbard /* 57473faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 57483faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 57493faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 57503faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 57513faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 57523faa52c0SJohn Hubbard * unless the page refcount overflowed: 57533faa52c0SJohn Hubbard */ 57543faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 57553faa52c0SJohn Hubbard page = NULL; 57563faa52c0SJohn Hubbard goto out; 57573faa52c0SJohn Hubbard } 5758e66f17ffSNaoya Horiguchi } else { 5759c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5760e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5761e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5762e66f17ffSNaoya Horiguchi goto retry; 5763e66f17ffSNaoya Horiguchi } 5764e66f17ffSNaoya Horiguchi /* 5765e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5766e66f17ffSNaoya Horiguchi * follow_page_mask(). 5767e66f17ffSNaoya Horiguchi */ 5768e66f17ffSNaoya Horiguchi } 5769e66f17ffSNaoya Horiguchi out: 5770e66f17ffSNaoya Horiguchi spin_unlock(ptl); 57719e5fc74cSSteve Capper return page; 57729e5fc74cSSteve Capper } 57739e5fc74cSSteve Capper 577461f77edaSNaoya Horiguchi struct page * __weak 57759e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5776e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 57779e5fc74cSSteve Capper { 57783faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5779e66f17ffSNaoya Horiguchi return NULL; 57809e5fc74cSSteve Capper 5781e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 57829e5fc74cSSteve Capper } 57839e5fc74cSSteve Capper 5784faaa5b62SAnshuman Khandual struct page * __weak 5785faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5786faaa5b62SAnshuman Khandual { 57873faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5788faaa5b62SAnshuman Khandual return NULL; 5789faaa5b62SAnshuman Khandual 5790faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5791faaa5b62SAnshuman Khandual } 5792faaa5b62SAnshuman Khandual 579331caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 579431caf665SNaoya Horiguchi { 5795bcc54222SNaoya Horiguchi bool ret = true; 5796bcc54222SNaoya Horiguchi 5797db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 57988f251a3dSMike Kravetz if (!PageHeadHuge(page) || 57998f251a3dSMike Kravetz !HPageMigratable(page) || 58000eb2df2bSMuchun Song !get_page_unless_zero(page)) { 5801bcc54222SNaoya Horiguchi ret = false; 5802bcc54222SNaoya Horiguchi goto unlock; 5803bcc54222SNaoya Horiguchi } 58048f251a3dSMike Kravetz ClearHPageMigratable(page); 580531caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5806bcc54222SNaoya Horiguchi unlock: 5807db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 5808bcc54222SNaoya Horiguchi return ret; 580931caf665SNaoya Horiguchi } 581031caf665SNaoya Horiguchi 581131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 581231caf665SNaoya Horiguchi { 5813db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 58148f251a3dSMike Kravetz SetHPageMigratable(page); 581531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 5816db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 581731caf665SNaoya Horiguchi put_page(page); 581831caf665SNaoya Horiguchi } 5819ab5ac90aSMichal Hocko 5820ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5821ab5ac90aSMichal Hocko { 5822ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5823ab5ac90aSMichal Hocko 5824ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5825ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5826ab5ac90aSMichal Hocko 5827ab5ac90aSMichal Hocko /* 5828ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5829ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5830ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5831ab5ac90aSMichal Hocko * the temporary status. 5832ab5ac90aSMichal Hocko * 5833ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5834ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5835ab5ac90aSMichal Hocko * the per-node's. 5836ab5ac90aSMichal Hocko */ 58379157c311SMike Kravetz if (HPageTemporary(newpage)) { 5838ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5839ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5840ab5ac90aSMichal Hocko 58419157c311SMike Kravetz SetHPageTemporary(oldpage); 58429157c311SMike Kravetz ClearHPageTemporary(newpage); 5843ab5ac90aSMichal Hocko 58445af1ab1dSMiaohe Lin /* 58455af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 58465af1ab1dSMiaohe Lin * when we do not cross the node. 58475af1ab1dSMiaohe Lin */ 58485af1ab1dSMiaohe Lin if (new_nid == old_nid) 58495af1ab1dSMiaohe Lin return; 5850db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 5851ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5852ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5853ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5854ab5ac90aSMichal Hocko } 5855db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 5856ab5ac90aSMichal Hocko } 5857ab5ac90aSMichal Hocko } 5858cf11e85fSRoman Gushchin 58596dfeaff9SPeter Xu /* 58606dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 58616dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 58626dfeaff9SPeter Xu */ 58636dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 58646dfeaff9SPeter Xu { 58656dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 58666dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 58676dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 58686dfeaff9SPeter Xu struct mmu_notifier_range range; 58696dfeaff9SPeter Xu unsigned long address, start, end; 58706dfeaff9SPeter Xu spinlock_t *ptl; 58716dfeaff9SPeter Xu pte_t *ptep; 58726dfeaff9SPeter Xu 58736dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 58746dfeaff9SPeter Xu return; 58756dfeaff9SPeter Xu 58766dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 58776dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 58786dfeaff9SPeter Xu 58796dfeaff9SPeter Xu if (start >= end) 58806dfeaff9SPeter Xu return; 58816dfeaff9SPeter Xu 58826dfeaff9SPeter Xu /* 58836dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 58846dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 58856dfeaff9SPeter Xu */ 58866dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 58876dfeaff9SPeter Xu start, end); 58886dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 58896dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 58906dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 58916dfeaff9SPeter Xu unsigned long tmp = address; 58926dfeaff9SPeter Xu 58936dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 58946dfeaff9SPeter Xu if (!ptep) 58956dfeaff9SPeter Xu continue; 58966dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 58976dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 58986dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 58996dfeaff9SPeter Xu spin_unlock(ptl); 59006dfeaff9SPeter Xu } 59016dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 59026dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 59036dfeaff9SPeter Xu /* 59046dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 59056dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 59066dfeaff9SPeter Xu */ 59076dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 59086dfeaff9SPeter Xu } 59096dfeaff9SPeter Xu 5910cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5911cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5912cf11e85fSRoman Gushchin 5913cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5914cf11e85fSRoman Gushchin { 5915cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5916cf11e85fSRoman Gushchin return 0; 5917cf11e85fSRoman Gushchin } 5918cf11e85fSRoman Gushchin 5919cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5920cf11e85fSRoman Gushchin 5921cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5922cf11e85fSRoman Gushchin { 5923cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5924cf11e85fSRoman Gushchin int nid; 5925cf11e85fSRoman Gushchin 5926cf11e85fSRoman Gushchin cma_reserve_called = true; 5927cf11e85fSRoman Gushchin 5928cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5929cf11e85fSRoman Gushchin return; 5930cf11e85fSRoman Gushchin 5931cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5932cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5933cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5934cf11e85fSRoman Gushchin return; 5935cf11e85fSRoman Gushchin } 5936cf11e85fSRoman Gushchin 5937cf11e85fSRoman Gushchin /* 5938cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5939cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5940cf11e85fSRoman Gushchin */ 5941cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5942cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5943cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5944cf11e85fSRoman Gushchin 5945cf11e85fSRoman Gushchin reserved = 0; 5946cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5947cf11e85fSRoman Gushchin int res; 59482281f797SBarry Song char name[CMA_MAX_NAME]; 5949cf11e85fSRoman Gushchin 5950cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5951cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5952cf11e85fSRoman Gushchin 59532281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5954cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 595529d0f41dSBarry Song 0, false, name, 5956cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5957cf11e85fSRoman Gushchin if (res) { 5958cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5959cf11e85fSRoman Gushchin res, nid); 5960cf11e85fSRoman Gushchin continue; 5961cf11e85fSRoman Gushchin } 5962cf11e85fSRoman Gushchin 5963cf11e85fSRoman Gushchin reserved += size; 5964cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5965cf11e85fSRoman Gushchin size / SZ_1M, nid); 5966cf11e85fSRoman Gushchin 5967cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5968cf11e85fSRoman Gushchin break; 5969cf11e85fSRoman Gushchin } 5970cf11e85fSRoman Gushchin } 5971cf11e85fSRoman Gushchin 5972cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5973cf11e85fSRoman Gushchin { 5974cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5975cf11e85fSRoman Gushchin return; 5976cf11e85fSRoman Gushchin 5977cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5978cf11e85fSRoman Gushchin } 5979cf11e85fSRoman Gushchin 5980cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5981