1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 33d6606683SLinus Torvalds 3463551ae0SDavid Gibson #include <asm/page.h> 35ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3624669e58SAneesh Kumar K.V #include <asm/tlb.h> 3763551ae0SDavid Gibson 3824669e58SAneesh Kumar K.V #include <linux/io.h> 3963551ae0SDavid Gibson #include <linux/hugetlb.h> 409dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 419a305230SLee Schermerhorn #include <linux/node.h> 42ab5ac90aSMichal Hocko #include <linux/page_owner.h> 437835e98bSNick Piggin #include "internal.h" 44f41f2ed4SMuchun Song #include "hugetlb_vmemmap.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 470f0953a1bSIngo Molnar * for all the existing 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; 10831a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1084bf50bab2SNaoya Horiguchi 10859487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1086bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 10878e3560d9SPavel Tatashin if (pin && !is_pinnable_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 1379ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1380ad2fa371SMuchun Song bool adjust_surplus) 1381ad2fa371SMuchun Song { 1382ad2fa371SMuchun Song int zeroed; 1383ad2fa371SMuchun Song int nid = page_to_nid(page); 1384ad2fa371SMuchun Song 1385ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1386ad2fa371SMuchun Song 1387ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1388ad2fa371SMuchun Song 1389ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1390ad2fa371SMuchun Song h->nr_huge_pages++; 1391ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1392ad2fa371SMuchun Song 1393ad2fa371SMuchun Song if (adjust_surplus) { 1394ad2fa371SMuchun Song h->surplus_huge_pages++; 1395ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1396ad2fa371SMuchun Song } 1397ad2fa371SMuchun Song 1398ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1399ad2fa371SMuchun Song set_page_private(page, 0); 1400ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1401ad2fa371SMuchun Song 1402ad2fa371SMuchun Song /* 1403ad2fa371SMuchun Song * This page is now managed by the hugetlb allocator and has 1404ad2fa371SMuchun Song * no users -- drop the last reference. 1405ad2fa371SMuchun Song */ 1406ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1407ad2fa371SMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 1408ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1409ad2fa371SMuchun Song enqueue_huge_page(h, page); 1410ad2fa371SMuchun Song } 1411ad2fa371SMuchun Song 1412b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 14136af2acb6SAdam Litke { 14146af2acb6SAdam Litke int i; 1415dbfee5aeSMike Kravetz struct page *subpage = page; 1416a5516438SAndi Kleen 14174eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1418944d9fecSLuiz Capitulino return; 141918229df5SAndy Whitcroft 1420ad2fa371SMuchun Song if (alloc_huge_page_vmemmap(h, page)) { 1421ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1422ad2fa371SMuchun Song /* 1423ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1424ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1425ad2fa371SMuchun Song * as a surplus page. 1426ad2fa371SMuchun Song */ 1427ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1428ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1429ad2fa371SMuchun Song return; 1430ad2fa371SMuchun Song } 1431ad2fa371SMuchun Song 1432dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1433dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1434dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 143532f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1436a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1437a7407a27SLuiz Capitulino 1 << PG_writeback); 14386af2acb6SAdam Litke } 1439944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1440944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1441944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1442944d9fecSLuiz Capitulino } else { 1443a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 14446af2acb6SAdam Litke } 1445944d9fecSLuiz Capitulino } 14466af2acb6SAdam Litke 1447b65d4adbSMuchun Song /* 1448b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1449b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1450b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1451b65d4adbSMuchun Song * the vmemmap pages. 1452b65d4adbSMuchun Song * 1453b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1454b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1455b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1456b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1457b65d4adbSMuchun Song */ 1458b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1459b65d4adbSMuchun Song 1460b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1461b65d4adbSMuchun Song { 1462b65d4adbSMuchun Song struct llist_node *node; 1463b65d4adbSMuchun Song 1464b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1465b65d4adbSMuchun Song 1466b65d4adbSMuchun Song while (node) { 1467b65d4adbSMuchun Song struct page *page; 1468b65d4adbSMuchun Song struct hstate *h; 1469b65d4adbSMuchun Song 1470b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1471b65d4adbSMuchun Song struct page, mapping); 1472b65d4adbSMuchun Song node = node->next; 1473b65d4adbSMuchun Song page->mapping = NULL; 1474b65d4adbSMuchun Song /* 1475b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1476b65d4adbSMuchun Song * is going to trigger because a previous call to 1477b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1478b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1479b65d4adbSMuchun Song */ 1480b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1481b65d4adbSMuchun Song 1482b65d4adbSMuchun Song __update_and_free_page(h, page); 1483b65d4adbSMuchun Song 1484b65d4adbSMuchun Song cond_resched(); 1485b65d4adbSMuchun Song } 1486b65d4adbSMuchun Song } 1487b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1488b65d4adbSMuchun Song 1489b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1490b65d4adbSMuchun Song { 1491b65d4adbSMuchun Song if (free_vmemmap_pages_per_hpage(h)) 1492b65d4adbSMuchun Song flush_work(&free_hpage_work); 1493b65d4adbSMuchun Song } 1494b65d4adbSMuchun Song 1495b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1496b65d4adbSMuchun Song bool atomic) 1497b65d4adbSMuchun Song { 1498ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1499b65d4adbSMuchun Song __update_and_free_page(h, page); 1500b65d4adbSMuchun Song return; 1501b65d4adbSMuchun Song } 1502b65d4adbSMuchun Song 1503b65d4adbSMuchun Song /* 1504b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1505b65d4adbSMuchun Song * 1506b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1507b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1508b65d4adbSMuchun Song * hasn't retrieved the list yet. 1509b65d4adbSMuchun Song */ 1510b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1511b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1512b65d4adbSMuchun Song } 1513b65d4adbSMuchun Song 151410c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 151510c6ec49SMike Kravetz { 151610c6ec49SMike Kravetz struct page *page, *t_page; 151710c6ec49SMike Kravetz 151810c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1519b65d4adbSMuchun Song update_and_free_page(h, page, false); 152010c6ec49SMike Kravetz cond_resched(); 152110c6ec49SMike Kravetz } 152210c6ec49SMike Kravetz } 152310c6ec49SMike Kravetz 1524e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1525e5ff2159SAndi Kleen { 1526e5ff2159SAndi Kleen struct hstate *h; 1527e5ff2159SAndi Kleen 1528e5ff2159SAndi Kleen for_each_hstate(h) { 1529e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1530e5ff2159SAndi Kleen return h; 1531e5ff2159SAndi Kleen } 1532e5ff2159SAndi Kleen return NULL; 1533e5ff2159SAndi Kleen } 1534e5ff2159SAndi Kleen 1535db71ef79SMike Kravetz void free_huge_page(struct page *page) 153627a85ef1SDavid Gibson { 1537a5516438SAndi Kleen /* 1538a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1539a5516438SAndi Kleen * compound page destructor. 1540a5516438SAndi Kleen */ 1541e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 15427893d1d5SAdam Litke int nid = page_to_nid(page); 1543d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 154407443a85SJoonsoo Kim bool restore_reserve; 1545db71ef79SMike Kravetz unsigned long flags; 154627a85ef1SDavid Gibson 1547b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1548b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 15498ace22bcSYongkai Wu 1550d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 15518ace22bcSYongkai Wu page->mapping = NULL; 1552d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1553d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 155427a85ef1SDavid Gibson 15551c5ecae3SMike Kravetz /* 1556d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 15570919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 15580919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 15590919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 15606c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 15610919e1b6SMike Kravetz * remove the reserved page from the subpool. 15620919e1b6SMike Kravetz */ 15630919e1b6SMike Kravetz if (!restore_reserve) { 15640919e1b6SMike Kravetz /* 15650919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 15660919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 15670919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 15680919e1b6SMike Kravetz * operation. 15691c5ecae3SMike Kravetz */ 15701c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 15711c5ecae3SMike Kravetz restore_reserve = true; 15720919e1b6SMike Kravetz } 15731c5ecae3SMike Kravetz 1574db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 15758f251a3dSMike Kravetz ClearHPageMigratable(page); 15766d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 15776d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 157808cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 157908cf9fafSMina Almasry pages_per_huge_page(h), page); 158007443a85SJoonsoo Kim if (restore_reserve) 158107443a85SJoonsoo Kim h->resv_huge_pages++; 158207443a85SJoonsoo Kim 15839157c311SMike Kravetz if (HPageTemporary(page)) { 15846eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1585db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1586b65d4adbSMuchun Song update_and_free_page(h, page, true); 1587ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 15880edaecfaSAneesh Kumar K.V /* remove the page from active list */ 15896eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1590db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1591b65d4adbSMuchun Song update_and_free_page(h, page, true); 15927893d1d5SAdam Litke } else { 15935d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1594a5516438SAndi Kleen enqueue_huge_page(h, page); 1595db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 159627a85ef1SDavid Gibson } 15971121828aSMike Kravetz } 159827a85ef1SDavid Gibson 1599d3d99fccSOscar Salvador /* 1600d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1601d3d99fccSOscar Salvador */ 1602d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1603d3d99fccSOscar Salvador { 1604d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1605d3d99fccSOscar Salvador h->nr_huge_pages++; 1606d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1607d3d99fccSOscar Salvador } 1608d3d99fccSOscar Salvador 1609f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1610b7ba30c6SAndi Kleen { 1611f41f2ed4SMuchun Song free_huge_page_vmemmap(h, page); 16120edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1613f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1614ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 16159dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 16161adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1617d3d99fccSOscar Salvador } 1618d3d99fccSOscar Salvador 1619d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1620d3d99fccSOscar Salvador { 1621f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1622db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1623d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1624db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1625b7ba30c6SAndi Kleen } 1626b7ba30c6SAndi Kleen 1627d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 162820a0307cSWu Fengguang { 162920a0307cSWu Fengguang int i; 163020a0307cSWu Fengguang int nr_pages = 1 << order; 163120a0307cSWu Fengguang struct page *p = page + 1; 163220a0307cSWu Fengguang 163320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 163420a0307cSWu Fengguang set_compound_order(page, order); 1635ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1636de09d31dSKirill A. Shutemov __SetPageHead(page); 163720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1638ef5a22beSAndrea Arcangeli /* 1639ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1640ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1641ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 16427c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1643ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1644ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1645ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1646ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1647ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1648ef5a22beSAndrea Arcangeli * after get_user_pages(). 1649ef5a22beSAndrea Arcangeli */ 1650ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 165158a84aa9SYouquan Song set_page_count(p, 0); 16521d798ca3SKirill A. Shutemov set_compound_head(p, page); 165320a0307cSWu Fengguang } 1654b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 165547e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 165620a0307cSWu Fengguang } 165720a0307cSWu Fengguang 16587795912cSAndrew Morton /* 16597795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 16607795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 16617795912cSAndrew Morton * details. 16627795912cSAndrew Morton */ 166320a0307cSWu Fengguang int PageHuge(struct page *page) 166420a0307cSWu Fengguang { 166520a0307cSWu Fengguang if (!PageCompound(page)) 166620a0307cSWu Fengguang return 0; 166720a0307cSWu Fengguang 166820a0307cSWu Fengguang page = compound_head(page); 1669f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 167020a0307cSWu Fengguang } 167143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 167243131e14SNaoya Horiguchi 167327c73ae7SAndrea Arcangeli /* 167427c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 167527c73ae7SAndrea Arcangeli * normal or transparent huge pages. 167627c73ae7SAndrea Arcangeli */ 167727c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 167827c73ae7SAndrea Arcangeli { 167927c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 168027c73ae7SAndrea Arcangeli return 0; 168127c73ae7SAndrea Arcangeli 1682d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 168327c73ae7SAndrea Arcangeli } 168427c73ae7SAndrea Arcangeli 1685c0d0381aSMike Kravetz /* 1686c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1687c0d0381aSMike Kravetz * 1688336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1689336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1690336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1691c0d0381aSMike Kravetz */ 1692c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1693c0d0381aSMike Kravetz { 1694336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1695c0d0381aSMike Kravetz 1696c0d0381aSMike Kravetz if (!mapping) 1697c0d0381aSMike Kravetz return mapping; 1698c0d0381aSMike Kravetz 1699c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1700c0d0381aSMike Kravetz return mapping; 1701c0d0381aSMike Kravetz 1702c0d0381aSMike Kravetz return NULL; 1703c0d0381aSMike Kravetz } 1704c0d0381aSMike Kravetz 1705fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 170613d60f4bSZhang Yi { 170713d60f4bSZhang Yi struct page *page_head = compound_head(page); 170813d60f4bSZhang Yi pgoff_t index = page_index(page_head); 170913d60f4bSZhang Yi unsigned long compound_idx; 171013d60f4bSZhang Yi 171113d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 171213d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 171313d60f4bSZhang Yi else 171413d60f4bSZhang Yi compound_idx = page - page_head; 171513d60f4bSZhang Yi 171613d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 171713d60f4bSZhang Yi } 171813d60f4bSZhang Yi 17190c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1720f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1721f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17221da177e4SLinus Torvalds { 1723af0fb9dfSMichal Hocko int order = huge_page_order(h); 17241da177e4SLinus Torvalds struct page *page; 1725f60858f9SMike Kravetz bool alloc_try_hard = true; 1726f96efd58SJoe Jin 1727f60858f9SMike Kravetz /* 1728f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1729f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1730f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1731f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1732f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1733f60858f9SMike Kravetz */ 1734f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1735f60858f9SMike Kravetz alloc_try_hard = false; 1736f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1737f60858f9SMike Kravetz if (alloc_try_hard) 1738f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1739af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1740af0fb9dfSMichal Hocko nid = numa_mem_id(); 174184172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1742af0fb9dfSMichal Hocko if (page) 1743af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1744af0fb9dfSMichal Hocko else 1745af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 174663b4613cSNishanth Aravamudan 1747f60858f9SMike Kravetz /* 1748f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1749f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1750f60858f9SMike Kravetz * normal 'try hard' allocations. 1751f60858f9SMike Kravetz */ 1752f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1753f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1754f60858f9SMike Kravetz 1755f60858f9SMike Kravetz /* 1756f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1757f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1758f60858f9SMike Kravetz * overall state change. 1759f60858f9SMike Kravetz */ 1760f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1761f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1762f60858f9SMike Kravetz 176363b4613cSNishanth Aravamudan return page; 176463b4613cSNishanth Aravamudan } 176563b4613cSNishanth Aravamudan 1766af0fb9dfSMichal Hocko /* 17670c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 17680c397daeSMichal Hocko * should use this function to get new hugetlb pages 17690c397daeSMichal Hocko */ 17700c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1771f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1772f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17730c397daeSMichal Hocko { 17740c397daeSMichal Hocko struct page *page; 17750c397daeSMichal Hocko 17760c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17770c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 17780c397daeSMichal Hocko else 17790c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1780f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 17810c397daeSMichal Hocko if (!page) 17820c397daeSMichal Hocko return NULL; 17830c397daeSMichal Hocko 17840c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17850c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 17860c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 17870c397daeSMichal Hocko 17880c397daeSMichal Hocko return page; 17890c397daeSMichal Hocko } 17900c397daeSMichal Hocko 17910c397daeSMichal Hocko /* 1792af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1793af0fb9dfSMichal Hocko * manner. 1794af0fb9dfSMichal Hocko */ 1795f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1796f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1797b2261026SJoonsoo Kim { 1798b2261026SJoonsoo Kim struct page *page; 1799b2261026SJoonsoo Kim int nr_nodes, node; 1800af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1801b2261026SJoonsoo Kim 1802b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1803f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1804f60858f9SMike Kravetz node_alloc_noretry); 1805af0fb9dfSMichal Hocko if (page) 1806b2261026SJoonsoo Kim break; 1807b2261026SJoonsoo Kim } 1808b2261026SJoonsoo Kim 1809af0fb9dfSMichal Hocko if (!page) 1810af0fb9dfSMichal Hocko return 0; 1811b2261026SJoonsoo Kim 1812af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1813af0fb9dfSMichal Hocko 1814af0fb9dfSMichal Hocko return 1; 1815b2261026SJoonsoo Kim } 1816b2261026SJoonsoo Kim 1817e8c5c824SLee Schermerhorn /* 181810c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 181910c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 182010c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 182110c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 1822e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1823e8c5c824SLee Schermerhorn */ 182410c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 182510c6ec49SMike Kravetz nodemask_t *nodes_allowed, 18266ae11b27SLee Schermerhorn bool acct_surplus) 1827e8c5c824SLee Schermerhorn { 1828b2261026SJoonsoo Kim int nr_nodes, node; 182910c6ec49SMike Kravetz struct page *page = NULL; 1830e8c5c824SLee Schermerhorn 18319487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1832b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1833685f3457SLee Schermerhorn /* 1834685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1835685f3457SLee Schermerhorn * nodes with surplus pages. 1836685f3457SLee Schermerhorn */ 1837b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1838b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 183910c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 1840e8c5c824SLee Schermerhorn struct page, lru); 18416eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 18429a76db09SLee Schermerhorn break; 1843e8c5c824SLee Schermerhorn } 1844b2261026SJoonsoo Kim } 1845e8c5c824SLee Schermerhorn 184610c6ec49SMike Kravetz return page; 1847e8c5c824SLee Schermerhorn } 1848e8c5c824SLee Schermerhorn 1849c8721bbbSNaoya Horiguchi /* 1850c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1851faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1852faf53defSNaoya Horiguchi * This function returns values like below: 1853faf53defSNaoya Horiguchi * 1854ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 1855ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 1856ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 1857ad2fa371SMuchun Song * is enabled. 1858faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1859faf53defSNaoya Horiguchi * (allocated or reserved.) 1860faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1861faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1862c8721bbbSNaoya Horiguchi */ 1863c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1864c8721bbbSNaoya Horiguchi { 18656bc9b564SNaoya Horiguchi int rc = -EBUSY; 1866082d5b6bSGerald Schaefer 18677ffddd49SMuchun Song retry: 1868faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1869faf53defSNaoya Horiguchi if (!PageHuge(page)) 1870faf53defSNaoya Horiguchi return 0; 1871faf53defSNaoya Horiguchi 1872db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1873faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1874faf53defSNaoya Horiguchi rc = 0; 1875faf53defSNaoya Horiguchi goto out; 1876faf53defSNaoya Horiguchi } 1877faf53defSNaoya Horiguchi 1878faf53defSNaoya Horiguchi if (!page_count(page)) { 18792247bb33SGerald Schaefer struct page *head = compound_head(page); 18802247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 18816bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1882082d5b6bSGerald Schaefer goto out; 18837ffddd49SMuchun Song 18847ffddd49SMuchun Song /* 18857ffddd49SMuchun Song * We should make sure that the page is already on the free list 18867ffddd49SMuchun Song * when it is dissolved. 18877ffddd49SMuchun Song */ 18886c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 1889db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18907ffddd49SMuchun Song cond_resched(); 18917ffddd49SMuchun Song 18927ffddd49SMuchun Song /* 18937ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 18947ffddd49SMuchun Song * encounter this race. In fact, we have a chance 18957ffddd49SMuchun Song * to successfully dissolve the page if we do a 18967ffddd49SMuchun Song * retry. Because the race window is quite small. 18977ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 18987ffddd49SMuchun Song * for increasing the success rate of dissolving page. 18997ffddd49SMuchun Song */ 19007ffddd49SMuchun Song goto retry; 19017ffddd49SMuchun Song } 19027ffddd49SMuchun Song 1903ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 1904ad2fa371SMuchun Song h->max_huge_pages--; 1905ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1906ad2fa371SMuchun Song 1907c3114a84SAnshuman Khandual /* 1908ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 1909ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 1910ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 1911ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 1912ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 1913ad2fa371SMuchun Song * appropriate action on failure. 1914ad2fa371SMuchun Song */ 1915ad2fa371SMuchun Song rc = alloc_huge_page_vmemmap(h, head); 1916ad2fa371SMuchun Song if (!rc) { 1917ad2fa371SMuchun Song /* 1918ad2fa371SMuchun Song * Move PageHWPoison flag from head page to the raw 1919ad2fa371SMuchun Song * error page, which makes any subpages rather than 1920ad2fa371SMuchun Song * the error page reusable. 1921c3114a84SAnshuman Khandual */ 1922c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1923c3114a84SAnshuman Khandual SetPageHWPoison(page); 1924c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1925c3114a84SAnshuman Khandual } 1926b65d4adbSMuchun Song update_and_free_page(h, head, false); 1927ad2fa371SMuchun Song } else { 1928ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1929ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 1930ad2fa371SMuchun Song h->max_huge_pages++; 1931ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1932ad2fa371SMuchun Song } 1933ad2fa371SMuchun Song 1934ad2fa371SMuchun Song return rc; 1935c8721bbbSNaoya Horiguchi } 1936082d5b6bSGerald Schaefer out: 1937db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1938082d5b6bSGerald Schaefer return rc; 1939c8721bbbSNaoya Horiguchi } 1940c8721bbbSNaoya Horiguchi 1941c8721bbbSNaoya Horiguchi /* 1942c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1943c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 19442247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 19452247bb33SGerald Schaefer * part of it lies within the given range. 1946082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1947082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1948c8721bbbSNaoya Horiguchi */ 1949082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1950c8721bbbSNaoya Horiguchi { 1951c8721bbbSNaoya Horiguchi unsigned long pfn; 1952eb03aa00SGerald Schaefer struct page *page; 1953082d5b6bSGerald Schaefer int rc = 0; 1954c8721bbbSNaoya Horiguchi 1955d0177639SLi Zhong if (!hugepages_supported()) 1956082d5b6bSGerald Schaefer return rc; 1957d0177639SLi Zhong 1958eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1959eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1960eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1961eb03aa00SGerald Schaefer if (rc) 1962082d5b6bSGerald Schaefer break; 1963eb03aa00SGerald Schaefer } 1964082d5b6bSGerald Schaefer 1965082d5b6bSGerald Schaefer return rc; 1966c8721bbbSNaoya Horiguchi } 1967c8721bbbSNaoya Horiguchi 1968ab5ac90aSMichal Hocko /* 1969ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1970ab5ac90aSMichal Hocko */ 19710c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1972aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 19737893d1d5SAdam Litke { 19749980d744SMichal Hocko struct page *page = NULL; 19757893d1d5SAdam Litke 1976bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1977aa888a74SAndi Kleen return NULL; 1978aa888a74SAndi Kleen 1979db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 19809980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 19819980d744SMichal Hocko goto out_unlock; 1982db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1983d1c3fb1fSNishanth Aravamudan 1984f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 19859980d744SMichal Hocko if (!page) 19860c397daeSMichal Hocko return NULL; 1987d1c3fb1fSNishanth Aravamudan 1988db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 19899980d744SMichal Hocko /* 19909980d744SMichal Hocko * We could have raced with the pool size change. 19919980d744SMichal Hocko * Double check that and simply deallocate the new page 19929980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 19939980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 19949980d744SMichal Hocko * codeflow 19959980d744SMichal Hocko */ 19969980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 19979157c311SMike Kravetz SetHPageTemporary(page); 1998db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19999980d744SMichal Hocko put_page(page); 20002bf753e6SKai Shen return NULL; 20019980d744SMichal Hocko } else { 20029980d744SMichal Hocko h->surplus_huge_pages++; 20034704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 20047893d1d5SAdam Litke } 20059980d744SMichal Hocko 20069980d744SMichal Hocko out_unlock: 2007db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20087893d1d5SAdam Litke 20097893d1d5SAdam Litke return page; 20107893d1d5SAdam Litke } 20117893d1d5SAdam Litke 2012bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2013ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2014ab5ac90aSMichal Hocko { 2015ab5ac90aSMichal Hocko struct page *page; 2016ab5ac90aSMichal Hocko 2017ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2018ab5ac90aSMichal Hocko return NULL; 2019ab5ac90aSMichal Hocko 2020f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2021ab5ac90aSMichal Hocko if (!page) 2022ab5ac90aSMichal Hocko return NULL; 2023ab5ac90aSMichal Hocko 2024ab5ac90aSMichal Hocko /* 2025ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2026ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2027ab5ac90aSMichal Hocko */ 20289157c311SMike Kravetz SetHPageTemporary(page); 2029ab5ac90aSMichal Hocko 2030ab5ac90aSMichal Hocko return page; 2031ab5ac90aSMichal Hocko } 2032ab5ac90aSMichal Hocko 2033e4e574b7SAdam Litke /* 2034099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2035099730d6SDave Hansen */ 2036e0ec90eeSDave Hansen static 20370c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2038099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2039099730d6SDave Hansen { 2040aaf14e40SMichal Hocko struct page *page; 2041aaf14e40SMichal Hocko struct mempolicy *mpol; 2042aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2043aaf14e40SMichal Hocko int nid; 2044aaf14e40SMichal Hocko nodemask_t *nodemask; 2045aaf14e40SMichal Hocko 2046aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 20470c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 2048aaf14e40SMichal Hocko mpol_cond_put(mpol); 2049aaf14e40SMichal Hocko 2050aaf14e40SMichal Hocko return page; 2051099730d6SDave Hansen } 2052099730d6SDave Hansen 2053ab5ac90aSMichal Hocko /* page migration callback function */ 20543e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2055d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 20564db9b2efSMichal Hocko { 2057db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 20584db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 20593e59fcb0SMichal Hocko struct page *page; 20603e59fcb0SMichal Hocko 20613e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 20623e59fcb0SMichal Hocko if (page) { 2063db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20643e59fcb0SMichal Hocko return page; 20654db9b2efSMichal Hocko } 20664db9b2efSMichal Hocko } 2067db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20684db9b2efSMichal Hocko 20690c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 20704db9b2efSMichal Hocko } 20714db9b2efSMichal Hocko 2072ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2073389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2074389c8178SMichal Hocko unsigned long address) 2075ebd63723SMichal Hocko { 2076ebd63723SMichal Hocko struct mempolicy *mpol; 2077ebd63723SMichal Hocko nodemask_t *nodemask; 2078ebd63723SMichal Hocko struct page *page; 2079ebd63723SMichal Hocko gfp_t gfp_mask; 2080ebd63723SMichal Hocko int node; 2081ebd63723SMichal Hocko 2082ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2083ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2084d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2085ebd63723SMichal Hocko mpol_cond_put(mpol); 2086ebd63723SMichal Hocko 2087ebd63723SMichal Hocko return page; 2088ebd63723SMichal Hocko } 2089ebd63723SMichal Hocko 2090bf50bab2SNaoya Horiguchi /* 209125985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2092e4e574b7SAdam Litke * of size 'delta'. 2093e4e574b7SAdam Litke */ 20940a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 20951b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2096e4e574b7SAdam Litke { 2097e4e574b7SAdam Litke struct list_head surplus_list; 2098e4e574b7SAdam Litke struct page *page, *tmp; 20990a4f3d1bSLiu Xiang int ret; 21000a4f3d1bSLiu Xiang long i; 21010a4f3d1bSLiu Xiang long needed, allocated; 210228073b02SHillf Danton bool alloc_ok = true; 2103e4e574b7SAdam Litke 21049487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2105a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2106ac09b3a1SAdam Litke if (needed <= 0) { 2107a5516438SAndi Kleen h->resv_huge_pages += delta; 2108e4e574b7SAdam Litke return 0; 2109ac09b3a1SAdam Litke } 2110e4e574b7SAdam Litke 2111e4e574b7SAdam Litke allocated = 0; 2112e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2113e4e574b7SAdam Litke 2114e4e574b7SAdam Litke ret = -ENOMEM; 2115e4e574b7SAdam Litke retry: 2116db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2117e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 21180c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2119aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 212028073b02SHillf Danton if (!page) { 212128073b02SHillf Danton alloc_ok = false; 212228073b02SHillf Danton break; 212328073b02SHillf Danton } 2124e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 212569ed779aSDavid Rientjes cond_resched(); 2126e4e574b7SAdam Litke } 212728073b02SHillf Danton allocated += i; 2128e4e574b7SAdam Litke 2129e4e574b7SAdam Litke /* 2130e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2131e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2132e4e574b7SAdam Litke */ 2133db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2134a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2135a5516438SAndi Kleen (h->free_huge_pages + allocated); 213628073b02SHillf Danton if (needed > 0) { 213728073b02SHillf Danton if (alloc_ok) 2138e4e574b7SAdam Litke goto retry; 213928073b02SHillf Danton /* 214028073b02SHillf Danton * We were not able to allocate enough pages to 214128073b02SHillf Danton * satisfy the entire reservation so we free what 214228073b02SHillf Danton * we've allocated so far. 214328073b02SHillf Danton */ 214428073b02SHillf Danton goto free; 214528073b02SHillf Danton } 2146e4e574b7SAdam Litke /* 2147e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 214825985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2149e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2150ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2151ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2152ac09b3a1SAdam Litke * before they are reserved. 2153e4e574b7SAdam Litke */ 2154e4e574b7SAdam Litke needed += allocated; 2155a5516438SAndi Kleen h->resv_huge_pages += delta; 2156e4e574b7SAdam Litke ret = 0; 2157a9869b83SNaoya Horiguchi 215819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 215919fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2160e558464bSMuchun Song int zeroed; 2161e558464bSMuchun Song 216219fc3f0aSAdam Litke if ((--needed) < 0) 216319fc3f0aSAdam Litke break; 2164a9869b83SNaoya Horiguchi /* 2165a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2166a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2167a9869b83SNaoya Horiguchi */ 2168e558464bSMuchun Song zeroed = put_page_testzero(page); 2169e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2170a5516438SAndi Kleen enqueue_huge_page(h, page); 217119fc3f0aSAdam Litke } 217228073b02SHillf Danton free: 2173db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 217419fc3f0aSAdam Litke 217519fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2176c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2177a9869b83SNaoya Horiguchi put_page(page); 2178db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2179e4e574b7SAdam Litke 2180e4e574b7SAdam Litke return ret; 2181e4e574b7SAdam Litke } 2182e4e574b7SAdam Litke 2183e4e574b7SAdam Litke /* 2184e5bbc8a6SMike Kravetz * This routine has two main purposes: 2185e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2186e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2187e5bbc8a6SMike Kravetz * to the associated reservation map. 2188e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2189e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2190e4e574b7SAdam Litke */ 2191a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2192a5516438SAndi Kleen unsigned long unused_resv_pages) 2193e4e574b7SAdam Litke { 2194e4e574b7SAdam Litke unsigned long nr_pages; 219510c6ec49SMike Kravetz struct page *page; 219610c6ec49SMike Kravetz LIST_HEAD(page_list); 219710c6ec49SMike Kravetz 21989487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 219910c6ec49SMike Kravetz /* Uncommit the reservation */ 220010c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2201e4e574b7SAdam Litke 2202aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2203bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2204e5bbc8a6SMike Kravetz goto out; 2205aa888a74SAndi Kleen 2206e5bbc8a6SMike Kravetz /* 2207e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2208e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2209e5bbc8a6SMike Kravetz */ 2210a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2211e4e574b7SAdam Litke 2212685f3457SLee Schermerhorn /* 2213685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 22149b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 22159b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 22169b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 221710c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 22189b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2219685f3457SLee Schermerhorn */ 2220685f3457SLee Schermerhorn while (nr_pages--) { 222110c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 222210c6ec49SMike Kravetz if (!page) 2223e5bbc8a6SMike Kravetz goto out; 222410c6ec49SMike Kravetz 222510c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2226e4e574b7SAdam Litke } 2227e5bbc8a6SMike Kravetz 2228e5bbc8a6SMike Kravetz out: 2229db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 223010c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2231db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2232e4e574b7SAdam Litke } 2233e4e574b7SAdam Litke 22345e911373SMike Kravetz 2235c37f9fb1SAndy Whitcroft /* 2236feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 22375e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2238cf3ad20bSMike Kravetz * 2239cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2240cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2241cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2242cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2243cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2244feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2245feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2246feba16e2SMike Kravetz * is called in such cases. 2247cf3ad20bSMike Kravetz * 2248cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2249cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2250cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2251cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2252cf3ad20bSMike Kravetz * take appropriate action. 225396b96a96SMike Kravetz * 225496b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 225596b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 225696b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 225796b96a96SMike Kravetz * reservation exists. 2258846be085SMike Kravetz * 2259846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2260846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2261846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2262846be085SMike Kravetz * exists. 2263c37f9fb1SAndy Whitcroft */ 22645e911373SMike Kravetz enum vma_resv_mode { 22655e911373SMike Kravetz VMA_NEEDS_RESV, 22665e911373SMike Kravetz VMA_COMMIT_RESV, 2267feba16e2SMike Kravetz VMA_END_RESV, 226896b96a96SMike Kravetz VMA_ADD_RESV, 2269846be085SMike Kravetz VMA_DEL_RESV, 22705e911373SMike Kravetz }; 2271cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2272cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 22735e911373SMike Kravetz enum vma_resv_mode mode) 2274c37f9fb1SAndy Whitcroft { 22754e35f483SJoonsoo Kim struct resv_map *resv; 22764e35f483SJoonsoo Kim pgoff_t idx; 2277cf3ad20bSMike Kravetz long ret; 22780db9d74eSMina Almasry long dummy_out_regions_needed; 2279c37f9fb1SAndy Whitcroft 22804e35f483SJoonsoo Kim resv = vma_resv_map(vma); 22814e35f483SJoonsoo Kim if (!resv) 2282c37f9fb1SAndy Whitcroft return 1; 2283c37f9fb1SAndy Whitcroft 22844e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 22855e911373SMike Kravetz switch (mode) { 22865e911373SMike Kravetz case VMA_NEEDS_RESV: 22870db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 22880db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 22890db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 22900db9d74eSMina Almasry * ever require 1 region. 22910db9d74eSMina Almasry */ 22920db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 22935e911373SMike Kravetz break; 22945e911373SMike Kravetz case VMA_COMMIT_RESV: 2295075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22960db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22970db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22985e911373SMike Kravetz break; 2299feba16e2SMike Kravetz case VMA_END_RESV: 23000db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 23015e911373SMike Kravetz ret = 0; 23025e911373SMike Kravetz break; 230396b96a96SMike Kravetz case VMA_ADD_RESV: 23040db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2305075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 23060db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 23070db9d74eSMina Almasry VM_BUG_ON(ret < 0); 23080db9d74eSMina Almasry } else { 23090db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 231096b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 231196b96a96SMike Kravetz } 231296b96a96SMike Kravetz break; 2313846be085SMike Kravetz case VMA_DEL_RESV: 2314846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2315846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2316846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2317846be085SMike Kravetz } else { 2318846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2319846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2320846be085SMike Kravetz VM_BUG_ON(ret < 0); 2321846be085SMike Kravetz } 2322846be085SMike Kravetz break; 23235e911373SMike Kravetz default: 23245e911373SMike Kravetz BUG(); 23255e911373SMike Kravetz } 232684afd99bSAndy Whitcroft 2327846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2328cf3ad20bSMike Kravetz return ret; 232967961f9dSMike Kravetz /* 2330bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2331bf3d12b9SMiaohe Lin * 233267961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 233367961f9dSMike Kravetz * However, a file associated with mapping could have been 233467961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 233567961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 233667961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 233767961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 233867961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 233967961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 234067961f9dSMike Kravetz * reservation has already been consumed. As a result, the 234167961f9dSMike Kravetz * return value of this routine is the opposite of the 234267961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 234367961f9dSMike Kravetz */ 2344bf3d12b9SMiaohe Lin if (ret > 0) 234567961f9dSMike Kravetz return 0; 2346bf3d12b9SMiaohe Lin if (ret == 0) 234767961f9dSMike Kravetz return 1; 2348bf3d12b9SMiaohe Lin return ret; 234984afd99bSAndy Whitcroft } 2350cf3ad20bSMike Kravetz 2351cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2352a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2353c37f9fb1SAndy Whitcroft { 23545e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2355cf3ad20bSMike Kravetz } 2356c37f9fb1SAndy Whitcroft 2357cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2358cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2359cf3ad20bSMike Kravetz { 23605e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 23615e911373SMike Kravetz } 23625e911373SMike Kravetz 2363feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 23645e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 23655e911373SMike Kravetz { 2366feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2367c37f9fb1SAndy Whitcroft } 2368c37f9fb1SAndy Whitcroft 236996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 237096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 237196b96a96SMike Kravetz { 237296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 237396b96a96SMike Kravetz } 237496b96a96SMike Kravetz 2375846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2376846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 237796b96a96SMike Kravetz { 2378846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2379846be085SMike Kravetz } 2380846be085SMike Kravetz 2381846be085SMike Kravetz /* 2382846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2383846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2384846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2385846be085SMike Kravetz * 2386846be085SMike Kravetz * It handles two specific cases: 2387846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2388846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2389846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2390846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2391846be085SMike Kravetz * 2392846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2393846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2394846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2395846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2396846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2397846be085SMike Kravetz * reserve map indicates there is a reservation present. 2398846be085SMike Kravetz * 2399846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2400846be085SMike Kravetz */ 2401846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2402846be085SMike Kravetz unsigned long address, struct page *page) 2403846be085SMike Kravetz { 240496b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 240596b96a96SMike Kravetz 2406846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2407846be085SMike Kravetz if (unlikely(rc < 0)) 240896b96a96SMike Kravetz /* 240996b96a96SMike Kravetz * Rare out of memory condition in reserve map 2410d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 241196b96a96SMike Kravetz * global reserve count will not be incremented 241296b96a96SMike Kravetz * by free_huge_page. This will make it appear 241396b96a96SMike Kravetz * as though the reservation for this page was 241496b96a96SMike Kravetz * consumed. This may prevent the task from 241596b96a96SMike Kravetz * faulting in the page at a later time. This 241696b96a96SMike Kravetz * is better than inconsistent global huge page 241796b96a96SMike Kravetz * accounting of reserve counts. 241896b96a96SMike Kravetz */ 2419d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2420846be085SMike Kravetz else if (rc) 2421846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2422846be085SMike Kravetz else 2423846be085SMike Kravetz vma_end_reservation(h, vma, address); 2424846be085SMike Kravetz } else { 2425846be085SMike Kravetz if (!rc) { 242696b96a96SMike Kravetz /* 2427846be085SMike Kravetz * This indicates there is an entry in the reserve map 2428846be085SMike Kravetz * added by alloc_huge_page. We know it was added 2429846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2430846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2431846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2432846be085SMike Kravetz * does not consume a reservation. 243396b96a96SMike Kravetz */ 2434846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2435846be085SMike Kravetz if (rc < 0) 2436846be085SMike Kravetz /* 2437846be085SMike Kravetz * VERY rare out of memory condition. Since 2438846be085SMike Kravetz * we can not delete the entry, set 2439846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2440846be085SMike Kravetz * count will be incremented when the page 2441846be085SMike Kravetz * is freed. This reserve will be consumed 2442846be085SMike Kravetz * on a subsequent allocation. 2443846be085SMike Kravetz */ 2444846be085SMike Kravetz SetHPageRestoreReserve(page); 2445846be085SMike Kravetz } else if (rc < 0) { 2446846be085SMike Kravetz /* 2447846be085SMike Kravetz * Rare out of memory condition from 2448846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2449846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2450846be085SMike Kravetz * this implies there is not an entry in the 2451846be085SMike Kravetz * reserve map. 2452846be085SMike Kravetz * 2453846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2454846be085SMike Kravetz * no reservation. We are done. 2455846be085SMike Kravetz */ 2456846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2457846be085SMike Kravetz /* 2458846be085SMike Kravetz * For private mappings, no entry indicates 2459846be085SMike Kravetz * a reservation is present. Since we can 2460846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2461846be085SMike Kravetz * on the page so reserve count will be 2462846be085SMike Kravetz * incremented when freed. This reserve will 2463846be085SMike Kravetz * be consumed on a subsequent allocation. 2464846be085SMike Kravetz */ 2465846be085SMike Kravetz SetHPageRestoreReserve(page); 246696b96a96SMike Kravetz } else 2467846be085SMike Kravetz /* 2468846be085SMike Kravetz * No reservation present, do nothing 2469846be085SMike Kravetz */ 247096b96a96SMike Kravetz vma_end_reservation(h, vma, address); 247196b96a96SMike Kravetz } 247296b96a96SMike Kravetz } 247396b96a96SMike Kravetz 2474369fa227SOscar Salvador /* 2475369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2476369fa227SOscar Salvador * @h: struct hstate old page belongs to 2477369fa227SOscar Salvador * @old_page: Old page to dissolve 2478ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2479369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2480369fa227SOscar Salvador */ 2481ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2482ae37c7ffSOscar Salvador struct list_head *list) 2483369fa227SOscar Salvador { 2484369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2485369fa227SOscar Salvador int nid = page_to_nid(old_page); 2486369fa227SOscar Salvador struct page *new_page; 2487369fa227SOscar Salvador int ret = 0; 2488369fa227SOscar Salvador 2489369fa227SOscar Salvador /* 2490369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2491f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2492f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2493f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2494f41f2ed4SMuchun Song * under the lock. 2495369fa227SOscar Salvador */ 2496369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2497369fa227SOscar Salvador if (!new_page) 2498369fa227SOscar Salvador return -ENOMEM; 2499f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2500369fa227SOscar Salvador 2501369fa227SOscar Salvador retry: 2502369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2503369fa227SOscar Salvador if (!PageHuge(old_page)) { 2504369fa227SOscar Salvador /* 2505369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2506369fa227SOscar Salvador */ 2507369fa227SOscar Salvador goto free_new; 2508369fa227SOscar Salvador } else if (page_count(old_page)) { 2509369fa227SOscar Salvador /* 2510ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2511ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2512369fa227SOscar Salvador */ 2513ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2514ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2515369fa227SOscar Salvador ret = -EBUSY; 2516ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2517369fa227SOscar Salvador goto free_new; 2518369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2519369fa227SOscar Salvador /* 2520369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2521369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2522369fa227SOscar Salvador * we retry. 2523369fa227SOscar Salvador */ 2524369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2525369fa227SOscar Salvador cond_resched(); 2526369fa227SOscar Salvador goto retry; 2527369fa227SOscar Salvador } else { 2528369fa227SOscar Salvador /* 2529369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2530369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2531369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2532369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2533369fa227SOscar Salvador * stable since this happens under the lock. 2534369fa227SOscar Salvador */ 2535369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2536369fa227SOscar Salvador 2537369fa227SOscar Salvador /* 2538369fa227SOscar Salvador * Reference count trick is needed because allocator gives us 2539369fa227SOscar Salvador * referenced page but the pool requires pages with 0 refcount. 2540369fa227SOscar Salvador */ 2541369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2542369fa227SOscar Salvador page_ref_dec(new_page); 2543369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2544369fa227SOscar Salvador 2545369fa227SOscar Salvador /* 2546369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2547369fa227SOscar Salvador */ 2548369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2549b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2550369fa227SOscar Salvador } 2551369fa227SOscar Salvador 2552369fa227SOscar Salvador return ret; 2553369fa227SOscar Salvador 2554369fa227SOscar Salvador free_new: 2555369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2556b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2557369fa227SOscar Salvador 2558369fa227SOscar Salvador return ret; 2559369fa227SOscar Salvador } 2560369fa227SOscar Salvador 2561ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2562369fa227SOscar Salvador { 2563369fa227SOscar Salvador struct hstate *h; 2564369fa227SOscar Salvador struct page *head; 2565ae37c7ffSOscar Salvador int ret = -EBUSY; 2566369fa227SOscar Salvador 2567369fa227SOscar Salvador /* 2568369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2569369fa227SOscar Salvador * to carefully check the state under the lock. 2570369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2571369fa227SOscar Salvador */ 2572369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2573369fa227SOscar Salvador if (PageHuge(page)) { 2574369fa227SOscar Salvador head = compound_head(page); 2575369fa227SOscar Salvador h = page_hstate(head); 2576369fa227SOscar Salvador } else { 2577369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2578369fa227SOscar Salvador return 0; 2579369fa227SOscar Salvador } 2580369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2581369fa227SOscar Salvador 2582369fa227SOscar Salvador /* 2583369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2584369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2585369fa227SOscar Salvador * of bailing out right away without further retrying. 2586369fa227SOscar Salvador */ 2587369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2588369fa227SOscar Salvador return -ENOMEM; 2589369fa227SOscar Salvador 2590ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2591ae37c7ffSOscar Salvador ret = 0; 2592ae37c7ffSOscar Salvador else if (!page_count(head)) 2593ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2594ae37c7ffSOscar Salvador 2595ae37c7ffSOscar Salvador return ret; 2596369fa227SOscar Salvador } 2597369fa227SOscar Salvador 259870c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 259904f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2600348ea204SAdam Litke { 260190481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2602a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2603348ea204SAdam Litke struct page *page; 2604d85f69b0SMike Kravetz long map_chg, map_commit; 2605d85f69b0SMike Kravetz long gbl_chg; 26066d76dcf4SAneesh Kumar K.V int ret, idx; 26076d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 260808cf9fafSMina Almasry bool deferred_reserve; 26092fc39cecSAdam Litke 26106d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2611a1e78772SMel Gorman /* 2612d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2613d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2614d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2615a1e78772SMel Gorman */ 2616d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2617d85f69b0SMike Kravetz if (map_chg < 0) 261876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2619d85f69b0SMike Kravetz 2620d85f69b0SMike Kravetz /* 2621d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2622d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2623d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2624d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2625d85f69b0SMike Kravetz * checked against any subpool limit. 2626d85f69b0SMike Kravetz */ 2627d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2628d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2629d85f69b0SMike Kravetz if (gbl_chg < 0) { 2630feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 263176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 26325e911373SMike Kravetz } 263390d8b7e6SAdam Litke 2634d85f69b0SMike Kravetz /* 2635d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2636d85f69b0SMike Kravetz * map, there could be reservations associated with the 2637d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2638d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2639d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2640d85f69b0SMike Kravetz * the subpool reservations. 2641d85f69b0SMike Kravetz */ 2642d85f69b0SMike Kravetz if (avoid_reserve) 2643d85f69b0SMike Kravetz gbl_chg = 1; 2644d85f69b0SMike Kravetz } 2645d85f69b0SMike Kravetz 264608cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 264708cf9fafSMina Almasry */ 26486501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 264908cf9fafSMina Almasry if (deferred_reserve) { 265008cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 265108cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 26528f34af6fSJianyu Zhan if (ret) 26538f34af6fSJianyu Zhan goto out_subpool_put; 265408cf9fafSMina Almasry } 265508cf9fafSMina Almasry 265608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 265708cf9fafSMina Almasry if (ret) 265808cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 26598f34af6fSJianyu Zhan 2660db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2661d85f69b0SMike Kravetz /* 2662d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2663d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2664d85f69b0SMike Kravetz * a reservation exists for the allocation. 2665d85f69b0SMike Kravetz */ 2666d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 266781a6fcaeSJoonsoo Kim if (!page) { 2668db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 26690c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 26708f34af6fSJianyu Zhan if (!page) 26718f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2672a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2673d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2674a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2675a88c7695SNaoya Horiguchi } 2676db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 267715a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 267881a6fcaeSJoonsoo Kim /* Fall through */ 2679a1e78772SMel Gorman } 268081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 268108cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 268208cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 268308cf9fafSMina Almasry */ 268408cf9fafSMina Almasry if (deferred_reserve) { 268508cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 268608cf9fafSMina Almasry h_cg, page); 268708cf9fafSMina Almasry } 268808cf9fafSMina Almasry 2689db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2690a1e78772SMel Gorman 2691d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2692a1e78772SMel Gorman 2693d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2694d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 269533039678SMike Kravetz /* 269633039678SMike Kravetz * The page was added to the reservation map between 269733039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 269833039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 269933039678SMike Kravetz * Adjust for the subpool count incremented above AND 270033039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 270133039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 270233039678SMike Kravetz * no longer applies. 270333039678SMike Kravetz */ 270433039678SMike Kravetz long rsv_adjust; 270533039678SMike Kravetz 270633039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 270733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 270879aa925bSMike Kravetz if (deferred_reserve) 270979aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 271079aa925bSMike Kravetz pages_per_huge_page(h), page); 271133039678SMike Kravetz } 27127893d1d5SAdam Litke return page; 27138f34af6fSJianyu Zhan 27148f34af6fSJianyu Zhan out_uncharge_cgroup: 27158f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 271608cf9fafSMina Almasry out_uncharge_cgroup_reservation: 271708cf9fafSMina Almasry if (deferred_reserve) 271808cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 271908cf9fafSMina Almasry h_cg); 27208f34af6fSJianyu Zhan out_subpool_put: 2721d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 27228f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2723feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 27248f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2725b45b5bd6SDavid Gibson } 2726b45b5bd6SDavid Gibson 2727e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2728e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2729e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2730aa888a74SAndi Kleen { 2731aa888a74SAndi Kleen struct huge_bootmem_page *m; 2732b2261026SJoonsoo Kim int nr_nodes, node; 2733aa888a74SAndi Kleen 2734b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2735aa888a74SAndi Kleen void *addr; 2736aa888a74SAndi Kleen 2737eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 27388b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 273997ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2740aa888a74SAndi Kleen if (addr) { 2741aa888a74SAndi Kleen /* 2742aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2743aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2744aa888a74SAndi Kleen * puts them into the mem_map). 2745aa888a74SAndi Kleen */ 2746aa888a74SAndi Kleen m = addr; 2747aa888a74SAndi Kleen goto found; 2748aa888a74SAndi Kleen } 2749aa888a74SAndi Kleen } 2750aa888a74SAndi Kleen return 0; 2751aa888a74SAndi Kleen 2752aa888a74SAndi Kleen found: 2753df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2754aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2755330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2756aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2757aa888a74SAndi Kleen m->hstate = h; 2758aa888a74SAndi Kleen return 1; 2759aa888a74SAndi Kleen } 2760aa888a74SAndi Kleen 2761d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2762d00181b9SKirill A. Shutemov unsigned int order) 276318229df5SAndy Whitcroft { 276418229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 276518229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 276618229df5SAndy Whitcroft else 276718229df5SAndy Whitcroft prep_compound_page(page, order); 276818229df5SAndy Whitcroft } 276918229df5SAndy Whitcroft 2770aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2771aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2772aa888a74SAndi Kleen { 2773aa888a74SAndi Kleen struct huge_bootmem_page *m; 2774aa888a74SAndi Kleen 2775aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 277640d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2777aa888a74SAndi Kleen struct hstate *h = m->hstate; 2778ee8f248dSBecky Bruce 2779aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2780c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2781ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2782aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2783af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2784af0fb9dfSMichal Hocko 2785b0320c7bSRafael Aquini /* 2786b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2787b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2788b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2789b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2790b0320c7bSRafael Aquini */ 2791bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2792c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2793520495feSCannon Matthews cond_resched(); 2794aa888a74SAndi Kleen } 2795aa888a74SAndi Kleen } 2796aa888a74SAndi Kleen 27978faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 27981da177e4SLinus Torvalds { 27991da177e4SLinus Torvalds unsigned long i; 2800f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2801f60858f9SMike Kravetz 2802f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2803f60858f9SMike Kravetz /* 2804f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2805f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2806f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2807f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2808f60858f9SMike Kravetz */ 2809f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2810f60858f9SMike Kravetz GFP_KERNEL); 2811f60858f9SMike Kravetz } else { 2812f60858f9SMike Kravetz /* allocations done at boot time */ 2813f60858f9SMike Kravetz node_alloc_noretry = NULL; 2814f60858f9SMike Kravetz } 2815f60858f9SMike Kravetz 2816f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2817f60858f9SMike Kravetz if (node_alloc_noretry) 2818f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 28191da177e4SLinus Torvalds 2820e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2821bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2822dbda8feaSBarry Song if (hugetlb_cma_size) { 2823cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 28247ecc9565SChen Wandun goto free; 2825cf11e85fSRoman Gushchin } 2826aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2827aa888a74SAndi Kleen break; 28280c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2829f60858f9SMike Kravetz &node_states[N_MEMORY], 2830f60858f9SMike Kravetz node_alloc_noretry)) 28311da177e4SLinus Torvalds break; 283269ed779aSDavid Rientjes cond_resched(); 28331da177e4SLinus Torvalds } 2834d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2835d715cf80SLiam R. Howlett char buf[32]; 2836d715cf80SLiam R. Howlett 2837c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2838d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2839d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 28408faa8b07SAndi Kleen h->max_huge_pages = i; 2841e5ff2159SAndi Kleen } 28427ecc9565SChen Wandun free: 2843f60858f9SMike Kravetz kfree(node_alloc_noretry); 2844d715cf80SLiam R. Howlett } 2845e5ff2159SAndi Kleen 2846e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2847e5ff2159SAndi Kleen { 2848e5ff2159SAndi Kleen struct hstate *h; 2849e5ff2159SAndi Kleen 2850e5ff2159SAndi Kleen for_each_hstate(h) { 2851641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2852641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2853641844f5SNaoya Horiguchi 28548faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2855bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 28568faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2857e5ff2159SAndi Kleen } 2858641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2859e5ff2159SAndi Kleen } 2860e5ff2159SAndi Kleen 2861e5ff2159SAndi Kleen static void __init report_hugepages(void) 2862e5ff2159SAndi Kleen { 2863e5ff2159SAndi Kleen struct hstate *h; 2864e5ff2159SAndi Kleen 2865e5ff2159SAndi Kleen for_each_hstate(h) { 28664abd32dbSAndi Kleen char buf[32]; 2867c6247f72SMatthew Wilcox 2868c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2869ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2870c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2871e5ff2159SAndi Kleen } 2872e5ff2159SAndi Kleen } 2873e5ff2159SAndi Kleen 28741da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 28756ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 28766ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 28771da177e4SLinus Torvalds { 28784415cc8dSChristoph Lameter int i; 28791121828aSMike Kravetz LIST_HEAD(page_list); 28804415cc8dSChristoph Lameter 28819487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2882bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2883aa888a74SAndi Kleen return; 2884aa888a74SAndi Kleen 28851121828aSMike Kravetz /* 28861121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 28871121828aSMike Kravetz */ 28886ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 288910c6ec49SMike Kravetz struct page *page, *next; 2890a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2891a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2892a5516438SAndi Kleen if (count >= h->nr_huge_pages) 28931121828aSMike Kravetz goto out; 28941da177e4SLinus Torvalds if (PageHighMem(page)) 28951da177e4SLinus Torvalds continue; 28966eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 28971121828aSMike Kravetz list_add(&page->lru, &page_list); 28981121828aSMike Kravetz } 28991121828aSMike Kravetz } 29001121828aSMike Kravetz 29011121828aSMike Kravetz out: 2902db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 290310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2904db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 29051da177e4SLinus Torvalds } 29061da177e4SLinus Torvalds #else 29076ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 29086ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 29091da177e4SLinus Torvalds { 29101da177e4SLinus Torvalds } 29111da177e4SLinus Torvalds #endif 29121da177e4SLinus Torvalds 291320a0307cSWu Fengguang /* 291420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 291520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 291620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 291720a0307cSWu Fengguang */ 29186ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 29196ae11b27SLee Schermerhorn int delta) 292020a0307cSWu Fengguang { 2921b2261026SJoonsoo Kim int nr_nodes, node; 292220a0307cSWu Fengguang 29239487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 292420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 292520a0307cSWu Fengguang 2926e8c5c824SLee Schermerhorn if (delta < 0) { 2927b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2928b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2929b2261026SJoonsoo Kim goto found; 2930b2261026SJoonsoo Kim } 2931b2261026SJoonsoo Kim } else { 2932b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2933b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2934b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2935b2261026SJoonsoo Kim goto found; 2936e8c5c824SLee Schermerhorn } 29379a76db09SLee Schermerhorn } 2938b2261026SJoonsoo Kim return 0; 293920a0307cSWu Fengguang 2940b2261026SJoonsoo Kim found: 294120a0307cSWu Fengguang h->surplus_huge_pages += delta; 2942b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2943b2261026SJoonsoo Kim return 1; 294420a0307cSWu Fengguang } 294520a0307cSWu Fengguang 2946a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2947fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 29486ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 29491da177e4SLinus Torvalds { 29507893d1d5SAdam Litke unsigned long min_count, ret; 295110c6ec49SMike Kravetz struct page *page; 295210c6ec49SMike Kravetz LIST_HEAD(page_list); 2953f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2954f60858f9SMike Kravetz 2955f60858f9SMike Kravetz /* 2956f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2957f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2958f60858f9SMike Kravetz * the requested huge pages. 2959f60858f9SMike Kravetz */ 2960f60858f9SMike Kravetz if (node_alloc_noretry) 2961f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2962f60858f9SMike Kravetz else 2963f60858f9SMike Kravetz return -ENOMEM; 29641da177e4SLinus Torvalds 296529383967SMike Kravetz /* 296629383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 296729383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 296829383967SMike Kravetz */ 296929383967SMike Kravetz mutex_lock(&h->resize_lock); 2970b65d4adbSMuchun Song flush_free_hpage_work(h); 2971db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 29724eb0716eSAlexandre Ghiti 29734eb0716eSAlexandre Ghiti /* 2974fd875dcaSMike Kravetz * Check for a node specific request. 2975fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2976fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2977fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2978fd875dcaSMike Kravetz */ 2979fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2980fd875dcaSMike Kravetz unsigned long old_count = count; 2981fd875dcaSMike Kravetz 2982fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2983fd875dcaSMike Kravetz /* 2984fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2985fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2986fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2987fd875dcaSMike Kravetz * largest possible value to align with their intention. 2988fd875dcaSMike Kravetz */ 2989fd875dcaSMike Kravetz if (count < old_count) 2990fd875dcaSMike Kravetz count = ULONG_MAX; 2991fd875dcaSMike Kravetz } 2992fd875dcaSMike Kravetz 2993fd875dcaSMike Kravetz /* 29944eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 29954eb0716eSAlexandre Ghiti * page range allocation. 29964eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 29974eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 29984eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 29994eb0716eSAlexandre Ghiti */ 30004eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 30014eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3002db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 300329383967SMike Kravetz mutex_unlock(&h->resize_lock); 3004f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 30054eb0716eSAlexandre Ghiti return -EINVAL; 30064eb0716eSAlexandre Ghiti } 30074eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 30084eb0716eSAlexandre Ghiti } 3009aa888a74SAndi Kleen 30107893d1d5SAdam Litke /* 30117893d1d5SAdam Litke * Increase the pool size 30127893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 30137893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3014d1c3fb1fSNishanth Aravamudan * 30150c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3016d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3017d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3018d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3019d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 30207893d1d5SAdam Litke */ 3021a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 30226ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 30237893d1d5SAdam Litke break; 30247893d1d5SAdam Litke } 30257893d1d5SAdam Litke 3026a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 30277893d1d5SAdam Litke /* 30287893d1d5SAdam Litke * If this allocation races such that we no longer need the 30297893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 30307893d1d5SAdam Litke * and reducing the surplus. 30317893d1d5SAdam Litke */ 3032db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3033649920c6SJia He 3034649920c6SJia He /* yield cpu to avoid soft lockup */ 3035649920c6SJia He cond_resched(); 3036649920c6SJia He 3037f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3038f60858f9SMike Kravetz node_alloc_noretry); 3039db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 30407893d1d5SAdam Litke if (!ret) 30417893d1d5SAdam Litke goto out; 30427893d1d5SAdam Litke 3043536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3044536240f2SMel Gorman if (signal_pending(current)) 3045536240f2SMel Gorman goto out; 30467893d1d5SAdam Litke } 30477893d1d5SAdam Litke 30487893d1d5SAdam Litke /* 30497893d1d5SAdam Litke * Decrease the pool size 30507893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 30517893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 30527893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 30537893d1d5SAdam Litke * to the desired size as pages become free. 3054d1c3fb1fSNishanth Aravamudan * 3055d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3056d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3057d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 30580c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3059d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3060d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3061d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 30627893d1d5SAdam Litke */ 3063a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 30646b0c880dSAdam Litke min_count = max(count, min_count); 30656ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 306610c6ec49SMike Kravetz 306710c6ec49SMike Kravetz /* 306810c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 306910c6ec49SMike Kravetz */ 3070a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 307110c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 307210c6ec49SMike Kravetz if (!page) 30731da177e4SLinus Torvalds break; 307410c6ec49SMike Kravetz 307510c6ec49SMike Kravetz list_add(&page->lru, &page_list); 30761da177e4SLinus Torvalds } 307710c6ec49SMike Kravetz /* free the pages after dropping lock */ 3078db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 307910c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3080b65d4adbSMuchun Song flush_free_hpage_work(h); 3081db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 308210c6ec49SMike Kravetz 3083a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 30846ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 30857893d1d5SAdam Litke break; 30867893d1d5SAdam Litke } 30877893d1d5SAdam Litke out: 30884eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3089db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 309029383967SMike Kravetz mutex_unlock(&h->resize_lock); 30914eb0716eSAlexandre Ghiti 3092f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3093f60858f9SMike Kravetz 30944eb0716eSAlexandre Ghiti return 0; 30951da177e4SLinus Torvalds } 30961da177e4SLinus Torvalds 3097a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3098a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3099a3437870SNishanth Aravamudan 3100a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 3101a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 3102a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 3103a3437870SNishanth Aravamudan 3104a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3105a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3106a3437870SNishanth Aravamudan 31079a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 31089a305230SLee Schermerhorn 31099a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3110a3437870SNishanth Aravamudan { 3111a3437870SNishanth Aravamudan int i; 31129a305230SLee Schermerhorn 3113a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 31149a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 31159a305230SLee Schermerhorn if (nidp) 31169a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3117a3437870SNishanth Aravamudan return &hstates[i]; 31189a305230SLee Schermerhorn } 31199a305230SLee Schermerhorn 31209a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3121a3437870SNishanth Aravamudan } 3122a3437870SNishanth Aravamudan 312306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3124a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3125a3437870SNishanth Aravamudan { 31269a305230SLee Schermerhorn struct hstate *h; 31279a305230SLee Schermerhorn unsigned long nr_huge_pages; 31289a305230SLee Schermerhorn int nid; 31299a305230SLee Schermerhorn 31309a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 31319a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 31329a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 31339a305230SLee Schermerhorn else 31349a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 31359a305230SLee Schermerhorn 3136ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3137a3437870SNishanth Aravamudan } 3138adbe8726SEric B Munson 3139238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3140238d3c13SDavid Rientjes struct hstate *h, int nid, 3141238d3c13SDavid Rientjes unsigned long count, size_t len) 3142a3437870SNishanth Aravamudan { 3143a3437870SNishanth Aravamudan int err; 31442d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3145a3437870SNishanth Aravamudan 31462d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 31472d0adf7eSOscar Salvador return -EINVAL; 3148adbe8726SEric B Munson 31499a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 31509a305230SLee Schermerhorn /* 31519a305230SLee Schermerhorn * global hstate attribute 31529a305230SLee Schermerhorn */ 31539a305230SLee Schermerhorn if (!(obey_mempolicy && 31542d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 31552d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 31562d0adf7eSOscar Salvador else 31572d0adf7eSOscar Salvador n_mask = &nodes_allowed; 31582d0adf7eSOscar Salvador } else { 31599a305230SLee Schermerhorn /* 3160fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3161fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 31629a305230SLee Schermerhorn */ 31632d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 31642d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3165fd875dcaSMike Kravetz } 31669a305230SLee Schermerhorn 31672d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 316806808b08SLee Schermerhorn 31694eb0716eSAlexandre Ghiti return err ? err : len; 317006808b08SLee Schermerhorn } 317106808b08SLee Schermerhorn 3172238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3173238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3174238d3c13SDavid Rientjes size_t len) 3175238d3c13SDavid Rientjes { 3176238d3c13SDavid Rientjes struct hstate *h; 3177238d3c13SDavid Rientjes unsigned long count; 3178238d3c13SDavid Rientjes int nid; 3179238d3c13SDavid Rientjes int err; 3180238d3c13SDavid Rientjes 3181238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3182238d3c13SDavid Rientjes if (err) 3183238d3c13SDavid Rientjes return err; 3184238d3c13SDavid Rientjes 3185238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3186238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3187238d3c13SDavid Rientjes } 3188238d3c13SDavid Rientjes 318906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 319006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 319106808b08SLee Schermerhorn { 319206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 319306808b08SLee Schermerhorn } 319406808b08SLee Schermerhorn 319506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 319606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 319706808b08SLee Schermerhorn { 3198238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3199a3437870SNishanth Aravamudan } 3200a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3201a3437870SNishanth Aravamudan 320206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 320306808b08SLee Schermerhorn 320406808b08SLee Schermerhorn /* 320506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 320606808b08SLee Schermerhorn * huge page alloc/free. 320706808b08SLee Schermerhorn */ 320806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3209ae7a927dSJoe Perches struct kobj_attribute *attr, 3210ae7a927dSJoe Perches char *buf) 321106808b08SLee Schermerhorn { 321206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 321306808b08SLee Schermerhorn } 321406808b08SLee Schermerhorn 321506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 321606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 321706808b08SLee Schermerhorn { 3218238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 321906808b08SLee Schermerhorn } 322006808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 322106808b08SLee Schermerhorn #endif 322206808b08SLee Schermerhorn 322306808b08SLee Schermerhorn 3224a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3225a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3226a3437870SNishanth Aravamudan { 32279a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3228ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3229a3437870SNishanth Aravamudan } 3230adbe8726SEric B Munson 3231a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3232a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3233a3437870SNishanth Aravamudan { 3234a3437870SNishanth Aravamudan int err; 3235a3437870SNishanth Aravamudan unsigned long input; 32369a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3237a3437870SNishanth Aravamudan 3238bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3239adbe8726SEric B Munson return -EINVAL; 3240adbe8726SEric B Munson 32413dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3242a3437870SNishanth Aravamudan if (err) 324373ae31e5SEric B Munson return err; 3244a3437870SNishanth Aravamudan 3245db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3246a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3247db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3248a3437870SNishanth Aravamudan 3249a3437870SNishanth Aravamudan return count; 3250a3437870SNishanth Aravamudan } 3251a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3252a3437870SNishanth Aravamudan 3253a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3254a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3255a3437870SNishanth Aravamudan { 32569a305230SLee Schermerhorn struct hstate *h; 32579a305230SLee Schermerhorn unsigned long free_huge_pages; 32589a305230SLee Schermerhorn int nid; 32599a305230SLee Schermerhorn 32609a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 32619a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 32629a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 32639a305230SLee Schermerhorn else 32649a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 32659a305230SLee Schermerhorn 3266ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3267a3437870SNishanth Aravamudan } 3268a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3269a3437870SNishanth Aravamudan 3270a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3271a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3272a3437870SNishanth Aravamudan { 32739a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3274ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3275a3437870SNishanth Aravamudan } 3276a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3277a3437870SNishanth Aravamudan 3278a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3279a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3280a3437870SNishanth Aravamudan { 32819a305230SLee Schermerhorn struct hstate *h; 32829a305230SLee Schermerhorn unsigned long surplus_huge_pages; 32839a305230SLee Schermerhorn int nid; 32849a305230SLee Schermerhorn 32859a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 32869a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 32879a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 32889a305230SLee Schermerhorn else 32899a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 32909a305230SLee Schermerhorn 3291ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3292a3437870SNishanth Aravamudan } 3293a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3294a3437870SNishanth Aravamudan 3295a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3296a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3297a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3298a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3299a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3300a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 330106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 330206808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 330306808b08SLee Schermerhorn #endif 3304a3437870SNishanth Aravamudan NULL, 3305a3437870SNishanth Aravamudan }; 3306a3437870SNishanth Aravamudan 330767e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3308a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3309a3437870SNishanth Aravamudan }; 3310a3437870SNishanth Aravamudan 3311094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 33129a305230SLee Schermerhorn struct kobject **hstate_kobjs, 331367e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3314a3437870SNishanth Aravamudan { 3315a3437870SNishanth Aravamudan int retval; 3316972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3317a3437870SNishanth Aravamudan 33189a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 33199a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3320a3437870SNishanth Aravamudan return -ENOMEM; 3321a3437870SNishanth Aravamudan 33229a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3323cc2205a6SMiaohe Lin if (retval) { 33249a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3325cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3326cc2205a6SMiaohe Lin } 3327a3437870SNishanth Aravamudan 3328a3437870SNishanth Aravamudan return retval; 3329a3437870SNishanth Aravamudan } 3330a3437870SNishanth Aravamudan 3331a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3332a3437870SNishanth Aravamudan { 3333a3437870SNishanth Aravamudan struct hstate *h; 3334a3437870SNishanth Aravamudan int err; 3335a3437870SNishanth Aravamudan 3336a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3337a3437870SNishanth Aravamudan if (!hugepages_kobj) 3338a3437870SNishanth Aravamudan return; 3339a3437870SNishanth Aravamudan 3340a3437870SNishanth Aravamudan for_each_hstate(h) { 33419a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 33429a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3343a3437870SNishanth Aravamudan if (err) 3344282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3345a3437870SNishanth Aravamudan } 3346a3437870SNishanth Aravamudan } 3347a3437870SNishanth Aravamudan 33489a305230SLee Schermerhorn #ifdef CONFIG_NUMA 33499a305230SLee Schermerhorn 33509a305230SLee Schermerhorn /* 33519a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 335210fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 335310fbcf4cSKay Sievers * index of a node device or _hstate == node id. 335410fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 33559a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 33569a305230SLee Schermerhorn */ 33579a305230SLee Schermerhorn struct node_hstate { 33589a305230SLee Schermerhorn struct kobject *hugepages_kobj; 33599a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 33609a305230SLee Schermerhorn }; 3361b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 33629a305230SLee Schermerhorn 33639a305230SLee Schermerhorn /* 336410fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 33659a305230SLee Schermerhorn */ 33669a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 33679a305230SLee Schermerhorn &nr_hugepages_attr.attr, 33689a305230SLee Schermerhorn &free_hugepages_attr.attr, 33699a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 33709a305230SLee Schermerhorn NULL, 33719a305230SLee Schermerhorn }; 33729a305230SLee Schermerhorn 337367e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 33749a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 33759a305230SLee Schermerhorn }; 33769a305230SLee Schermerhorn 33779a305230SLee Schermerhorn /* 337810fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 33799a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 33809a305230SLee Schermerhorn */ 33819a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 33829a305230SLee Schermerhorn { 33839a305230SLee Schermerhorn int nid; 33849a305230SLee Schermerhorn 33859a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 33869a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 33879a305230SLee Schermerhorn int i; 33889a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 33899a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 33909a305230SLee Schermerhorn if (nidp) 33919a305230SLee Schermerhorn *nidp = nid; 33929a305230SLee Schermerhorn return &hstates[i]; 33939a305230SLee Schermerhorn } 33949a305230SLee Schermerhorn } 33959a305230SLee Schermerhorn 33969a305230SLee Schermerhorn BUG(); 33979a305230SLee Schermerhorn return NULL; 33989a305230SLee Schermerhorn } 33999a305230SLee Schermerhorn 34009a305230SLee Schermerhorn /* 340110fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 34029a305230SLee Schermerhorn * No-op if no hstate attributes attached. 34039a305230SLee Schermerhorn */ 34043cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 34059a305230SLee Schermerhorn { 34069a305230SLee Schermerhorn struct hstate *h; 340710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 34089a305230SLee Schermerhorn 34099a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 34109b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 34119a305230SLee Schermerhorn 3412972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3413972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3414972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3415972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3416972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3417972dc4deSAneesh Kumar K.V } 34189a305230SLee Schermerhorn } 34199a305230SLee Schermerhorn 34209a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 34219a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 34229a305230SLee Schermerhorn } 34239a305230SLee Schermerhorn 34249a305230SLee Schermerhorn 34259a305230SLee Schermerhorn /* 342610fbcf4cSKay Sievers * Register hstate attributes for a single node device. 34279a305230SLee Schermerhorn * No-op if attributes already registered. 34289a305230SLee Schermerhorn */ 34293cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 34309a305230SLee Schermerhorn { 34319a305230SLee Schermerhorn struct hstate *h; 343210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 34339a305230SLee Schermerhorn int err; 34349a305230SLee Schermerhorn 34359a305230SLee Schermerhorn if (nhs->hugepages_kobj) 34369a305230SLee Schermerhorn return; /* already allocated */ 34379a305230SLee Schermerhorn 34389a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 343910fbcf4cSKay Sievers &node->dev.kobj); 34409a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 34419a305230SLee Schermerhorn return; 34429a305230SLee Schermerhorn 34439a305230SLee Schermerhorn for_each_hstate(h) { 34449a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 34459a305230SLee Schermerhorn nhs->hstate_kobjs, 34469a305230SLee Schermerhorn &per_node_hstate_attr_group); 34479a305230SLee Schermerhorn if (err) { 3448282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 344910fbcf4cSKay Sievers h->name, node->dev.id); 34509a305230SLee Schermerhorn hugetlb_unregister_node(node); 34519a305230SLee Schermerhorn break; 34529a305230SLee Schermerhorn } 34539a305230SLee Schermerhorn } 34549a305230SLee Schermerhorn } 34559a305230SLee Schermerhorn 34569a305230SLee Schermerhorn /* 34579b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 345810fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 345910fbcf4cSKay Sievers * registered their associated device by this time. 34609a305230SLee Schermerhorn */ 34617d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 34629a305230SLee Schermerhorn { 34639a305230SLee Schermerhorn int nid; 34649a305230SLee Schermerhorn 34658cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 34668732794bSWen Congyang struct node *node = node_devices[nid]; 346710fbcf4cSKay Sievers if (node->dev.id == nid) 34689a305230SLee Schermerhorn hugetlb_register_node(node); 34699a305230SLee Schermerhorn } 34709a305230SLee Schermerhorn 34719a305230SLee Schermerhorn /* 347210fbcf4cSKay Sievers * Let the node device driver know we're here so it can 34739a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 34749a305230SLee Schermerhorn */ 34759a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 34769a305230SLee Schermerhorn hugetlb_unregister_node); 34779a305230SLee Schermerhorn } 34789a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 34799a305230SLee Schermerhorn 34809a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 34819a305230SLee Schermerhorn { 34829a305230SLee Schermerhorn BUG(); 34839a305230SLee Schermerhorn if (nidp) 34849a305230SLee Schermerhorn *nidp = -1; 34859a305230SLee Schermerhorn return NULL; 34869a305230SLee Schermerhorn } 34879a305230SLee Schermerhorn 34889a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 34899a305230SLee Schermerhorn 34909a305230SLee Schermerhorn #endif 34919a305230SLee Schermerhorn 3492a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3493a3437870SNishanth Aravamudan { 34948382d914SDavidlohr Bueso int i; 34958382d914SDavidlohr Bueso 3496d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3497d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3498d6995da3SMike Kravetz 3499c2833a5bSMike Kravetz if (!hugepages_supported()) { 3500c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3501c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 35020ef89d25SBenjamin Herrenschmidt return 0; 3503d715cf80SLiam R. Howlett } 3504d715cf80SLiam R. Howlett 3505282f4214SMike Kravetz /* 3506282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3507282f4214SMike Kravetz * architectures depend on setup being done here. 3508282f4214SMike Kravetz */ 3509a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3510282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3511282f4214SMike Kravetz /* 3512282f4214SMike Kravetz * If we did not parse a default huge page size, set 3513282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3514282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3515282f4214SMike Kravetz * specified, set that here as well. 3516282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3517282f4214SMike Kravetz * setting. A warning will be printed in this case. 3518282f4214SMike Kravetz */ 3519282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3520f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3521282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3522282f4214SMike Kravetz char buf[32]; 3523282f4214SMike Kravetz 3524282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3525282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3526282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3527282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3528282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3529282f4214SMike Kravetz default_hstate_max_huge_pages); 3530282f4214SMike Kravetz } 3531282f4214SMike Kravetz default_hstate.max_huge_pages = 3532282f4214SMike Kravetz default_hstate_max_huge_pages; 3533282f4214SMike Kravetz } 3534f8b74815SVaishali Thakkar } 3535a3437870SNishanth Aravamudan 3536cf11e85fSRoman Gushchin hugetlb_cma_check(); 3537a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3538aa888a74SAndi Kleen gather_bootmem_prealloc(); 3539a3437870SNishanth Aravamudan report_hugepages(); 3540a3437870SNishanth Aravamudan 3541a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 35429a305230SLee Schermerhorn hugetlb_register_all_nodes(); 35437179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 35449a305230SLee Schermerhorn 35458382d914SDavidlohr Bueso #ifdef CONFIG_SMP 35468382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 35478382d914SDavidlohr Bueso #else 35488382d914SDavidlohr Bueso num_fault_mutexes = 1; 35498382d914SDavidlohr Bueso #endif 3550c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 35516da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 35526da2ec56SKees Cook GFP_KERNEL); 3553c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 35548382d914SDavidlohr Bueso 35558382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3556c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3557a3437870SNishanth Aravamudan return 0; 3558a3437870SNishanth Aravamudan } 35593e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3560a3437870SNishanth Aravamudan 3561ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3562ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 35639fee021dSVaishali Thakkar { 3564ae94da89SMike Kravetz return size == HPAGE_SIZE; 35659fee021dSVaishali Thakkar } 35669fee021dSVaishali Thakkar 3567d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3568a3437870SNishanth Aravamudan { 3569a3437870SNishanth Aravamudan struct hstate *h; 35708faa8b07SAndi Kleen unsigned long i; 35718faa8b07SAndi Kleen 3572a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3573a3437870SNishanth Aravamudan return; 3574a3437870SNishanth Aravamudan } 357547d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3576a3437870SNishanth Aravamudan BUG_ON(order == 0); 357747d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 357829383967SMike Kravetz mutex_init(&h->resize_lock); 3579a3437870SNishanth Aravamudan h->order = order; 3580aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 35818faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 35828faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 35830edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 358454f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 358554f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3586a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3587a3437870SNishanth Aravamudan huge_page_size(h)/1024); 358877490587SMuchun Song hugetlb_vmemmap_init(h); 35898faa8b07SAndi Kleen 3590a3437870SNishanth Aravamudan parsed_hstate = h; 3591a3437870SNishanth Aravamudan } 3592a3437870SNishanth Aravamudan 3593282f4214SMike Kravetz /* 3594282f4214SMike Kravetz * hugepages command line processing 3595282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3596282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3597282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3598282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3599282f4214SMike Kravetz */ 3600282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3601a3437870SNishanth Aravamudan { 3602a3437870SNishanth Aravamudan unsigned long *mhp; 36038faa8b07SAndi Kleen static unsigned long *last_mhp; 3604a3437870SNishanth Aravamudan 36059fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3606282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 36079fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3608282f4214SMike Kravetz return 0; 36099fee021dSVaishali Thakkar } 3610282f4214SMike Kravetz 3611a3437870SNishanth Aravamudan /* 3612282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3613282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3614282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3615282f4214SMike Kravetz * default_hugepagesz. 3616a3437870SNishanth Aravamudan */ 36179fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3618a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3619a3437870SNishanth Aravamudan else 3620a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3621a3437870SNishanth Aravamudan 36228faa8b07SAndi Kleen if (mhp == last_mhp) { 3623282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3624282f4214SMike Kravetz return 0; 36258faa8b07SAndi Kleen } 36268faa8b07SAndi Kleen 3627a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3628a3437870SNishanth Aravamudan *mhp = 0; 3629a3437870SNishanth Aravamudan 36308faa8b07SAndi Kleen /* 36318faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 363204adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 36338faa8b07SAndi Kleen * use the bootmem allocator. 36348faa8b07SAndi Kleen */ 363504adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 36368faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 36378faa8b07SAndi Kleen 36388faa8b07SAndi Kleen last_mhp = mhp; 36398faa8b07SAndi Kleen 3640a3437870SNishanth Aravamudan return 1; 3641a3437870SNishanth Aravamudan } 3642282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3643e11bfbfcSNick Piggin 3644282f4214SMike Kravetz /* 3645282f4214SMike Kravetz * hugepagesz command line processing 3646282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3647282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3648282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3649282f4214SMike Kravetz * hugepagesz argument was valid. 3650282f4214SMike Kravetz */ 3651359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3652e11bfbfcSNick Piggin { 3653359f2544SMike Kravetz unsigned long size; 3654282f4214SMike Kravetz struct hstate *h; 3655282f4214SMike Kravetz 3656282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3657359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3658359f2544SMike Kravetz 3659359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3660282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3661359f2544SMike Kravetz return 0; 3662359f2544SMike Kravetz } 3663359f2544SMike Kravetz 3664282f4214SMike Kravetz h = size_to_hstate(size); 3665282f4214SMike Kravetz if (h) { 3666282f4214SMike Kravetz /* 3667282f4214SMike Kravetz * hstate for this size already exists. This is normally 3668282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3669282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3670282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3671282f4214SMike Kravetz * previously specified. 3672282f4214SMike Kravetz */ 3673282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3674282f4214SMike Kravetz default_hstate.max_huge_pages) { 3675282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 367638237830SMike Kravetz return 0; 367738237830SMike Kravetz } 367838237830SMike Kravetz 3679282f4214SMike Kravetz /* 3680282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3681282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3682282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3683282f4214SMike Kravetz */ 3684282f4214SMike Kravetz parsed_hstate = h; 3685282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3686e11bfbfcSNick Piggin return 1; 3687e11bfbfcSNick Piggin } 3688282f4214SMike Kravetz 3689359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3690282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3691359f2544SMike Kravetz return 1; 3692359f2544SMike Kravetz } 3693359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3694359f2544SMike Kravetz 3695282f4214SMike Kravetz /* 3696282f4214SMike Kravetz * default_hugepagesz command line input 3697282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3698282f4214SMike Kravetz */ 3699ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3700e11bfbfcSNick Piggin { 3701ae94da89SMike Kravetz unsigned long size; 3702ae94da89SMike Kravetz 3703282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3704282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3705282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3706282f4214SMike Kravetz return 0; 3707282f4214SMike Kravetz } 3708282f4214SMike Kravetz 3709282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3710282f4214SMike Kravetz 3711282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3712282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3713282f4214SMike Kravetz return 0; 3714282f4214SMike Kravetz } 3715282f4214SMike Kravetz 3716282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3717282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3718282f4214SMike Kravetz parsed_default_hugepagesz = true; 3719282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3720282f4214SMike Kravetz 3721282f4214SMike Kravetz /* 3722282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3723282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3724282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3725282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3726282f4214SMike Kravetz * allocated here from bootmem allocator. 3727282f4214SMike Kravetz */ 3728282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3729282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3730282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3731282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3732282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3733282f4214SMike Kravetz } 3734282f4214SMike Kravetz 3735e11bfbfcSNick Piggin return 1; 3736e11bfbfcSNick Piggin } 3737ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3738a3437870SNishanth Aravamudan 37398ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 37408a213460SNishanth Aravamudan { 37418a213460SNishanth Aravamudan int node; 37428a213460SNishanth Aravamudan unsigned int nr = 0; 37438ca39e68SMuchun Song nodemask_t *mpol_allowed; 37448ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 37458ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 37468a213460SNishanth Aravamudan 37478ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 37488ca39e68SMuchun Song 37498ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3750c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 37518a213460SNishanth Aravamudan nr += array[node]; 37528ca39e68SMuchun Song } 37538a213460SNishanth Aravamudan 37548a213460SNishanth Aravamudan return nr; 37558a213460SNishanth Aravamudan } 37568a213460SNishanth Aravamudan 37578a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 375817743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 375917743798SMuchun Song void *buffer, size_t *length, 376017743798SMuchun Song loff_t *ppos, unsigned long *out) 376117743798SMuchun Song { 376217743798SMuchun Song struct ctl_table dup_table; 376317743798SMuchun Song 376417743798SMuchun Song /* 376517743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 376617743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 376717743798SMuchun Song */ 376817743798SMuchun Song dup_table = *table; 376917743798SMuchun Song dup_table.data = out; 377017743798SMuchun Song 377117743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 377217743798SMuchun Song } 377317743798SMuchun Song 377406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 377506808b08SLee Schermerhorn struct ctl_table *table, int write, 377632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 37771da177e4SLinus Torvalds { 3778e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3779238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 378008d4a246SMichal Hocko int ret; 3781e5ff2159SAndi Kleen 3782457c1b27SNishanth Aravamudan if (!hugepages_supported()) 378386613628SJan Stancek return -EOPNOTSUPP; 3784457c1b27SNishanth Aravamudan 378517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 378617743798SMuchun Song &tmp); 378708d4a246SMichal Hocko if (ret) 378808d4a246SMichal Hocko goto out; 3789e5ff2159SAndi Kleen 3790238d3c13SDavid Rientjes if (write) 3791238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3792238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 379308d4a246SMichal Hocko out: 379408d4a246SMichal Hocko return ret; 37951da177e4SLinus Torvalds } 3796396faf03SMel Gorman 379706808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 379832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 379906808b08SLee Schermerhorn { 380006808b08SLee Schermerhorn 380106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 380206808b08SLee Schermerhorn buffer, length, ppos); 380306808b08SLee Schermerhorn } 380406808b08SLee Schermerhorn 380506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 380606808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 380732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 380806808b08SLee Schermerhorn { 380906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 381006808b08SLee Schermerhorn buffer, length, ppos); 381106808b08SLee Schermerhorn } 381206808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 381306808b08SLee Schermerhorn 3814a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 381532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3816a3d0c6aaSNishanth Aravamudan { 3817a5516438SAndi Kleen struct hstate *h = &default_hstate; 3818e5ff2159SAndi Kleen unsigned long tmp; 381908d4a246SMichal Hocko int ret; 3820e5ff2159SAndi Kleen 3821457c1b27SNishanth Aravamudan if (!hugepages_supported()) 382286613628SJan Stancek return -EOPNOTSUPP; 3823457c1b27SNishanth Aravamudan 3824e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3825e5ff2159SAndi Kleen 3826bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3827adbe8726SEric B Munson return -EINVAL; 3828adbe8726SEric B Munson 382917743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 383017743798SMuchun Song &tmp); 383108d4a246SMichal Hocko if (ret) 383208d4a246SMichal Hocko goto out; 3833e5ff2159SAndi Kleen 3834e5ff2159SAndi Kleen if (write) { 3835db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3836e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3837db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3838e5ff2159SAndi Kleen } 383908d4a246SMichal Hocko out: 384008d4a246SMichal Hocko return ret; 3841a3d0c6aaSNishanth Aravamudan } 3842a3d0c6aaSNishanth Aravamudan 38431da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 38441da177e4SLinus Torvalds 3845e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 38461da177e4SLinus Torvalds { 3847fcb2b0c5SRoman Gushchin struct hstate *h; 3848fcb2b0c5SRoman Gushchin unsigned long total = 0; 3849fcb2b0c5SRoman Gushchin 3850457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3851457c1b27SNishanth Aravamudan return; 3852fcb2b0c5SRoman Gushchin 3853fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3854fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3855fcb2b0c5SRoman Gushchin 3856aca78307SMiaohe Lin total += huge_page_size(h) * count; 3857fcb2b0c5SRoman Gushchin 3858fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3859e1759c21SAlexey Dobriyan seq_printf(m, 38601da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 38611da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3862b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 38637893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 38644f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3865fcb2b0c5SRoman Gushchin count, 3866a5516438SAndi Kleen h->free_huge_pages, 3867a5516438SAndi Kleen h->resv_huge_pages, 3868a5516438SAndi Kleen h->surplus_huge_pages, 3869aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3870fcb2b0c5SRoman Gushchin } 3871fcb2b0c5SRoman Gushchin 3872aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 38731da177e4SLinus Torvalds } 38741da177e4SLinus Torvalds 38757981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 38761da177e4SLinus Torvalds { 3877a5516438SAndi Kleen struct hstate *h = &default_hstate; 38787981593bSJoe Perches 3879457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3880457c1b27SNishanth Aravamudan return 0; 38817981593bSJoe Perches 38827981593bSJoe Perches return sysfs_emit_at(buf, len, 38831da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3884a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3885a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3886a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3887a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3888a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 38891da177e4SLinus Torvalds } 38901da177e4SLinus Torvalds 3891949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3892949f7ec5SDavid Rientjes { 3893949f7ec5SDavid Rientjes struct hstate *h; 3894949f7ec5SDavid Rientjes int nid; 3895949f7ec5SDavid Rientjes 3896457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3897457c1b27SNishanth Aravamudan return; 3898457c1b27SNishanth Aravamudan 3899949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3900949f7ec5SDavid Rientjes for_each_hstate(h) 3901949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3902949f7ec5SDavid Rientjes nid, 3903949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3904949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3905949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3906aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3907949f7ec5SDavid Rientjes } 3908949f7ec5SDavid Rientjes 39095d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 39105d317b2bSNaoya Horiguchi { 39115d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 39125d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 39135d317b2bSNaoya Horiguchi } 39145d317b2bSNaoya Horiguchi 39151da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 39161da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 39171da177e4SLinus Torvalds { 3918d0028588SWanpeng Li struct hstate *h; 3919d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3920d0028588SWanpeng Li 3921d0028588SWanpeng Li for_each_hstate(h) 3922d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3923d0028588SWanpeng Li return nr_total_pages; 39241da177e4SLinus Torvalds } 39251da177e4SLinus Torvalds 3926a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3927fc1b8a73SMel Gorman { 3928fc1b8a73SMel Gorman int ret = -ENOMEM; 3929fc1b8a73SMel Gorman 39300aa7f354SMiaohe Lin if (!delta) 39310aa7f354SMiaohe Lin return 0; 39320aa7f354SMiaohe Lin 3933db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3934fc1b8a73SMel Gorman /* 3935fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3936fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3937fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3938fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3939fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3940fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3941fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3942fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3943fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3944fc1b8a73SMel Gorman * 3945fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3946fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3947fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3948fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3949fc1b8a73SMel Gorman * semantics that cpuset has. 39508ca39e68SMuchun Song * 39518ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 39528ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 39538ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 39548ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 39558ca39e68SMuchun Song * above. 3956fc1b8a73SMel Gorman */ 3957fc1b8a73SMel Gorman if (delta > 0) { 3958a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3959fc1b8a73SMel Gorman goto out; 3960fc1b8a73SMel Gorman 39618ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3962a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3963fc1b8a73SMel Gorman goto out; 3964fc1b8a73SMel Gorman } 3965fc1b8a73SMel Gorman } 3966fc1b8a73SMel Gorman 3967fc1b8a73SMel Gorman ret = 0; 3968fc1b8a73SMel Gorman if (delta < 0) 3969a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3970fc1b8a73SMel Gorman 3971fc1b8a73SMel Gorman out: 3972db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3973fc1b8a73SMel Gorman return ret; 3974fc1b8a73SMel Gorman } 3975fc1b8a73SMel Gorman 397684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 397784afd99bSAndy Whitcroft { 3978f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 397984afd99bSAndy Whitcroft 398084afd99bSAndy Whitcroft /* 398184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 398284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 398384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 398425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 398584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 398684afd99bSAndy Whitcroft * new reference here without additional locking. 398784afd99bSAndy Whitcroft */ 39884e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3989f522c3acSJoonsoo Kim kref_get(&resv->refs); 399084afd99bSAndy Whitcroft } 399184afd99bSAndy Whitcroft 3992a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3993a1e78772SMel Gorman { 3994a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3995f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 399690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 39974e35f483SJoonsoo Kim unsigned long reserve, start, end; 39981c5ecae3SMike Kravetz long gbl_reserve; 399984afd99bSAndy Whitcroft 40004e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 40014e35f483SJoonsoo Kim return; 40024e35f483SJoonsoo Kim 4003a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4004a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 400584afd99bSAndy Whitcroft 40064e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4007e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 40087251ff78SAdam Litke if (reserve) { 40091c5ecae3SMike Kravetz /* 40101c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 40111c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 40121c5ecae3SMike Kravetz */ 40131c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 40141c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 40157251ff78SAdam Litke } 4016e9fe92aeSMina Almasry 4017e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4018a1e78772SMel Gorman } 4019a1e78772SMel Gorman 402031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 402131383c68SDan Williams { 402231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 402331383c68SDan Williams return -EINVAL; 402431383c68SDan Williams return 0; 402531383c68SDan Williams } 402631383c68SDan Williams 402705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 402805ea8860SDan Williams { 4029aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 403005ea8860SDan Williams } 403105ea8860SDan Williams 40321da177e4SLinus Torvalds /* 40331da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 40341da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 40356c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 40361da177e4SLinus Torvalds * this far. 40371da177e4SLinus Torvalds */ 4038b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 40391da177e4SLinus Torvalds { 40401da177e4SLinus Torvalds BUG(); 4041d0217ac0SNick Piggin return 0; 40421da177e4SLinus Torvalds } 40431da177e4SLinus Torvalds 4044eec3636aSJane Chu /* 4045eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4046eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4047eec3636aSJane Chu * This is because under System V memory model, mappings created via 4048eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4049eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4050eec3636aSJane Chu */ 4051f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4052d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 405384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4054a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4055dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 405605ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 40571da177e4SLinus Torvalds }; 40581da177e4SLinus Torvalds 40591e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 40601e8f889bSDavid Gibson int writable) 406163551ae0SDavid Gibson { 406263551ae0SDavid Gibson pte_t entry; 4063*79c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 406463551ae0SDavid Gibson 40651e8f889bSDavid Gibson if (writable) { 4066106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4067106c992aSGerald Schaefer vma->vm_page_prot))); 406863551ae0SDavid Gibson } else { 4069106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4070106c992aSGerald Schaefer vma->vm_page_prot)); 407163551ae0SDavid Gibson } 407263551ae0SDavid Gibson entry = pte_mkyoung(entry); 407363551ae0SDavid Gibson entry = pte_mkhuge(entry); 4074*79c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 407563551ae0SDavid Gibson 407663551ae0SDavid Gibson return entry; 407763551ae0SDavid Gibson } 407863551ae0SDavid Gibson 40791e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 40801e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 40811e8f889bSDavid Gibson { 40821e8f889bSDavid Gibson pte_t entry; 40831e8f889bSDavid Gibson 4084106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 408532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 40864b3073e1SRussell King update_mmu_cache(vma, address, ptep); 40871e8f889bSDavid Gibson } 40881e8f889bSDavid Gibson 4089d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 40904a705fefSNaoya Horiguchi { 40914a705fefSNaoya Horiguchi swp_entry_t swp; 40924a705fefSNaoya Horiguchi 40934a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4094d5ed7444SAneesh Kumar K.V return false; 40954a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4096d79d176aSBaoquan He if (is_migration_entry(swp)) 4097d5ed7444SAneesh Kumar K.V return true; 40984a705fefSNaoya Horiguchi else 4099d5ed7444SAneesh Kumar K.V return false; 41004a705fefSNaoya Horiguchi } 41014a705fefSNaoya Horiguchi 41023e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 41034a705fefSNaoya Horiguchi { 41044a705fefSNaoya Horiguchi swp_entry_t swp; 41054a705fefSNaoya Horiguchi 41064a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 41073e5c3600SBaoquan He return false; 41084a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4109d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 41103e5c3600SBaoquan He return true; 41114a705fefSNaoya Horiguchi else 41123e5c3600SBaoquan He return false; 41134a705fefSNaoya Horiguchi } 41141e8f889bSDavid Gibson 41154eae4efaSPeter Xu static void 41164eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 41174eae4efaSPeter Xu struct page *new_page) 41184eae4efaSPeter Xu { 41194eae4efaSPeter Xu __SetPageUptodate(new_page); 41204eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 41214eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 41224eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 41234eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 41244eae4efaSPeter Xu SetHPageMigratable(new_page); 41254eae4efaSPeter Xu } 41264eae4efaSPeter Xu 412763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 412863551ae0SDavid Gibson struct vm_area_struct *vma) 412963551ae0SDavid Gibson { 41305e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 413163551ae0SDavid Gibson struct page *ptepage; 41321c59827dSHugh Dickins unsigned long addr; 4133ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 4134a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4135a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 41364eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4137c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4138ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4139e8569dd2SAndreas Sandberg int ret = 0; 41401e8f889bSDavid Gibson 4141ac46d4f3SJérôme Glisse if (cow) { 41427269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 41436f4f13e8SJérôme Glisse vma->vm_start, 4144ac46d4f3SJérôme Glisse vma->vm_end); 4145ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4146c0d0381aSMike Kravetz } else { 4147c0d0381aSMike Kravetz /* 4148c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4149c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4150c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4151c0d0381aSMike Kravetz * huge_pmd_unshare. 4152c0d0381aSMike Kravetz */ 4153c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4154ac46d4f3SJérôme Glisse } 4155e8569dd2SAndreas Sandberg 4156a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 4157cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 41587868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4159c74df32cSHugh Dickins if (!src_pte) 4160c74df32cSHugh Dickins continue; 4161aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 4162e8569dd2SAndreas Sandberg if (!dst_pte) { 4163e8569dd2SAndreas Sandberg ret = -ENOMEM; 4164e8569dd2SAndreas Sandberg break; 4165e8569dd2SAndreas Sandberg } 4166c5c99429SLarry Woodman 41675e41540cSMike Kravetz /* 41685e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 41695e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 41705e41540cSMike Kravetz * 41715e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 41725e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 41735e41540cSMike Kravetz * Check here before taking page table lock, and once again 41745e41540cSMike Kravetz * after taking the lock below. 41755e41540cSMike Kravetz */ 41765e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 41775e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 4178c5c99429SLarry Woodman continue; 4179c5c99429SLarry Woodman 4180cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4181cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4182cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 41834a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 41845e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 41854eae4efaSPeter Xu again: 41865e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 41875e41540cSMike Kravetz /* 41885e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 41895e41540cSMike Kravetz * unlikely case dst entry !none as this implies 41905e41540cSMike Kravetz * sharing with another vma. 41915e41540cSMike Kravetz */ 41924a705fefSNaoya Horiguchi ; 41934a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 41944a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 41954a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 41964a705fefSNaoya Horiguchi 41974a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 41984a705fefSNaoya Horiguchi /* 41994a705fefSNaoya Horiguchi * COW mappings require pages in both 42004a705fefSNaoya Horiguchi * parent and child to be set to read. 42014a705fefSNaoya Horiguchi */ 42024a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 42034a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4204e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4205e5251fd4SPunit Agrawal entry, sz); 42064a705fefSNaoya Horiguchi } 4207e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 42084a705fefSNaoya Horiguchi } else { 42094eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 42104eae4efaSPeter Xu ptepage = pte_page(entry); 42114eae4efaSPeter Xu get_page(ptepage); 42124eae4efaSPeter Xu 42134eae4efaSPeter Xu /* 42144eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 42154eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 42164eae4efaSPeter Xu * COW earlier during fork. 42174eae4efaSPeter Xu * 42184eae4efaSPeter Xu * When pre-allocating the page or copying data, we 42194eae4efaSPeter Xu * need to be without the pgtable locks since we could 42204eae4efaSPeter Xu * sleep during the process. 42214eae4efaSPeter Xu */ 42224eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 42234eae4efaSPeter Xu pte_t src_pte_old = entry; 42244eae4efaSPeter Xu struct page *new; 42254eae4efaSPeter Xu 42264eae4efaSPeter Xu spin_unlock(src_ptl); 42274eae4efaSPeter Xu spin_unlock(dst_ptl); 42284eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 42294eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 42304eae4efaSPeter Xu if (IS_ERR(new)) { 42314eae4efaSPeter Xu put_page(ptepage); 42324eae4efaSPeter Xu ret = PTR_ERR(new); 42334eae4efaSPeter Xu break; 42344eae4efaSPeter Xu } 42354eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 42364eae4efaSPeter Xu npages); 42374eae4efaSPeter Xu put_page(ptepage); 42384eae4efaSPeter Xu 42394eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 42404eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 42414eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 42424eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 42434eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 42444eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4245846be085SMike Kravetz restore_reserve_on_error(h, vma, addr, 4246846be085SMike Kravetz new); 42474eae4efaSPeter Xu put_page(new); 42484eae4efaSPeter Xu /* dst_entry won't change as in child */ 42494eae4efaSPeter Xu goto again; 42504eae4efaSPeter Xu } 42514eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 42524eae4efaSPeter Xu spin_unlock(src_ptl); 42534eae4efaSPeter Xu spin_unlock(dst_ptl); 42544eae4efaSPeter Xu continue; 42554eae4efaSPeter Xu } 42564eae4efaSPeter Xu 425734ee645eSJoerg Roedel if (cow) { 42580f10851eSJérôme Glisse /* 42590f10851eSJérôme Glisse * No need to notify as we are downgrading page 42600f10851eSJérôme Glisse * table protection not changing it to point 42610f10851eSJérôme Glisse * to a new page. 42620f10851eSJérôme Glisse * 4263ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 42640f10851eSJérôme Glisse */ 42657f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 426684894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 426734ee645eSJoerg Roedel } 42684eae4efaSPeter Xu 426953f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 427063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 42714eae4efaSPeter Xu hugetlb_count_add(npages, dst); 42721c59827dSHugh Dickins } 4273cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4274cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 427563551ae0SDavid Gibson } 427663551ae0SDavid Gibson 4277e8569dd2SAndreas Sandberg if (cow) 4278ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4279c0d0381aSMike Kravetz else 4280c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4281e8569dd2SAndreas Sandberg 4282e8569dd2SAndreas Sandberg return ret; 428363551ae0SDavid Gibson } 428463551ae0SDavid Gibson 428524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 428624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 428724669e58SAneesh Kumar K.V struct page *ref_page) 428863551ae0SDavid Gibson { 428963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 429063551ae0SDavid Gibson unsigned long address; 4291c7546f8fSDavid Gibson pte_t *ptep; 429263551ae0SDavid Gibson pte_t pte; 4293cb900f41SKirill A. Shutemov spinlock_t *ptl; 429463551ae0SDavid Gibson struct page *page; 4295a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4296a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4297ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4298a5516438SAndi Kleen 429963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4300a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4301a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 430263551ae0SDavid Gibson 430307e32661SAneesh Kumar K.V /* 430407e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 430507e32661SAneesh Kumar K.V * to huge page. 430607e32661SAneesh Kumar K.V */ 4307ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 430824669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4309dff11abeSMike Kravetz 4310dff11abeSMike Kravetz /* 4311dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4312dff11abeSMike Kravetz */ 43136f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 43146f4f13e8SJérôme Glisse end); 4315ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4316ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4317569f48b8SHillf Danton address = start; 4318569f48b8SHillf Danton for (; address < end; address += sz) { 43197868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4320c7546f8fSDavid Gibson if (!ptep) 4321c7546f8fSDavid Gibson continue; 4322c7546f8fSDavid Gibson 4323cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 432434ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 432531d49da5SAneesh Kumar K.V spin_unlock(ptl); 4326dff11abeSMike Kravetz /* 4327dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 4328dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 4329dff11abeSMike Kravetz */ 433031d49da5SAneesh Kumar K.V continue; 433131d49da5SAneesh Kumar K.V } 433239dde65cSChen, Kenneth W 43336629326bSHillf Danton pte = huge_ptep_get(ptep); 433431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 433531d49da5SAneesh Kumar K.V spin_unlock(ptl); 433631d49da5SAneesh Kumar K.V continue; 433731d49da5SAneesh Kumar K.V } 43386629326bSHillf Danton 43396629326bSHillf Danton /* 43409fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 43419fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 43426629326bSHillf Danton */ 43439fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 43449386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 434531d49da5SAneesh Kumar K.V spin_unlock(ptl); 434631d49da5SAneesh Kumar K.V continue; 43478c4894c6SNaoya Horiguchi } 43486629326bSHillf Danton 43496629326bSHillf Danton page = pte_page(pte); 435004f2cbe3SMel Gorman /* 435104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 435204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 435304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 435404f2cbe3SMel Gorman */ 435504f2cbe3SMel Gorman if (ref_page) { 435631d49da5SAneesh Kumar K.V if (page != ref_page) { 435731d49da5SAneesh Kumar K.V spin_unlock(ptl); 435831d49da5SAneesh Kumar K.V continue; 435931d49da5SAneesh Kumar K.V } 436004f2cbe3SMel Gorman /* 436104f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 436204f2cbe3SMel Gorman * future faults in this VMA will fail rather than 436304f2cbe3SMel Gorman * looking like data was lost 436404f2cbe3SMel Gorman */ 436504f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 436604f2cbe3SMel Gorman } 436704f2cbe3SMel Gorman 4368c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4369b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4370106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 43716649a386SKen Chen set_page_dirty(page); 43729e81130bSHillf Danton 43735d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4374d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 437531d49da5SAneesh Kumar K.V 4376cb900f41SKirill A. Shutemov spin_unlock(ptl); 4377e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 437824669e58SAneesh Kumar K.V /* 437931d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 438024669e58SAneesh Kumar K.V */ 438131d49da5SAneesh Kumar K.V if (ref_page) 438231d49da5SAneesh Kumar K.V break; 4383fe1668aeSChen, Kenneth W } 4384ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 438524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 43861da177e4SLinus Torvalds } 438763551ae0SDavid Gibson 4388d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4389d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4390d833352aSMel Gorman unsigned long end, struct page *ref_page) 4391d833352aSMel Gorman { 4392d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4393d833352aSMel Gorman 4394d833352aSMel Gorman /* 4395d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4396d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4397d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4398d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4399d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4400c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4401d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4402c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4403d833352aSMel Gorman */ 4404d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4405d833352aSMel Gorman } 4406d833352aSMel Gorman 4407502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 440804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4409502717f4SChen, Kenneth W { 441024669e58SAneesh Kumar K.V struct mmu_gather tlb; 4411dff11abeSMike Kravetz 4412a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 441324669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4414ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4415502717f4SChen, Kenneth W } 4416502717f4SChen, Kenneth W 441704f2cbe3SMel Gorman /* 441804f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4419578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 442004f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 442104f2cbe3SMel Gorman * same region. 442204f2cbe3SMel Gorman */ 44232f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 44242a4b3dedSHarvey Harrison struct page *page, unsigned long address) 442504f2cbe3SMel Gorman { 44267526674dSAdam Litke struct hstate *h = hstate_vma(vma); 442704f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 442804f2cbe3SMel Gorman struct address_space *mapping; 442904f2cbe3SMel Gorman pgoff_t pgoff; 443004f2cbe3SMel Gorman 443104f2cbe3SMel Gorman /* 443204f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 443304f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 443404f2cbe3SMel Gorman */ 44357526674dSAdam Litke address = address & huge_page_mask(h); 443636e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 443736e4f20aSMichal Hocko vma->vm_pgoff; 443893c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 443904f2cbe3SMel Gorman 44404eb2b1dcSMel Gorman /* 44414eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 44424eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 44434eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 44444eb2b1dcSMel Gorman */ 444583cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 44466b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 444704f2cbe3SMel Gorman /* Do not unmap the current VMA */ 444804f2cbe3SMel Gorman if (iter_vma == vma) 444904f2cbe3SMel Gorman continue; 445004f2cbe3SMel Gorman 445104f2cbe3SMel Gorman /* 44522f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 44532f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 44542f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 44552f84a899SMel Gorman */ 44562f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 44572f84a899SMel Gorman continue; 44582f84a899SMel Gorman 44592f84a899SMel Gorman /* 446004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 446104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 446204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 446304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 446404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 446504f2cbe3SMel Gorman */ 446604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 446724669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 446824669e58SAneesh Kumar K.V address + huge_page_size(h), page); 446904f2cbe3SMel Gorman } 447083cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 447104f2cbe3SMel Gorman } 447204f2cbe3SMel Gorman 44730fe6e20bSNaoya Horiguchi /* 44740fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4475ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4476ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4477ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 44780fe6e20bSNaoya Horiguchi */ 44792b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4480974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4481cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 44821e8f889bSDavid Gibson { 44833999f52eSAneesh Kumar K.V pte_t pte; 4484a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44851e8f889bSDavid Gibson struct page *old_page, *new_page; 44862b740303SSouptick Joarder int outside_reserve = 0; 44872b740303SSouptick Joarder vm_fault_t ret = 0; 4488974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4489ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 44901e8f889bSDavid Gibson 44913999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 44921e8f889bSDavid Gibson old_page = pte_page(pte); 44931e8f889bSDavid Gibson 449404f2cbe3SMel Gorman retry_avoidcopy: 44951e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 44961e8f889bSDavid Gibson * and just make the page writable */ 449737a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 44985a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 44995b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 450083c54070SNick Piggin return 0; 45011e8f889bSDavid Gibson } 45021e8f889bSDavid Gibson 450304f2cbe3SMel Gorman /* 450404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 450504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 450604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 450704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 450804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 450904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 451004f2cbe3SMel Gorman * of the full address range. 451104f2cbe3SMel Gorman */ 45125944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 451304f2cbe3SMel Gorman old_page != pagecache_page) 451404f2cbe3SMel Gorman outside_reserve = 1; 451504f2cbe3SMel Gorman 451609cbfeafSKirill A. Shutemov get_page(old_page); 4517b76c8cfbSLarry Woodman 4518ad4404a2SDavidlohr Bueso /* 4519ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4520ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4521ad4404a2SDavidlohr Bueso */ 4522cb900f41SKirill A. Shutemov spin_unlock(ptl); 45235b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 45241e8f889bSDavid Gibson 45252fc39cecSAdam Litke if (IS_ERR(new_page)) { 452604f2cbe3SMel Gorman /* 452704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 452804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 452904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 453004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 453104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 453204f2cbe3SMel Gorman */ 453304f2cbe3SMel Gorman if (outside_reserve) { 4534e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4535e7dd91c4SMike Kravetz pgoff_t idx; 4536e7dd91c4SMike Kravetz u32 hash; 4537e7dd91c4SMike Kravetz 453809cbfeafSKirill A. Shutemov put_page(old_page); 453904f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4540e7dd91c4SMike Kravetz /* 4541e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4542e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4543e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4544e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4545e7dd91c4SMike Kravetz * PMD sharing. 4546e7dd91c4SMike Kravetz * 4547e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4548e7dd91c4SMike Kravetz */ 4549e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4550e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4551e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4552e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4553e7dd91c4SMike Kravetz 45545b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4555e7dd91c4SMike Kravetz 4556e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4557e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4558cb900f41SKirill A. Shutemov spin_lock(ptl); 45595b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4560a9af0c5dSNaoya Horiguchi if (likely(ptep && 4561a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 456204f2cbe3SMel Gorman goto retry_avoidcopy; 4563a734bcc8SHillf Danton /* 4564cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4565cb900f41SKirill A. Shutemov * lock, and our job is done. 4566a734bcc8SHillf Danton */ 4567a734bcc8SHillf Danton return 0; 456804f2cbe3SMel Gorman } 456904f2cbe3SMel Gorman 45702b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4571ad4404a2SDavidlohr Bueso goto out_release_old; 45721e8f889bSDavid Gibson } 45731e8f889bSDavid Gibson 45740fe6e20bSNaoya Horiguchi /* 45750fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 45760fe6e20bSNaoya Horiguchi * anon_vma prepared. 45770fe6e20bSNaoya Horiguchi */ 457844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4579ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4580ad4404a2SDavidlohr Bueso goto out_release_all; 458144e2aa93SDean Nelson } 45820fe6e20bSNaoya Horiguchi 4583974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 458447ad8475SAndrea Arcangeli pages_per_huge_page(h)); 45850ed361deSNick Piggin __SetPageUptodate(new_page); 45861e8f889bSDavid Gibson 45877269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 45886f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4589ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4590ad4404a2SDavidlohr Bueso 4591b76c8cfbSLarry Woodman /* 4592cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4593b76c8cfbSLarry Woodman * before the page tables are altered 4594b76c8cfbSLarry Woodman */ 4595cb900f41SKirill A. Shutemov spin_lock(ptl); 45965b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4597a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4598d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 459907443a85SJoonsoo Kim 46001e8f889bSDavid Gibson /* Break COW */ 46015b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4602ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 46035b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 46041e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4605d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 46065b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 46078f251a3dSMike Kravetz SetHPageMigratable(new_page); 46081e8f889bSDavid Gibson /* Make the old page be freed below */ 46091e8f889bSDavid Gibson new_page = old_page; 46101e8f889bSDavid Gibson } 4611cb900f41SKirill A. Shutemov spin_unlock(ptl); 4612ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4613ad4404a2SDavidlohr Bueso out_release_all: 46145b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 461509cbfeafSKirill A. Shutemov put_page(new_page); 4616ad4404a2SDavidlohr Bueso out_release_old: 461709cbfeafSKirill A. Shutemov put_page(old_page); 46188312034fSJoonsoo Kim 4619ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4620ad4404a2SDavidlohr Bueso return ret; 46211e8f889bSDavid Gibson } 46221e8f889bSDavid Gibson 462304f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4624a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4625a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 462604f2cbe3SMel Gorman { 462704f2cbe3SMel Gorman struct address_space *mapping; 4628e7c4b0bfSAndy Whitcroft pgoff_t idx; 462904f2cbe3SMel Gorman 463004f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4631a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 463204f2cbe3SMel Gorman 463304f2cbe3SMel Gorman return find_lock_page(mapping, idx); 463404f2cbe3SMel Gorman } 463504f2cbe3SMel Gorman 46363ae77f43SHugh Dickins /* 46373ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 46383ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 46393ae77f43SHugh Dickins */ 46403ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 46412a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 46422a15efc9SHugh Dickins { 46432a15efc9SHugh Dickins struct address_space *mapping; 46442a15efc9SHugh Dickins pgoff_t idx; 46452a15efc9SHugh Dickins struct page *page; 46462a15efc9SHugh Dickins 46472a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 46482a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 46492a15efc9SHugh Dickins 46502a15efc9SHugh Dickins page = find_get_page(mapping, idx); 46512a15efc9SHugh Dickins if (page) 46522a15efc9SHugh Dickins put_page(page); 46532a15efc9SHugh Dickins return page != NULL; 46542a15efc9SHugh Dickins } 46552a15efc9SHugh Dickins 4656ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4657ab76ad54SMike Kravetz pgoff_t idx) 4658ab76ad54SMike Kravetz { 4659ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4660ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4661ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4662ab76ad54SMike Kravetz 4663ab76ad54SMike Kravetz if (err) 4664ab76ad54SMike Kravetz return err; 4665d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4666ab76ad54SMike Kravetz 466722146c3cSMike Kravetz /* 466822146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 466922146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 467022146c3cSMike Kravetz */ 467122146c3cSMike Kravetz set_page_dirty(page); 467222146c3cSMike Kravetz 4673ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4674ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4675ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4676ab76ad54SMike Kravetz return 0; 4677ab76ad54SMike Kravetz } 4678ab76ad54SMike Kravetz 46797677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 46807677f7fdSAxel Rasmussen struct address_space *mapping, 46817677f7fdSAxel Rasmussen pgoff_t idx, 46827677f7fdSAxel Rasmussen unsigned int flags, 46837677f7fdSAxel Rasmussen unsigned long haddr, 46847677f7fdSAxel Rasmussen unsigned long reason) 46857677f7fdSAxel Rasmussen { 46867677f7fdSAxel Rasmussen vm_fault_t ret; 46877677f7fdSAxel Rasmussen u32 hash; 46887677f7fdSAxel Rasmussen struct vm_fault vmf = { 46897677f7fdSAxel Rasmussen .vma = vma, 46907677f7fdSAxel Rasmussen .address = haddr, 46917677f7fdSAxel Rasmussen .flags = flags, 46927677f7fdSAxel Rasmussen 46937677f7fdSAxel Rasmussen /* 46947677f7fdSAxel Rasmussen * Hard to debug if it ends up being 46957677f7fdSAxel Rasmussen * used by a callee that assumes 46967677f7fdSAxel Rasmussen * something about the other 46977677f7fdSAxel Rasmussen * uninitialized fields... same as in 46987677f7fdSAxel Rasmussen * memory.c 46997677f7fdSAxel Rasmussen */ 47007677f7fdSAxel Rasmussen }; 47017677f7fdSAxel Rasmussen 47027677f7fdSAxel Rasmussen /* 47037677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 47047677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 47057677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 47067677f7fdSAxel Rasmussen */ 47077677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 47087677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 47097677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 47107677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 47117677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 47127677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 47137677f7fdSAxel Rasmussen 47147677f7fdSAxel Rasmussen return ret; 47157677f7fdSAxel Rasmussen } 47167677f7fdSAxel Rasmussen 47172b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 47182b740303SSouptick Joarder struct vm_area_struct *vma, 47198382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4720788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4721ac9b9c66SHugh Dickins { 4722a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 47232b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4724409eb8c2SHillf Danton int anon_rmap = 0; 47254c887265SAdam Litke unsigned long size; 47264c887265SAdam Litke struct page *page; 47271e8f889bSDavid Gibson pte_t new_pte; 4728cb900f41SKirill A. Shutemov spinlock_t *ptl; 4729285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4730cb6acd01SMike Kravetz bool new_page = false; 47314c887265SAdam Litke 473204f2cbe3SMel Gorman /* 473304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 473404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 473525985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 473604f2cbe3SMel Gorman */ 473704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4738910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 473904f2cbe3SMel Gorman current->pid); 474004f2cbe3SMel Gorman return ret; 474104f2cbe3SMel Gorman } 474204f2cbe3SMel Gorman 47434c887265SAdam Litke /* 474487bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 474587bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 474687bf91d3SMike Kravetz * once for faults beyond end of file. 47474c887265SAdam Litke */ 4748a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4749ebed4bfcSHugh Dickins if (idx >= size) 4750ebed4bfcSHugh Dickins goto out; 47511a1aad8aSMike Kravetz 475287bf91d3SMike Kravetz retry: 475387bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 475487bf91d3SMike Kravetz if (!page) { 47557677f7fdSAxel Rasmussen /* Check for page in userfault range */ 47561a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 47577677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 47587677f7fdSAxel Rasmussen flags, haddr, 47597677f7fdSAxel Rasmussen VM_UFFD_MISSING); 47601a1aad8aSMike Kravetz goto out; 47611a1aad8aSMike Kravetz } 47621a1aad8aSMike Kravetz 4763285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 47642fc39cecSAdam Litke if (IS_ERR(page)) { 47654643d67eSMike Kravetz /* 47664643d67eSMike Kravetz * Returning error will result in faulting task being 47674643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 47684643d67eSMike Kravetz * tasks from racing to fault in the same page which 47694643d67eSMike Kravetz * could result in false unable to allocate errors. 47704643d67eSMike Kravetz * Page migration does not take the fault mutex, but 47714643d67eSMike Kravetz * does a clear then write of pte's under page table 47724643d67eSMike Kravetz * lock. Page fault code could race with migration, 47734643d67eSMike Kravetz * notice the clear pte and try to allocate a page 47744643d67eSMike Kravetz * here. Before returning error, get ptl and make 47754643d67eSMike Kravetz * sure there really is no pte entry. 47764643d67eSMike Kravetz */ 47774643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 47784643d67eSMike Kravetz ret = 0; 4779d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 47802b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 4781d83e6c8aSMiaohe Lin spin_unlock(ptl); 47826bda666aSChristoph Lameter goto out; 47836bda666aSChristoph Lameter } 478447ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 47850ed361deSNick Piggin __SetPageUptodate(page); 4786cb6acd01SMike Kravetz new_page = true; 4787ac9b9c66SHugh Dickins 4788f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4789ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 47906bda666aSChristoph Lameter if (err) { 47916bda666aSChristoph Lameter put_page(page); 47926bda666aSChristoph Lameter if (err == -EEXIST) 47936bda666aSChristoph Lameter goto retry; 47946bda666aSChristoph Lameter goto out; 47956bda666aSChristoph Lameter } 479623be7468SMel Gorman } else { 47976bda666aSChristoph Lameter lock_page(page); 47980fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 47990fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 48000fe6e20bSNaoya Horiguchi goto backout_unlocked; 480123be7468SMel Gorman } 4802409eb8c2SHillf Danton anon_rmap = 1; 48030fe6e20bSNaoya Horiguchi } 48040fe6e20bSNaoya Horiguchi } else { 480557303d80SAndy Whitcroft /* 4806998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4807998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4808998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4809fd6a03edSNaoya Horiguchi */ 4810fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 48110eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4812972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4813fd6a03edSNaoya Horiguchi goto backout_unlocked; 48146bda666aSChristoph Lameter } 48157677f7fdSAxel Rasmussen 48167677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 48177677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 48187677f7fdSAxel Rasmussen unlock_page(page); 48197677f7fdSAxel Rasmussen put_page(page); 48207677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 48217677f7fdSAxel Rasmussen flags, haddr, 48227677f7fdSAxel Rasmussen VM_UFFD_MINOR); 48237677f7fdSAxel Rasmussen goto out; 48247677f7fdSAxel Rasmussen } 4825998b4382SNaoya Horiguchi } 48261e8f889bSDavid Gibson 482757303d80SAndy Whitcroft /* 482857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 482957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 483057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 483157303d80SAndy Whitcroft * the spinlock. 483257303d80SAndy Whitcroft */ 48335e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4834285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 48352b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 48362b26736cSAndy Whitcroft goto backout_unlocked; 48372b26736cSAndy Whitcroft } 48385e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4839285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 48405e911373SMike Kravetz } 484157303d80SAndy Whitcroft 48428bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 484383c54070SNick Piggin ret = 0; 48447f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 48454c887265SAdam Litke goto backout; 48464c887265SAdam Litke 484707443a85SJoonsoo Kim if (anon_rmap) { 4848d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4849285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4850ac714904SChoi Gi-yong } else 485153f9263bSKirill A. Shutemov page_dup_rmap(page, true); 48521e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 48531e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4854285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 48551e8f889bSDavid Gibson 48565d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4857788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 48581e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4859974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 48601e8f889bSDavid Gibson } 48611e8f889bSDavid Gibson 4862cb900f41SKirill A. Shutemov spin_unlock(ptl); 4863cb6acd01SMike Kravetz 4864cb6acd01SMike Kravetz /* 48658f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 48668f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 48678f251a3dSMike Kravetz * been isolated for migration. 4868cb6acd01SMike Kravetz */ 4869cb6acd01SMike Kravetz if (new_page) 48708f251a3dSMike Kravetz SetHPageMigratable(page); 4871cb6acd01SMike Kravetz 48724c887265SAdam Litke unlock_page(page); 48734c887265SAdam Litke out: 4874ac9b9c66SHugh Dickins return ret; 48754c887265SAdam Litke 48764c887265SAdam Litke backout: 4877cb900f41SKirill A. Shutemov spin_unlock(ptl); 48782b26736cSAndy Whitcroft backout_unlocked: 48794c887265SAdam Litke unlock_page(page); 4880285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 48814c887265SAdam Litke put_page(page); 48824c887265SAdam Litke goto out; 4883ac9b9c66SHugh Dickins } 4884ac9b9c66SHugh Dickins 48858382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4886188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 48878382d914SDavidlohr Bueso { 48888382d914SDavidlohr Bueso unsigned long key[2]; 48898382d914SDavidlohr Bueso u32 hash; 48908382d914SDavidlohr Bueso 48918382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 48928382d914SDavidlohr Bueso key[1] = idx; 48938382d914SDavidlohr Bueso 489455254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 48958382d914SDavidlohr Bueso 48968382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 48978382d914SDavidlohr Bueso } 48988382d914SDavidlohr Bueso #else 48998382d914SDavidlohr Bueso /* 49006c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 49018382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 49028382d914SDavidlohr Bueso */ 4903188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 49048382d914SDavidlohr Bueso { 49058382d914SDavidlohr Bueso return 0; 49068382d914SDavidlohr Bueso } 49078382d914SDavidlohr Bueso #endif 49088382d914SDavidlohr Bueso 49092b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4910788c7df4SHugh Dickins unsigned long address, unsigned int flags) 491186e5216fSAdam Litke { 49128382d914SDavidlohr Bueso pte_t *ptep, entry; 4913cb900f41SKirill A. Shutemov spinlock_t *ptl; 49142b740303SSouptick Joarder vm_fault_t ret; 49158382d914SDavidlohr Bueso u32 hash; 49168382d914SDavidlohr Bueso pgoff_t idx; 49170fe6e20bSNaoya Horiguchi struct page *page = NULL; 491857303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4919a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49208382d914SDavidlohr Bueso struct address_space *mapping; 49210f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4922285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 492386e5216fSAdam Litke 4924285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4925fd6a03edSNaoya Horiguchi if (ptep) { 4926c0d0381aSMike Kravetz /* 4927c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4928c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4929c0d0381aSMike Kravetz * not actually modifying content here. 4930c0d0381aSMike Kravetz */ 4931fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4932290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4933cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4934290408d4SNaoya Horiguchi return 0; 4935290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4936aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4937972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4938b43a9990SMike Kravetz } 49398382d914SDavidlohr Bueso 4940c0d0381aSMike Kravetz /* 4941c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 494287bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 494387bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 494487bf91d3SMike Kravetz * and making the ptep no longer valid. 494587bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4946c0d0381aSMike Kravetz * 4947c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4948c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4949c0d0381aSMike Kravetz * something has changed. 4950c0d0381aSMike Kravetz */ 4951ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4952c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4953aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 4954c0d0381aSMike Kravetz if (!ptep) { 4955c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4956c0d0381aSMike Kravetz return VM_FAULT_OOM; 4957c0d0381aSMike Kravetz } 4958ddeaab32SMike Kravetz 49593935baa9SDavid Gibson /* 49603935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 49613935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 49623935baa9SDavid Gibson * the same page in the page cache. 49633935baa9SDavid Gibson */ 4964c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4965188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4966c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 49678382d914SDavidlohr Bueso 49687f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 49697f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 49708382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4971b4d1d99fSDavid Gibson goto out_mutex; 49723935baa9SDavid Gibson } 497386e5216fSAdam Litke 497483c54070SNick Piggin ret = 0; 49751e8f889bSDavid Gibson 497657303d80SAndy Whitcroft /* 49770f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 49780f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 49797c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 49807c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 49817c8de358SEthon Paul * properly handle it. 49820f792cf9SNaoya Horiguchi */ 49830f792cf9SNaoya Horiguchi if (!pte_present(entry)) 49840f792cf9SNaoya Horiguchi goto out_mutex; 49850f792cf9SNaoya Horiguchi 49860f792cf9SNaoya Horiguchi /* 498757303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 498857303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 498957303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 499057303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 499157303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 499257303d80SAndy Whitcroft * consumed. 499357303d80SAndy Whitcroft */ 4994106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4995285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 49962b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4997b4d1d99fSDavid Gibson goto out_mutex; 49982b26736cSAndy Whitcroft } 49995e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5000285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 500157303d80SAndy Whitcroft 5002f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 500357303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 5004285b8dcaSHuang Ying vma, haddr); 500557303d80SAndy Whitcroft } 500657303d80SAndy Whitcroft 50070f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 50080fe6e20bSNaoya Horiguchi 50091e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 5010b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5011cb900f41SKirill A. Shutemov goto out_ptl; 5012b4d1d99fSDavid Gibson 50130f792cf9SNaoya Horiguchi /* 50140f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 50150f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 50160f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 50170f792cf9SNaoya Horiguchi */ 50180f792cf9SNaoya Horiguchi page = pte_page(entry); 50190f792cf9SNaoya Horiguchi if (page != pagecache_page) 50200f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 50210f792cf9SNaoya Horiguchi need_wait_lock = 1; 50220f792cf9SNaoya Horiguchi goto out_ptl; 50230f792cf9SNaoya Horiguchi } 50240f792cf9SNaoya Horiguchi 50250f792cf9SNaoya Horiguchi get_page(page); 5026b4d1d99fSDavid Gibson 5027788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 5028106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5029974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 5030cb900f41SKirill A. Shutemov pagecache_page, ptl); 50310f792cf9SNaoya Horiguchi goto out_put_page; 5032b4d1d99fSDavid Gibson } 5033106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5034b4d1d99fSDavid Gibson } 5035b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5036285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5037788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5038285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 50390f792cf9SNaoya Horiguchi out_put_page: 50400f792cf9SNaoya Horiguchi if (page != pagecache_page) 50410f792cf9SNaoya Horiguchi unlock_page(page); 50420f792cf9SNaoya Horiguchi put_page(page); 5043cb900f41SKirill A. Shutemov out_ptl: 5044cb900f41SKirill A. Shutemov spin_unlock(ptl); 504557303d80SAndy Whitcroft 504657303d80SAndy Whitcroft if (pagecache_page) { 504757303d80SAndy Whitcroft unlock_page(pagecache_page); 504857303d80SAndy Whitcroft put_page(pagecache_page); 504957303d80SAndy Whitcroft } 5050b4d1d99fSDavid Gibson out_mutex: 5051c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5052c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 50530f792cf9SNaoya Horiguchi /* 50540f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 50550f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 50560f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 50570f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 50580f792cf9SNaoya Horiguchi * here without taking refcount. 50590f792cf9SNaoya Horiguchi */ 50600f792cf9SNaoya Horiguchi if (need_wait_lock) 50610f792cf9SNaoya Horiguchi wait_on_page_locked(page); 50621e8f889bSDavid Gibson return ret; 506386e5216fSAdam Litke } 506486e5216fSAdam Litke 5065714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 50668fb5debcSMike Kravetz /* 50678fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 50688fb5debcSMike Kravetz * modifications for huge pages. 50698fb5debcSMike Kravetz */ 50708fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 50718fb5debcSMike Kravetz pte_t *dst_pte, 50728fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 50738fb5debcSMike Kravetz unsigned long dst_addr, 50748fb5debcSMike Kravetz unsigned long src_addr, 5075f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 50768fb5debcSMike Kravetz struct page **pagep) 50778fb5debcSMike Kravetz { 5078f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 50791e392147SAndrea Arcangeli struct address_space *mapping; 50801e392147SAndrea Arcangeli pgoff_t idx; 50811e392147SAndrea Arcangeli unsigned long size; 50821c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 50838fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 50848fb5debcSMike Kravetz pte_t _dst_pte; 50858fb5debcSMike Kravetz spinlock_t *ptl; 50868fb5debcSMike Kravetz int ret; 50878fb5debcSMike Kravetz struct page *page; 5088f6191471SAxel Rasmussen int writable; 50898fb5debcSMike Kravetz 5090f6191471SAxel Rasmussen mapping = dst_vma->vm_file->f_mapping; 5091f6191471SAxel Rasmussen idx = vma_hugecache_offset(h, dst_vma, dst_addr); 5092f6191471SAxel Rasmussen 5093f6191471SAxel Rasmussen if (is_continue) { 5094f6191471SAxel Rasmussen ret = -EFAULT; 5095f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5096f6191471SAxel Rasmussen if (!page) 5097f6191471SAxel Rasmussen goto out; 5098f6191471SAxel Rasmussen } else if (!*pagep) { 5099d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5100d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5101d84cf06eSMina Almasry */ 5102d84cf06eSMina Almasry if (vm_shared && 5103d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5104d84cf06eSMina Almasry ret = -EEXIST; 51058fb5debcSMike Kravetz goto out; 5106d84cf06eSMina Almasry } 5107d84cf06eSMina Almasry 5108d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5109d84cf06eSMina Almasry if (IS_ERR(page)) { 5110d84cf06eSMina Almasry ret = -ENOMEM; 5111d84cf06eSMina Almasry goto out; 5112d84cf06eSMina Almasry } 51138fb5debcSMike Kravetz 51148fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 51158fb5debcSMike Kravetz (const void __user *) src_addr, 5116810a56b9SMike Kravetz pages_per_huge_page(h), false); 51178fb5debcSMike Kravetz 5118c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 51198fb5debcSMike Kravetz if (unlikely(ret)) { 51209e368259SAndrea Arcangeli ret = -ENOENT; 51218fb5debcSMike Kravetz *pagep = page; 51228fb5debcSMike Kravetz /* don't free the page */ 51238fb5debcSMike Kravetz goto out; 51248fb5debcSMike Kravetz } 51258fb5debcSMike Kravetz } else { 51268fb5debcSMike Kravetz page = *pagep; 51278fb5debcSMike Kravetz *pagep = NULL; 51288fb5debcSMike Kravetz } 51298fb5debcSMike Kravetz 51308fb5debcSMike Kravetz /* 51318fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 51328fb5debcSMike Kravetz * preceding stores to the page contents become visible before 51338fb5debcSMike Kravetz * the set_pte_at() write. 51348fb5debcSMike Kravetz */ 51358fb5debcSMike Kravetz __SetPageUptodate(page); 51368fb5debcSMike Kravetz 5137f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 5138f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 51391e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 51401e392147SAndrea Arcangeli ret = -EFAULT; 51411e392147SAndrea Arcangeli if (idx >= size) 51421e392147SAndrea Arcangeli goto out_release_nounlock; 51431c9e8defSMike Kravetz 51441e392147SAndrea Arcangeli /* 51451e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 51461e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 51471e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 51481e392147SAndrea Arcangeli * the caller. 51491e392147SAndrea Arcangeli */ 51501c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 51511c9e8defSMike Kravetz if (ret) 51521c9e8defSMike Kravetz goto out_release_nounlock; 51531c9e8defSMike Kravetz } 51541c9e8defSMike Kravetz 51558fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 51568fb5debcSMike Kravetz spin_lock(ptl); 51578fb5debcSMike Kravetz 51581e392147SAndrea Arcangeli /* 51591e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 51601e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 51611e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 51621e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 51631e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 51641e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 51651e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 51661e392147SAndrea Arcangeli */ 51671e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 51681e392147SAndrea Arcangeli ret = -EFAULT; 51691e392147SAndrea Arcangeli if (idx >= size) 51701e392147SAndrea Arcangeli goto out_release_unlock; 51711e392147SAndrea Arcangeli 51728fb5debcSMike Kravetz ret = -EEXIST; 51738fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 51748fb5debcSMike Kravetz goto out_release_unlock; 51758fb5debcSMike Kravetz 51761c9e8defSMike Kravetz if (vm_shared) { 51771c9e8defSMike Kravetz page_dup_rmap(page, true); 51781c9e8defSMike Kravetz } else { 5179d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 51808fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 51811c9e8defSMike Kravetz } 51828fb5debcSMike Kravetz 5183f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 5184f6191471SAxel Rasmussen if (is_continue && !vm_shared) 5185f6191471SAxel Rasmussen writable = 0; 5186f6191471SAxel Rasmussen else 5187f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 5188f6191471SAxel Rasmussen 5189f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 5190f6191471SAxel Rasmussen if (writable) 51918fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 51928fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 51938fb5debcSMike Kravetz 51948fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 51958fb5debcSMike Kravetz 51968fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 51978fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 51988fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 51998fb5debcSMike Kravetz 52008fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 52018fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 52028fb5debcSMike Kravetz 52038fb5debcSMike Kravetz spin_unlock(ptl); 5204f6191471SAxel Rasmussen if (!is_continue) 52058f251a3dSMike Kravetz SetHPageMigratable(page); 5206f6191471SAxel Rasmussen if (vm_shared || is_continue) 52071c9e8defSMike Kravetz unlock_page(page); 52088fb5debcSMike Kravetz ret = 0; 52098fb5debcSMike Kravetz out: 52108fb5debcSMike Kravetz return ret; 52118fb5debcSMike Kravetz out_release_unlock: 52128fb5debcSMike Kravetz spin_unlock(ptl); 5213f6191471SAxel Rasmussen if (vm_shared || is_continue) 52141c9e8defSMike Kravetz unlock_page(page); 52155af10dfdSAndrea Arcangeli out_release_nounlock: 5216846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 52178fb5debcSMike Kravetz put_page(page); 52188fb5debcSMike Kravetz goto out; 52198fb5debcSMike Kravetz } 5220714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 52218fb5debcSMike Kravetz 522282e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 522382e5d378SJoao Martins int refs, struct page **pages, 522482e5d378SJoao Martins struct vm_area_struct **vmas) 522582e5d378SJoao Martins { 522682e5d378SJoao Martins int nr; 522782e5d378SJoao Martins 522882e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 522982e5d378SJoao Martins if (likely(pages)) 523082e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 523182e5d378SJoao Martins if (vmas) 523282e5d378SJoao Martins vmas[nr] = vma; 523382e5d378SJoao Martins } 523482e5d378SJoao Martins } 523582e5d378SJoao Martins 523628a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 523763551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 523828a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 52394f6da934SPeter Xu long i, unsigned int flags, int *locked) 524063551ae0SDavid Gibson { 5241d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5242d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 524328a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5244a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52450fa5bc40SJoao Martins int err = -EFAULT, refs; 524663551ae0SDavid Gibson 524763551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 524863551ae0SDavid Gibson pte_t *pte; 5249cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 52502a15efc9SHugh Dickins int absent; 525163551ae0SDavid Gibson struct page *page; 525263551ae0SDavid Gibson 52534c887265SAdam Litke /* 525402057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 525502057967SDavid Rientjes * potentially allocating memory. 525602057967SDavid Rientjes */ 5257fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 525802057967SDavid Rientjes remainder = 0; 525902057967SDavid Rientjes break; 526002057967SDavid Rientjes } 526102057967SDavid Rientjes 526202057967SDavid Rientjes /* 52634c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 52642a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 52654c887265SAdam Litke * first, for the page indexing below to work. 5266cb900f41SKirill A. Shutemov * 5267cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 52684c887265SAdam Litke */ 52697868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 52707868a208SPunit Agrawal huge_page_size(h)); 5271cb900f41SKirill A. Shutemov if (pte) 5272cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 52732a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 527463551ae0SDavid Gibson 52752a15efc9SHugh Dickins /* 52762a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 52773ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 52783ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 52793ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 52803ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 52812a15efc9SHugh Dickins */ 52823ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 52833ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5284cb900f41SKirill A. Shutemov if (pte) 5285cb900f41SKirill A. Shutemov spin_unlock(ptl); 52862a15efc9SHugh Dickins remainder = 0; 52872a15efc9SHugh Dickins break; 52882a15efc9SHugh Dickins } 52892a15efc9SHugh Dickins 52909cc3a5bdSNaoya Horiguchi /* 52919cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 52929cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 52939cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 52949cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 52959cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 52969cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 52979cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 52989cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 52999cc3a5bdSNaoya Horiguchi */ 53009cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5301106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5302106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 53032b740303SSouptick Joarder vm_fault_t ret; 530487ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 53054c887265SAdam Litke 5306cb900f41SKirill A. Shutemov if (pte) 5307cb900f41SKirill A. Shutemov spin_unlock(ptl); 530887ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 530987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 53104f6da934SPeter Xu if (locked) 531171335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 531271335f37SPeter Xu FAULT_FLAG_KILLABLE; 531387ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 531487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 531587ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 531687ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 53174426e945SPeter Xu /* 53184426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 53194426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 53204426e945SPeter Xu */ 532187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 532287ffc118SAndrea Arcangeli } 532387ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 532487ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 53252be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 53261c59827dSHugh Dickins remainder = 0; 53271c59827dSHugh Dickins break; 53281c59827dSHugh Dickins } 532987ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 53304f6da934SPeter Xu if (locked && 53311ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 53324f6da934SPeter Xu *locked = 0; 533387ffc118SAndrea Arcangeli *nr_pages = 0; 533487ffc118SAndrea Arcangeli /* 533587ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 533687ffc118SAndrea Arcangeli * error, it will return zero 533787ffc118SAndrea Arcangeli * instead. 533887ffc118SAndrea Arcangeli * 533987ffc118SAndrea Arcangeli * No need to update "position" as the 534087ffc118SAndrea Arcangeli * caller will not check it after 534187ffc118SAndrea Arcangeli * *nr_pages is set to 0. 534287ffc118SAndrea Arcangeli */ 534387ffc118SAndrea Arcangeli return i; 534487ffc118SAndrea Arcangeli } 534587ffc118SAndrea Arcangeli continue; 534687ffc118SAndrea Arcangeli } 534763551ae0SDavid Gibson 5348a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 53497f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 53508fde12caSLinus Torvalds 53518fde12caSLinus Torvalds /* 5352acbfb087SZhigang Lu * If subpage information not requested, update counters 5353acbfb087SZhigang Lu * and skip the same_page loop below. 5354acbfb087SZhigang Lu */ 5355acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 5356acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 5357acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 5358acbfb087SZhigang Lu vaddr += huge_page_size(h); 5359acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 5360acbfb087SZhigang Lu i += pages_per_huge_page(h); 5361acbfb087SZhigang Lu spin_unlock(ptl); 5362acbfb087SZhigang Lu continue; 5363acbfb087SZhigang Lu } 5364acbfb087SZhigang Lu 536582e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 536682e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 53670fa5bc40SJoao Martins 536882e5d378SJoao Martins if (pages || vmas) 536982e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 537082e5d378SJoao Martins vma, refs, 537182e5d378SJoao Martins likely(pages) ? pages + i : NULL, 537282e5d378SJoao Martins vmas ? vmas + i : NULL); 537363551ae0SDavid Gibson 537482e5d378SJoao Martins if (pages) { 53750fa5bc40SJoao Martins /* 53760fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 53770fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 53780fa5bc40SJoao Martins * checked that the huge page is present in the page 53790fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 53800fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 53810fa5bc40SJoao Martins * head page and tail pages from being rearranged in 53820fa5bc40SJoao Martins * any way. So this page must be available at this 53830fa5bc40SJoao Martins * point, unless the page refcount overflowed: 53840fa5bc40SJoao Martins */ 538582e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 53860fa5bc40SJoao Martins refs, 53870fa5bc40SJoao Martins flags))) { 53880fa5bc40SJoao Martins spin_unlock(ptl); 53890fa5bc40SJoao Martins remainder = 0; 53900fa5bc40SJoao Martins err = -ENOMEM; 53910fa5bc40SJoao Martins break; 53920fa5bc40SJoao Martins } 5393d5d4b0aaSChen, Kenneth W } 539482e5d378SJoao Martins 539582e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 539682e5d378SJoao Martins remainder -= refs; 539782e5d378SJoao Martins i += refs; 539882e5d378SJoao Martins 5399cb900f41SKirill A. Shutemov spin_unlock(ptl); 540063551ae0SDavid Gibson } 540128a35716SMichel Lespinasse *nr_pages = remainder; 540287ffc118SAndrea Arcangeli /* 540387ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 540487ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 540587ffc118SAndrea Arcangeli * branch. 540687ffc118SAndrea Arcangeli */ 540763551ae0SDavid Gibson *position = vaddr; 540863551ae0SDavid Gibson 54092be7cfedSDaniel Jordan return i ? i : err; 541063551ae0SDavid Gibson } 54118f860591SZhang, Yanmin 54127da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 54138f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 54148f860591SZhang, Yanmin { 54158f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 54168f860591SZhang, Yanmin unsigned long start = address; 54178f860591SZhang, Yanmin pte_t *ptep; 54188f860591SZhang, Yanmin pte_t pte; 5419a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 54207da4d641SPeter Zijlstra unsigned long pages = 0; 5421dff11abeSMike Kravetz bool shared_pmd = false; 5422ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5423dff11abeSMike Kravetz 5424dff11abeSMike Kravetz /* 5425dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5426ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5427dff11abeSMike Kravetz * range if PMD sharing is possible. 5428dff11abeSMike Kravetz */ 54297269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 54307269f999SJérôme Glisse 0, vma, mm, start, end); 5431ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 54328f860591SZhang, Yanmin 54338f860591SZhang, Yanmin BUG_ON(address >= end); 5434ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 54358f860591SZhang, Yanmin 5436ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 543783cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5438a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5439cb900f41SKirill A. Shutemov spinlock_t *ptl; 54407868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 54418f860591SZhang, Yanmin if (!ptep) 54428f860591SZhang, Yanmin continue; 5443cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 544434ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 54457da4d641SPeter Zijlstra pages++; 5446cb900f41SKirill A. Shutemov spin_unlock(ptl); 5447dff11abeSMike Kravetz shared_pmd = true; 544839dde65cSChen, Kenneth W continue; 54497da4d641SPeter Zijlstra } 5450a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5451a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5452a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5453a8bda28dSNaoya Horiguchi continue; 5454a8bda28dSNaoya Horiguchi } 5455a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5456a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5457a8bda28dSNaoya Horiguchi 5458a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5459a8bda28dSNaoya Horiguchi pte_t newpte; 5460a8bda28dSNaoya Horiguchi 5461a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5462a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5463e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5464e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5465a8bda28dSNaoya Horiguchi pages++; 5466a8bda28dSNaoya Horiguchi } 5467a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5468a8bda28dSNaoya Horiguchi continue; 5469a8bda28dSNaoya Horiguchi } 5470a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5471023bdd00SAneesh Kumar K.V pte_t old_pte; 5472*79c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 5473023bdd00SAneesh Kumar K.V 5474023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5475023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5476*79c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 5477023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 54787da4d641SPeter Zijlstra pages++; 54798f860591SZhang, Yanmin } 5480cb900f41SKirill A. Shutemov spin_unlock(ptl); 54818f860591SZhang, Yanmin } 5482d833352aSMel Gorman /* 5483c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5484d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5485c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5486dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5487dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5488d833352aSMel Gorman */ 5489dff11abeSMike Kravetz if (shared_pmd) 5490ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5491dff11abeSMike Kravetz else 54925491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 54930f10851eSJérôme Glisse /* 54940f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 54950f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 54960f10851eSJérôme Glisse * 5497ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 54980f10851eSJérôme Glisse */ 549983cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5500ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 55017da4d641SPeter Zijlstra 55027da4d641SPeter Zijlstra return pages << h->order; 55038f860591SZhang, Yanmin } 55048f860591SZhang, Yanmin 550533b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 550633b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5507a1e78772SMel Gorman long from, long to, 55085a6fe125SMel Gorman struct vm_area_struct *vma, 5509ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5510e4e574b7SAdam Litke { 551133b8f84aSMike Kravetz long chg, add = -1; 5512a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 551390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 55149119a41eSJoonsoo Kim struct resv_map *resv_map; 5515075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 55160db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5517e4e574b7SAdam Litke 551863489f8eSMike Kravetz /* This should never happen */ 551963489f8eSMike Kravetz if (from > to) { 552063489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 552133b8f84aSMike Kravetz return false; 552263489f8eSMike Kravetz } 552363489f8eSMike Kravetz 5524a1e78772SMel Gorman /* 552517c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 552617c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 552790481622SDavid Gibson * without using reserves 552817c9d12eSMel Gorman */ 5529ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 553033b8f84aSMike Kravetz return true; 553117c9d12eSMel Gorman 553217c9d12eSMel Gorman /* 5533a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5534a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5535a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5536a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5537a1e78772SMel Gorman */ 55389119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5539f27a5136SMike Kravetz /* 5540f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5541f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5542f27a5136SMike Kravetz * hugetlbfs_get_inode). 5543f27a5136SMike Kravetz */ 55444e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 55459119a41eSJoonsoo Kim 55460db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 55479119a41eSJoonsoo Kim 55489119a41eSJoonsoo Kim } else { 5549e9fe92aeSMina Almasry /* Private mapping. */ 55509119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 55515a6fe125SMel Gorman if (!resv_map) 555233b8f84aSMike Kravetz return false; 55535a6fe125SMel Gorman 555417c9d12eSMel Gorman chg = to - from; 555517c9d12eSMel Gorman 55565a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 55575a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 55585a6fe125SMel Gorman } 55595a6fe125SMel Gorman 556033b8f84aSMike Kravetz if (chg < 0) 5561c50ac050SDave Hansen goto out_err; 556217c9d12eSMel Gorman 556333b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 556433b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5565075a61d0SMina Almasry goto out_err; 5566075a61d0SMina Almasry 5567075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5568075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5569075a61d0SMina Almasry * of the resv_map. 5570075a61d0SMina Almasry */ 5571075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5572075a61d0SMina Almasry } 5573075a61d0SMina Almasry 55741c5ecae3SMike Kravetz /* 55751c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 55761c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 55771c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 55781c5ecae3SMike Kravetz */ 55791c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 558033b8f84aSMike Kravetz if (gbl_reserve < 0) 5581075a61d0SMina Almasry goto out_uncharge_cgroup; 558217c9d12eSMel Gorman 558317c9d12eSMel Gorman /* 558417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 558590481622SDavid Gibson * Hand the pages back to the subpool if there are not 558617c9d12eSMel Gorman */ 558733b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5588075a61d0SMina Almasry goto out_put_pages; 558917c9d12eSMel Gorman 559017c9d12eSMel Gorman /* 559117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 559217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 559317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 559417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 559517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 559617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 559717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 559817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 559917c9d12eSMel Gorman * else has to be done for private mappings here 560017c9d12eSMel Gorman */ 560133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5602075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 560333039678SMike Kravetz 56040db9d74eSMina Almasry if (unlikely(add < 0)) { 56050db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5606075a61d0SMina Almasry goto out_put_pages; 56070db9d74eSMina Almasry } else if (unlikely(chg > add)) { 560833039678SMike Kravetz /* 560933039678SMike Kravetz * pages in this range were added to the reserve 561033039678SMike Kravetz * map between region_chg and region_add. This 561133039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 561233039678SMike Kravetz * the subpool and reserve counts modified above 561333039678SMike Kravetz * based on the difference. 561433039678SMike Kravetz */ 561533039678SMike Kravetz long rsv_adjust; 561633039678SMike Kravetz 5617d85aecf2SMiaohe Lin /* 5618d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5619d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5620d85aecf2SMiaohe Lin */ 5621075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5622075a61d0SMina Almasry hstate_index(h), 5623075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5624075a61d0SMina Almasry 562533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 562633039678SMike Kravetz chg - add); 562733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5628d85aecf2SMiaohe Lin } else if (h_cg) { 5629d85aecf2SMiaohe Lin /* 5630d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5631d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5632d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5633d85aecf2SMiaohe Lin * done. 5634d85aecf2SMiaohe Lin */ 5635d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 563633039678SMike Kravetz } 563733039678SMike Kravetz } 563833b8f84aSMike Kravetz return true; 563933b8f84aSMike Kravetz 5640075a61d0SMina Almasry out_put_pages: 5641075a61d0SMina Almasry /* put back original number of pages, chg */ 5642075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5643075a61d0SMina Almasry out_uncharge_cgroup: 5644075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5645075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5646c50ac050SDave Hansen out_err: 56475e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 56480db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 56490db9d74eSMina Almasry * region_add failed or didn't run. 56500db9d74eSMina Almasry */ 56510db9d74eSMina Almasry if (chg >= 0 && add < 0) 56520db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5653f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5654f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 565533b8f84aSMike Kravetz return false; 5656a43a8c39SChen, Kenneth W } 5657a43a8c39SChen, Kenneth W 5658b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5659b5cec28dSMike Kravetz long freed) 5660a43a8c39SChen, Kenneth W { 5661a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 56624e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 56639119a41eSJoonsoo Kim long chg = 0; 566490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 56651c5ecae3SMike Kravetz long gbl_reserve; 566645c682a6SKen Chen 5667f27a5136SMike Kravetz /* 5668f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5669f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5670f27a5136SMike Kravetz */ 5671b5cec28dSMike Kravetz if (resv_map) { 5672b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5673b5cec28dSMike Kravetz /* 5674b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5675b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5676b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5677b5cec28dSMike Kravetz */ 5678b5cec28dSMike Kravetz if (chg < 0) 5679b5cec28dSMike Kravetz return chg; 5680b5cec28dSMike Kravetz } 5681b5cec28dSMike Kravetz 568245c682a6SKen Chen spin_lock(&inode->i_lock); 5683e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 568445c682a6SKen Chen spin_unlock(&inode->i_lock); 568545c682a6SKen Chen 56861c5ecae3SMike Kravetz /* 56871c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 56881c5ecae3SMike Kravetz * reservations to be released may be adjusted. 5689dddf31a4SMiaohe Lin * 5690dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 5691dddf31a4SMiaohe Lin * won't go negative. 56921c5ecae3SMike Kravetz */ 56931c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 56941c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5695b5cec28dSMike Kravetz 5696b5cec28dSMike Kravetz return 0; 5697a43a8c39SChen, Kenneth W } 569893f70f90SNaoya Horiguchi 56993212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 57003212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 57013212b535SSteve Capper struct vm_area_struct *vma, 57023212b535SSteve Capper unsigned long addr, pgoff_t idx) 57033212b535SSteve Capper { 57043212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 57053212b535SSteve Capper svma->vm_start; 57063212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 57073212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 57083212b535SSteve Capper 57093212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5710de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5711de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 57123212b535SSteve Capper 57133212b535SSteve Capper /* 57143212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 57153212b535SSteve Capper * page table page. 57163212b535SSteve Capper */ 57173212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 57183212b535SSteve Capper vm_flags != svm_flags || 571907e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 57203212b535SSteve Capper return 0; 57213212b535SSteve Capper 57223212b535SSteve Capper return saddr; 57233212b535SSteve Capper } 57243212b535SSteve Capper 572531aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 57263212b535SSteve Capper { 57273212b535SSteve Capper unsigned long base = addr & PUD_MASK; 57283212b535SSteve Capper unsigned long end = base + PUD_SIZE; 57293212b535SSteve Capper 57303212b535SSteve Capper /* 57313212b535SSteve Capper * check on proper vm_flags and page table alignment 57323212b535SSteve Capper */ 5733017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 573431aafb45SNicholas Krause return true; 573531aafb45SNicholas Krause return false; 57363212b535SSteve Capper } 57373212b535SSteve Capper 5738c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5739c1991e07SPeter Xu { 5740c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5741c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5742c1991e07SPeter Xu return false; 5743c1991e07SPeter Xu #endif 5744c1991e07SPeter Xu return vma_shareable(vma, addr); 5745c1991e07SPeter Xu } 5746c1991e07SPeter Xu 57473212b535SSteve Capper /* 5748017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5749017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5750017b1660SMike Kravetz * shared pmd mappings. 5751017b1660SMike Kravetz */ 5752017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5753017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5754017b1660SMike Kravetz { 5755a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5756a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5757017b1660SMike Kravetz 5758a1ba9da8SLi Xinhai /* 5759f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 5760f0953a1bSIngo Molnar * must be at least partially within in. 5761a1ba9da8SLi Xinhai */ 5762a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5763a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5764017b1660SMike Kravetz return; 5765017b1660SMike Kravetz 576675802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5767a1ba9da8SLi Xinhai if (*start > v_start) 5768a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5769017b1660SMike Kravetz 5770a1ba9da8SLi Xinhai if (*end < v_end) 5771a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5772017b1660SMike Kravetz } 5773017b1660SMike Kravetz 5774017b1660SMike Kravetz /* 57753212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 57763212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 57773212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5778c0d0381aSMike Kravetz * code much cleaner. 5779c0d0381aSMike Kravetz * 57800bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 57810bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 57820bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 57830bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 57840bf7b64eSMike Kravetz * 57850bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 57860bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 57870bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 57880bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 57890bf7b64eSMike Kravetz * only required for subsequent processing. 57903212b535SSteve Capper */ 5791aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5792aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 57933212b535SSteve Capper { 57943212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 57953212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 57963212b535SSteve Capper vma->vm_pgoff; 57973212b535SSteve Capper struct vm_area_struct *svma; 57983212b535SSteve Capper unsigned long saddr; 57993212b535SSteve Capper pte_t *spte = NULL; 58003212b535SSteve Capper pte_t *pte; 5801cb900f41SKirill A. Shutemov spinlock_t *ptl; 58023212b535SSteve Capper 58030bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 58043212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 58053212b535SSteve Capper if (svma == vma) 58063212b535SSteve Capper continue; 58073212b535SSteve Capper 58083212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 58093212b535SSteve Capper if (saddr) { 58107868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 58117868a208SPunit Agrawal vma_mmu_pagesize(svma)); 58123212b535SSteve Capper if (spte) { 58133212b535SSteve Capper get_page(virt_to_page(spte)); 58143212b535SSteve Capper break; 58153212b535SSteve Capper } 58163212b535SSteve Capper } 58173212b535SSteve Capper } 58183212b535SSteve Capper 58193212b535SSteve Capper if (!spte) 58203212b535SSteve Capper goto out; 58213212b535SSteve Capper 58228bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5823dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 58243212b535SSteve Capper pud_populate(mm, pud, 58253212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5826c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5827dc6c9a35SKirill A. Shutemov } else { 58283212b535SSteve Capper put_page(virt_to_page(spte)); 5829dc6c9a35SKirill A. Shutemov } 5830cb900f41SKirill A. Shutemov spin_unlock(ptl); 58313212b535SSteve Capper out: 58323212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 58333212b535SSteve Capper return pte; 58343212b535SSteve Capper } 58353212b535SSteve Capper 58363212b535SSteve Capper /* 58373212b535SSteve Capper * unmap huge page backed by shared pte. 58383212b535SSteve Capper * 58393212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 58403212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 58413212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 58423212b535SSteve Capper * 5843c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 58443212b535SSteve Capper * 58453212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 58463212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 58473212b535SSteve Capper */ 584834ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 584934ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 58503212b535SSteve Capper { 58513212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5852c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5853c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 58543212b535SSteve Capper 585534ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 58563212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 58573212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 58583212b535SSteve Capper return 0; 58593212b535SSteve Capper 58603212b535SSteve Capper pud_clear(pud); 58613212b535SSteve Capper put_page(virt_to_page(ptep)); 5862dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 58633212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 58643212b535SSteve Capper return 1; 58653212b535SSteve Capper } 5866c1991e07SPeter Xu 58679e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5868aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5869aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 58709e5fc74cSSteve Capper { 58719e5fc74cSSteve Capper return NULL; 58729e5fc74cSSteve Capper } 5873e81f2d22SZhang Zhen 587434ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 587534ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5876e81f2d22SZhang Zhen { 5877e81f2d22SZhang Zhen return 0; 5878e81f2d22SZhang Zhen } 5879017b1660SMike Kravetz 5880017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5881017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5882017b1660SMike Kravetz { 5883017b1660SMike Kravetz } 5884c1991e07SPeter Xu 5885c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5886c1991e07SPeter Xu { 5887c1991e07SPeter Xu return false; 5888c1991e07SPeter Xu } 58893212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 58903212b535SSteve Capper 58919e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5892aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 58939e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 58949e5fc74cSSteve Capper { 58959e5fc74cSSteve Capper pgd_t *pgd; 5896c2febafcSKirill A. Shutemov p4d_t *p4d; 58979e5fc74cSSteve Capper pud_t *pud; 58989e5fc74cSSteve Capper pte_t *pte = NULL; 58999e5fc74cSSteve Capper 59009e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5901f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5902f4f0a3d8SKirill A. Shutemov if (!p4d) 5903f4f0a3d8SKirill A. Shutemov return NULL; 5904c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 59059e5fc74cSSteve Capper if (pud) { 59069e5fc74cSSteve Capper if (sz == PUD_SIZE) { 59079e5fc74cSSteve Capper pte = (pte_t *)pud; 59089e5fc74cSSteve Capper } else { 59099e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5910c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5911aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 59129e5fc74cSSteve Capper else 59139e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 59149e5fc74cSSteve Capper } 59159e5fc74cSSteve Capper } 59164e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 59179e5fc74cSSteve Capper 59189e5fc74cSSteve Capper return pte; 59199e5fc74cSSteve Capper } 59209e5fc74cSSteve Capper 59219b19df29SPunit Agrawal /* 59229b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 59239b19df29SPunit Agrawal * entry at address @addr 59249b19df29SPunit Agrawal * 59258ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 59268ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 59279b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 59289b19df29SPunit Agrawal * table. 59299b19df29SPunit Agrawal */ 59307868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 59317868a208SPunit Agrawal unsigned long addr, unsigned long sz) 59329e5fc74cSSteve Capper { 59339e5fc74cSSteve Capper pgd_t *pgd; 5934c2febafcSKirill A. Shutemov p4d_t *p4d; 59358ac0b81aSLi Xinhai pud_t *pud; 59368ac0b81aSLi Xinhai pmd_t *pmd; 59379e5fc74cSSteve Capper 59389e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5939c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5940c2febafcSKirill A. Shutemov return NULL; 5941c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5942c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5943c2febafcSKirill A. Shutemov return NULL; 59449b19df29SPunit Agrawal 5945c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 59468ac0b81aSLi Xinhai if (sz == PUD_SIZE) 59478ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 59489e5fc74cSSteve Capper return (pte_t *)pud; 59498ac0b81aSLi Xinhai if (!pud_present(*pud)) 59508ac0b81aSLi Xinhai return NULL; 59518ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 59529b19df29SPunit Agrawal 59539e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 59548ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 59559e5fc74cSSteve Capper return (pte_t *)pmd; 59569e5fc74cSSteve Capper } 59579e5fc74cSSteve Capper 595861f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 595961f77edaSNaoya Horiguchi 596061f77edaSNaoya Horiguchi /* 596161f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 596261f77edaSNaoya Horiguchi * behavior. 596361f77edaSNaoya Horiguchi */ 596461f77edaSNaoya Horiguchi struct page * __weak 596561f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 596661f77edaSNaoya Horiguchi int write) 596761f77edaSNaoya Horiguchi { 596861f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 596961f77edaSNaoya Horiguchi } 597061f77edaSNaoya Horiguchi 597161f77edaSNaoya Horiguchi struct page * __weak 59724dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 59734dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 59744dc71451SAneesh Kumar K.V { 59754dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 59764dc71451SAneesh Kumar K.V return NULL; 59774dc71451SAneesh Kumar K.V } 59784dc71451SAneesh Kumar K.V 59794dc71451SAneesh Kumar K.V struct page * __weak 59809e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5981e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 59829e5fc74cSSteve Capper { 5983e66f17ffSNaoya Horiguchi struct page *page = NULL; 5984e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5985c9d398faSNaoya Horiguchi pte_t pte; 59863faa52c0SJohn Hubbard 59873faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 59883faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 59893faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 59903faa52c0SJohn Hubbard return NULL; 59913faa52c0SJohn Hubbard 5992e66f17ffSNaoya Horiguchi retry: 5993e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5994e66f17ffSNaoya Horiguchi spin_lock(ptl); 5995e66f17ffSNaoya Horiguchi /* 5996e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5997e66f17ffSNaoya Horiguchi * unmapped from other threads. 5998e66f17ffSNaoya Horiguchi */ 5999e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 6000e66f17ffSNaoya Horiguchi goto out; 6001c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 6002c9d398faSNaoya Horiguchi if (pte_present(pte)) { 600397534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 60043faa52c0SJohn Hubbard /* 60053faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 60063faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 60073faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 60083faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 60093faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 60103faa52c0SJohn Hubbard * unless the page refcount overflowed: 60113faa52c0SJohn Hubbard */ 60123faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 60133faa52c0SJohn Hubbard page = NULL; 60143faa52c0SJohn Hubbard goto out; 60153faa52c0SJohn Hubbard } 6016e66f17ffSNaoya Horiguchi } else { 6017c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 6018e66f17ffSNaoya Horiguchi spin_unlock(ptl); 6019e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 6020e66f17ffSNaoya Horiguchi goto retry; 6021e66f17ffSNaoya Horiguchi } 6022e66f17ffSNaoya Horiguchi /* 6023e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 6024e66f17ffSNaoya Horiguchi * follow_page_mask(). 6025e66f17ffSNaoya Horiguchi */ 6026e66f17ffSNaoya Horiguchi } 6027e66f17ffSNaoya Horiguchi out: 6028e66f17ffSNaoya Horiguchi spin_unlock(ptl); 60299e5fc74cSSteve Capper return page; 60309e5fc74cSSteve Capper } 60319e5fc74cSSteve Capper 603261f77edaSNaoya Horiguchi struct page * __weak 60339e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 6034e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 60359e5fc74cSSteve Capper { 60363faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6037e66f17ffSNaoya Horiguchi return NULL; 60389e5fc74cSSteve Capper 6039e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 60409e5fc74cSSteve Capper } 60419e5fc74cSSteve Capper 6042faaa5b62SAnshuman Khandual struct page * __weak 6043faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 6044faaa5b62SAnshuman Khandual { 60453faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6046faaa5b62SAnshuman Khandual return NULL; 6047faaa5b62SAnshuman Khandual 6048faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 6049faaa5b62SAnshuman Khandual } 6050faaa5b62SAnshuman Khandual 605131caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 605231caf665SNaoya Horiguchi { 6053bcc54222SNaoya Horiguchi bool ret = true; 6054bcc54222SNaoya Horiguchi 6055db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 60568f251a3dSMike Kravetz if (!PageHeadHuge(page) || 60578f251a3dSMike Kravetz !HPageMigratable(page) || 60580eb2df2bSMuchun Song !get_page_unless_zero(page)) { 6059bcc54222SNaoya Horiguchi ret = false; 6060bcc54222SNaoya Horiguchi goto unlock; 6061bcc54222SNaoya Horiguchi } 60628f251a3dSMike Kravetz ClearHPageMigratable(page); 606331caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 6064bcc54222SNaoya Horiguchi unlock: 6065db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6066bcc54222SNaoya Horiguchi return ret; 606731caf665SNaoya Horiguchi } 606831caf665SNaoya Horiguchi 606925182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 607025182f05SNaoya Horiguchi { 607125182f05SNaoya Horiguchi int ret = 0; 607225182f05SNaoya Horiguchi 607325182f05SNaoya Horiguchi *hugetlb = false; 607425182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 607525182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 607625182f05SNaoya Horiguchi *hugetlb = true; 607725182f05SNaoya Horiguchi if (HPageFreed(page) || HPageMigratable(page)) 607825182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 60790ed950d1SNaoya Horiguchi else 60800ed950d1SNaoya Horiguchi ret = -EBUSY; 608125182f05SNaoya Horiguchi } 608225182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 608325182f05SNaoya Horiguchi return ret; 608425182f05SNaoya Horiguchi } 608525182f05SNaoya Horiguchi 608631caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 608731caf665SNaoya Horiguchi { 6088db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 60898f251a3dSMike Kravetz SetHPageMigratable(page); 609031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 6091db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 609231caf665SNaoya Horiguchi put_page(page); 609331caf665SNaoya Horiguchi } 6094ab5ac90aSMichal Hocko 6095ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 6096ab5ac90aSMichal Hocko { 6097ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 6098ab5ac90aSMichal Hocko 6099ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 6100ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 6101ab5ac90aSMichal Hocko 6102ab5ac90aSMichal Hocko /* 6103ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 6104ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 6105ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 6106ab5ac90aSMichal Hocko * the temporary status. 6107ab5ac90aSMichal Hocko * 6108ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 6109ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 6110ab5ac90aSMichal Hocko * the per-node's. 6111ab5ac90aSMichal Hocko */ 61129157c311SMike Kravetz if (HPageTemporary(newpage)) { 6113ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 6114ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 6115ab5ac90aSMichal Hocko 61169157c311SMike Kravetz SetHPageTemporary(oldpage); 61179157c311SMike Kravetz ClearHPageTemporary(newpage); 6118ab5ac90aSMichal Hocko 61195af1ab1dSMiaohe Lin /* 61205af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 61215af1ab1dSMiaohe Lin * when we do not cross the node. 61225af1ab1dSMiaohe Lin */ 61235af1ab1dSMiaohe Lin if (new_nid == old_nid) 61245af1ab1dSMiaohe Lin return; 6125db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 6126ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 6127ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 6128ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 6129ab5ac90aSMichal Hocko } 6130db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6131ab5ac90aSMichal Hocko } 6132ab5ac90aSMichal Hocko } 6133cf11e85fSRoman Gushchin 61346dfeaff9SPeter Xu /* 61356dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 61366dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 61376dfeaff9SPeter Xu */ 61386dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 61396dfeaff9SPeter Xu { 61406dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 61416dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 61426dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 61436dfeaff9SPeter Xu struct mmu_notifier_range range; 61446dfeaff9SPeter Xu unsigned long address, start, end; 61456dfeaff9SPeter Xu spinlock_t *ptl; 61466dfeaff9SPeter Xu pte_t *ptep; 61476dfeaff9SPeter Xu 61486dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 61496dfeaff9SPeter Xu return; 61506dfeaff9SPeter Xu 61516dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 61526dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 61536dfeaff9SPeter Xu 61546dfeaff9SPeter Xu if (start >= end) 61556dfeaff9SPeter Xu return; 61566dfeaff9SPeter Xu 61576dfeaff9SPeter Xu /* 61586dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 61596dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 61606dfeaff9SPeter Xu */ 61616dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 61626dfeaff9SPeter Xu start, end); 61636dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 61646dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 61656dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 61666dfeaff9SPeter Xu unsigned long tmp = address; 61676dfeaff9SPeter Xu 61686dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 61696dfeaff9SPeter Xu if (!ptep) 61706dfeaff9SPeter Xu continue; 61716dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 61726dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 61736dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 61746dfeaff9SPeter Xu spin_unlock(ptl); 61756dfeaff9SPeter Xu } 61766dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 61776dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 61786dfeaff9SPeter Xu /* 61796dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 61806dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 61816dfeaff9SPeter Xu */ 61826dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 61836dfeaff9SPeter Xu } 61846dfeaff9SPeter Xu 6185cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 6186cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 6187cf11e85fSRoman Gushchin 6188cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 6189cf11e85fSRoman Gushchin { 6190cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 6191cf11e85fSRoman Gushchin return 0; 6192cf11e85fSRoman Gushchin } 6193cf11e85fSRoman Gushchin 6194cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 6195cf11e85fSRoman Gushchin 6196cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 6197cf11e85fSRoman Gushchin { 6198cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 6199cf11e85fSRoman Gushchin int nid; 6200cf11e85fSRoman Gushchin 6201cf11e85fSRoman Gushchin cma_reserve_called = true; 6202cf11e85fSRoman Gushchin 6203cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 6204cf11e85fSRoman Gushchin return; 6205cf11e85fSRoman Gushchin 6206cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 6207cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 6208cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 6209cf11e85fSRoman Gushchin return; 6210cf11e85fSRoman Gushchin } 6211cf11e85fSRoman Gushchin 6212cf11e85fSRoman Gushchin /* 6213cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 6214cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 6215cf11e85fSRoman Gushchin */ 6216cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 6217cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 6218cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 6219cf11e85fSRoman Gushchin 6220cf11e85fSRoman Gushchin reserved = 0; 6221cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 6222cf11e85fSRoman Gushchin int res; 62232281f797SBarry Song char name[CMA_MAX_NAME]; 6224cf11e85fSRoman Gushchin 6225cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 6226cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 6227cf11e85fSRoman Gushchin 62282281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 6229cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 623029d0f41dSBarry Song 0, false, name, 6231cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 6232cf11e85fSRoman Gushchin if (res) { 6233cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 6234cf11e85fSRoman Gushchin res, nid); 6235cf11e85fSRoman Gushchin continue; 6236cf11e85fSRoman Gushchin } 6237cf11e85fSRoman Gushchin 6238cf11e85fSRoman Gushchin reserved += size; 6239cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 6240cf11e85fSRoman Gushchin size / SZ_1M, nid); 6241cf11e85fSRoman Gushchin 6242cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 6243cf11e85fSRoman Gushchin break; 6244cf11e85fSRoman Gushchin } 6245cf11e85fSRoman Gushchin } 6246cf11e85fSRoman Gushchin 6247cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 6248cf11e85fSRoman Gushchin { 6249cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 6250cf11e85fSRoman Gushchin return; 6251cf11e85fSRoman Gushchin 6252cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 6253cf11e85fSRoman Gushchin } 6254cf11e85fSRoman Gushchin 6255cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 6256