1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 33d6606683SLinus Torvalds 3463551ae0SDavid Gibson #include <asm/page.h> 35ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3624669e58SAneesh Kumar K.V #include <asm/tlb.h> 3763551ae0SDavid Gibson 3824669e58SAneesh Kumar K.V #include <linux/io.h> 3963551ae0SDavid Gibson #include <linux/hugetlb.h> 409dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 419a305230SLee Schermerhorn #include <linux/node.h> 421a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 451da177e4SLinus Torvalds 46c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 47e5ff2159SAndi Kleen unsigned int default_hstate_idx; 48e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 49cf11e85fSRoman Gushchin 50dbda8feaSBarry Song #ifdef CONFIG_CMA 51cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 52dbda8feaSBarry Song #endif 53dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 54cf11e85fSRoman Gushchin 55641844f5SNaoya Horiguchi /* 56641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 57641844f5SNaoya Horiguchi * at boot time. 58641844f5SNaoya Horiguchi */ 59641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 60e5ff2159SAndi Kleen 6153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6253ba51d2SJon Tollefson 63e5ff2159SAndi Kleen /* for command line parsing */ 64e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 65e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 669fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 67282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 68e5ff2159SAndi Kleen 693935baa9SDavid Gibson /* 7031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 723935baa9SDavid Gibson */ 73c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 740bd0f9fbSEric Paris 758382d914SDavidlohr Bueso /* 768382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 778382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 788382d914SDavidlohr Bueso */ 798382d914SDavidlohr Bueso static int num_fault_mutexes; 80c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 818382d914SDavidlohr Bueso 827ca02d0aSMike Kravetz /* Forward declaration */ 837ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 847ca02d0aSMike Kravetz 8590481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 8690481622SDavid Gibson { 8790481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 8890481622SDavid Gibson 8990481622SDavid Gibson spin_unlock(&spool->lock); 9090481622SDavid Gibson 9190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 927c8de358SEthon Paul * remain, give up any reservations based on minimum size and 937ca02d0aSMike Kravetz * free the subpool */ 947ca02d0aSMike Kravetz if (free) { 957ca02d0aSMike Kravetz if (spool->min_hpages != -1) 967ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 977ca02d0aSMike Kravetz -spool->min_hpages); 9890481622SDavid Gibson kfree(spool); 9990481622SDavid Gibson } 1007ca02d0aSMike Kravetz } 10190481622SDavid Gibson 1027ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1037ca02d0aSMike Kravetz long min_hpages) 10490481622SDavid Gibson { 10590481622SDavid Gibson struct hugepage_subpool *spool; 10690481622SDavid Gibson 107c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 10890481622SDavid Gibson if (!spool) 10990481622SDavid Gibson return NULL; 11090481622SDavid Gibson 11190481622SDavid Gibson spin_lock_init(&spool->lock); 11290481622SDavid Gibson spool->count = 1; 1137ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1147ca02d0aSMike Kravetz spool->hstate = h; 1157ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1167ca02d0aSMike Kravetz 1177ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1187ca02d0aSMike Kravetz kfree(spool); 1197ca02d0aSMike Kravetz return NULL; 1207ca02d0aSMike Kravetz } 1217ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 12290481622SDavid Gibson 12390481622SDavid Gibson return spool; 12490481622SDavid Gibson } 12590481622SDavid Gibson 12690481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 12790481622SDavid Gibson { 12890481622SDavid Gibson spin_lock(&spool->lock); 12990481622SDavid Gibson BUG_ON(!spool->count); 13090481622SDavid Gibson spool->count--; 13190481622SDavid Gibson unlock_or_release_subpool(spool); 13290481622SDavid Gibson } 13390481622SDavid Gibson 1341c5ecae3SMike Kravetz /* 1351c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1361c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1379e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1381c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1391c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1407c8de358SEthon Paul * a subpool minimum size must be maintained. 1411c5ecae3SMike Kravetz */ 1421c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 14390481622SDavid Gibson long delta) 14490481622SDavid Gibson { 1451c5ecae3SMike Kravetz long ret = delta; 14690481622SDavid Gibson 14790481622SDavid Gibson if (!spool) 1481c5ecae3SMike Kravetz return ret; 14990481622SDavid Gibson 15090481622SDavid Gibson spin_lock(&spool->lock); 15190481622SDavid Gibson 1521c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1531c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1541c5ecae3SMike Kravetz spool->used_hpages += delta; 1551c5ecae3SMike Kravetz else { 1561c5ecae3SMike Kravetz ret = -ENOMEM; 1571c5ecae3SMike Kravetz goto unlock_ret; 1581c5ecae3SMike Kravetz } 1591c5ecae3SMike Kravetz } 1601c5ecae3SMike Kravetz 16109a95e29SMike Kravetz /* minimum size accounting */ 16209a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1631c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1641c5ecae3SMike Kravetz /* 1651c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1661c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1671c5ecae3SMike Kravetz */ 1681c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1691c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1701c5ecae3SMike Kravetz } else { 1711c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1721c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1731c5ecae3SMike Kravetz } 1741c5ecae3SMike Kravetz } 1751c5ecae3SMike Kravetz 1761c5ecae3SMike Kravetz unlock_ret: 1771c5ecae3SMike Kravetz spin_unlock(&spool->lock); 17890481622SDavid Gibson return ret; 17990481622SDavid Gibson } 18090481622SDavid Gibson 1811c5ecae3SMike Kravetz /* 1821c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1831c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1841c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1851c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1861c5ecae3SMike Kravetz */ 1871c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 18890481622SDavid Gibson long delta) 18990481622SDavid Gibson { 1901c5ecae3SMike Kravetz long ret = delta; 1911c5ecae3SMike Kravetz 19290481622SDavid Gibson if (!spool) 1931c5ecae3SMike Kravetz return delta; 19490481622SDavid Gibson 19590481622SDavid Gibson spin_lock(&spool->lock); 1961c5ecae3SMike Kravetz 1971c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 19890481622SDavid Gibson spool->used_hpages -= delta; 1991c5ecae3SMike Kravetz 20009a95e29SMike Kravetz /* minimum size accounting */ 20109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2021c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2031c5ecae3SMike Kravetz ret = 0; 2041c5ecae3SMike Kravetz else 2051c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2061c5ecae3SMike Kravetz 2071c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2081c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2091c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2101c5ecae3SMike Kravetz } 2111c5ecae3SMike Kravetz 2121c5ecae3SMike Kravetz /* 2131c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2141c5ecae3SMike Kravetz * quota reference, free it now. 2151c5ecae3SMike Kravetz */ 21690481622SDavid Gibson unlock_or_release_subpool(spool); 2171c5ecae3SMike Kravetz 2181c5ecae3SMike Kravetz return ret; 21990481622SDavid Gibson } 22090481622SDavid Gibson 22190481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 22290481622SDavid Gibson { 22390481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 22490481622SDavid Gibson } 22590481622SDavid Gibson 22690481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 22790481622SDavid Gibson { 228496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 22990481622SDavid Gibson } 23090481622SDavid Gibson 2310db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2320db9d74eSMina Almasry * it for use. 2330db9d74eSMina Almasry */ 2340db9d74eSMina Almasry static struct file_region * 2350db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2360db9d74eSMina Almasry { 2370db9d74eSMina Almasry struct file_region *nrg = NULL; 2380db9d74eSMina Almasry 2390db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2400db9d74eSMina Almasry 2410db9d74eSMina Almasry resv->region_cache_count--; 2420db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2430db9d74eSMina Almasry VM_BUG_ON(!nrg); 2440db9d74eSMina Almasry list_del(&nrg->link); 2450db9d74eSMina Almasry 2460db9d74eSMina Almasry nrg->from = from; 2470db9d74eSMina Almasry nrg->to = to; 2480db9d74eSMina Almasry 2490db9d74eSMina Almasry return nrg; 2500db9d74eSMina Almasry } 2510db9d74eSMina Almasry 252075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 253075a61d0SMina Almasry struct file_region *rg) 254075a61d0SMina Almasry { 255075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 256075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 257075a61d0SMina Almasry nrg->css = rg->css; 258075a61d0SMina Almasry if (rg->css) 259075a61d0SMina Almasry css_get(rg->css); 260075a61d0SMina Almasry #endif 261075a61d0SMina Almasry } 262075a61d0SMina Almasry 263075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 264075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 265075a61d0SMina Almasry struct hstate *h, 266075a61d0SMina Almasry struct resv_map *resv, 267075a61d0SMina Almasry struct file_region *nrg) 268075a61d0SMina Almasry { 269075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 270075a61d0SMina Almasry if (h_cg) { 271075a61d0SMina Almasry nrg->reservation_counter = 272075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 273075a61d0SMina Almasry nrg->css = &h_cg->css; 274075a61d0SMina Almasry if (!resv->pages_per_hpage) 275075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 276075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 277075a61d0SMina Almasry * a resv_map. 278075a61d0SMina Almasry */ 279075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 280075a61d0SMina Almasry } else { 281075a61d0SMina Almasry nrg->reservation_counter = NULL; 282075a61d0SMina Almasry nrg->css = NULL; 283075a61d0SMina Almasry } 284075a61d0SMina Almasry #endif 285075a61d0SMina Almasry } 286075a61d0SMina Almasry 287a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 288a9b3f867SMina Almasry struct file_region *org) 289a9b3f867SMina Almasry { 290a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 291a9b3f867SMina Almasry return rg && org && 292a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 293a9b3f867SMina Almasry rg->css == org->css; 294a9b3f867SMina Almasry 295a9b3f867SMina Almasry #else 296a9b3f867SMina Almasry return true; 297a9b3f867SMina Almasry #endif 298a9b3f867SMina Almasry } 299a9b3f867SMina Almasry 300a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 301a9b3f867SMina Almasry { 302a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 303a9b3f867SMina Almasry 304a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 305a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 306a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 307a9b3f867SMina Almasry prg->to = rg->to; 308a9b3f867SMina Almasry 309a9b3f867SMina Almasry list_del(&rg->link); 310a9b3f867SMina Almasry kfree(rg); 311a9b3f867SMina Almasry 312a9b3f867SMina Almasry coalesce_file_region(resv, prg); 313a9b3f867SMina Almasry return; 314a9b3f867SMina Almasry } 315a9b3f867SMina Almasry 316a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 317a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 318a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 319a9b3f867SMina Almasry nrg->from = rg->from; 320a9b3f867SMina Almasry 321a9b3f867SMina Almasry list_del(&rg->link); 322a9b3f867SMina Almasry kfree(rg); 323a9b3f867SMina Almasry 324a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 325a9b3f867SMina Almasry return; 326a9b3f867SMina Almasry } 327a9b3f867SMina Almasry } 328a9b3f867SMina Almasry 329d75c6af9SMina Almasry /* Must be called with resv->lock held. Calling this with count_only == true 330d75c6af9SMina Almasry * will count the number of pages to be added but will not modify the linked 3310db9d74eSMina Almasry * list. If regions_needed != NULL and count_only == true, then regions_needed 3320db9d74eSMina Almasry * will indicate the number of file_regions needed in the cache to carry out to 3330db9d74eSMina Almasry * add the regions for this range. 334d75c6af9SMina Almasry */ 335d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 336075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 337075a61d0SMina Almasry struct hstate *h, long *regions_needed, 338075a61d0SMina Almasry bool count_only) 339d75c6af9SMina Almasry { 3400db9d74eSMina Almasry long add = 0; 341d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3420db9d74eSMina Almasry long last_accounted_offset = f; 343d75c6af9SMina Almasry struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; 344d75c6af9SMina Almasry 3450db9d74eSMina Almasry if (regions_needed) 3460db9d74eSMina Almasry *regions_needed = 0; 347d75c6af9SMina Almasry 3480db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3490db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3500db9d74eSMina Almasry * bounds checking. 3510db9d74eSMina Almasry */ 3520db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 3530db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 3540db9d74eSMina Almasry if (rg->from < f) { 3550db9d74eSMina Almasry /* If this region ends after the last accounted offset, 3560db9d74eSMina Almasry * then we need to update last_accounted_offset. 3570db9d74eSMina Almasry */ 3580db9d74eSMina Almasry if (rg->to > last_accounted_offset) 3590db9d74eSMina Almasry last_accounted_offset = rg->to; 3600db9d74eSMina Almasry continue; 3610db9d74eSMina Almasry } 362d75c6af9SMina Almasry 3630db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 3640db9d74eSMina Almasry * finished. 3650db9d74eSMina Almasry */ 366d75c6af9SMina Almasry if (rg->from > t) 367d75c6af9SMina Almasry break; 368d75c6af9SMina Almasry 3690db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 3700db9d74eSMina Almasry * update last_accounted_offset. 371d75c6af9SMina Almasry */ 3720db9d74eSMina Almasry if (rg->from > last_accounted_offset) { 3730db9d74eSMina Almasry add += rg->from - last_accounted_offset; 3740db9d74eSMina Almasry if (!count_only) { 3750db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3760db9d74eSMina Almasry resv, last_accounted_offset, rg->from); 377075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, 378075a61d0SMina Almasry resv, nrg); 3790db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 380a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 3810db9d74eSMina Almasry } else if (regions_needed) 3820db9d74eSMina Almasry *regions_needed += 1; 383d75c6af9SMina Almasry } 384d75c6af9SMina Almasry 3850db9d74eSMina Almasry last_accounted_offset = rg->to; 3860db9d74eSMina Almasry } 3870db9d74eSMina Almasry 3880db9d74eSMina Almasry /* Handle the case where our range extends beyond 3890db9d74eSMina Almasry * last_accounted_offset. 3900db9d74eSMina Almasry */ 3910db9d74eSMina Almasry if (last_accounted_offset < t) { 3920db9d74eSMina Almasry add += t - last_accounted_offset; 3930db9d74eSMina Almasry if (!count_only) { 3940db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3950db9d74eSMina Almasry resv, last_accounted_offset, t); 396075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); 3970db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 398a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 3990db9d74eSMina Almasry } else if (regions_needed) 4000db9d74eSMina Almasry *regions_needed += 1; 4010db9d74eSMina Almasry } 4020db9d74eSMina Almasry 4030db9d74eSMina Almasry VM_BUG_ON(add < 0); 4040db9d74eSMina Almasry return add; 4050db9d74eSMina Almasry } 4060db9d74eSMina Almasry 4070db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4080db9d74eSMina Almasry */ 4090db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4100db9d74eSMina Almasry int regions_needed) 4110db9d74eSMina Almasry __must_hold(&resv->lock) 4120db9d74eSMina Almasry { 4130db9d74eSMina Almasry struct list_head allocated_regions; 4140db9d74eSMina Almasry int to_allocate = 0, i = 0; 4150db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4160db9d74eSMina Almasry 4170db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4180db9d74eSMina Almasry 4190db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4200db9d74eSMina Almasry 4210db9d74eSMina Almasry /* 4220db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4230db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4240db9d74eSMina Almasry * 4250db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4260db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4270db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4280db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4290db9d74eSMina Almasry */ 4300db9d74eSMina Almasry while (resv->region_cache_count < 4310db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4320db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4330db9d74eSMina Almasry resv->region_cache_count; 4340db9d74eSMina Almasry 4350db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4360db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4370db9d74eSMina Almasry * needing to allocate for regions_needed. 4380db9d74eSMina Almasry */ 4390db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4400db9d74eSMina Almasry 4410db9d74eSMina Almasry spin_unlock(&resv->lock); 4420db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4430db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4440db9d74eSMina Almasry if (!trg) 4450db9d74eSMina Almasry goto out_of_memory; 4460db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4470db9d74eSMina Almasry } 4480db9d74eSMina Almasry 4490db9d74eSMina Almasry spin_lock(&resv->lock); 4500db9d74eSMina Almasry 4510db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 4520db9d74eSMina Almasry list_del(&rg->link); 4530db9d74eSMina Almasry list_add(&rg->link, &resv->region_cache); 4540db9d74eSMina Almasry resv->region_cache_count++; 4550db9d74eSMina Almasry } 4560db9d74eSMina Almasry } 4570db9d74eSMina Almasry 4580db9d74eSMina Almasry return 0; 4590db9d74eSMina Almasry 4600db9d74eSMina Almasry out_of_memory: 4610db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 462d75c6af9SMina Almasry list_del(&rg->link); 463d75c6af9SMina Almasry kfree(rg); 464d75c6af9SMina Almasry } 4650db9d74eSMina Almasry return -ENOMEM; 466d75c6af9SMina Almasry } 467d75c6af9SMina Almasry 4681dd308a7SMike Kravetz /* 4691dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 4700db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 4710db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 4720db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 4730db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 4740db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 475cf3ad20bSMike Kravetz * 4760db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 4770db9d74eSMina Almasry * 4780db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 4790db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 4807c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 4810db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 4820db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 4830db9d74eSMina Almasry * 1 page will only require at most 1 entry. 4841dd308a7SMike Kravetz */ 4850db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 486075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 487075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 48896822904SAndy Whitcroft { 4890db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 49096822904SAndy Whitcroft 4917b24d861SDavidlohr Bueso spin_lock(&resv->lock); 4920db9d74eSMina Almasry retry: 4930db9d74eSMina Almasry 4940db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 495075a61d0SMina Almasry add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed, 496075a61d0SMina Almasry true); 49796822904SAndy Whitcroft 4985e911373SMike Kravetz /* 4990db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5000db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5010db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5020db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5030db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5040db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5050db9d74eSMina Almasry * operation. 5065e911373SMike Kravetz */ 5070db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5080db9d74eSMina Almasry resv->region_cache_count < 5090db9d74eSMina Almasry resv->adds_in_progress + 5100db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5110db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5120db9d74eSMina Almasry * allocate file_region entries. 5130db9d74eSMina Almasry */ 5140db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5155e911373SMike Kravetz 5160db9d74eSMina Almasry if (allocate_file_region_entries( 5170db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5180db9d74eSMina Almasry return -ENOMEM; 5195e911373SMike Kravetz } 5205e911373SMike Kravetz 5210db9d74eSMina Almasry goto retry; 5220db9d74eSMina Almasry } 523cf3ad20bSMike Kravetz 524075a61d0SMina Almasry add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false); 5250db9d74eSMina Almasry 5260db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5270db9d74eSMina Almasry 5287b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 529cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 530cf3ad20bSMike Kravetz return add; 53196822904SAndy Whitcroft } 53296822904SAndy Whitcroft 5331dd308a7SMike Kravetz /* 5341dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5351dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5361dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5371dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5381dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5391dd308a7SMike Kravetz * not change the number of huge pages represented by the 5400db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5410db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5420db9d74eSMina Almasry * file_region structure is added. 5430db9d74eSMina Almasry * 5440db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5450db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5460db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5475e911373SMike Kravetz * 5485e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5495e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5505e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5515e911373SMike Kravetz * is needed and can not be allocated. 5521dd308a7SMike Kravetz */ 5530db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5540db9d74eSMina Almasry long *out_regions_needed) 55596822904SAndy Whitcroft { 55696822904SAndy Whitcroft long chg = 0; 55796822904SAndy Whitcroft 5587b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5595e911373SMike Kravetz 5600db9d74eSMina Almasry /* Count how many hugepages in this range are NOT respresented. */ 561075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 562075a61d0SMina Almasry out_regions_needed, true); 5635e911373SMike Kravetz 5640db9d74eSMina Almasry if (*out_regions_needed == 0) 5650db9d74eSMina Almasry *out_regions_needed = 1; 5665e911373SMike Kravetz 5670db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5685e911373SMike Kravetz return -ENOMEM; 5695e911373SMike Kravetz 5700db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 57196822904SAndy Whitcroft 5727b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 57396822904SAndy Whitcroft return chg; 57496822904SAndy Whitcroft } 57596822904SAndy Whitcroft 5761dd308a7SMike Kravetz /* 5775e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 5785e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 5795e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 5805e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 5810db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 5820db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 5830db9d74eSMina Almasry * the adds_in_progress counter. 5845e911373SMike Kravetz * 5855e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 5865e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 5875e911373SMike Kravetz * arguments will match the associated region_chg call. 5885e911373SMike Kravetz */ 5890db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 5900db9d74eSMina Almasry long regions_needed) 5915e911373SMike Kravetz { 5925e911373SMike Kravetz spin_lock(&resv->lock); 5935e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 5940db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 5955e911373SMike Kravetz spin_unlock(&resv->lock); 5965e911373SMike Kravetz } 5975e911373SMike Kravetz 5985e911373SMike Kravetz /* 599feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 600feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 601feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 602feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 603feba16e2SMike Kravetz * 604feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 605feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 606feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 607feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 608feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 609feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 610feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6111dd308a7SMike Kravetz */ 612feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 61396822904SAndy Whitcroft { 6141406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 61596822904SAndy Whitcroft struct file_region *rg, *trg; 616feba16e2SMike Kravetz struct file_region *nrg = NULL; 617feba16e2SMike Kravetz long del = 0; 61896822904SAndy Whitcroft 619feba16e2SMike Kravetz retry: 6207b24d861SDavidlohr Bueso spin_lock(&resv->lock); 621feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 622dbe409e4SMike Kravetz /* 623dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 624dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 625dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 626dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 627dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 628dbe409e4SMike Kravetz */ 629dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 630feba16e2SMike Kravetz continue; 631dbe409e4SMike Kravetz 632feba16e2SMike Kravetz if (rg->from >= t) 63396822904SAndy Whitcroft break; 63496822904SAndy Whitcroft 635feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 636feba16e2SMike Kravetz /* 637feba16e2SMike Kravetz * Check for an entry in the cache before dropping 638feba16e2SMike Kravetz * lock and attempting allocation. 639feba16e2SMike Kravetz */ 640feba16e2SMike Kravetz if (!nrg && 641feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 642feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 643feba16e2SMike Kravetz struct file_region, 644feba16e2SMike Kravetz link); 645feba16e2SMike Kravetz list_del(&nrg->link); 646feba16e2SMike Kravetz resv->region_cache_count--; 64796822904SAndy Whitcroft } 64896822904SAndy Whitcroft 649feba16e2SMike Kravetz if (!nrg) { 650feba16e2SMike Kravetz spin_unlock(&resv->lock); 651feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 652feba16e2SMike Kravetz if (!nrg) 653feba16e2SMike Kravetz return -ENOMEM; 654feba16e2SMike Kravetz goto retry; 655feba16e2SMike Kravetz } 656feba16e2SMike Kravetz 657feba16e2SMike Kravetz del += t - f; 658feba16e2SMike Kravetz 659feba16e2SMike Kravetz /* New entry for end of split region */ 660feba16e2SMike Kravetz nrg->from = t; 661feba16e2SMike Kravetz nrg->to = rg->to; 662075a61d0SMina Almasry 663075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 664075a61d0SMina Almasry 665feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 666feba16e2SMike Kravetz 667feba16e2SMike Kravetz /* Original entry is trimmed */ 668feba16e2SMike Kravetz rg->to = f; 669feba16e2SMike Kravetz 670075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region( 671075a61d0SMina Almasry resv, rg, nrg->to - nrg->from); 672075a61d0SMina Almasry 673feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 674feba16e2SMike Kravetz nrg = NULL; 67596822904SAndy Whitcroft break; 676feba16e2SMike Kravetz } 677feba16e2SMike Kravetz 678feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 679feba16e2SMike Kravetz del += rg->to - rg->from; 680075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 681075a61d0SMina Almasry rg->to - rg->from); 68296822904SAndy Whitcroft list_del(&rg->link); 68396822904SAndy Whitcroft kfree(rg); 684feba16e2SMike Kravetz continue; 68596822904SAndy Whitcroft } 6867b24d861SDavidlohr Bueso 687feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 688feba16e2SMike Kravetz del += t - rg->from; 689feba16e2SMike Kravetz rg->from = t; 690075a61d0SMina Almasry 691075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 692075a61d0SMina Almasry t - rg->from); 693feba16e2SMike Kravetz } else { /* Trim end of region */ 694feba16e2SMike Kravetz del += rg->to - f; 695feba16e2SMike Kravetz rg->to = f; 696075a61d0SMina Almasry 697075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 698075a61d0SMina Almasry rg->to - f); 699feba16e2SMike Kravetz } 700feba16e2SMike Kravetz } 701feba16e2SMike Kravetz 7027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 703feba16e2SMike Kravetz kfree(nrg); 704feba16e2SMike Kravetz return del; 70596822904SAndy Whitcroft } 70696822904SAndy Whitcroft 7071dd308a7SMike Kravetz /* 708b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 709b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 710b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 711b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 712b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 713b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 714b5cec28dSMike Kravetz * counts. 715b5cec28dSMike Kravetz */ 71672e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 717b5cec28dSMike Kravetz { 718b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 719b5cec28dSMike Kravetz long rsv_adjust; 720b5cec28dSMike Kravetz 721b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 72272e2936cSzhong jiang if (rsv_adjust) { 723b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 724b5cec28dSMike Kravetz 725b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 726b5cec28dSMike Kravetz } 727b5cec28dSMike Kravetz } 728b5cec28dSMike Kravetz 729b5cec28dSMike Kravetz /* 7301dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7311dd308a7SMike Kravetz * that intersect with the range [f, t). 7321dd308a7SMike Kravetz */ 7331406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 73484afd99bSAndy Whitcroft { 7351406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 73684afd99bSAndy Whitcroft struct file_region *rg; 73784afd99bSAndy Whitcroft long chg = 0; 73884afd99bSAndy Whitcroft 7397b24d861SDavidlohr Bueso spin_lock(&resv->lock); 74084afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 74184afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 742f2135a4aSWang Sheng-Hui long seg_from; 743f2135a4aSWang Sheng-Hui long seg_to; 74484afd99bSAndy Whitcroft 74584afd99bSAndy Whitcroft if (rg->to <= f) 74684afd99bSAndy Whitcroft continue; 74784afd99bSAndy Whitcroft if (rg->from >= t) 74884afd99bSAndy Whitcroft break; 74984afd99bSAndy Whitcroft 75084afd99bSAndy Whitcroft seg_from = max(rg->from, f); 75184afd99bSAndy Whitcroft seg_to = min(rg->to, t); 75284afd99bSAndy Whitcroft 75384afd99bSAndy Whitcroft chg += seg_to - seg_from; 75484afd99bSAndy Whitcroft } 7557b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 75684afd99bSAndy Whitcroft 75784afd99bSAndy Whitcroft return chg; 75884afd99bSAndy Whitcroft } 75984afd99bSAndy Whitcroft 76096822904SAndy Whitcroft /* 761e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 762e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 763e7c4b0bfSAndy Whitcroft */ 764a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 765a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 766e7c4b0bfSAndy Whitcroft { 767a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 768a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 769e7c4b0bfSAndy Whitcroft } 770e7c4b0bfSAndy Whitcroft 7710fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 7720fe6e20bSNaoya Horiguchi unsigned long address) 7730fe6e20bSNaoya Horiguchi { 7740fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 7750fe6e20bSNaoya Horiguchi } 776dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 7770fe6e20bSNaoya Horiguchi 77884afd99bSAndy Whitcroft /* 77908fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 78008fba699SMel Gorman * cases this will be same size as used by the page table entries. 78108fba699SMel Gorman */ 78208fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 78308fba699SMel Gorman { 78405ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 78505ea8860SDan Williams return vma->vm_ops->pagesize(vma); 78608fba699SMel Gorman return PAGE_SIZE; 78708fba699SMel Gorman } 788f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 78908fba699SMel Gorman 79008fba699SMel Gorman /* 7913340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 7923340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 79309135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 79409135cc5SDan Williams * version of this symbol is required. 7953340289dSMel Gorman */ 79609135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 7973340289dSMel Gorman { 7983340289dSMel Gorman return vma_kernel_pagesize(vma); 7993340289dSMel Gorman } 8003340289dSMel Gorman 8013340289dSMel Gorman /* 80284afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 80384afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 80484afd99bSAndy Whitcroft * alignment. 80584afd99bSAndy Whitcroft */ 80684afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 80784afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 80804f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 80984afd99bSAndy Whitcroft 810a1e78772SMel Gorman /* 811a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 812a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 813a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 814a1e78772SMel Gorman * 815a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 816a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 817a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 818a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 81984afd99bSAndy Whitcroft * 82084afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 82184afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 82284afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 82384afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 82484afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 82584afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 82684afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 82784afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 828a1e78772SMel Gorman */ 829e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 830e7c4b0bfSAndy Whitcroft { 831e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 832e7c4b0bfSAndy Whitcroft } 833e7c4b0bfSAndy Whitcroft 834e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 835e7c4b0bfSAndy Whitcroft unsigned long value) 836e7c4b0bfSAndy Whitcroft { 837e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 838e7c4b0bfSAndy Whitcroft } 839e7c4b0bfSAndy Whitcroft 840e9fe92aeSMina Almasry static void 841e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 842e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 843e9fe92aeSMina Almasry struct hstate *h) 844e9fe92aeSMina Almasry { 845e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 846e9fe92aeSMina Almasry if (!h_cg || !h) { 847e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 848e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 849e9fe92aeSMina Almasry resv_map->css = NULL; 850e9fe92aeSMina Almasry } else { 851e9fe92aeSMina Almasry resv_map->reservation_counter = 852e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 853e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 854e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 855e9fe92aeSMina Almasry } 856e9fe92aeSMina Almasry #endif 857e9fe92aeSMina Almasry } 858e9fe92aeSMina Almasry 8599119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 86084afd99bSAndy Whitcroft { 86184afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8625e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8635e911373SMike Kravetz 8645e911373SMike Kravetz if (!resv_map || !rg) { 8655e911373SMike Kravetz kfree(resv_map); 8665e911373SMike Kravetz kfree(rg); 86784afd99bSAndy Whitcroft return NULL; 8685e911373SMike Kravetz } 86984afd99bSAndy Whitcroft 87084afd99bSAndy Whitcroft kref_init(&resv_map->refs); 8717b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 87284afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 87384afd99bSAndy Whitcroft 8745e911373SMike Kravetz resv_map->adds_in_progress = 0; 875e9fe92aeSMina Almasry /* 876e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 877e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 878e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 879e9fe92aeSMina Almasry * reservations are to be un-charged from here. 880e9fe92aeSMina Almasry */ 881e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 8825e911373SMike Kravetz 8835e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 8845e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 8855e911373SMike Kravetz resv_map->region_cache_count = 1; 8865e911373SMike Kravetz 88784afd99bSAndy Whitcroft return resv_map; 88884afd99bSAndy Whitcroft } 88984afd99bSAndy Whitcroft 8909119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 89184afd99bSAndy Whitcroft { 89284afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 8935e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 8945e911373SMike Kravetz struct file_region *rg, *trg; 89584afd99bSAndy Whitcroft 89684afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 897feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 8985e911373SMike Kravetz 8995e911373SMike Kravetz /* ... and any entries left in the cache */ 9005e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9015e911373SMike Kravetz list_del(&rg->link); 9025e911373SMike Kravetz kfree(rg); 9035e911373SMike Kravetz } 9045e911373SMike Kravetz 9055e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9065e911373SMike Kravetz 90784afd99bSAndy Whitcroft kfree(resv_map); 90884afd99bSAndy Whitcroft } 90984afd99bSAndy Whitcroft 9104e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9114e35f483SJoonsoo Kim { 912f27a5136SMike Kravetz /* 913f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 914f27a5136SMike Kravetz * address space within the inode. This original address space 915f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 916f27a5136SMike Kravetz * address space embedded within the inode. 917f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 918f27a5136SMike Kravetz * this may not be true for device special inodes. 919f27a5136SMike Kravetz */ 920f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9214e35f483SJoonsoo Kim } 9224e35f483SJoonsoo Kim 92384afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 924a1e78772SMel Gorman { 92581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9264e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9274e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9284e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9294e35f483SJoonsoo Kim 9304e35f483SJoonsoo Kim return inode_resv_map(inode); 9314e35f483SJoonsoo Kim 9324e35f483SJoonsoo Kim } else { 93384afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 93484afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9354e35f483SJoonsoo Kim } 936a1e78772SMel Gorman } 937a1e78772SMel Gorman 93884afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 939a1e78772SMel Gorman { 94081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 94181d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 942a1e78772SMel Gorman 94384afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 94484afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 94504f2cbe3SMel Gorman } 94604f2cbe3SMel Gorman 94704f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 94804f2cbe3SMel Gorman { 94981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 95081d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 951e7c4b0bfSAndy Whitcroft 952e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 95304f2cbe3SMel Gorman } 95404f2cbe3SMel Gorman 95504f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 95604f2cbe3SMel Gorman { 95781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 958e7c4b0bfSAndy Whitcroft 959e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 960a1e78772SMel Gorman } 961a1e78772SMel Gorman 96204f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 963a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 964a1e78772SMel Gorman { 96581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 966f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 967a1e78772SMel Gorman vma->vm_private_data = (void *)0; 968a1e78772SMel Gorman } 969a1e78772SMel Gorman 970a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 971559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 972a1e78772SMel Gorman { 973af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 974af0ed73eSJoonsoo Kim /* 975af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 976af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 977af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 978af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 979af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 980af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 981af0ed73eSJoonsoo Kim * properly, so add work-around here. 982af0ed73eSJoonsoo Kim */ 983af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 984559ec2f8SNicholas Krause return true; 985af0ed73eSJoonsoo Kim else 986559ec2f8SNicholas Krause return false; 987af0ed73eSJoonsoo Kim } 988a63884e9SJoonsoo Kim 989a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 9901fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 9911fb1b0e9SMike Kravetz /* 9921fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 9931fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 9941fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 9957c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 9961fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 9971fb1b0e9SMike Kravetz */ 9981fb1b0e9SMike Kravetz if (chg) 9991fb1b0e9SMike Kravetz return false; 10001fb1b0e9SMike Kravetz else 1001559ec2f8SNicholas Krause return true; 10021fb1b0e9SMike Kravetz } 1003a63884e9SJoonsoo Kim 1004a63884e9SJoonsoo Kim /* 1005a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1006a63884e9SJoonsoo Kim * private mappings. 1007a63884e9SJoonsoo Kim */ 100867961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 100967961f9dSMike Kravetz /* 101067961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 101167961f9dSMike Kravetz * could have been performed on the private mapping. 101267961f9dSMike Kravetz * Examine the value of chg to determine if reserves 101367961f9dSMike Kravetz * actually exist or were previously consumed. 101467961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 101567961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 101667961f9dSMike Kravetz * private mappings has different (opposite) semantics 101767961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 101867961f9dSMike Kravetz * has already taken this difference in semantics into 101967961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 102067961f9dSMike Kravetz * as in the shared case above. Code could easily be 102167961f9dSMike Kravetz * combined, but keeping it separate draws attention to 102267961f9dSMike Kravetz * subtle differences. 102367961f9dSMike Kravetz */ 102467961f9dSMike Kravetz if (chg) 102567961f9dSMike Kravetz return false; 102667961f9dSMike Kravetz else 1027559ec2f8SNicholas Krause return true; 102867961f9dSMike Kravetz } 1029a63884e9SJoonsoo Kim 1030559ec2f8SNicholas Krause return false; 1031a1e78772SMel Gorman } 1032a1e78772SMel Gorman 1033a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10341da177e4SLinus Torvalds { 10351da177e4SLinus Torvalds int nid = page_to_nid(page); 10360edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1037a5516438SAndi Kleen h->free_huge_pages++; 1038a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10391da177e4SLinus Torvalds } 10401da177e4SLinus Torvalds 104194310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1042bf50bab2SNaoya Horiguchi { 1043bf50bab2SNaoya Horiguchi struct page *page; 1044bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1045bf50bab2SNaoya Horiguchi 1046bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1047bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1048bbe88753SJoonsoo Kim continue; 1049bbe88753SJoonsoo Kim 1050243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 1051c8721bbbSNaoya Horiguchi break; 1052bbe88753SJoonsoo Kim } 1053bbe88753SJoonsoo Kim 1054c8721bbbSNaoya Horiguchi /* 1055c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 1056c8721bbbSNaoya Horiguchi * the allocation fails. 1057c8721bbbSNaoya Horiguchi */ 1058c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 1059bf50bab2SNaoya Horiguchi return NULL; 10600edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1061a9869b83SNaoya Horiguchi set_page_refcounted(page); 1062bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1063bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1064bf50bab2SNaoya Horiguchi return page; 1065bf50bab2SNaoya Horiguchi } 1066bf50bab2SNaoya Horiguchi 10673e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10683e59fcb0SMichal Hocko nodemask_t *nmask) 106994310cbcSAnshuman Khandual { 10703e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 10713e59fcb0SMichal Hocko struct zonelist *zonelist; 10723e59fcb0SMichal Hocko struct zone *zone; 10733e59fcb0SMichal Hocko struct zoneref *z; 107498fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 10753e59fcb0SMichal Hocko 10763e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 10773e59fcb0SMichal Hocko 10783e59fcb0SMichal Hocko retry_cpuset: 10793e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 10803e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 108194310cbcSAnshuman Khandual struct page *page; 108294310cbcSAnshuman Khandual 10833e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 10843e59fcb0SMichal Hocko continue; 10853e59fcb0SMichal Hocko /* 10863e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 10873e59fcb0SMichal Hocko * zone aware 10883e59fcb0SMichal Hocko */ 10893e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 10903e59fcb0SMichal Hocko continue; 10913e59fcb0SMichal Hocko node = zone_to_nid(zone); 109294310cbcSAnshuman Khandual 109394310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 109494310cbcSAnshuman Khandual if (page) 109594310cbcSAnshuman Khandual return page; 109694310cbcSAnshuman Khandual } 10973e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 10983e59fcb0SMichal Hocko goto retry_cpuset; 10993e59fcb0SMichal Hocko 110094310cbcSAnshuman Khandual return NULL; 110194310cbcSAnshuman Khandual } 110294310cbcSAnshuman Khandual 1103a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1104a5516438SAndi Kleen struct vm_area_struct *vma, 1105af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1106af0ed73eSJoonsoo Kim long chg) 11071da177e4SLinus Torvalds { 11083e59fcb0SMichal Hocko struct page *page; 1109480eccf9SLee Schermerhorn struct mempolicy *mpol; 111004ec6264SVlastimil Babka gfp_t gfp_mask; 11113e59fcb0SMichal Hocko nodemask_t *nodemask; 111204ec6264SVlastimil Babka int nid; 11131da177e4SLinus Torvalds 1114a1e78772SMel Gorman /* 1115a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1116a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1117a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1118a1e78772SMel Gorman */ 1119af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1120a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1121c0ff7453SMiao Xie goto err; 1122a1e78772SMel Gorman 112304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1124a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11256eab04a8SJustin P. Mattock goto err; 112604f2cbe3SMel Gorman 112704ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 112804ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11293e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11303e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 113107443a85SJoonsoo Kim SetPagePrivate(page); 1132a63884e9SJoonsoo Kim h->resv_huge_pages--; 1133bf50bab2SNaoya Horiguchi } 1134cc9a6c87SMel Gorman 1135cc9a6c87SMel Gorman mpol_cond_put(mpol); 1136cc9a6c87SMel Gorman return page; 1137cc9a6c87SMel Gorman 1138c0ff7453SMiao Xie err: 1139cc9a6c87SMel Gorman return NULL; 11401da177e4SLinus Torvalds } 11411da177e4SLinus Torvalds 11421cac6f2cSLuiz Capitulino /* 11431cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11441cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11451cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11461cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11471cac6f2cSLuiz Capitulino * node for alloc or free. 11481cac6f2cSLuiz Capitulino */ 11491cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11501cac6f2cSLuiz Capitulino { 11510edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11521cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11531cac6f2cSLuiz Capitulino 11541cac6f2cSLuiz Capitulino return nid; 11551cac6f2cSLuiz Capitulino } 11561cac6f2cSLuiz Capitulino 11571cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11581cac6f2cSLuiz Capitulino { 11591cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11601cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11611cac6f2cSLuiz Capitulino return nid; 11621cac6f2cSLuiz Capitulino } 11631cac6f2cSLuiz Capitulino 11641cac6f2cSLuiz Capitulino /* 11651cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11661cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11671cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11681cac6f2cSLuiz Capitulino * mask. 11691cac6f2cSLuiz Capitulino */ 11701cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 11711cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 11721cac6f2cSLuiz Capitulino { 11731cac6f2cSLuiz Capitulino int nid; 11741cac6f2cSLuiz Capitulino 11751cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11761cac6f2cSLuiz Capitulino 11771cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 11781cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 11791cac6f2cSLuiz Capitulino 11801cac6f2cSLuiz Capitulino return nid; 11811cac6f2cSLuiz Capitulino } 11821cac6f2cSLuiz Capitulino 11831cac6f2cSLuiz Capitulino /* 11841cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 11851cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 11861cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 11871cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 11881cac6f2cSLuiz Capitulino */ 11891cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 11901cac6f2cSLuiz Capitulino { 11911cac6f2cSLuiz Capitulino int nid; 11921cac6f2cSLuiz Capitulino 11931cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11941cac6f2cSLuiz Capitulino 11951cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 11961cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 11971cac6f2cSLuiz Capitulino 11981cac6f2cSLuiz Capitulino return nid; 11991cac6f2cSLuiz Capitulino } 12001cac6f2cSLuiz Capitulino 12011cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12021cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12031cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12041cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12051cac6f2cSLuiz Capitulino nr_nodes--) 12061cac6f2cSLuiz Capitulino 12071cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12081cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12091cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12101cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12111cac6f2cSLuiz Capitulino nr_nodes--) 12121cac6f2cSLuiz Capitulino 1213e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1214944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1215d00181b9SKirill A. Shutemov unsigned int order) 1216944d9fecSLuiz Capitulino { 1217944d9fecSLuiz Capitulino int i; 1218944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1219944d9fecSLuiz Capitulino struct page *p = page + 1; 1220944d9fecSLuiz Capitulino 1221c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 122247e29d32SJohn Hubbard if (hpage_pincount_available(page)) 122347e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 122447e29d32SJohn Hubbard 1225944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12261d798ca3SKirill A. Shutemov clear_compound_head(p); 1227944d9fecSLuiz Capitulino set_page_refcounted(p); 1228944d9fecSLuiz Capitulino } 1229944d9fecSLuiz Capitulino 1230944d9fecSLuiz Capitulino set_compound_order(page, 0); 1231944d9fecSLuiz Capitulino __ClearPageHead(page); 1232944d9fecSLuiz Capitulino } 1233944d9fecSLuiz Capitulino 1234d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1235944d9fecSLuiz Capitulino { 1236cf11e85fSRoman Gushchin /* 1237cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1238cf11e85fSRoman Gushchin * cma_release() returns false. 1239cf11e85fSRoman Gushchin */ 1240dbda8feaSBarry Song #ifdef CONFIG_CMA 1241dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1242cf11e85fSRoman Gushchin return; 1243dbda8feaSBarry Song #endif 1244cf11e85fSRoman Gushchin 1245944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1246944d9fecSLuiz Capitulino } 1247944d9fecSLuiz Capitulino 12484eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1249d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1250d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1251944d9fecSLuiz Capitulino { 12525e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1253944d9fecSLuiz Capitulino 1254dbda8feaSBarry Song #ifdef CONFIG_CMA 1255dbda8feaSBarry Song { 1256cf11e85fSRoman Gushchin struct page *page; 1257cf11e85fSRoman Gushchin int node; 1258cf11e85fSRoman Gushchin 1259cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1260cf11e85fSRoman Gushchin if (!hugetlb_cma[node]) 1261cf11e85fSRoman Gushchin continue; 1262cf11e85fSRoman Gushchin 1263cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1264cf11e85fSRoman Gushchin huge_page_order(h), true); 1265cf11e85fSRoman Gushchin if (page) 1266cf11e85fSRoman Gushchin return page; 1267cf11e85fSRoman Gushchin } 1268cf11e85fSRoman Gushchin } 1269dbda8feaSBarry Song #endif 1270cf11e85fSRoman Gushchin 12715e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1272944d9fecSLuiz Capitulino } 1273944d9fecSLuiz Capitulino 1274944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1275d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 12764eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 12774eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 12784eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 12794eb0716eSAlexandre Ghiti { 12804eb0716eSAlexandre Ghiti return NULL; 12814eb0716eSAlexandre Ghiti } 12824eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1283944d9fecSLuiz Capitulino 1284e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1285d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 12864eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 12874eb0716eSAlexandre Ghiti { 12884eb0716eSAlexandre Ghiti return NULL; 12894eb0716eSAlexandre Ghiti } 1290d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1291944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1292d00181b9SKirill A. Shutemov unsigned int order) { } 1293944d9fecSLuiz Capitulino #endif 1294944d9fecSLuiz Capitulino 1295a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 12966af2acb6SAdam Litke { 12976af2acb6SAdam Litke int i; 1298a5516438SAndi Kleen 12994eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1300944d9fecSLuiz Capitulino return; 130118229df5SAndy Whitcroft 1302a5516438SAndi Kleen h->nr_huge_pages--; 1303a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1304a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 130532f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 130632f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1307a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1308a7407a27SLuiz Capitulino 1 << PG_writeback); 13096af2acb6SAdam Litke } 1310309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13111adc4d41SMina Almasry VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1312f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13136af2acb6SAdam Litke set_page_refcounted(page); 1314944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1315cf11e85fSRoman Gushchin /* 1316cf11e85fSRoman Gushchin * Temporarily drop the hugetlb_lock, because 1317cf11e85fSRoman Gushchin * we might block in free_gigantic_page(). 1318cf11e85fSRoman Gushchin */ 1319cf11e85fSRoman Gushchin spin_unlock(&hugetlb_lock); 1320944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1321944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1322cf11e85fSRoman Gushchin spin_lock(&hugetlb_lock); 1323944d9fecSLuiz Capitulino } else { 1324a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13256af2acb6SAdam Litke } 1326944d9fecSLuiz Capitulino } 13276af2acb6SAdam Litke 1328e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1329e5ff2159SAndi Kleen { 1330e5ff2159SAndi Kleen struct hstate *h; 1331e5ff2159SAndi Kleen 1332e5ff2159SAndi Kleen for_each_hstate(h) { 1333e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1334e5ff2159SAndi Kleen return h; 1335e5ff2159SAndi Kleen } 1336e5ff2159SAndi Kleen return NULL; 1337e5ff2159SAndi Kleen } 1338e5ff2159SAndi Kleen 1339bcc54222SNaoya Horiguchi /* 1340bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1341bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1342bcc54222SNaoya Horiguchi * 1343bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1344bcc54222SNaoya Horiguchi */ 1345bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1346bcc54222SNaoya Horiguchi { 1347bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1348bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1349bcc54222SNaoya Horiguchi } 1350bcc54222SNaoya Horiguchi 1351bcc54222SNaoya Horiguchi /* never called for tail page */ 1352bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1353bcc54222SNaoya Horiguchi { 1354bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1355bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1356bcc54222SNaoya Horiguchi } 1357bcc54222SNaoya Horiguchi 1358bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1359bcc54222SNaoya Horiguchi { 1360bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1361bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1362bcc54222SNaoya Horiguchi } 1363bcc54222SNaoya Horiguchi 1364ab5ac90aSMichal Hocko /* 1365ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1366ab5ac90aSMichal Hocko * code 1367ab5ac90aSMichal Hocko */ 1368ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1369ab5ac90aSMichal Hocko { 1370ab5ac90aSMichal Hocko if (!PageHuge(page)) 1371ab5ac90aSMichal Hocko return false; 1372ab5ac90aSMichal Hocko 1373ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1374ab5ac90aSMichal Hocko } 1375ab5ac90aSMichal Hocko 1376ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1377ab5ac90aSMichal Hocko { 1378ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1379ab5ac90aSMichal Hocko } 1380ab5ac90aSMichal Hocko 1381ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1382ab5ac90aSMichal Hocko { 1383ab5ac90aSMichal Hocko page[2].mapping = NULL; 1384ab5ac90aSMichal Hocko } 1385ab5ac90aSMichal Hocko 1386c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 138727a85ef1SDavid Gibson { 1388a5516438SAndi Kleen /* 1389a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1390a5516438SAndi Kleen * compound page destructor. 1391a5516438SAndi Kleen */ 1392e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 13937893d1d5SAdam Litke int nid = page_to_nid(page); 139490481622SDavid Gibson struct hugepage_subpool *spool = 139590481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 139607443a85SJoonsoo Kim bool restore_reserve; 139727a85ef1SDavid Gibson 1398b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1399b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14008ace22bcSYongkai Wu 14018ace22bcSYongkai Wu set_page_private(page, 0); 14028ace22bcSYongkai Wu page->mapping = NULL; 140307443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 140416c794b4SJoonsoo Kim ClearPagePrivate(page); 140527a85ef1SDavid Gibson 14061c5ecae3SMike Kravetz /* 14070919e1b6SMike Kravetz * If PagePrivate() was set on page, page allocation consumed a 14080919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14090919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14100919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14110919e1b6SMike Kravetz * reservtion, do not call hugepage_subpool_put_pages() as this will 14120919e1b6SMike Kravetz * remove the reserved page from the subpool. 14130919e1b6SMike Kravetz */ 14140919e1b6SMike Kravetz if (!restore_reserve) { 14150919e1b6SMike Kravetz /* 14160919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14170919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14180919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14190919e1b6SMike Kravetz * operation. 14201c5ecae3SMike Kravetz */ 14211c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14221c5ecae3SMike Kravetz restore_reserve = true; 14230919e1b6SMike Kravetz } 14241c5ecae3SMike Kravetz 142527a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1426bcc54222SNaoya Horiguchi clear_page_huge_active(page); 14276d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14286d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 142908cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 143008cf9fafSMina Almasry pages_per_huge_page(h), page); 143107443a85SJoonsoo Kim if (restore_reserve) 143207443a85SJoonsoo Kim h->resv_huge_pages++; 143307443a85SJoonsoo Kim 1434ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1435ab5ac90aSMichal Hocko list_del(&page->lru); 1436ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1437ab5ac90aSMichal Hocko update_and_free_page(h, page); 1438ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14390edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14400edaecfaSAneesh Kumar K.V list_del(&page->lru); 1441a5516438SAndi Kleen update_and_free_page(h, page); 1442a5516438SAndi Kleen h->surplus_huge_pages--; 1443a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 14447893d1d5SAdam Litke } else { 14455d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1446a5516438SAndi Kleen enqueue_huge_page(h, page); 14477893d1d5SAdam Litke } 144827a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 144927a85ef1SDavid Gibson } 145027a85ef1SDavid Gibson 1451c77c0a8aSWaiman Long /* 1452c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1453c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1454c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1455c77c0a8aSWaiman Long * 1456c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1457c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1458c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1459c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1460c77c0a8aSWaiman Long */ 1461c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1462c77c0a8aSWaiman Long 1463c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1464c77c0a8aSWaiman Long { 1465c77c0a8aSWaiman Long struct llist_node *node; 1466c77c0a8aSWaiman Long struct page *page; 1467c77c0a8aSWaiman Long 1468c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1469c77c0a8aSWaiman Long 1470c77c0a8aSWaiman Long while (node) { 1471c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1472c77c0a8aSWaiman Long struct page, mapping); 1473c77c0a8aSWaiman Long node = node->next; 1474c77c0a8aSWaiman Long __free_huge_page(page); 1475c77c0a8aSWaiman Long } 1476c77c0a8aSWaiman Long } 1477c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1478c77c0a8aSWaiman Long 1479c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1480c77c0a8aSWaiman Long { 1481c77c0a8aSWaiman Long /* 1482c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1483c77c0a8aSWaiman Long */ 1484c77c0a8aSWaiman Long if (!in_task()) { 1485c77c0a8aSWaiman Long /* 1486c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1487c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1488c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1489c77c0a8aSWaiman Long */ 1490c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1491c77c0a8aSWaiman Long &hpage_freelist)) 1492c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1493c77c0a8aSWaiman Long return; 1494c77c0a8aSWaiman Long } 1495c77c0a8aSWaiman Long 1496c77c0a8aSWaiman Long __free_huge_page(page); 1497c77c0a8aSWaiman Long } 1498c77c0a8aSWaiman Long 1499a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1500b7ba30c6SAndi Kleen { 15010edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1502f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1503b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 15049dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15051adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1506a5516438SAndi Kleen h->nr_huge_pages++; 1507a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1508b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1509b7ba30c6SAndi Kleen } 1510b7ba30c6SAndi Kleen 1511d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 151220a0307cSWu Fengguang { 151320a0307cSWu Fengguang int i; 151420a0307cSWu Fengguang int nr_pages = 1 << order; 151520a0307cSWu Fengguang struct page *p = page + 1; 151620a0307cSWu Fengguang 151720a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 151820a0307cSWu Fengguang set_compound_order(page, order); 1519ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1520de09d31dSKirill A. Shutemov __SetPageHead(page); 152120a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1522ef5a22beSAndrea Arcangeli /* 1523ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1524ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1525ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15267c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1527ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1528ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1529ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1530ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1531ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1532ef5a22beSAndrea Arcangeli * after get_user_pages(). 1533ef5a22beSAndrea Arcangeli */ 1534ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 153558a84aa9SYouquan Song set_page_count(p, 0); 15361d798ca3SKirill A. Shutemov set_compound_head(p, page); 153720a0307cSWu Fengguang } 1538b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 153947e29d32SJohn Hubbard 154047e29d32SJohn Hubbard if (hpage_pincount_available(page)) 154147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 154220a0307cSWu Fengguang } 154320a0307cSWu Fengguang 15447795912cSAndrew Morton /* 15457795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15467795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15477795912cSAndrew Morton * details. 15487795912cSAndrew Morton */ 154920a0307cSWu Fengguang int PageHuge(struct page *page) 155020a0307cSWu Fengguang { 155120a0307cSWu Fengguang if (!PageCompound(page)) 155220a0307cSWu Fengguang return 0; 155320a0307cSWu Fengguang 155420a0307cSWu Fengguang page = compound_head(page); 1555f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 155620a0307cSWu Fengguang } 155743131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 155843131e14SNaoya Horiguchi 155927c73ae7SAndrea Arcangeli /* 156027c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 156127c73ae7SAndrea Arcangeli * normal or transparent huge pages. 156227c73ae7SAndrea Arcangeli */ 156327c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 156427c73ae7SAndrea Arcangeli { 156527c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 156627c73ae7SAndrea Arcangeli return 0; 156727c73ae7SAndrea Arcangeli 1568d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 156927c73ae7SAndrea Arcangeli } 157027c73ae7SAndrea Arcangeli 1571c0d0381aSMike Kravetz /* 1572c0d0381aSMike Kravetz * Find address_space associated with hugetlbfs page. 1573c0d0381aSMike Kravetz * Upon entry page is locked and page 'was' mapped although mapped state 1574c0d0381aSMike Kravetz * could change. If necessary, use anon_vma to find vma and associated 1575c0d0381aSMike Kravetz * address space. The returned mapping may be stale, but it can not be 1576c0d0381aSMike Kravetz * invalid as page lock (which is held) is required to destroy mapping. 1577c0d0381aSMike Kravetz */ 1578c0d0381aSMike Kravetz static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) 1579c0d0381aSMike Kravetz { 1580c0d0381aSMike Kravetz struct anon_vma *anon_vma; 1581c0d0381aSMike Kravetz pgoff_t pgoff_start, pgoff_end; 1582c0d0381aSMike Kravetz struct anon_vma_chain *avc; 1583c0d0381aSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1584c0d0381aSMike Kravetz 1585c0d0381aSMike Kravetz /* Simple file based mapping */ 1586c0d0381aSMike Kravetz if (mapping) 1587c0d0381aSMike Kravetz return mapping; 1588c0d0381aSMike Kravetz 1589c0d0381aSMike Kravetz /* 1590c0d0381aSMike Kravetz * Even anonymous hugetlbfs mappings are associated with an 1591c0d0381aSMike Kravetz * underlying hugetlbfs file (see hugetlb_file_setup in mmap 1592c0d0381aSMike Kravetz * code). Find a vma associated with the anonymous vma, and 1593c0d0381aSMike Kravetz * use the file pointer to get address_space. 1594c0d0381aSMike Kravetz */ 1595c0d0381aSMike Kravetz anon_vma = page_lock_anon_vma_read(hpage); 1596c0d0381aSMike Kravetz if (!anon_vma) 1597c0d0381aSMike Kravetz return mapping; /* NULL */ 1598c0d0381aSMike Kravetz 1599c0d0381aSMike Kravetz /* Use first found vma */ 1600c0d0381aSMike Kravetz pgoff_start = page_to_pgoff(hpage); 16011139d336SMike Kravetz pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1; 1602c0d0381aSMike Kravetz anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 1603c0d0381aSMike Kravetz pgoff_start, pgoff_end) { 1604c0d0381aSMike Kravetz struct vm_area_struct *vma = avc->vma; 1605c0d0381aSMike Kravetz 1606c0d0381aSMike Kravetz mapping = vma->vm_file->f_mapping; 1607c0d0381aSMike Kravetz break; 1608c0d0381aSMike Kravetz } 1609c0d0381aSMike Kravetz 1610c0d0381aSMike Kravetz anon_vma_unlock_read(anon_vma); 1611c0d0381aSMike Kravetz return mapping; 1612c0d0381aSMike Kravetz } 1613c0d0381aSMike Kravetz 1614c0d0381aSMike Kravetz /* 1615c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1616c0d0381aSMike Kravetz * 1617c0d0381aSMike Kravetz * Upon entry, the page is locked which allows us to find the mapping 1618c0d0381aSMike Kravetz * even in the case of an anon page. However, locking order dictates 1619c0d0381aSMike Kravetz * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs 1620c0d0381aSMike Kravetz * specific. So, we first try to lock the sema while still holding the 1621c0d0381aSMike Kravetz * page lock. If this works, great! If not, then we need to drop the 1622c0d0381aSMike Kravetz * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of 1623c0d0381aSMike Kravetz * course, need to revalidate state along the way. 1624c0d0381aSMike Kravetz */ 1625c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1626c0d0381aSMike Kravetz { 1627c0d0381aSMike Kravetz struct address_space *mapping, *mapping2; 1628c0d0381aSMike Kravetz 1629c0d0381aSMike Kravetz mapping = _get_hugetlb_page_mapping(hpage); 1630c0d0381aSMike Kravetz retry: 1631c0d0381aSMike Kravetz if (!mapping) 1632c0d0381aSMike Kravetz return mapping; 1633c0d0381aSMike Kravetz 1634c0d0381aSMike Kravetz /* 1635c0d0381aSMike Kravetz * If no contention, take lock and return 1636c0d0381aSMike Kravetz */ 1637c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1638c0d0381aSMike Kravetz return mapping; 1639c0d0381aSMike Kravetz 1640c0d0381aSMike Kravetz /* 1641c0d0381aSMike Kravetz * Must drop page lock and wait on mapping sema. 1642c0d0381aSMike Kravetz * Note: Once page lock is dropped, mapping could become invalid. 1643c0d0381aSMike Kravetz * As a hack, increase map count until we lock page again. 1644c0d0381aSMike Kravetz */ 1645c0d0381aSMike Kravetz atomic_inc(&hpage->_mapcount); 1646c0d0381aSMike Kravetz unlock_page(hpage); 1647c0d0381aSMike Kravetz i_mmap_lock_write(mapping); 1648c0d0381aSMike Kravetz lock_page(hpage); 1649c0d0381aSMike Kravetz atomic_add_negative(-1, &hpage->_mapcount); 1650c0d0381aSMike Kravetz 1651c0d0381aSMike Kravetz /* verify page is still mapped */ 1652c0d0381aSMike Kravetz if (!page_mapped(hpage)) { 1653c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1654c0d0381aSMike Kravetz return NULL; 1655c0d0381aSMike Kravetz } 1656c0d0381aSMike Kravetz 1657c0d0381aSMike Kravetz /* 1658c0d0381aSMike Kravetz * Get address space again and verify it is the same one 1659c0d0381aSMike Kravetz * we locked. If not, drop lock and retry. 1660c0d0381aSMike Kravetz */ 1661c0d0381aSMike Kravetz mapping2 = _get_hugetlb_page_mapping(hpage); 1662c0d0381aSMike Kravetz if (mapping2 != mapping) { 1663c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1664c0d0381aSMike Kravetz mapping = mapping2; 1665c0d0381aSMike Kravetz goto retry; 1666c0d0381aSMike Kravetz } 1667c0d0381aSMike Kravetz 1668c0d0381aSMike Kravetz return mapping; 1669c0d0381aSMike Kravetz } 1670c0d0381aSMike Kravetz 167113d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 167213d60f4bSZhang Yi { 167313d60f4bSZhang Yi struct page *page_head = compound_head(page); 167413d60f4bSZhang Yi pgoff_t index = page_index(page_head); 167513d60f4bSZhang Yi unsigned long compound_idx; 167613d60f4bSZhang Yi 167713d60f4bSZhang Yi if (!PageHuge(page_head)) 167813d60f4bSZhang Yi return page_index(page); 167913d60f4bSZhang Yi 168013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 168113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 168213d60f4bSZhang Yi else 168313d60f4bSZhang Yi compound_idx = page - page_head; 168413d60f4bSZhang Yi 168513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 168613d60f4bSZhang Yi } 168713d60f4bSZhang Yi 16880c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1689f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1690f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16911da177e4SLinus Torvalds { 1692af0fb9dfSMichal Hocko int order = huge_page_order(h); 16931da177e4SLinus Torvalds struct page *page; 1694f60858f9SMike Kravetz bool alloc_try_hard = true; 1695f96efd58SJoe Jin 1696f60858f9SMike Kravetz /* 1697f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1698f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1699f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1700f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1701f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1702f60858f9SMike Kravetz */ 1703f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1704f60858f9SMike Kravetz alloc_try_hard = false; 1705f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1706f60858f9SMike Kravetz if (alloc_try_hard) 1707f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1708af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1709af0fb9dfSMichal Hocko nid = numa_mem_id(); 1710af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1711af0fb9dfSMichal Hocko if (page) 1712af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1713af0fb9dfSMichal Hocko else 1714af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 171563b4613cSNishanth Aravamudan 1716f60858f9SMike Kravetz /* 1717f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1718f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1719f60858f9SMike Kravetz * normal 'try hard' allocations. 1720f60858f9SMike Kravetz */ 1721f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1722f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1723f60858f9SMike Kravetz 1724f60858f9SMike Kravetz /* 1725f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1726f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1727f60858f9SMike Kravetz * overall state change. 1728f60858f9SMike Kravetz */ 1729f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1730f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1731f60858f9SMike Kravetz 173263b4613cSNishanth Aravamudan return page; 173363b4613cSNishanth Aravamudan } 173463b4613cSNishanth Aravamudan 1735af0fb9dfSMichal Hocko /* 17360c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 17370c397daeSMichal Hocko * should use this function to get new hugetlb pages 17380c397daeSMichal Hocko */ 17390c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1740f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1741f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17420c397daeSMichal Hocko { 17430c397daeSMichal Hocko struct page *page; 17440c397daeSMichal Hocko 17450c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17460c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 17470c397daeSMichal Hocko else 17480c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1749f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 17500c397daeSMichal Hocko if (!page) 17510c397daeSMichal Hocko return NULL; 17520c397daeSMichal Hocko 17530c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17540c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 17550c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 17560c397daeSMichal Hocko 17570c397daeSMichal Hocko return page; 17580c397daeSMichal Hocko } 17590c397daeSMichal Hocko 17600c397daeSMichal Hocko /* 1761af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1762af0fb9dfSMichal Hocko * manner. 1763af0fb9dfSMichal Hocko */ 1764f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1765f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1766b2261026SJoonsoo Kim { 1767b2261026SJoonsoo Kim struct page *page; 1768b2261026SJoonsoo Kim int nr_nodes, node; 1769af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1770b2261026SJoonsoo Kim 1771b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1772f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1773f60858f9SMike Kravetz node_alloc_noretry); 1774af0fb9dfSMichal Hocko if (page) 1775b2261026SJoonsoo Kim break; 1776b2261026SJoonsoo Kim } 1777b2261026SJoonsoo Kim 1778af0fb9dfSMichal Hocko if (!page) 1779af0fb9dfSMichal Hocko return 0; 1780b2261026SJoonsoo Kim 1781af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1782af0fb9dfSMichal Hocko 1783af0fb9dfSMichal Hocko return 1; 1784b2261026SJoonsoo Kim } 1785b2261026SJoonsoo Kim 1786e8c5c824SLee Schermerhorn /* 1787e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1788e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1789e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1790e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1791e8c5c824SLee Schermerhorn */ 17926ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17936ae11b27SLee Schermerhorn bool acct_surplus) 1794e8c5c824SLee Schermerhorn { 1795b2261026SJoonsoo Kim int nr_nodes, node; 1796e8c5c824SLee Schermerhorn int ret = 0; 1797e8c5c824SLee Schermerhorn 1798b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1799685f3457SLee Schermerhorn /* 1800685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1801685f3457SLee Schermerhorn * nodes with surplus pages. 1802685f3457SLee Schermerhorn */ 1803b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1804b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1805e8c5c824SLee Schermerhorn struct page *page = 1806b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1807e8c5c824SLee Schermerhorn struct page, lru); 1808e8c5c824SLee Schermerhorn list_del(&page->lru); 1809e8c5c824SLee Schermerhorn h->free_huge_pages--; 1810b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1811685f3457SLee Schermerhorn if (acct_surplus) { 1812685f3457SLee Schermerhorn h->surplus_huge_pages--; 1813b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1814685f3457SLee Schermerhorn } 1815e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1816e8c5c824SLee Schermerhorn ret = 1; 18179a76db09SLee Schermerhorn break; 1818e8c5c824SLee Schermerhorn } 1819b2261026SJoonsoo Kim } 1820e8c5c824SLee Schermerhorn 1821e8c5c824SLee Schermerhorn return ret; 1822e8c5c824SLee Schermerhorn } 1823e8c5c824SLee Schermerhorn 1824c8721bbbSNaoya Horiguchi /* 1825c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1826faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1827faf53defSNaoya Horiguchi * This function returns values like below: 1828faf53defSNaoya Horiguchi * 1829faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1830faf53defSNaoya Horiguchi * (allocated or reserved.) 1831faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1832faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1833c8721bbbSNaoya Horiguchi */ 1834c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1835c8721bbbSNaoya Horiguchi { 18366bc9b564SNaoya Horiguchi int rc = -EBUSY; 1837082d5b6bSGerald Schaefer 1838faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1839faf53defSNaoya Horiguchi if (!PageHuge(page)) 1840faf53defSNaoya Horiguchi return 0; 1841faf53defSNaoya Horiguchi 1842c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1843faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1844faf53defSNaoya Horiguchi rc = 0; 1845faf53defSNaoya Horiguchi goto out; 1846faf53defSNaoya Horiguchi } 1847faf53defSNaoya Horiguchi 1848faf53defSNaoya Horiguchi if (!page_count(page)) { 18492247bb33SGerald Schaefer struct page *head = compound_head(page); 18502247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 18512247bb33SGerald Schaefer int nid = page_to_nid(head); 18526bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1853082d5b6bSGerald Schaefer goto out; 1854c3114a84SAnshuman Khandual /* 1855c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1856c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1857c3114a84SAnshuman Khandual */ 1858c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1859c3114a84SAnshuman Khandual SetPageHWPoison(page); 1860c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1861c3114a84SAnshuman Khandual } 18622247bb33SGerald Schaefer list_del(&head->lru); 1863c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1864c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1865c1470b33Szhong jiang h->max_huge_pages--; 18662247bb33SGerald Schaefer update_and_free_page(h, head); 18676bc9b564SNaoya Horiguchi rc = 0; 1868c8721bbbSNaoya Horiguchi } 1869082d5b6bSGerald Schaefer out: 1870c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1871082d5b6bSGerald Schaefer return rc; 1872c8721bbbSNaoya Horiguchi } 1873c8721bbbSNaoya Horiguchi 1874c8721bbbSNaoya Horiguchi /* 1875c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1876c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18772247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18782247bb33SGerald Schaefer * part of it lies within the given range. 1879082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1880082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1881c8721bbbSNaoya Horiguchi */ 1882082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1883c8721bbbSNaoya Horiguchi { 1884c8721bbbSNaoya Horiguchi unsigned long pfn; 1885eb03aa00SGerald Schaefer struct page *page; 1886082d5b6bSGerald Schaefer int rc = 0; 1887c8721bbbSNaoya Horiguchi 1888d0177639SLi Zhong if (!hugepages_supported()) 1889082d5b6bSGerald Schaefer return rc; 1890d0177639SLi Zhong 1891eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1892eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1893eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1894eb03aa00SGerald Schaefer if (rc) 1895082d5b6bSGerald Schaefer break; 1896eb03aa00SGerald Schaefer } 1897082d5b6bSGerald Schaefer 1898082d5b6bSGerald Schaefer return rc; 1899c8721bbbSNaoya Horiguchi } 1900c8721bbbSNaoya Horiguchi 1901ab5ac90aSMichal Hocko /* 1902ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1903ab5ac90aSMichal Hocko */ 19040c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1905aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 19067893d1d5SAdam Litke { 19079980d744SMichal Hocko struct page *page = NULL; 19087893d1d5SAdam Litke 1909bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1910aa888a74SAndi Kleen return NULL; 1911aa888a74SAndi Kleen 1912d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 19139980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 19149980d744SMichal Hocko goto out_unlock; 1915d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1916d1c3fb1fSNishanth Aravamudan 1917f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 19189980d744SMichal Hocko if (!page) 19190c397daeSMichal Hocko return NULL; 1920d1c3fb1fSNishanth Aravamudan 19217893d1d5SAdam Litke spin_lock(&hugetlb_lock); 19229980d744SMichal Hocko /* 19239980d744SMichal Hocko * We could have raced with the pool size change. 19249980d744SMichal Hocko * Double check that and simply deallocate the new page 19259980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 19269980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 19279980d744SMichal Hocko * codeflow 19289980d744SMichal Hocko */ 19299980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 19309980d744SMichal Hocko SetPageHugeTemporary(page); 19312bf753e6SKai Shen spin_unlock(&hugetlb_lock); 19329980d744SMichal Hocko put_page(page); 19332bf753e6SKai Shen return NULL; 19349980d744SMichal Hocko } else { 19359980d744SMichal Hocko h->surplus_huge_pages++; 19364704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 19377893d1d5SAdam Litke } 19389980d744SMichal Hocko 19399980d744SMichal Hocko out_unlock: 1940d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 19417893d1d5SAdam Litke 19427893d1d5SAdam Litke return page; 19437893d1d5SAdam Litke } 19447893d1d5SAdam Litke 1945bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1946ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1947ab5ac90aSMichal Hocko { 1948ab5ac90aSMichal Hocko struct page *page; 1949ab5ac90aSMichal Hocko 1950ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1951ab5ac90aSMichal Hocko return NULL; 1952ab5ac90aSMichal Hocko 1953f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1954ab5ac90aSMichal Hocko if (!page) 1955ab5ac90aSMichal Hocko return NULL; 1956ab5ac90aSMichal Hocko 1957ab5ac90aSMichal Hocko /* 1958ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1959ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1960ab5ac90aSMichal Hocko */ 1961ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1962ab5ac90aSMichal Hocko 1963ab5ac90aSMichal Hocko return page; 1964ab5ac90aSMichal Hocko } 1965ab5ac90aSMichal Hocko 1966e4e574b7SAdam Litke /* 1967099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1968099730d6SDave Hansen */ 1969e0ec90eeSDave Hansen static 19700c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1971099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1972099730d6SDave Hansen { 1973aaf14e40SMichal Hocko struct page *page; 1974aaf14e40SMichal Hocko struct mempolicy *mpol; 1975aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1976aaf14e40SMichal Hocko int nid; 1977aaf14e40SMichal Hocko nodemask_t *nodemask; 1978aaf14e40SMichal Hocko 1979aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19800c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1981aaf14e40SMichal Hocko mpol_cond_put(mpol); 1982aaf14e40SMichal Hocko 1983aaf14e40SMichal Hocko return page; 1984099730d6SDave Hansen } 1985099730d6SDave Hansen 1986ab5ac90aSMichal Hocko /* page migration callback function */ 19873e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1988d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19894db9b2efSMichal Hocko { 19904db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 19914db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19923e59fcb0SMichal Hocko struct page *page; 19933e59fcb0SMichal Hocko 19943e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19953e59fcb0SMichal Hocko if (page) { 19963e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 19973e59fcb0SMichal Hocko return page; 19984db9b2efSMichal Hocko } 19994db9b2efSMichal Hocko } 20004db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 20014db9b2efSMichal Hocko 20020c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 20034db9b2efSMichal Hocko } 20044db9b2efSMichal Hocko 2005ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2006389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2007389c8178SMichal Hocko unsigned long address) 2008ebd63723SMichal Hocko { 2009ebd63723SMichal Hocko struct mempolicy *mpol; 2010ebd63723SMichal Hocko nodemask_t *nodemask; 2011ebd63723SMichal Hocko struct page *page; 2012ebd63723SMichal Hocko gfp_t gfp_mask; 2013ebd63723SMichal Hocko int node; 2014ebd63723SMichal Hocko 2015ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2016ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2017d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2018ebd63723SMichal Hocko mpol_cond_put(mpol); 2019ebd63723SMichal Hocko 2020ebd63723SMichal Hocko return page; 2021ebd63723SMichal Hocko } 2022ebd63723SMichal Hocko 2023bf50bab2SNaoya Horiguchi /* 202425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2025e4e574b7SAdam Litke * of size 'delta'. 2026e4e574b7SAdam Litke */ 2027a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 20281b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2029e4e574b7SAdam Litke { 2030e4e574b7SAdam Litke struct list_head surplus_list; 2031e4e574b7SAdam Litke struct page *page, *tmp; 2032e4e574b7SAdam Litke int ret, i; 2033e4e574b7SAdam Litke int needed, allocated; 203428073b02SHillf Danton bool alloc_ok = true; 2035e4e574b7SAdam Litke 2036a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2037ac09b3a1SAdam Litke if (needed <= 0) { 2038a5516438SAndi Kleen h->resv_huge_pages += delta; 2039e4e574b7SAdam Litke return 0; 2040ac09b3a1SAdam Litke } 2041e4e574b7SAdam Litke 2042e4e574b7SAdam Litke allocated = 0; 2043e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2044e4e574b7SAdam Litke 2045e4e574b7SAdam Litke ret = -ENOMEM; 2046e4e574b7SAdam Litke retry: 2047e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 2048e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 20490c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2050aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 205128073b02SHillf Danton if (!page) { 205228073b02SHillf Danton alloc_ok = false; 205328073b02SHillf Danton break; 205428073b02SHillf Danton } 2055e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 205669ed779aSDavid Rientjes cond_resched(); 2057e4e574b7SAdam Litke } 205828073b02SHillf Danton allocated += i; 2059e4e574b7SAdam Litke 2060e4e574b7SAdam Litke /* 2061e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2062e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2063e4e574b7SAdam Litke */ 2064e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 2065a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2066a5516438SAndi Kleen (h->free_huge_pages + allocated); 206728073b02SHillf Danton if (needed > 0) { 206828073b02SHillf Danton if (alloc_ok) 2069e4e574b7SAdam Litke goto retry; 207028073b02SHillf Danton /* 207128073b02SHillf Danton * We were not able to allocate enough pages to 207228073b02SHillf Danton * satisfy the entire reservation so we free what 207328073b02SHillf Danton * we've allocated so far. 207428073b02SHillf Danton */ 207528073b02SHillf Danton goto free; 207628073b02SHillf Danton } 2077e4e574b7SAdam Litke /* 2078e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 207925985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2080e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2081ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2082ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2083ac09b3a1SAdam Litke * before they are reserved. 2084e4e574b7SAdam Litke */ 2085e4e574b7SAdam Litke needed += allocated; 2086a5516438SAndi Kleen h->resv_huge_pages += delta; 2087e4e574b7SAdam Litke ret = 0; 2088a9869b83SNaoya Horiguchi 208919fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 209019fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 209119fc3f0aSAdam Litke if ((--needed) < 0) 209219fc3f0aSAdam Litke break; 2093a9869b83SNaoya Horiguchi /* 2094a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2095a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2096a9869b83SNaoya Horiguchi */ 2097a9869b83SNaoya Horiguchi put_page_testzero(page); 2098309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 2099a5516438SAndi Kleen enqueue_huge_page(h, page); 210019fc3f0aSAdam Litke } 210128073b02SHillf Danton free: 2102b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 210319fc3f0aSAdam Litke 210419fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2105c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2106a9869b83SNaoya Horiguchi put_page(page); 210719fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2108e4e574b7SAdam Litke 2109e4e574b7SAdam Litke return ret; 2110e4e574b7SAdam Litke } 2111e4e574b7SAdam Litke 2112e4e574b7SAdam Litke /* 2113e5bbc8a6SMike Kravetz * This routine has two main purposes: 2114e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2115e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2116e5bbc8a6SMike Kravetz * to the associated reservation map. 2117e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2118e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2119e5bbc8a6SMike Kravetz * 2120e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2121e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2122e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2123e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2124e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2125e4e574b7SAdam Litke */ 2126a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2127a5516438SAndi Kleen unsigned long unused_resv_pages) 2128e4e574b7SAdam Litke { 2129e4e574b7SAdam Litke unsigned long nr_pages; 2130e4e574b7SAdam Litke 2131aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2132bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2133e5bbc8a6SMike Kravetz goto out; 2134aa888a74SAndi Kleen 2135e5bbc8a6SMike Kravetz /* 2136e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2137e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2138e5bbc8a6SMike Kravetz */ 2139a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2140e4e574b7SAdam Litke 2141685f3457SLee Schermerhorn /* 2142685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 21439b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 21449b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 21459b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 21469e7ee400SRandy Dunlap * free_pool_huge_page() will balance the freed pages across the 21479b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2148e5bbc8a6SMike Kravetz * 2149e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2150e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2151e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2152685f3457SLee Schermerhorn */ 2153685f3457SLee Schermerhorn while (nr_pages--) { 2154e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2155e5bbc8a6SMike Kravetz unused_resv_pages--; 21568cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2157e5bbc8a6SMike Kravetz goto out; 21587848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2159e4e574b7SAdam Litke } 2160e5bbc8a6SMike Kravetz 2161e5bbc8a6SMike Kravetz out: 2162e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2163e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2164e4e574b7SAdam Litke } 2165e4e574b7SAdam Litke 21665e911373SMike Kravetz 2167c37f9fb1SAndy Whitcroft /* 2168feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21695e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2170cf3ad20bSMike Kravetz * 2171cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2172cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2173cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2174cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2175cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2176feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2177feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2178feba16e2SMike Kravetz * is called in such cases. 2179cf3ad20bSMike Kravetz * 2180cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2181cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2182cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2183cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2184cf3ad20bSMike Kravetz * take appropriate action. 218596b96a96SMike Kravetz * 218696b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 218796b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 218896b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 218996b96a96SMike Kravetz * reservation exists. 2190c37f9fb1SAndy Whitcroft */ 21915e911373SMike Kravetz enum vma_resv_mode { 21925e911373SMike Kravetz VMA_NEEDS_RESV, 21935e911373SMike Kravetz VMA_COMMIT_RESV, 2194feba16e2SMike Kravetz VMA_END_RESV, 219596b96a96SMike Kravetz VMA_ADD_RESV, 21965e911373SMike Kravetz }; 2197cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2198cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21995e911373SMike Kravetz enum vma_resv_mode mode) 2200c37f9fb1SAndy Whitcroft { 22014e35f483SJoonsoo Kim struct resv_map *resv; 22024e35f483SJoonsoo Kim pgoff_t idx; 2203cf3ad20bSMike Kravetz long ret; 22040db9d74eSMina Almasry long dummy_out_regions_needed; 2205c37f9fb1SAndy Whitcroft 22064e35f483SJoonsoo Kim resv = vma_resv_map(vma); 22074e35f483SJoonsoo Kim if (!resv) 2208c37f9fb1SAndy Whitcroft return 1; 2209c37f9fb1SAndy Whitcroft 22104e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 22115e911373SMike Kravetz switch (mode) { 22125e911373SMike Kravetz case VMA_NEEDS_RESV: 22130db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 22140db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 22150db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 22160db9d74eSMina Almasry * ever require 1 region. 22170db9d74eSMina Almasry */ 22180db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 22195e911373SMike Kravetz break; 22205e911373SMike Kravetz case VMA_COMMIT_RESV: 2221075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22220db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22230db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22245e911373SMike Kravetz break; 2225feba16e2SMike Kravetz case VMA_END_RESV: 22260db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 22275e911373SMike Kravetz ret = 0; 22285e911373SMike Kravetz break; 222996b96a96SMike Kravetz case VMA_ADD_RESV: 22300db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2231075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22320db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22330db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22340db9d74eSMina Almasry } else { 22350db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 223696b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 223796b96a96SMike Kravetz } 223896b96a96SMike Kravetz break; 22395e911373SMike Kravetz default: 22405e911373SMike Kravetz BUG(); 22415e911373SMike Kravetz } 224284afd99bSAndy Whitcroft 22434e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2244cf3ad20bSMike Kravetz return ret; 224567961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 224667961f9dSMike Kravetz /* 224767961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 224867961f9dSMike Kravetz * However, a file associated with mapping could have been 224967961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 225067961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 225167961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 225267961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 225367961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 225467961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 225567961f9dSMike Kravetz * reservation has already been consumed. As a result, the 225667961f9dSMike Kravetz * return value of this routine is the opposite of the 225767961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 225867961f9dSMike Kravetz */ 225967961f9dSMike Kravetz if (ret) 226067961f9dSMike Kravetz return 0; 226167961f9dSMike Kravetz else 226267961f9dSMike Kravetz return 1; 226367961f9dSMike Kravetz } 22644e35f483SJoonsoo Kim else 2265cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 226684afd99bSAndy Whitcroft } 2267cf3ad20bSMike Kravetz 2268cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2269a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2270c37f9fb1SAndy Whitcroft { 22715e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2272cf3ad20bSMike Kravetz } 2273c37f9fb1SAndy Whitcroft 2274cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2275cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2276cf3ad20bSMike Kravetz { 22775e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22785e911373SMike Kravetz } 22795e911373SMike Kravetz 2280feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22815e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22825e911373SMike Kravetz { 2283feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2284c37f9fb1SAndy Whitcroft } 2285c37f9fb1SAndy Whitcroft 228696b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 228796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 228896b96a96SMike Kravetz { 228996b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 229096b96a96SMike Kravetz } 229196b96a96SMike Kravetz 229296b96a96SMike Kravetz /* 229396b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 229496b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 229596b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 229696b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 229796b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 229896b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 229996b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 230096b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 230196b96a96SMike Kravetz * to be made by free_huge_page. 230296b96a96SMike Kravetz */ 230396b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 230496b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 230596b96a96SMike Kravetz struct page *page) 230696b96a96SMike Kravetz { 230796b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 230896b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 230996b96a96SMike Kravetz 231096b96a96SMike Kravetz if (unlikely(rc < 0)) { 231196b96a96SMike Kravetz /* 231296b96a96SMike Kravetz * Rare out of memory condition in reserve map 231396b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 231496b96a96SMike Kravetz * global reserve count will not be incremented 231596b96a96SMike Kravetz * by free_huge_page. This will make it appear 231696b96a96SMike Kravetz * as though the reservation for this page was 231796b96a96SMike Kravetz * consumed. This may prevent the task from 231896b96a96SMike Kravetz * faulting in the page at a later time. This 231996b96a96SMike Kravetz * is better than inconsistent global huge page 232096b96a96SMike Kravetz * accounting of reserve counts. 232196b96a96SMike Kravetz */ 232296b96a96SMike Kravetz ClearPagePrivate(page); 232396b96a96SMike Kravetz } else if (rc) { 232496b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 232596b96a96SMike Kravetz if (unlikely(rc < 0)) 232696b96a96SMike Kravetz /* 232796b96a96SMike Kravetz * See above comment about rare out of 232896b96a96SMike Kravetz * memory condition. 232996b96a96SMike Kravetz */ 233096b96a96SMike Kravetz ClearPagePrivate(page); 233196b96a96SMike Kravetz } else 233296b96a96SMike Kravetz vma_end_reservation(h, vma, address); 233396b96a96SMike Kravetz } 233496b96a96SMike Kravetz } 233596b96a96SMike Kravetz 233670c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 233704f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2338348ea204SAdam Litke { 233990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2340a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2341348ea204SAdam Litke struct page *page; 2342d85f69b0SMike Kravetz long map_chg, map_commit; 2343d85f69b0SMike Kravetz long gbl_chg; 23446d76dcf4SAneesh Kumar K.V int ret, idx; 23456d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 234608cf9fafSMina Almasry bool deferred_reserve; 23472fc39cecSAdam Litke 23486d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2349a1e78772SMel Gorman /* 2350d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2351d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2352d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2353a1e78772SMel Gorman */ 2354d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2355d85f69b0SMike Kravetz if (map_chg < 0) 235676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2357d85f69b0SMike Kravetz 2358d85f69b0SMike Kravetz /* 2359d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2360d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2361d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2362d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2363d85f69b0SMike Kravetz * checked against any subpool limit. 2364d85f69b0SMike Kravetz */ 2365d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2366d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2367d85f69b0SMike Kravetz if (gbl_chg < 0) { 2368feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 236976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 23705e911373SMike Kravetz } 237190d8b7e6SAdam Litke 2372d85f69b0SMike Kravetz /* 2373d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2374d85f69b0SMike Kravetz * map, there could be reservations associated with the 2375d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2376d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2377d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2378d85f69b0SMike Kravetz * the subpool reservations. 2379d85f69b0SMike Kravetz */ 2380d85f69b0SMike Kravetz if (avoid_reserve) 2381d85f69b0SMike Kravetz gbl_chg = 1; 2382d85f69b0SMike Kravetz } 2383d85f69b0SMike Kravetz 238408cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 238508cf9fafSMina Almasry */ 238608cf9fafSMina Almasry deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); 238708cf9fafSMina Almasry if (deferred_reserve) { 238808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 238908cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 23908f34af6fSJianyu Zhan if (ret) 23918f34af6fSJianyu Zhan goto out_subpool_put; 239208cf9fafSMina Almasry } 239308cf9fafSMina Almasry 239408cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 239508cf9fafSMina Almasry if (ret) 239608cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 23978f34af6fSJianyu Zhan 2398a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2399d85f69b0SMike Kravetz /* 2400d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2401d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2402d85f69b0SMike Kravetz * a reservation exists for the allocation. 2403d85f69b0SMike Kravetz */ 2404d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 240581a6fcaeSJoonsoo Kim if (!page) { 240694ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 24070c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 24088f34af6fSJianyu Zhan if (!page) 24098f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2410a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2411a88c7695SNaoya Horiguchi SetPagePrivate(page); 2412a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2413a88c7695SNaoya Horiguchi } 241479dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 241579dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 241681a6fcaeSJoonsoo Kim /* Fall through */ 2417a1e78772SMel Gorman } 241881a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 241908cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 242008cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 242108cf9fafSMina Almasry */ 242208cf9fafSMina Almasry if (deferred_reserve) { 242308cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 242408cf9fafSMina Almasry h_cg, page); 242508cf9fafSMina Almasry } 242608cf9fafSMina Almasry 242781a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2428a1e78772SMel Gorman 242990481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2430a1e78772SMel Gorman 2431d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2432d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 243333039678SMike Kravetz /* 243433039678SMike Kravetz * The page was added to the reservation map between 243533039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 243633039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 243733039678SMike Kravetz * Adjust for the subpool count incremented above AND 243833039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 243933039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 244033039678SMike Kravetz * no longer applies. 244133039678SMike Kravetz */ 244233039678SMike Kravetz long rsv_adjust; 244333039678SMike Kravetz 244433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 244533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 244633039678SMike Kravetz } 24477893d1d5SAdam Litke return page; 24488f34af6fSJianyu Zhan 24498f34af6fSJianyu Zhan out_uncharge_cgroup: 24508f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 245108cf9fafSMina Almasry out_uncharge_cgroup_reservation: 245208cf9fafSMina Almasry if (deferred_reserve) 245308cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 245408cf9fafSMina Almasry h_cg); 24558f34af6fSJianyu Zhan out_subpool_put: 2456d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 24578f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2458feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 24598f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2460b45b5bd6SDavid Gibson } 2461b45b5bd6SDavid Gibson 2462e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2463e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2464e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2465aa888a74SAndi Kleen { 2466aa888a74SAndi Kleen struct huge_bootmem_page *m; 2467b2261026SJoonsoo Kim int nr_nodes, node; 2468aa888a74SAndi Kleen 2469b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2470aa888a74SAndi Kleen void *addr; 2471aa888a74SAndi Kleen 2472eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 24738b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 247497ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2475aa888a74SAndi Kleen if (addr) { 2476aa888a74SAndi Kleen /* 2477aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2478aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2479aa888a74SAndi Kleen * puts them into the mem_map). 2480aa888a74SAndi Kleen */ 2481aa888a74SAndi Kleen m = addr; 2482aa888a74SAndi Kleen goto found; 2483aa888a74SAndi Kleen } 2484aa888a74SAndi Kleen } 2485aa888a74SAndi Kleen return 0; 2486aa888a74SAndi Kleen 2487aa888a74SAndi Kleen found: 2488df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2489aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2490330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2491aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2492aa888a74SAndi Kleen m->hstate = h; 2493aa888a74SAndi Kleen return 1; 2494aa888a74SAndi Kleen } 2495aa888a74SAndi Kleen 2496d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2497d00181b9SKirill A. Shutemov unsigned int order) 249818229df5SAndy Whitcroft { 249918229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 250018229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 250118229df5SAndy Whitcroft else 250218229df5SAndy Whitcroft prep_compound_page(page, order); 250318229df5SAndy Whitcroft } 250418229df5SAndy Whitcroft 2505aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2506aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2507aa888a74SAndi Kleen { 2508aa888a74SAndi Kleen struct huge_bootmem_page *m; 2509aa888a74SAndi Kleen 2510aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 251140d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2512aa888a74SAndi Kleen struct hstate *h = m->hstate; 2513ee8f248dSBecky Bruce 2514aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 251518229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2516ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2517aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2518af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2519af0fb9dfSMichal Hocko 2520b0320c7bSRafael Aquini /* 2521b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2522b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2523b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2524b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2525b0320c7bSRafael Aquini */ 2526bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 25273dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2528520495feSCannon Matthews cond_resched(); 2529aa888a74SAndi Kleen } 2530aa888a74SAndi Kleen } 2531aa888a74SAndi Kleen 25328faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 25331da177e4SLinus Torvalds { 25341da177e4SLinus Torvalds unsigned long i; 2535f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2536f60858f9SMike Kravetz 2537f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2538f60858f9SMike Kravetz /* 2539f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2540f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2541f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2542f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2543f60858f9SMike Kravetz */ 2544f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2545f60858f9SMike Kravetz GFP_KERNEL); 2546f60858f9SMike Kravetz } else { 2547f60858f9SMike Kravetz /* allocations done at boot time */ 2548f60858f9SMike Kravetz node_alloc_noretry = NULL; 2549f60858f9SMike Kravetz } 2550f60858f9SMike Kravetz 2551f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2552f60858f9SMike Kravetz if (node_alloc_noretry) 2553f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 25541da177e4SLinus Torvalds 2555e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2556bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2557dbda8feaSBarry Song if (hugetlb_cma_size) { 2558cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 2559cf11e85fSRoman Gushchin break; 2560cf11e85fSRoman Gushchin } 2561aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2562aa888a74SAndi Kleen break; 25630c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2564f60858f9SMike Kravetz &node_states[N_MEMORY], 2565f60858f9SMike Kravetz node_alloc_noretry)) 25661da177e4SLinus Torvalds break; 256769ed779aSDavid Rientjes cond_resched(); 25681da177e4SLinus Torvalds } 2569d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2570d715cf80SLiam R. Howlett char buf[32]; 2571d715cf80SLiam R. Howlett 2572c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2573d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2574d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 25758faa8b07SAndi Kleen h->max_huge_pages = i; 2576e5ff2159SAndi Kleen } 2577f60858f9SMike Kravetz 2578f60858f9SMike Kravetz kfree(node_alloc_noretry); 2579d715cf80SLiam R. Howlett } 2580e5ff2159SAndi Kleen 2581e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2582e5ff2159SAndi Kleen { 2583e5ff2159SAndi Kleen struct hstate *h; 2584e5ff2159SAndi Kleen 2585e5ff2159SAndi Kleen for_each_hstate(h) { 2586641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2587641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2588641844f5SNaoya Horiguchi 25898faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2590bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 25918faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2592e5ff2159SAndi Kleen } 2593641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2594e5ff2159SAndi Kleen } 2595e5ff2159SAndi Kleen 2596e5ff2159SAndi Kleen static void __init report_hugepages(void) 2597e5ff2159SAndi Kleen { 2598e5ff2159SAndi Kleen struct hstate *h; 2599e5ff2159SAndi Kleen 2600e5ff2159SAndi Kleen for_each_hstate(h) { 26014abd32dbSAndi Kleen char buf[32]; 2602c6247f72SMatthew Wilcox 2603c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2604ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2605c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2606e5ff2159SAndi Kleen } 2607e5ff2159SAndi Kleen } 2608e5ff2159SAndi Kleen 26091da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 26106ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 26116ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26121da177e4SLinus Torvalds { 26134415cc8dSChristoph Lameter int i; 26144415cc8dSChristoph Lameter 2615bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2616aa888a74SAndi Kleen return; 2617aa888a74SAndi Kleen 26186ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 26191da177e4SLinus Torvalds struct page *page, *next; 2620a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2621a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2622a5516438SAndi Kleen if (count >= h->nr_huge_pages) 26236b0c880dSAdam Litke return; 26241da177e4SLinus Torvalds if (PageHighMem(page)) 26251da177e4SLinus Torvalds continue; 26261da177e4SLinus Torvalds list_del(&page->lru); 2627e5ff2159SAndi Kleen update_and_free_page(h, page); 2628a5516438SAndi Kleen h->free_huge_pages--; 2629a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 26301da177e4SLinus Torvalds } 26311da177e4SLinus Torvalds } 26321da177e4SLinus Torvalds } 26331da177e4SLinus Torvalds #else 26346ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 26356ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26361da177e4SLinus Torvalds { 26371da177e4SLinus Torvalds } 26381da177e4SLinus Torvalds #endif 26391da177e4SLinus Torvalds 264020a0307cSWu Fengguang /* 264120a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 264220a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 264320a0307cSWu Fengguang * Returns 1 if an adjustment was made. 264420a0307cSWu Fengguang */ 26456ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 26466ae11b27SLee Schermerhorn int delta) 264720a0307cSWu Fengguang { 2648b2261026SJoonsoo Kim int nr_nodes, node; 264920a0307cSWu Fengguang 265020a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 265120a0307cSWu Fengguang 2652e8c5c824SLee Schermerhorn if (delta < 0) { 2653b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2654b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2655b2261026SJoonsoo Kim goto found; 2656b2261026SJoonsoo Kim } 2657b2261026SJoonsoo Kim } else { 2658b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2659b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2660b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2661b2261026SJoonsoo Kim goto found; 2662e8c5c824SLee Schermerhorn } 26639a76db09SLee Schermerhorn } 2664b2261026SJoonsoo Kim return 0; 266520a0307cSWu Fengguang 2666b2261026SJoonsoo Kim found: 266720a0307cSWu Fengguang h->surplus_huge_pages += delta; 2668b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2669b2261026SJoonsoo Kim return 1; 267020a0307cSWu Fengguang } 267120a0307cSWu Fengguang 2672a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2673fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 26746ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26751da177e4SLinus Torvalds { 26767893d1d5SAdam Litke unsigned long min_count, ret; 2677f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2678f60858f9SMike Kravetz 2679f60858f9SMike Kravetz /* 2680f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2681f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2682f60858f9SMike Kravetz * the requested huge pages. 2683f60858f9SMike Kravetz */ 2684f60858f9SMike Kravetz if (node_alloc_noretry) 2685f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2686f60858f9SMike Kravetz else 2687f60858f9SMike Kravetz return -ENOMEM; 26881da177e4SLinus Torvalds 26894eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 26904eb0716eSAlexandre Ghiti 26914eb0716eSAlexandre Ghiti /* 2692fd875dcaSMike Kravetz * Check for a node specific request. 2693fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2694fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2695fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2696fd875dcaSMike Kravetz */ 2697fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2698fd875dcaSMike Kravetz unsigned long old_count = count; 2699fd875dcaSMike Kravetz 2700fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2701fd875dcaSMike Kravetz /* 2702fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2703fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2704fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2705fd875dcaSMike Kravetz * largest possible value to align with their intention. 2706fd875dcaSMike Kravetz */ 2707fd875dcaSMike Kravetz if (count < old_count) 2708fd875dcaSMike Kravetz count = ULONG_MAX; 2709fd875dcaSMike Kravetz } 2710fd875dcaSMike Kravetz 2711fd875dcaSMike Kravetz /* 27124eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 27134eb0716eSAlexandre Ghiti * page range allocation. 27144eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 27154eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 27164eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 27174eb0716eSAlexandre Ghiti */ 27184eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 27194eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 27204eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2721f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 27224eb0716eSAlexandre Ghiti return -EINVAL; 27234eb0716eSAlexandre Ghiti } 27244eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 27254eb0716eSAlexandre Ghiti } 2726aa888a74SAndi Kleen 27277893d1d5SAdam Litke /* 27287893d1d5SAdam Litke * Increase the pool size 27297893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 27307893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2731d1c3fb1fSNishanth Aravamudan * 27320c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2733d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2734d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2735d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2736d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 27377893d1d5SAdam Litke */ 2738a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 27396ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 27407893d1d5SAdam Litke break; 27417893d1d5SAdam Litke } 27427893d1d5SAdam Litke 2743a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 27447893d1d5SAdam Litke /* 27457893d1d5SAdam Litke * If this allocation races such that we no longer need the 27467893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 27477893d1d5SAdam Litke * and reducing the surplus. 27487893d1d5SAdam Litke */ 27497893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2750649920c6SJia He 2751649920c6SJia He /* yield cpu to avoid soft lockup */ 2752649920c6SJia He cond_resched(); 2753649920c6SJia He 2754f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2755f60858f9SMike Kravetz node_alloc_noretry); 27567893d1d5SAdam Litke spin_lock(&hugetlb_lock); 27577893d1d5SAdam Litke if (!ret) 27587893d1d5SAdam Litke goto out; 27597893d1d5SAdam Litke 2760536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2761536240f2SMel Gorman if (signal_pending(current)) 2762536240f2SMel Gorman goto out; 27637893d1d5SAdam Litke } 27647893d1d5SAdam Litke 27657893d1d5SAdam Litke /* 27667893d1d5SAdam Litke * Decrease the pool size 27677893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27687893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27697893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27707893d1d5SAdam Litke * to the desired size as pages become free. 2771d1c3fb1fSNishanth Aravamudan * 2772d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2773d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2774d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27750c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2776d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2777d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2778d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27797893d1d5SAdam Litke */ 2780a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27816b0c880dSAdam Litke min_count = max(count, min_count); 27826ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2783a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 27846ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 27851da177e4SLinus Torvalds break; 278655f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 27871da177e4SLinus Torvalds } 2788a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 27896ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 27907893d1d5SAdam Litke break; 27917893d1d5SAdam Litke } 27927893d1d5SAdam Litke out: 27934eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 27941da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 27954eb0716eSAlexandre Ghiti 2796f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2797f60858f9SMike Kravetz 27984eb0716eSAlexandre Ghiti return 0; 27991da177e4SLinus Torvalds } 28001da177e4SLinus Torvalds 2801a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2802a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2803a3437870SNishanth Aravamudan 2804a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2805a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2806a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2807a3437870SNishanth Aravamudan 2808a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2809a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2810a3437870SNishanth Aravamudan 28119a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 28129a305230SLee Schermerhorn 28139a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2814a3437870SNishanth Aravamudan { 2815a3437870SNishanth Aravamudan int i; 28169a305230SLee Schermerhorn 2817a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 28189a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 28199a305230SLee Schermerhorn if (nidp) 28209a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2821a3437870SNishanth Aravamudan return &hstates[i]; 28229a305230SLee Schermerhorn } 28239a305230SLee Schermerhorn 28249a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2825a3437870SNishanth Aravamudan } 2826a3437870SNishanth Aravamudan 282706808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2828a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2829a3437870SNishanth Aravamudan { 28309a305230SLee Schermerhorn struct hstate *h; 28319a305230SLee Schermerhorn unsigned long nr_huge_pages; 28329a305230SLee Schermerhorn int nid; 28339a305230SLee Schermerhorn 28349a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 28359a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 28369a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 28379a305230SLee Schermerhorn else 28389a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 28399a305230SLee Schermerhorn 28409a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2841a3437870SNishanth Aravamudan } 2842adbe8726SEric B Munson 2843238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2844238d3c13SDavid Rientjes struct hstate *h, int nid, 2845238d3c13SDavid Rientjes unsigned long count, size_t len) 2846a3437870SNishanth Aravamudan { 2847a3437870SNishanth Aravamudan int err; 28482d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2849a3437870SNishanth Aravamudan 28502d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 28512d0adf7eSOscar Salvador return -EINVAL; 2852adbe8726SEric B Munson 28539a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 28549a305230SLee Schermerhorn /* 28559a305230SLee Schermerhorn * global hstate attribute 28569a305230SLee Schermerhorn */ 28579a305230SLee Schermerhorn if (!(obey_mempolicy && 28582d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 28592d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 28602d0adf7eSOscar Salvador else 28612d0adf7eSOscar Salvador n_mask = &nodes_allowed; 28622d0adf7eSOscar Salvador } else { 28639a305230SLee Schermerhorn /* 2864fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2865fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28669a305230SLee Schermerhorn */ 28672d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28682d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2869fd875dcaSMike Kravetz } 28709a305230SLee Schermerhorn 28712d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 287206808b08SLee Schermerhorn 28734eb0716eSAlexandre Ghiti return err ? err : len; 287406808b08SLee Schermerhorn } 287506808b08SLee Schermerhorn 2876238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2877238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2878238d3c13SDavid Rientjes size_t len) 2879238d3c13SDavid Rientjes { 2880238d3c13SDavid Rientjes struct hstate *h; 2881238d3c13SDavid Rientjes unsigned long count; 2882238d3c13SDavid Rientjes int nid; 2883238d3c13SDavid Rientjes int err; 2884238d3c13SDavid Rientjes 2885238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2886238d3c13SDavid Rientjes if (err) 2887238d3c13SDavid Rientjes return err; 2888238d3c13SDavid Rientjes 2889238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2890238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2891238d3c13SDavid Rientjes } 2892238d3c13SDavid Rientjes 289306808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 289406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 289506808b08SLee Schermerhorn { 289606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 289706808b08SLee Schermerhorn } 289806808b08SLee Schermerhorn 289906808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 290006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 290106808b08SLee Schermerhorn { 2902238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2903a3437870SNishanth Aravamudan } 2904a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2905a3437870SNishanth Aravamudan 290606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 290706808b08SLee Schermerhorn 290806808b08SLee Schermerhorn /* 290906808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 291006808b08SLee Schermerhorn * huge page alloc/free. 291106808b08SLee Schermerhorn */ 291206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 291306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 291406808b08SLee Schermerhorn { 291506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 291606808b08SLee Schermerhorn } 291706808b08SLee Schermerhorn 291806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 291906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 292006808b08SLee Schermerhorn { 2921238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 292206808b08SLee Schermerhorn } 292306808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 292406808b08SLee Schermerhorn #endif 292506808b08SLee Schermerhorn 292606808b08SLee Schermerhorn 2927a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2928a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2929a3437870SNishanth Aravamudan { 29309a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2931a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2932a3437870SNishanth Aravamudan } 2933adbe8726SEric B Munson 2934a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2935a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2936a3437870SNishanth Aravamudan { 2937a3437870SNishanth Aravamudan int err; 2938a3437870SNishanth Aravamudan unsigned long input; 29399a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2940a3437870SNishanth Aravamudan 2941bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2942adbe8726SEric B Munson return -EINVAL; 2943adbe8726SEric B Munson 29443dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2945a3437870SNishanth Aravamudan if (err) 294673ae31e5SEric B Munson return err; 2947a3437870SNishanth Aravamudan 2948a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2949a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2950a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2951a3437870SNishanth Aravamudan 2952a3437870SNishanth Aravamudan return count; 2953a3437870SNishanth Aravamudan } 2954a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2955a3437870SNishanth Aravamudan 2956a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2957a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2958a3437870SNishanth Aravamudan { 29599a305230SLee Schermerhorn struct hstate *h; 29609a305230SLee Schermerhorn unsigned long free_huge_pages; 29619a305230SLee Schermerhorn int nid; 29629a305230SLee Schermerhorn 29639a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29649a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29659a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29669a305230SLee Schermerhorn else 29679a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29689a305230SLee Schermerhorn 29699a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2970a3437870SNishanth Aravamudan } 2971a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2972a3437870SNishanth Aravamudan 2973a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2974a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2975a3437870SNishanth Aravamudan { 29769a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2977a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2978a3437870SNishanth Aravamudan } 2979a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2980a3437870SNishanth Aravamudan 2981a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2982a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2983a3437870SNishanth Aravamudan { 29849a305230SLee Schermerhorn struct hstate *h; 29859a305230SLee Schermerhorn unsigned long surplus_huge_pages; 29869a305230SLee Schermerhorn int nid; 29879a305230SLee Schermerhorn 29889a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29899a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29909a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 29919a305230SLee Schermerhorn else 29929a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 29939a305230SLee Schermerhorn 29949a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2995a3437870SNishanth Aravamudan } 2996a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2997a3437870SNishanth Aravamudan 2998a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2999a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3000a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3001a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3002a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3003a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 300406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 300506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 300606808b08SLee Schermerhorn #endif 3007a3437870SNishanth Aravamudan NULL, 3008a3437870SNishanth Aravamudan }; 3009a3437870SNishanth Aravamudan 301067e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3011a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3012a3437870SNishanth Aravamudan }; 3013a3437870SNishanth Aravamudan 3014094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 30159a305230SLee Schermerhorn struct kobject **hstate_kobjs, 301667e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3017a3437870SNishanth Aravamudan { 3018a3437870SNishanth Aravamudan int retval; 3019972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3020a3437870SNishanth Aravamudan 30219a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 30229a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3023a3437870SNishanth Aravamudan return -ENOMEM; 3024a3437870SNishanth Aravamudan 30259a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3026a3437870SNishanth Aravamudan if (retval) 30279a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3028a3437870SNishanth Aravamudan 3029a3437870SNishanth Aravamudan return retval; 3030a3437870SNishanth Aravamudan } 3031a3437870SNishanth Aravamudan 3032a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3033a3437870SNishanth Aravamudan { 3034a3437870SNishanth Aravamudan struct hstate *h; 3035a3437870SNishanth Aravamudan int err; 3036a3437870SNishanth Aravamudan 3037a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3038a3437870SNishanth Aravamudan if (!hugepages_kobj) 3039a3437870SNishanth Aravamudan return; 3040a3437870SNishanth Aravamudan 3041a3437870SNishanth Aravamudan for_each_hstate(h) { 30429a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 30439a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3044a3437870SNishanth Aravamudan if (err) 3045282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3046a3437870SNishanth Aravamudan } 3047a3437870SNishanth Aravamudan } 3048a3437870SNishanth Aravamudan 30499a305230SLee Schermerhorn #ifdef CONFIG_NUMA 30509a305230SLee Schermerhorn 30519a305230SLee Schermerhorn /* 30529a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 305310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 305410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 305510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 30569a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 30579a305230SLee Schermerhorn */ 30589a305230SLee Schermerhorn struct node_hstate { 30599a305230SLee Schermerhorn struct kobject *hugepages_kobj; 30609a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30619a305230SLee Schermerhorn }; 3062b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30639a305230SLee Schermerhorn 30649a305230SLee Schermerhorn /* 306510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30669a305230SLee Schermerhorn */ 30679a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30689a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30699a305230SLee Schermerhorn &free_hugepages_attr.attr, 30709a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30719a305230SLee Schermerhorn NULL, 30729a305230SLee Schermerhorn }; 30739a305230SLee Schermerhorn 307467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30759a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30769a305230SLee Schermerhorn }; 30779a305230SLee Schermerhorn 30789a305230SLee Schermerhorn /* 307910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30809a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30819a305230SLee Schermerhorn */ 30829a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 30839a305230SLee Schermerhorn { 30849a305230SLee Schermerhorn int nid; 30859a305230SLee Schermerhorn 30869a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 30879a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 30889a305230SLee Schermerhorn int i; 30899a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 30909a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 30919a305230SLee Schermerhorn if (nidp) 30929a305230SLee Schermerhorn *nidp = nid; 30939a305230SLee Schermerhorn return &hstates[i]; 30949a305230SLee Schermerhorn } 30959a305230SLee Schermerhorn } 30969a305230SLee Schermerhorn 30979a305230SLee Schermerhorn BUG(); 30989a305230SLee Schermerhorn return NULL; 30999a305230SLee Schermerhorn } 31009a305230SLee Schermerhorn 31019a305230SLee Schermerhorn /* 310210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 31039a305230SLee Schermerhorn * No-op if no hstate attributes attached. 31049a305230SLee Schermerhorn */ 31053cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 31069a305230SLee Schermerhorn { 31079a305230SLee Schermerhorn struct hstate *h; 310810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 31099a305230SLee Schermerhorn 31109a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31119b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 31129a305230SLee Schermerhorn 3113972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3114972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3115972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3116972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3117972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3118972dc4deSAneesh Kumar K.V } 31199a305230SLee Schermerhorn } 31209a305230SLee Schermerhorn 31219a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 31229a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 31239a305230SLee Schermerhorn } 31249a305230SLee Schermerhorn 31259a305230SLee Schermerhorn 31269a305230SLee Schermerhorn /* 312710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 31289a305230SLee Schermerhorn * No-op if attributes already registered. 31299a305230SLee Schermerhorn */ 31303cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 31319a305230SLee Schermerhorn { 31329a305230SLee Schermerhorn struct hstate *h; 313310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 31349a305230SLee Schermerhorn int err; 31359a305230SLee Schermerhorn 31369a305230SLee Schermerhorn if (nhs->hugepages_kobj) 31379a305230SLee Schermerhorn return; /* already allocated */ 31389a305230SLee Schermerhorn 31399a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 314010fbcf4cSKay Sievers &node->dev.kobj); 31419a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31429a305230SLee Schermerhorn return; 31439a305230SLee Schermerhorn 31449a305230SLee Schermerhorn for_each_hstate(h) { 31459a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 31469a305230SLee Schermerhorn nhs->hstate_kobjs, 31479a305230SLee Schermerhorn &per_node_hstate_attr_group); 31489a305230SLee Schermerhorn if (err) { 3149282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 315010fbcf4cSKay Sievers h->name, node->dev.id); 31519a305230SLee Schermerhorn hugetlb_unregister_node(node); 31529a305230SLee Schermerhorn break; 31539a305230SLee Schermerhorn } 31549a305230SLee Schermerhorn } 31559a305230SLee Schermerhorn } 31569a305230SLee Schermerhorn 31579a305230SLee Schermerhorn /* 31589b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 315910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 316010fbcf4cSKay Sievers * registered their associated device by this time. 31619a305230SLee Schermerhorn */ 31627d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31639a305230SLee Schermerhorn { 31649a305230SLee Schermerhorn int nid; 31659a305230SLee Schermerhorn 31668cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31678732794bSWen Congyang struct node *node = node_devices[nid]; 316810fbcf4cSKay Sievers if (node->dev.id == nid) 31699a305230SLee Schermerhorn hugetlb_register_node(node); 31709a305230SLee Schermerhorn } 31719a305230SLee Schermerhorn 31729a305230SLee Schermerhorn /* 317310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31749a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31759a305230SLee Schermerhorn */ 31769a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31779a305230SLee Schermerhorn hugetlb_unregister_node); 31789a305230SLee Schermerhorn } 31799a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31809a305230SLee Schermerhorn 31819a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31829a305230SLee Schermerhorn { 31839a305230SLee Schermerhorn BUG(); 31849a305230SLee Schermerhorn if (nidp) 31859a305230SLee Schermerhorn *nidp = -1; 31869a305230SLee Schermerhorn return NULL; 31879a305230SLee Schermerhorn } 31889a305230SLee Schermerhorn 31899a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 31909a305230SLee Schermerhorn 31919a305230SLee Schermerhorn #endif 31929a305230SLee Schermerhorn 3193a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3194a3437870SNishanth Aravamudan { 31958382d914SDavidlohr Bueso int i; 31968382d914SDavidlohr Bueso 3197c2833a5bSMike Kravetz if (!hugepages_supported()) { 3198c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3199c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 32000ef89d25SBenjamin Herrenschmidt return 0; 3201d715cf80SLiam R. Howlett } 3202d715cf80SLiam R. Howlett 3203282f4214SMike Kravetz /* 3204282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3205282f4214SMike Kravetz * architectures depend on setup being done here. 3206282f4214SMike Kravetz */ 3207a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3208282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3209282f4214SMike Kravetz /* 3210282f4214SMike Kravetz * If we did not parse a default huge page size, set 3211282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3212282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3213282f4214SMike Kravetz * specified, set that here as well. 3214282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3215282f4214SMike Kravetz * setting. A warning will be printed in this case. 3216282f4214SMike Kravetz */ 3217282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3218f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3219282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3220282f4214SMike Kravetz char buf[32]; 3221282f4214SMike Kravetz 3222282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3223282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3224282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3225282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3226282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3227282f4214SMike Kravetz default_hstate_max_huge_pages); 3228282f4214SMike Kravetz } 3229282f4214SMike Kravetz default_hstate.max_huge_pages = 3230282f4214SMike Kravetz default_hstate_max_huge_pages; 3231282f4214SMike Kravetz } 3232f8b74815SVaishali Thakkar } 3233a3437870SNishanth Aravamudan 3234cf11e85fSRoman Gushchin hugetlb_cma_check(); 3235a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3236aa888a74SAndi Kleen gather_bootmem_prealloc(); 3237a3437870SNishanth Aravamudan report_hugepages(); 3238a3437870SNishanth Aravamudan 3239a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 32409a305230SLee Schermerhorn hugetlb_register_all_nodes(); 32417179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 32429a305230SLee Schermerhorn 32438382d914SDavidlohr Bueso #ifdef CONFIG_SMP 32448382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 32458382d914SDavidlohr Bueso #else 32468382d914SDavidlohr Bueso num_fault_mutexes = 1; 32478382d914SDavidlohr Bueso #endif 3248c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 32496da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 32506da2ec56SKees Cook GFP_KERNEL); 3251c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 32528382d914SDavidlohr Bueso 32538382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3254c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3255a3437870SNishanth Aravamudan return 0; 3256a3437870SNishanth Aravamudan } 32573e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3258a3437870SNishanth Aravamudan 3259ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3260ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32619fee021dSVaishali Thakkar { 3262ae94da89SMike Kravetz return size == HPAGE_SIZE; 32639fee021dSVaishali Thakkar } 32649fee021dSVaishali Thakkar 3265d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3266a3437870SNishanth Aravamudan { 3267a3437870SNishanth Aravamudan struct hstate *h; 32688faa8b07SAndi Kleen unsigned long i; 32698faa8b07SAndi Kleen 3270a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3271a3437870SNishanth Aravamudan return; 3272a3437870SNishanth Aravamudan } 327347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3274a3437870SNishanth Aravamudan BUG_ON(order == 0); 327547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3276a3437870SNishanth Aravamudan h->order = order; 3277a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 32788faa8b07SAndi Kleen h->nr_huge_pages = 0; 32798faa8b07SAndi Kleen h->free_huge_pages = 0; 32808faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32818faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 32820edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 328354f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 328454f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3285a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3286a3437870SNishanth Aravamudan huge_page_size(h)/1024); 32878faa8b07SAndi Kleen 3288a3437870SNishanth Aravamudan parsed_hstate = h; 3289a3437870SNishanth Aravamudan } 3290a3437870SNishanth Aravamudan 3291282f4214SMike Kravetz /* 3292282f4214SMike Kravetz * hugepages command line processing 3293282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3294282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3295282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3296282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3297282f4214SMike Kravetz */ 3298282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3299a3437870SNishanth Aravamudan { 3300a3437870SNishanth Aravamudan unsigned long *mhp; 33018faa8b07SAndi Kleen static unsigned long *last_mhp; 3302a3437870SNishanth Aravamudan 33039fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3304282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 33059fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3306282f4214SMike Kravetz return 0; 33079fee021dSVaishali Thakkar } 3308282f4214SMike Kravetz 3309a3437870SNishanth Aravamudan /* 3310282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3311282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3312282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3313282f4214SMike Kravetz * default_hugepagesz. 3314a3437870SNishanth Aravamudan */ 33159fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3316a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3317a3437870SNishanth Aravamudan else 3318a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3319a3437870SNishanth Aravamudan 33208faa8b07SAndi Kleen if (mhp == last_mhp) { 3321282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3322282f4214SMike Kravetz return 0; 33238faa8b07SAndi Kleen } 33248faa8b07SAndi Kleen 3325a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3326a3437870SNishanth Aravamudan *mhp = 0; 3327a3437870SNishanth Aravamudan 33288faa8b07SAndi Kleen /* 33298faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 33308faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 33318faa8b07SAndi Kleen * use the bootmem allocator. 33328faa8b07SAndi Kleen */ 333347d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 33348faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 33358faa8b07SAndi Kleen 33368faa8b07SAndi Kleen last_mhp = mhp; 33378faa8b07SAndi Kleen 3338a3437870SNishanth Aravamudan return 1; 3339a3437870SNishanth Aravamudan } 3340282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3341e11bfbfcSNick Piggin 3342282f4214SMike Kravetz /* 3343282f4214SMike Kravetz * hugepagesz command line processing 3344282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3345282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3346282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3347282f4214SMike Kravetz * hugepagesz argument was valid. 3348282f4214SMike Kravetz */ 3349359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3350e11bfbfcSNick Piggin { 3351359f2544SMike Kravetz unsigned long size; 3352282f4214SMike Kravetz struct hstate *h; 3353282f4214SMike Kravetz 3354282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3355359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3356359f2544SMike Kravetz 3357359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3358282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3359359f2544SMike Kravetz return 0; 3360359f2544SMike Kravetz } 3361359f2544SMike Kravetz 3362282f4214SMike Kravetz h = size_to_hstate(size); 3363282f4214SMike Kravetz if (h) { 3364282f4214SMike Kravetz /* 3365282f4214SMike Kravetz * hstate for this size already exists. This is normally 3366282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3367282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3368282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3369282f4214SMike Kravetz * previously specified. 3370282f4214SMike Kravetz */ 3371282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3372282f4214SMike Kravetz default_hstate.max_huge_pages) { 3373282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 337438237830SMike Kravetz return 0; 337538237830SMike Kravetz } 337638237830SMike Kravetz 3377282f4214SMike Kravetz /* 3378282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3379282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3380282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3381282f4214SMike Kravetz */ 3382282f4214SMike Kravetz parsed_hstate = h; 3383282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3384e11bfbfcSNick Piggin return 1; 3385e11bfbfcSNick Piggin } 3386282f4214SMike Kravetz 3387359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3388282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3389359f2544SMike Kravetz return 1; 3390359f2544SMike Kravetz } 3391359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3392359f2544SMike Kravetz 3393282f4214SMike Kravetz /* 3394282f4214SMike Kravetz * default_hugepagesz command line input 3395282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3396282f4214SMike Kravetz */ 3397ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3398e11bfbfcSNick Piggin { 3399ae94da89SMike Kravetz unsigned long size; 3400ae94da89SMike Kravetz 3401282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3402282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3403282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3404282f4214SMike Kravetz return 0; 3405282f4214SMike Kravetz } 3406282f4214SMike Kravetz 3407282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3408282f4214SMike Kravetz 3409282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3410282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3411282f4214SMike Kravetz return 0; 3412282f4214SMike Kravetz } 3413282f4214SMike Kravetz 3414282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3415282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3416282f4214SMike Kravetz parsed_default_hugepagesz = true; 3417282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3418282f4214SMike Kravetz 3419282f4214SMike Kravetz /* 3420282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3421282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3422282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3423282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3424282f4214SMike Kravetz * allocated here from bootmem allocator. 3425282f4214SMike Kravetz */ 3426282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3427282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3428282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3429282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3430282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3431282f4214SMike Kravetz } 3432282f4214SMike Kravetz 3433e11bfbfcSNick Piggin return 1; 3434e11bfbfcSNick Piggin } 3435ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3436a3437870SNishanth Aravamudan 34378ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 34388a213460SNishanth Aravamudan { 34398a213460SNishanth Aravamudan int node; 34408a213460SNishanth Aravamudan unsigned int nr = 0; 34418ca39e68SMuchun Song nodemask_t *mpol_allowed; 34428ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 34438ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 34448a213460SNishanth Aravamudan 34458ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 34468ca39e68SMuchun Song 34478ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 34488ca39e68SMuchun Song if (!mpol_allowed || 34498ca39e68SMuchun Song (mpol_allowed && node_isset(node, *mpol_allowed))) 34508a213460SNishanth Aravamudan nr += array[node]; 34518ca39e68SMuchun Song } 34528a213460SNishanth Aravamudan 34538a213460SNishanth Aravamudan return nr; 34548a213460SNishanth Aravamudan } 34558a213460SNishanth Aravamudan 34568a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 345706808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 345806808b08SLee Schermerhorn struct ctl_table *table, int write, 345932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34601da177e4SLinus Torvalds { 3461e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3462238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 346308d4a246SMichal Hocko int ret; 3464e5ff2159SAndi Kleen 3465457c1b27SNishanth Aravamudan if (!hugepages_supported()) 346686613628SJan Stancek return -EOPNOTSUPP; 3467457c1b27SNishanth Aravamudan 3468e5ff2159SAndi Kleen table->data = &tmp; 3469e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 347008d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 347108d4a246SMichal Hocko if (ret) 347208d4a246SMichal Hocko goto out; 3473e5ff2159SAndi Kleen 3474238d3c13SDavid Rientjes if (write) 3475238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3476238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 347708d4a246SMichal Hocko out: 347808d4a246SMichal Hocko return ret; 34791da177e4SLinus Torvalds } 3480396faf03SMel Gorman 348106808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 348232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 348306808b08SLee Schermerhorn { 348406808b08SLee Schermerhorn 348506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 348606808b08SLee Schermerhorn buffer, length, ppos); 348706808b08SLee Schermerhorn } 348806808b08SLee Schermerhorn 348906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 349006808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 349132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 349206808b08SLee Schermerhorn { 349306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 349406808b08SLee Schermerhorn buffer, length, ppos); 349506808b08SLee Schermerhorn } 349606808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 349706808b08SLee Schermerhorn 3498a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 349932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3500a3d0c6aaSNishanth Aravamudan { 3501a5516438SAndi Kleen struct hstate *h = &default_hstate; 3502e5ff2159SAndi Kleen unsigned long tmp; 350308d4a246SMichal Hocko int ret; 3504e5ff2159SAndi Kleen 3505457c1b27SNishanth Aravamudan if (!hugepages_supported()) 350686613628SJan Stancek return -EOPNOTSUPP; 3507457c1b27SNishanth Aravamudan 3508e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3509e5ff2159SAndi Kleen 3510bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3511adbe8726SEric B Munson return -EINVAL; 3512adbe8726SEric B Munson 3513e5ff2159SAndi Kleen table->data = &tmp; 3514e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 351508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 351608d4a246SMichal Hocko if (ret) 351708d4a246SMichal Hocko goto out; 3518e5ff2159SAndi Kleen 3519e5ff2159SAndi Kleen if (write) { 3520064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3521e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3522a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3523e5ff2159SAndi Kleen } 352408d4a246SMichal Hocko out: 352508d4a246SMichal Hocko return ret; 3526a3d0c6aaSNishanth Aravamudan } 3527a3d0c6aaSNishanth Aravamudan 35281da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 35291da177e4SLinus Torvalds 3530e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 35311da177e4SLinus Torvalds { 3532fcb2b0c5SRoman Gushchin struct hstate *h; 3533fcb2b0c5SRoman Gushchin unsigned long total = 0; 3534fcb2b0c5SRoman Gushchin 3535457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3536457c1b27SNishanth Aravamudan return; 3537fcb2b0c5SRoman Gushchin 3538fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3539fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3540fcb2b0c5SRoman Gushchin 3541fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3542fcb2b0c5SRoman Gushchin 3543fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3544e1759c21SAlexey Dobriyan seq_printf(m, 35451da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 35461da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3547b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35487893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35494f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3550fcb2b0c5SRoman Gushchin count, 3551a5516438SAndi Kleen h->free_huge_pages, 3552a5516438SAndi Kleen h->resv_huge_pages, 3553a5516438SAndi Kleen h->surplus_huge_pages, 3554fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3555fcb2b0c5SRoman Gushchin } 3556fcb2b0c5SRoman Gushchin 3557fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 35581da177e4SLinus Torvalds } 35591da177e4SLinus Torvalds 35601da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 35611da177e4SLinus Torvalds { 3562a5516438SAndi Kleen struct hstate *h = &default_hstate; 3563457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3564457c1b27SNishanth Aravamudan return 0; 35651da177e4SLinus Torvalds return sprintf(buf, 35661da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3567a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3568a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3569a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3570a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3571a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35721da177e4SLinus Torvalds } 35731da177e4SLinus Torvalds 3574949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3575949f7ec5SDavid Rientjes { 3576949f7ec5SDavid Rientjes struct hstate *h; 3577949f7ec5SDavid Rientjes int nid; 3578949f7ec5SDavid Rientjes 3579457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3580457c1b27SNishanth Aravamudan return; 3581457c1b27SNishanth Aravamudan 3582949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3583949f7ec5SDavid Rientjes for_each_hstate(h) 3584949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3585949f7ec5SDavid Rientjes nid, 3586949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3587949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3588949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3589949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3590949f7ec5SDavid Rientjes } 3591949f7ec5SDavid Rientjes 35925d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 35935d317b2bSNaoya Horiguchi { 35945d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 35955d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 35965d317b2bSNaoya Horiguchi } 35975d317b2bSNaoya Horiguchi 35981da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 35991da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 36001da177e4SLinus Torvalds { 3601d0028588SWanpeng Li struct hstate *h; 3602d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3603d0028588SWanpeng Li 3604d0028588SWanpeng Li for_each_hstate(h) 3605d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3606d0028588SWanpeng Li return nr_total_pages; 36071da177e4SLinus Torvalds } 36081da177e4SLinus Torvalds 3609a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3610fc1b8a73SMel Gorman { 3611fc1b8a73SMel Gorman int ret = -ENOMEM; 3612fc1b8a73SMel Gorman 3613fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3614fc1b8a73SMel Gorman /* 3615fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3616fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3617fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3618fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3619fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3620fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3621fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3622fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3623fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3624fc1b8a73SMel Gorman * 3625fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3626fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3627fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3628fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3629fc1b8a73SMel Gorman * semantics that cpuset has. 36308ca39e68SMuchun Song * 36318ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 36328ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 36338ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 36348ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 36358ca39e68SMuchun Song * above. 3636fc1b8a73SMel Gorman */ 3637fc1b8a73SMel Gorman if (delta > 0) { 3638a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3639fc1b8a73SMel Gorman goto out; 3640fc1b8a73SMel Gorman 36418ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3642a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3643fc1b8a73SMel Gorman goto out; 3644fc1b8a73SMel Gorman } 3645fc1b8a73SMel Gorman } 3646fc1b8a73SMel Gorman 3647fc1b8a73SMel Gorman ret = 0; 3648fc1b8a73SMel Gorman if (delta < 0) 3649a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3650fc1b8a73SMel Gorman 3651fc1b8a73SMel Gorman out: 3652fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3653fc1b8a73SMel Gorman return ret; 3654fc1b8a73SMel Gorman } 3655fc1b8a73SMel Gorman 365684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 365784afd99bSAndy Whitcroft { 3658f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 365984afd99bSAndy Whitcroft 366084afd99bSAndy Whitcroft /* 366184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 366284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 366384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 366425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 366584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 366684afd99bSAndy Whitcroft * new reference here without additional locking. 366784afd99bSAndy Whitcroft */ 36684e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3669f522c3acSJoonsoo Kim kref_get(&resv->refs); 367084afd99bSAndy Whitcroft } 367184afd99bSAndy Whitcroft 3672a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3673a1e78772SMel Gorman { 3674a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3675f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 367690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36774e35f483SJoonsoo Kim unsigned long reserve, start, end; 36781c5ecae3SMike Kravetz long gbl_reserve; 367984afd99bSAndy Whitcroft 36804e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36814e35f483SJoonsoo Kim return; 36824e35f483SJoonsoo Kim 3683a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3684a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 368584afd99bSAndy Whitcroft 36864e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3687e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36887251ff78SAdam Litke if (reserve) { 36891c5ecae3SMike Kravetz /* 36901c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36911c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36921c5ecae3SMike Kravetz */ 36931c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36941c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36957251ff78SAdam Litke } 3696e9fe92aeSMina Almasry 3697e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3698a1e78772SMel Gorman } 3699a1e78772SMel Gorman 370031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 370131383c68SDan Williams { 370231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 370331383c68SDan Williams return -EINVAL; 370431383c68SDan Williams return 0; 370531383c68SDan Williams } 370631383c68SDan Williams 370705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 370805ea8860SDan Williams { 370905ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 371005ea8860SDan Williams 371105ea8860SDan Williams return 1UL << huge_page_shift(hstate); 371205ea8860SDan Williams } 371305ea8860SDan Williams 37141da177e4SLinus Torvalds /* 37151da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 37161da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 37171da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 37181da177e4SLinus Torvalds * this far. 37191da177e4SLinus Torvalds */ 3720b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 37211da177e4SLinus Torvalds { 37221da177e4SLinus Torvalds BUG(); 3723d0217ac0SNick Piggin return 0; 37241da177e4SLinus Torvalds } 37251da177e4SLinus Torvalds 3726eec3636aSJane Chu /* 3727eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3728eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3729eec3636aSJane Chu * This is because under System V memory model, mappings created via 3730eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3731eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3732eec3636aSJane Chu */ 3733f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3734d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 373584afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3736a1e78772SMel Gorman .close = hugetlb_vm_op_close, 373731383c68SDan Williams .split = hugetlb_vm_op_split, 373805ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 37391da177e4SLinus Torvalds }; 37401da177e4SLinus Torvalds 37411e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 37421e8f889bSDavid Gibson int writable) 374363551ae0SDavid Gibson { 374463551ae0SDavid Gibson pte_t entry; 374563551ae0SDavid Gibson 37461e8f889bSDavid Gibson if (writable) { 3747106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3748106c992aSGerald Schaefer vma->vm_page_prot))); 374963551ae0SDavid Gibson } else { 3750106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3751106c992aSGerald Schaefer vma->vm_page_prot)); 375263551ae0SDavid Gibson } 375363551ae0SDavid Gibson entry = pte_mkyoung(entry); 375463551ae0SDavid Gibson entry = pte_mkhuge(entry); 3755d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 375663551ae0SDavid Gibson 375763551ae0SDavid Gibson return entry; 375863551ae0SDavid Gibson } 375963551ae0SDavid Gibson 37601e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37611e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37621e8f889bSDavid Gibson { 37631e8f889bSDavid Gibson pte_t entry; 37641e8f889bSDavid Gibson 3765106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 376632f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37674b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37681e8f889bSDavid Gibson } 37691e8f889bSDavid Gibson 3770d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37714a705fefSNaoya Horiguchi { 37724a705fefSNaoya Horiguchi swp_entry_t swp; 37734a705fefSNaoya Horiguchi 37744a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3775d5ed7444SAneesh Kumar K.V return false; 37764a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 37774a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3778d5ed7444SAneesh Kumar K.V return true; 37794a705fefSNaoya Horiguchi else 3780d5ed7444SAneesh Kumar K.V return false; 37814a705fefSNaoya Horiguchi } 37824a705fefSNaoya Horiguchi 37834a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 37844a705fefSNaoya Horiguchi { 37854a705fefSNaoya Horiguchi swp_entry_t swp; 37864a705fefSNaoya Horiguchi 37874a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37884a705fefSNaoya Horiguchi return 0; 37894a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 37904a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 37914a705fefSNaoya Horiguchi return 1; 37924a705fefSNaoya Horiguchi else 37934a705fefSNaoya Horiguchi return 0; 37944a705fefSNaoya Horiguchi } 37951e8f889bSDavid Gibson 379663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 379763551ae0SDavid Gibson struct vm_area_struct *vma) 379863551ae0SDavid Gibson { 37995e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 380063551ae0SDavid Gibson struct page *ptepage; 38011c59827dSHugh Dickins unsigned long addr; 38021e8f889bSDavid Gibson int cow; 3803a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3804a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3805c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3806ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3807e8569dd2SAndreas Sandberg int ret = 0; 38081e8f889bSDavid Gibson 38091e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 381063551ae0SDavid Gibson 3811ac46d4f3SJérôme Glisse if (cow) { 38127269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 38136f4f13e8SJérôme Glisse vma->vm_start, 3814ac46d4f3SJérôme Glisse vma->vm_end); 3815ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3816c0d0381aSMike Kravetz } else { 3817c0d0381aSMike Kravetz /* 3818c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3819c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3820c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3821c0d0381aSMike Kravetz * huge_pmd_unshare. 3822c0d0381aSMike Kravetz */ 3823c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3824ac46d4f3SJérôme Glisse } 3825e8569dd2SAndreas Sandberg 3826a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3827cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 38287868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3829c74df32cSHugh Dickins if (!src_pte) 3830c74df32cSHugh Dickins continue; 3831a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3832e8569dd2SAndreas Sandberg if (!dst_pte) { 3833e8569dd2SAndreas Sandberg ret = -ENOMEM; 3834e8569dd2SAndreas Sandberg break; 3835e8569dd2SAndreas Sandberg } 3836c5c99429SLarry Woodman 38375e41540cSMike Kravetz /* 38385e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38395e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38405e41540cSMike Kravetz * 38415e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38425e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38435e41540cSMike Kravetz * Check here before taking page table lock, and once again 38445e41540cSMike Kravetz * after taking the lock below. 38455e41540cSMike Kravetz */ 38465e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38475e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3848c5c99429SLarry Woodman continue; 3849c5c99429SLarry Woodman 3850cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3851cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3852cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38534a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38545e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38555e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38565e41540cSMike Kravetz /* 38575e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38585e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38595e41540cSMike Kravetz * sharing with another vma. 38605e41540cSMike Kravetz */ 38614a705fefSNaoya Horiguchi ; 38624a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38634a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38644a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38654a705fefSNaoya Horiguchi 38664a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38674a705fefSNaoya Horiguchi /* 38684a705fefSNaoya Horiguchi * COW mappings require pages in both 38694a705fefSNaoya Horiguchi * parent and child to be set to read. 38704a705fefSNaoya Horiguchi */ 38714a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38724a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3873e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3874e5251fd4SPunit Agrawal entry, sz); 38754a705fefSNaoya Horiguchi } 3876e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38774a705fefSNaoya Horiguchi } else { 387834ee645eSJoerg Roedel if (cow) { 38790f10851eSJérôme Glisse /* 38800f10851eSJérôme Glisse * No need to notify as we are downgrading page 38810f10851eSJérôme Glisse * table protection not changing it to point 38820f10851eSJérôme Glisse * to a new page. 38830f10851eSJérôme Glisse * 3884ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 38850f10851eSJérôme Glisse */ 38867f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 388734ee645eSJoerg Roedel } 38880253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 388963551ae0SDavid Gibson ptepage = pte_page(entry); 389063551ae0SDavid Gibson get_page(ptepage); 389153f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 389263551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 38935d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 38941c59827dSHugh Dickins } 3895cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3896cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 389763551ae0SDavid Gibson } 389863551ae0SDavid Gibson 3899e8569dd2SAndreas Sandberg if (cow) 3900ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3901c0d0381aSMike Kravetz else 3902c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3903e8569dd2SAndreas Sandberg 3904e8569dd2SAndreas Sandberg return ret; 390563551ae0SDavid Gibson } 390663551ae0SDavid Gibson 390724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 390824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 390924669e58SAneesh Kumar K.V struct page *ref_page) 391063551ae0SDavid Gibson { 391163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 391263551ae0SDavid Gibson unsigned long address; 3913c7546f8fSDavid Gibson pte_t *ptep; 391463551ae0SDavid Gibson pte_t pte; 3915cb900f41SKirill A. Shutemov spinlock_t *ptl; 391663551ae0SDavid Gibson struct page *page; 3917a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3918a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3919ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3920a5516438SAndi Kleen 392163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3922a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3923a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 392463551ae0SDavid Gibson 392507e32661SAneesh Kumar K.V /* 392607e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 392707e32661SAneesh Kumar K.V * to huge page. 392807e32661SAneesh Kumar K.V */ 3929ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 393024669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3931dff11abeSMike Kravetz 3932dff11abeSMike Kravetz /* 3933dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3934dff11abeSMike Kravetz */ 39356f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39366f4f13e8SJérôme Glisse end); 3937ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3938ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3939569f48b8SHillf Danton address = start; 3940569f48b8SHillf Danton for (; address < end; address += sz) { 39417868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3942c7546f8fSDavid Gibson if (!ptep) 3943c7546f8fSDavid Gibson continue; 3944c7546f8fSDavid Gibson 3945cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 394634ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 394731d49da5SAneesh Kumar K.V spin_unlock(ptl); 3948dff11abeSMike Kravetz /* 3949dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3950dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3951dff11abeSMike Kravetz */ 395231d49da5SAneesh Kumar K.V continue; 395331d49da5SAneesh Kumar K.V } 395439dde65cSChen, Kenneth W 39556629326bSHillf Danton pte = huge_ptep_get(ptep); 395631d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 395731d49da5SAneesh Kumar K.V spin_unlock(ptl); 395831d49da5SAneesh Kumar K.V continue; 395931d49da5SAneesh Kumar K.V } 39606629326bSHillf Danton 39616629326bSHillf Danton /* 39629fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39639fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39646629326bSHillf Danton */ 39659fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39669386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 396731d49da5SAneesh Kumar K.V spin_unlock(ptl); 396831d49da5SAneesh Kumar K.V continue; 39698c4894c6SNaoya Horiguchi } 39706629326bSHillf Danton 39716629326bSHillf Danton page = pte_page(pte); 397204f2cbe3SMel Gorman /* 397304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 397404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 397504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 397604f2cbe3SMel Gorman */ 397704f2cbe3SMel Gorman if (ref_page) { 397831d49da5SAneesh Kumar K.V if (page != ref_page) { 397931d49da5SAneesh Kumar K.V spin_unlock(ptl); 398031d49da5SAneesh Kumar K.V continue; 398131d49da5SAneesh Kumar K.V } 398204f2cbe3SMel Gorman /* 398304f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 398404f2cbe3SMel Gorman * future faults in this VMA will fail rather than 398504f2cbe3SMel Gorman * looking like data was lost 398604f2cbe3SMel Gorman */ 398704f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 398804f2cbe3SMel Gorman } 398904f2cbe3SMel Gorman 3990c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3991b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3992106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 39936649a386SKen Chen set_page_dirty(page); 39949e81130bSHillf Danton 39955d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3996d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 399731d49da5SAneesh Kumar K.V 3998cb900f41SKirill A. Shutemov spin_unlock(ptl); 3999e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 400024669e58SAneesh Kumar K.V /* 400131d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 400224669e58SAneesh Kumar K.V */ 400331d49da5SAneesh Kumar K.V if (ref_page) 400431d49da5SAneesh Kumar K.V break; 4005fe1668aeSChen, Kenneth W } 4006ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 400724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 40081da177e4SLinus Torvalds } 400963551ae0SDavid Gibson 4010d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4011d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4012d833352aSMel Gorman unsigned long end, struct page *ref_page) 4013d833352aSMel Gorman { 4014d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4015d833352aSMel Gorman 4016d833352aSMel Gorman /* 4017d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4018d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4019d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4020d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4021d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4022c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4023d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4024c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4025d833352aSMel Gorman */ 4026d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4027d833352aSMel Gorman } 4028d833352aSMel Gorman 4029502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 403004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4031502717f4SChen, Kenneth W { 403224669e58SAneesh Kumar K.V struct mm_struct *mm; 403324669e58SAneesh Kumar K.V struct mmu_gather tlb; 4034dff11abeSMike Kravetz unsigned long tlb_start = start; 4035dff11abeSMike Kravetz unsigned long tlb_end = end; 4036dff11abeSMike Kravetz 4037dff11abeSMike Kravetz /* 4038dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 4039dff11abeSMike Kravetz * start/end for worst case tlb flushing. 4040dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 4041dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 4042dff11abeSMike Kravetz * the largest possible range. 4043dff11abeSMike Kravetz */ 4044dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 404524669e58SAneesh Kumar K.V 404624669e58SAneesh Kumar K.V mm = vma->vm_mm; 404724669e58SAneesh Kumar K.V 4048dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 404924669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4050dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 4051502717f4SChen, Kenneth W } 4052502717f4SChen, Kenneth W 405304f2cbe3SMel Gorman /* 405404f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 405504f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 405604f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 405704f2cbe3SMel Gorman * same region. 405804f2cbe3SMel Gorman */ 40592f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40602a4b3dedSHarvey Harrison struct page *page, unsigned long address) 406104f2cbe3SMel Gorman { 40627526674dSAdam Litke struct hstate *h = hstate_vma(vma); 406304f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 406404f2cbe3SMel Gorman struct address_space *mapping; 406504f2cbe3SMel Gorman pgoff_t pgoff; 406604f2cbe3SMel Gorman 406704f2cbe3SMel Gorman /* 406804f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 406904f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 407004f2cbe3SMel Gorman */ 40717526674dSAdam Litke address = address & huge_page_mask(h); 407236e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 407336e4f20aSMichal Hocko vma->vm_pgoff; 407493c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 407504f2cbe3SMel Gorman 40764eb2b1dcSMel Gorman /* 40774eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40784eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40794eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40804eb2b1dcSMel Gorman */ 408183cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40826b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 408304f2cbe3SMel Gorman /* Do not unmap the current VMA */ 408404f2cbe3SMel Gorman if (iter_vma == vma) 408504f2cbe3SMel Gorman continue; 408604f2cbe3SMel Gorman 408704f2cbe3SMel Gorman /* 40882f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40892f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 40902f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 40912f84a899SMel Gorman */ 40922f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 40932f84a899SMel Gorman continue; 40942f84a899SMel Gorman 40952f84a899SMel Gorman /* 409604f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 409704f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 409804f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 409904f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 410004f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 410104f2cbe3SMel Gorman */ 410204f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 410324669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 410424669e58SAneesh Kumar K.V address + huge_page_size(h), page); 410504f2cbe3SMel Gorman } 410683cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 410704f2cbe3SMel Gorman } 410804f2cbe3SMel Gorman 41090fe6e20bSNaoya Horiguchi /* 41100fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4111ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4112ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4113ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 41140fe6e20bSNaoya Horiguchi */ 41152b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4116974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4117cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 41181e8f889bSDavid Gibson { 41193999f52eSAneesh Kumar K.V pte_t pte; 4120a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41211e8f889bSDavid Gibson struct page *old_page, *new_page; 41222b740303SSouptick Joarder int outside_reserve = 0; 41232b740303SSouptick Joarder vm_fault_t ret = 0; 4124974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4125ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 41261e8f889bSDavid Gibson 41273999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 41281e8f889bSDavid Gibson old_page = pte_page(pte); 41291e8f889bSDavid Gibson 413004f2cbe3SMel Gorman retry_avoidcopy: 41311e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 41321e8f889bSDavid Gibson * and just make the page writable */ 413337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 41345a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 41355b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 413683c54070SNick Piggin return 0; 41371e8f889bSDavid Gibson } 41381e8f889bSDavid Gibson 413904f2cbe3SMel Gorman /* 414004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 414104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 414204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 414304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 414404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 414504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 414604f2cbe3SMel Gorman * of the full address range. 414704f2cbe3SMel Gorman */ 41485944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 414904f2cbe3SMel Gorman old_page != pagecache_page) 415004f2cbe3SMel Gorman outside_reserve = 1; 415104f2cbe3SMel Gorman 415209cbfeafSKirill A. Shutemov get_page(old_page); 4153b76c8cfbSLarry Woodman 4154ad4404a2SDavidlohr Bueso /* 4155ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4156ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4157ad4404a2SDavidlohr Bueso */ 4158cb900f41SKirill A. Shutemov spin_unlock(ptl); 41595b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41601e8f889bSDavid Gibson 41612fc39cecSAdam Litke if (IS_ERR(new_page)) { 416204f2cbe3SMel Gorman /* 416304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 416404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 416504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 416604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 416704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 416804f2cbe3SMel Gorman */ 416904f2cbe3SMel Gorman if (outside_reserve) { 417009cbfeafSKirill A. Shutemov put_page(old_page); 417104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 41725b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 417304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4174cb900f41SKirill A. Shutemov spin_lock(ptl); 41755b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4176a9af0c5dSNaoya Horiguchi if (likely(ptep && 4177a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 417804f2cbe3SMel Gorman goto retry_avoidcopy; 4179a734bcc8SHillf Danton /* 4180cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4181cb900f41SKirill A. Shutemov * lock, and our job is done. 4182a734bcc8SHillf Danton */ 4183a734bcc8SHillf Danton return 0; 418404f2cbe3SMel Gorman } 418504f2cbe3SMel Gorman 41862b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4187ad4404a2SDavidlohr Bueso goto out_release_old; 41881e8f889bSDavid Gibson } 41891e8f889bSDavid Gibson 41900fe6e20bSNaoya Horiguchi /* 41910fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 41920fe6e20bSNaoya Horiguchi * anon_vma prepared. 41930fe6e20bSNaoya Horiguchi */ 419444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4195ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4196ad4404a2SDavidlohr Bueso goto out_release_all; 419744e2aa93SDean Nelson } 41980fe6e20bSNaoya Horiguchi 4199974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 420047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 42010ed361deSNick Piggin __SetPageUptodate(new_page); 42021e8f889bSDavid Gibson 42037269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 42046f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4205ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4206ad4404a2SDavidlohr Bueso 4207b76c8cfbSLarry Woodman /* 4208cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4209b76c8cfbSLarry Woodman * before the page tables are altered 4210b76c8cfbSLarry Woodman */ 4211cb900f41SKirill A. Shutemov spin_lock(ptl); 42125b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4213a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 421407443a85SJoonsoo Kim ClearPagePrivate(new_page); 421507443a85SJoonsoo Kim 42161e8f889bSDavid Gibson /* Break COW */ 42175b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4218ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 42195b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 42201e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4221d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 42225b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 4223cb6acd01SMike Kravetz set_page_huge_active(new_page); 42241e8f889bSDavid Gibson /* Make the old page be freed below */ 42251e8f889bSDavid Gibson new_page = old_page; 42261e8f889bSDavid Gibson } 4227cb900f41SKirill A. Shutemov spin_unlock(ptl); 4228ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4229ad4404a2SDavidlohr Bueso out_release_all: 42305b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 423109cbfeafSKirill A. Shutemov put_page(new_page); 4232ad4404a2SDavidlohr Bueso out_release_old: 423309cbfeafSKirill A. Shutemov put_page(old_page); 42348312034fSJoonsoo Kim 4235ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4236ad4404a2SDavidlohr Bueso return ret; 42371e8f889bSDavid Gibson } 42381e8f889bSDavid Gibson 423904f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4240a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4241a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 424204f2cbe3SMel Gorman { 424304f2cbe3SMel Gorman struct address_space *mapping; 4244e7c4b0bfSAndy Whitcroft pgoff_t idx; 424504f2cbe3SMel Gorman 424604f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4247a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 424804f2cbe3SMel Gorman 424904f2cbe3SMel Gorman return find_lock_page(mapping, idx); 425004f2cbe3SMel Gorman } 425104f2cbe3SMel Gorman 42523ae77f43SHugh Dickins /* 42533ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42543ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42553ae77f43SHugh Dickins */ 42563ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42572a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42582a15efc9SHugh Dickins { 42592a15efc9SHugh Dickins struct address_space *mapping; 42602a15efc9SHugh Dickins pgoff_t idx; 42612a15efc9SHugh Dickins struct page *page; 42622a15efc9SHugh Dickins 42632a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42642a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42652a15efc9SHugh Dickins 42662a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42672a15efc9SHugh Dickins if (page) 42682a15efc9SHugh Dickins put_page(page); 42692a15efc9SHugh Dickins return page != NULL; 42702a15efc9SHugh Dickins } 42712a15efc9SHugh Dickins 4272ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4273ab76ad54SMike Kravetz pgoff_t idx) 4274ab76ad54SMike Kravetz { 4275ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4276ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4277ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4278ab76ad54SMike Kravetz 4279ab76ad54SMike Kravetz if (err) 4280ab76ad54SMike Kravetz return err; 4281ab76ad54SMike Kravetz ClearPagePrivate(page); 4282ab76ad54SMike Kravetz 428322146c3cSMike Kravetz /* 428422146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 428522146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 428622146c3cSMike Kravetz */ 428722146c3cSMike Kravetz set_page_dirty(page); 428822146c3cSMike Kravetz 4289ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4290ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4291ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4292ab76ad54SMike Kravetz return 0; 4293ab76ad54SMike Kravetz } 4294ab76ad54SMike Kravetz 42952b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 42962b740303SSouptick Joarder struct vm_area_struct *vma, 42978382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4298788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4299ac9b9c66SHugh Dickins { 4300a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43012b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4302409eb8c2SHillf Danton int anon_rmap = 0; 43034c887265SAdam Litke unsigned long size; 43044c887265SAdam Litke struct page *page; 43051e8f889bSDavid Gibson pte_t new_pte; 4306cb900f41SKirill A. Shutemov spinlock_t *ptl; 4307285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4308cb6acd01SMike Kravetz bool new_page = false; 43094c887265SAdam Litke 431004f2cbe3SMel Gorman /* 431104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 431204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 431325985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 431404f2cbe3SMel Gorman */ 431504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4316910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 431704f2cbe3SMel Gorman current->pid); 431804f2cbe3SMel Gorman return ret; 431904f2cbe3SMel Gorman } 432004f2cbe3SMel Gorman 43214c887265SAdam Litke /* 432287bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 432387bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 432487bf91d3SMike Kravetz * once for faults beyond end of file. 43254c887265SAdam Litke */ 4326a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4327ebed4bfcSHugh Dickins if (idx >= size) 4328ebed4bfcSHugh Dickins goto out; 43291a1aad8aSMike Kravetz 433087bf91d3SMike Kravetz retry: 433187bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 433287bf91d3SMike Kravetz if (!page) { 43331a1aad8aSMike Kravetz /* 43341a1aad8aSMike Kravetz * Check for page in userfault range 43351a1aad8aSMike Kravetz */ 43361a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43371a1aad8aSMike Kravetz u32 hash; 43381a1aad8aSMike Kravetz struct vm_fault vmf = { 43391a1aad8aSMike Kravetz .vma = vma, 4340285b8dcaSHuang Ying .address = haddr, 43411a1aad8aSMike Kravetz .flags = flags, 43421a1aad8aSMike Kravetz /* 43431a1aad8aSMike Kravetz * Hard to debug if it ends up being 43441a1aad8aSMike Kravetz * used by a callee that assumes 43451a1aad8aSMike Kravetz * something about the other 43461a1aad8aSMike Kravetz * uninitialized fields... same as in 43471a1aad8aSMike Kravetz * memory.c 43481a1aad8aSMike Kravetz */ 43491a1aad8aSMike Kravetz }; 43501a1aad8aSMike Kravetz 43511a1aad8aSMike Kravetz /* 4352c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4353c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4354c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43551a1aad8aSMike Kravetz */ 4356188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43571a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4358c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43591a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4360c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 43611a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43621a1aad8aSMike Kravetz goto out; 43631a1aad8aSMike Kravetz } 43641a1aad8aSMike Kravetz 4365285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43662fc39cecSAdam Litke if (IS_ERR(page)) { 43674643d67eSMike Kravetz /* 43684643d67eSMike Kravetz * Returning error will result in faulting task being 43694643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43704643d67eSMike Kravetz * tasks from racing to fault in the same page which 43714643d67eSMike Kravetz * could result in false unable to allocate errors. 43724643d67eSMike Kravetz * Page migration does not take the fault mutex, but 43734643d67eSMike Kravetz * does a clear then write of pte's under page table 43744643d67eSMike Kravetz * lock. Page fault code could race with migration, 43754643d67eSMike Kravetz * notice the clear pte and try to allocate a page 43764643d67eSMike Kravetz * here. Before returning error, get ptl and make 43774643d67eSMike Kravetz * sure there really is no pte entry. 43784643d67eSMike Kravetz */ 43794643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 43804643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 43814643d67eSMike Kravetz ret = 0; 43824643d67eSMike Kravetz spin_unlock(ptl); 43834643d67eSMike Kravetz goto out; 43844643d67eSMike Kravetz } 43854643d67eSMike Kravetz spin_unlock(ptl); 43862b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 43876bda666aSChristoph Lameter goto out; 43886bda666aSChristoph Lameter } 438947ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 43900ed361deSNick Piggin __SetPageUptodate(page); 4391cb6acd01SMike Kravetz new_page = true; 4392ac9b9c66SHugh Dickins 4393f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4394ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 43956bda666aSChristoph Lameter if (err) { 43966bda666aSChristoph Lameter put_page(page); 43976bda666aSChristoph Lameter if (err == -EEXIST) 43986bda666aSChristoph Lameter goto retry; 43996bda666aSChristoph Lameter goto out; 44006bda666aSChristoph Lameter } 440123be7468SMel Gorman } else { 44026bda666aSChristoph Lameter lock_page(page); 44030fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 44040fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 44050fe6e20bSNaoya Horiguchi goto backout_unlocked; 440623be7468SMel Gorman } 4407409eb8c2SHillf Danton anon_rmap = 1; 44080fe6e20bSNaoya Horiguchi } 44090fe6e20bSNaoya Horiguchi } else { 441057303d80SAndy Whitcroft /* 4411998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4412998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4413998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4414fd6a03edSNaoya Horiguchi */ 4415fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 4416aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 4417972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4418fd6a03edSNaoya Horiguchi goto backout_unlocked; 44196bda666aSChristoph Lameter } 4420998b4382SNaoya Horiguchi } 44211e8f889bSDavid Gibson 442257303d80SAndy Whitcroft /* 442357303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 442457303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 442557303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 442657303d80SAndy Whitcroft * the spinlock. 442757303d80SAndy Whitcroft */ 44285e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4429285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44302b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44312b26736cSAndy Whitcroft goto backout_unlocked; 44322b26736cSAndy Whitcroft } 44335e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4434285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44355e911373SMike Kravetz } 443657303d80SAndy Whitcroft 44378bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 443883c54070SNick Piggin ret = 0; 44397f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44404c887265SAdam Litke goto backout; 44414c887265SAdam Litke 444207443a85SJoonsoo Kim if (anon_rmap) { 444307443a85SJoonsoo Kim ClearPagePrivate(page); 4444285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4445ac714904SChoi Gi-yong } else 444653f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44471e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44481e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4449285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44501e8f889bSDavid Gibson 44515d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4452788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44531e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4454974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44551e8f889bSDavid Gibson } 44561e8f889bSDavid Gibson 4457cb900f41SKirill A. Shutemov spin_unlock(ptl); 4458cb6acd01SMike Kravetz 4459cb6acd01SMike Kravetz /* 4460cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 4461cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 4462cb6acd01SMike Kravetz * isolated for migration. 4463cb6acd01SMike Kravetz */ 4464cb6acd01SMike Kravetz if (new_page) 4465cb6acd01SMike Kravetz set_page_huge_active(page); 4466cb6acd01SMike Kravetz 44674c887265SAdam Litke unlock_page(page); 44684c887265SAdam Litke out: 4469ac9b9c66SHugh Dickins return ret; 44704c887265SAdam Litke 44714c887265SAdam Litke backout: 4472cb900f41SKirill A. Shutemov spin_unlock(ptl); 44732b26736cSAndy Whitcroft backout_unlocked: 44744c887265SAdam Litke unlock_page(page); 4475285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 44764c887265SAdam Litke put_page(page); 44774c887265SAdam Litke goto out; 4478ac9b9c66SHugh Dickins } 4479ac9b9c66SHugh Dickins 44808382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4481188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44828382d914SDavidlohr Bueso { 44838382d914SDavidlohr Bueso unsigned long key[2]; 44848382d914SDavidlohr Bueso u32 hash; 44858382d914SDavidlohr Bueso 44868382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 44878382d914SDavidlohr Bueso key[1] = idx; 44888382d914SDavidlohr Bueso 448955254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 44908382d914SDavidlohr Bueso 44918382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 44928382d914SDavidlohr Bueso } 44938382d914SDavidlohr Bueso #else 44948382d914SDavidlohr Bueso /* 44958382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 44968382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 44978382d914SDavidlohr Bueso */ 4498188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44998382d914SDavidlohr Bueso { 45008382d914SDavidlohr Bueso return 0; 45018382d914SDavidlohr Bueso } 45028382d914SDavidlohr Bueso #endif 45038382d914SDavidlohr Bueso 45042b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4505788c7df4SHugh Dickins unsigned long address, unsigned int flags) 450686e5216fSAdam Litke { 45078382d914SDavidlohr Bueso pte_t *ptep, entry; 4508cb900f41SKirill A. Shutemov spinlock_t *ptl; 45092b740303SSouptick Joarder vm_fault_t ret; 45108382d914SDavidlohr Bueso u32 hash; 45118382d914SDavidlohr Bueso pgoff_t idx; 45120fe6e20bSNaoya Horiguchi struct page *page = NULL; 451357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4514a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45158382d914SDavidlohr Bueso struct address_space *mapping; 45160f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4517285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 451886e5216fSAdam Litke 4519285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4520fd6a03edSNaoya Horiguchi if (ptep) { 4521c0d0381aSMike Kravetz /* 4522c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4523c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4524c0d0381aSMike Kravetz * not actually modifying content here. 4525c0d0381aSMike Kravetz */ 4526fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4527290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4528cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4529290408d4SNaoya Horiguchi return 0; 4530290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4531aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4532972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4533b43a9990SMike Kravetz } 45348382d914SDavidlohr Bueso 4535c0d0381aSMike Kravetz /* 4536c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 453787bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 453887bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 453987bf91d3SMike Kravetz * and making the ptep no longer valid. 454087bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4541c0d0381aSMike Kravetz * 4542c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4543c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4544c0d0381aSMike Kravetz * something has changed. 4545c0d0381aSMike Kravetz */ 4546ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4547c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4548c0d0381aSMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4549c0d0381aSMike Kravetz if (!ptep) { 4550c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4551c0d0381aSMike Kravetz return VM_FAULT_OOM; 4552c0d0381aSMike Kravetz } 4553ddeaab32SMike Kravetz 45543935baa9SDavid Gibson /* 45553935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45563935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45573935baa9SDavid Gibson * the same page in the page cache. 45583935baa9SDavid Gibson */ 4559c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4560188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4561c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45628382d914SDavidlohr Bueso 45637f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45647f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45658382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4566b4d1d99fSDavid Gibson goto out_mutex; 45673935baa9SDavid Gibson } 456886e5216fSAdam Litke 456983c54070SNick Piggin ret = 0; 45701e8f889bSDavid Gibson 457157303d80SAndy Whitcroft /* 45720f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45730f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45747c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45757c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 45767c8de358SEthon Paul * properly handle it. 45770f792cf9SNaoya Horiguchi */ 45780f792cf9SNaoya Horiguchi if (!pte_present(entry)) 45790f792cf9SNaoya Horiguchi goto out_mutex; 45800f792cf9SNaoya Horiguchi 45810f792cf9SNaoya Horiguchi /* 458257303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 458357303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 458457303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 458557303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 458657303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 458757303d80SAndy Whitcroft * consumed. 458857303d80SAndy Whitcroft */ 4589106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4590285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 45912b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4592b4d1d99fSDavid Gibson goto out_mutex; 45932b26736cSAndy Whitcroft } 45945e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4595285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 459657303d80SAndy Whitcroft 4597f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 459857303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4599285b8dcaSHuang Ying vma, haddr); 460057303d80SAndy Whitcroft } 460157303d80SAndy Whitcroft 46020f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 46030fe6e20bSNaoya Horiguchi 46041e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4605b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4606cb900f41SKirill A. Shutemov goto out_ptl; 4607b4d1d99fSDavid Gibson 46080f792cf9SNaoya Horiguchi /* 46090f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 46100f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 46110f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 46120f792cf9SNaoya Horiguchi */ 46130f792cf9SNaoya Horiguchi page = pte_page(entry); 46140f792cf9SNaoya Horiguchi if (page != pagecache_page) 46150f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 46160f792cf9SNaoya Horiguchi need_wait_lock = 1; 46170f792cf9SNaoya Horiguchi goto out_ptl; 46180f792cf9SNaoya Horiguchi } 46190f792cf9SNaoya Horiguchi 46200f792cf9SNaoya Horiguchi get_page(page); 4621b4d1d99fSDavid Gibson 4622788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4623106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4624974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4625cb900f41SKirill A. Shutemov pagecache_page, ptl); 46260f792cf9SNaoya Horiguchi goto out_put_page; 4627b4d1d99fSDavid Gibson } 4628106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4629b4d1d99fSDavid Gibson } 4630b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4631285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4632788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4633285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46340f792cf9SNaoya Horiguchi out_put_page: 46350f792cf9SNaoya Horiguchi if (page != pagecache_page) 46360f792cf9SNaoya Horiguchi unlock_page(page); 46370f792cf9SNaoya Horiguchi put_page(page); 4638cb900f41SKirill A. Shutemov out_ptl: 4639cb900f41SKirill A. Shutemov spin_unlock(ptl); 464057303d80SAndy Whitcroft 464157303d80SAndy Whitcroft if (pagecache_page) { 464257303d80SAndy Whitcroft unlock_page(pagecache_page); 464357303d80SAndy Whitcroft put_page(pagecache_page); 464457303d80SAndy Whitcroft } 4645b4d1d99fSDavid Gibson out_mutex: 4646c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4647c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46480f792cf9SNaoya Horiguchi /* 46490f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46500f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46510f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46520f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46530f792cf9SNaoya Horiguchi * here without taking refcount. 46540f792cf9SNaoya Horiguchi */ 46550f792cf9SNaoya Horiguchi if (need_wait_lock) 46560f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46571e8f889bSDavid Gibson return ret; 465886e5216fSAdam Litke } 465986e5216fSAdam Litke 46608fb5debcSMike Kravetz /* 46618fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46628fb5debcSMike Kravetz * modifications for huge pages. 46638fb5debcSMike Kravetz */ 46648fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46658fb5debcSMike Kravetz pte_t *dst_pte, 46668fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46678fb5debcSMike Kravetz unsigned long dst_addr, 46688fb5debcSMike Kravetz unsigned long src_addr, 46698fb5debcSMike Kravetz struct page **pagep) 46708fb5debcSMike Kravetz { 46711e392147SAndrea Arcangeli struct address_space *mapping; 46721e392147SAndrea Arcangeli pgoff_t idx; 46731e392147SAndrea Arcangeli unsigned long size; 46741c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46758fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 46768fb5debcSMike Kravetz pte_t _dst_pte; 46778fb5debcSMike Kravetz spinlock_t *ptl; 46788fb5debcSMike Kravetz int ret; 46798fb5debcSMike Kravetz struct page *page; 46808fb5debcSMike Kravetz 46818fb5debcSMike Kravetz if (!*pagep) { 46828fb5debcSMike Kravetz ret = -ENOMEM; 46838fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 46848fb5debcSMike Kravetz if (IS_ERR(page)) 46858fb5debcSMike Kravetz goto out; 46868fb5debcSMike Kravetz 46878fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 46888fb5debcSMike Kravetz (const void __user *) src_addr, 4689810a56b9SMike Kravetz pages_per_huge_page(h), false); 46908fb5debcSMike Kravetz 4691c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 46928fb5debcSMike Kravetz if (unlikely(ret)) { 46939e368259SAndrea Arcangeli ret = -ENOENT; 46948fb5debcSMike Kravetz *pagep = page; 46958fb5debcSMike Kravetz /* don't free the page */ 46968fb5debcSMike Kravetz goto out; 46978fb5debcSMike Kravetz } 46988fb5debcSMike Kravetz } else { 46998fb5debcSMike Kravetz page = *pagep; 47008fb5debcSMike Kravetz *pagep = NULL; 47018fb5debcSMike Kravetz } 47028fb5debcSMike Kravetz 47038fb5debcSMike Kravetz /* 47048fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 47058fb5debcSMike Kravetz * preceding stores to the page contents become visible before 47068fb5debcSMike Kravetz * the set_pte_at() write. 47078fb5debcSMike Kravetz */ 47088fb5debcSMike Kravetz __SetPageUptodate(page); 47098fb5debcSMike Kravetz 47101e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 47111e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 47121e392147SAndrea Arcangeli 47131c9e8defSMike Kravetz /* 47141c9e8defSMike Kravetz * If shared, add to page cache 47151c9e8defSMike Kravetz */ 47161c9e8defSMike Kravetz if (vm_shared) { 47171e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47181e392147SAndrea Arcangeli ret = -EFAULT; 47191e392147SAndrea Arcangeli if (idx >= size) 47201e392147SAndrea Arcangeli goto out_release_nounlock; 47211c9e8defSMike Kravetz 47221e392147SAndrea Arcangeli /* 47231e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 47241e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 47251e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47261e392147SAndrea Arcangeli * the caller. 47271e392147SAndrea Arcangeli */ 47281c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47291c9e8defSMike Kravetz if (ret) 47301c9e8defSMike Kravetz goto out_release_nounlock; 47311c9e8defSMike Kravetz } 47321c9e8defSMike Kravetz 47338fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47348fb5debcSMike Kravetz spin_lock(ptl); 47358fb5debcSMike Kravetz 47361e392147SAndrea Arcangeli /* 47371e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47381e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47391e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47401e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47411e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47421e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47431e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47441e392147SAndrea Arcangeli */ 47451e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47461e392147SAndrea Arcangeli ret = -EFAULT; 47471e392147SAndrea Arcangeli if (idx >= size) 47481e392147SAndrea Arcangeli goto out_release_unlock; 47491e392147SAndrea Arcangeli 47508fb5debcSMike Kravetz ret = -EEXIST; 47518fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47528fb5debcSMike Kravetz goto out_release_unlock; 47538fb5debcSMike Kravetz 47541c9e8defSMike Kravetz if (vm_shared) { 47551c9e8defSMike Kravetz page_dup_rmap(page, true); 47561c9e8defSMike Kravetz } else { 47578fb5debcSMike Kravetz ClearPagePrivate(page); 47588fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47591c9e8defSMike Kravetz } 47608fb5debcSMike Kravetz 47618fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47628fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47638fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47648fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47658fb5debcSMike Kravetz 47668fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47678fb5debcSMike Kravetz 47688fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47698fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47708fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47718fb5debcSMike Kravetz 47728fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47738fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47748fb5debcSMike Kravetz 47758fb5debcSMike Kravetz spin_unlock(ptl); 4776cb6acd01SMike Kravetz set_page_huge_active(page); 47771c9e8defSMike Kravetz if (vm_shared) 47781c9e8defSMike Kravetz unlock_page(page); 47798fb5debcSMike Kravetz ret = 0; 47808fb5debcSMike Kravetz out: 47818fb5debcSMike Kravetz return ret; 47828fb5debcSMike Kravetz out_release_unlock: 47838fb5debcSMike Kravetz spin_unlock(ptl); 47841c9e8defSMike Kravetz if (vm_shared) 47851c9e8defSMike Kravetz unlock_page(page); 47865af10dfdSAndrea Arcangeli out_release_nounlock: 47878fb5debcSMike Kravetz put_page(page); 47888fb5debcSMike Kravetz goto out; 47898fb5debcSMike Kravetz } 47908fb5debcSMike Kravetz 479128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 479263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 479328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 47944f6da934SPeter Xu long i, unsigned int flags, int *locked) 479563551ae0SDavid Gibson { 4796d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4797d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 479828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4799a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48002be7cfedSDaniel Jordan int err = -EFAULT; 480163551ae0SDavid Gibson 480263551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 480363551ae0SDavid Gibson pte_t *pte; 4804cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 48052a15efc9SHugh Dickins int absent; 480663551ae0SDavid Gibson struct page *page; 480763551ae0SDavid Gibson 48084c887265SAdam Litke /* 480902057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 481002057967SDavid Rientjes * potentially allocating memory. 481102057967SDavid Rientjes */ 4812fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 481302057967SDavid Rientjes remainder = 0; 481402057967SDavid Rientjes break; 481502057967SDavid Rientjes } 481602057967SDavid Rientjes 481702057967SDavid Rientjes /* 48184c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48192a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48204c887265SAdam Litke * first, for the page indexing below to work. 4821cb900f41SKirill A. Shutemov * 4822cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48234c887265SAdam Litke */ 48247868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48257868a208SPunit Agrawal huge_page_size(h)); 4826cb900f41SKirill A. Shutemov if (pte) 4827cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48282a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 482963551ae0SDavid Gibson 48302a15efc9SHugh Dickins /* 48312a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48323ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48333ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48343ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48353ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48362a15efc9SHugh Dickins */ 48373ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48383ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4839cb900f41SKirill A. Shutemov if (pte) 4840cb900f41SKirill A. Shutemov spin_unlock(ptl); 48412a15efc9SHugh Dickins remainder = 0; 48422a15efc9SHugh Dickins break; 48432a15efc9SHugh Dickins } 48442a15efc9SHugh Dickins 48459cc3a5bdSNaoya Horiguchi /* 48469cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48479cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48489cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 48499cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 48509cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 48519cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 48529cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 48539cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 48549cc3a5bdSNaoya Horiguchi */ 48559cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4856106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4857106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 48582b740303SSouptick Joarder vm_fault_t ret; 485987ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 48604c887265SAdam Litke 4861cb900f41SKirill A. Shutemov if (pte) 4862cb900f41SKirill A. Shutemov spin_unlock(ptl); 486387ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 486487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 48654f6da934SPeter Xu if (locked) 486671335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 486771335f37SPeter Xu FAULT_FLAG_KILLABLE; 486887ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 486987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 487087ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 487187ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 48724426e945SPeter Xu /* 48734426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 48744426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 48754426e945SPeter Xu */ 487687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 487787ffc118SAndrea Arcangeli } 487887ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 487987ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 48802be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 48811c59827dSHugh Dickins remainder = 0; 48821c59827dSHugh Dickins break; 48831c59827dSHugh Dickins } 488487ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 48854f6da934SPeter Xu if (locked && 48861ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 48874f6da934SPeter Xu *locked = 0; 488887ffc118SAndrea Arcangeli *nr_pages = 0; 488987ffc118SAndrea Arcangeli /* 489087ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 489187ffc118SAndrea Arcangeli * error, it will return zero 489287ffc118SAndrea Arcangeli * instead. 489387ffc118SAndrea Arcangeli * 489487ffc118SAndrea Arcangeli * No need to update "position" as the 489587ffc118SAndrea Arcangeli * caller will not check it after 489687ffc118SAndrea Arcangeli * *nr_pages is set to 0. 489787ffc118SAndrea Arcangeli */ 489887ffc118SAndrea Arcangeli return i; 489987ffc118SAndrea Arcangeli } 490087ffc118SAndrea Arcangeli continue; 490187ffc118SAndrea Arcangeli } 490263551ae0SDavid Gibson 4903a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 49047f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 49058fde12caSLinus Torvalds 49068fde12caSLinus Torvalds /* 4907acbfb087SZhigang Lu * If subpage information not requested, update counters 4908acbfb087SZhigang Lu * and skip the same_page loop below. 4909acbfb087SZhigang Lu */ 4910acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4911acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4912acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4913acbfb087SZhigang Lu vaddr += huge_page_size(h); 4914acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4915acbfb087SZhigang Lu i += pages_per_huge_page(h); 4916acbfb087SZhigang Lu spin_unlock(ptl); 4917acbfb087SZhigang Lu continue; 4918acbfb087SZhigang Lu } 4919acbfb087SZhigang Lu 4920d5d4b0aaSChen, Kenneth W same_page: 4921d6692183SChen, Kenneth W if (pages) { 492269d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 49233faa52c0SJohn Hubbard /* 49243faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: 49253faa52c0SJohn Hubbard * a) we hold the ptl lock, and b) we've just checked 49263faa52c0SJohn Hubbard * that the huge page is present in the page tables. If 49273faa52c0SJohn Hubbard * the huge page is present, then the tail pages must 49283faa52c0SJohn Hubbard * also be present. The ptl prevents the head page and 49293faa52c0SJohn Hubbard * tail pages from being rearranged in any way. So this 49303faa52c0SJohn Hubbard * page must be available at this point, unless the page 49313faa52c0SJohn Hubbard * refcount overflowed: 49323faa52c0SJohn Hubbard */ 49333faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { 49343faa52c0SJohn Hubbard spin_unlock(ptl); 49353faa52c0SJohn Hubbard remainder = 0; 49363faa52c0SJohn Hubbard err = -ENOMEM; 49373faa52c0SJohn Hubbard break; 49383faa52c0SJohn Hubbard } 4939d6692183SChen, Kenneth W } 494063551ae0SDavid Gibson 494163551ae0SDavid Gibson if (vmas) 494263551ae0SDavid Gibson vmas[i] = vma; 494363551ae0SDavid Gibson 494463551ae0SDavid Gibson vaddr += PAGE_SIZE; 4945d5d4b0aaSChen, Kenneth W ++pfn_offset; 494663551ae0SDavid Gibson --remainder; 494763551ae0SDavid Gibson ++i; 4948d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4949a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4950d5d4b0aaSChen, Kenneth W /* 4951d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4952d5d4b0aaSChen, Kenneth W * of this compound page. 4953d5d4b0aaSChen, Kenneth W */ 4954d5d4b0aaSChen, Kenneth W goto same_page; 4955d5d4b0aaSChen, Kenneth W } 4956cb900f41SKirill A. Shutemov spin_unlock(ptl); 495763551ae0SDavid Gibson } 495828a35716SMichel Lespinasse *nr_pages = remainder; 495987ffc118SAndrea Arcangeli /* 496087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 496187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 496287ffc118SAndrea Arcangeli * branch. 496387ffc118SAndrea Arcangeli */ 496463551ae0SDavid Gibson *position = vaddr; 496563551ae0SDavid Gibson 49662be7cfedSDaniel Jordan return i ? i : err; 496763551ae0SDavid Gibson } 49688f860591SZhang, Yanmin 49695491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 49705491ae7bSAneesh Kumar K.V /* 49715491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 49725491ae7bSAneesh Kumar K.V * implement this. 49735491ae7bSAneesh Kumar K.V */ 49745491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 49755491ae7bSAneesh Kumar K.V #endif 49765491ae7bSAneesh Kumar K.V 49777da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 49788f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 49798f860591SZhang, Yanmin { 49808f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 49818f860591SZhang, Yanmin unsigned long start = address; 49828f860591SZhang, Yanmin pte_t *ptep; 49838f860591SZhang, Yanmin pte_t pte; 4984a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49857da4d641SPeter Zijlstra unsigned long pages = 0; 4986dff11abeSMike Kravetz bool shared_pmd = false; 4987ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4988dff11abeSMike Kravetz 4989dff11abeSMike Kravetz /* 4990dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4991ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4992dff11abeSMike Kravetz * range if PMD sharing is possible. 4993dff11abeSMike Kravetz */ 49947269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 49957269f999SJérôme Glisse 0, vma, mm, start, end); 4996ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 49978f860591SZhang, Yanmin 49988f860591SZhang, Yanmin BUG_ON(address >= end); 4999ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 50008f860591SZhang, Yanmin 5001ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 500283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5003a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5004cb900f41SKirill A. Shutemov spinlock_t *ptl; 50057868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 50068f860591SZhang, Yanmin if (!ptep) 50078f860591SZhang, Yanmin continue; 5008cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 500934ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 50107da4d641SPeter Zijlstra pages++; 5011cb900f41SKirill A. Shutemov spin_unlock(ptl); 5012dff11abeSMike Kravetz shared_pmd = true; 501339dde65cSChen, Kenneth W continue; 50147da4d641SPeter Zijlstra } 5015a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5016a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5017a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5018a8bda28dSNaoya Horiguchi continue; 5019a8bda28dSNaoya Horiguchi } 5020a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5021a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5022a8bda28dSNaoya Horiguchi 5023a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5024a8bda28dSNaoya Horiguchi pte_t newpte; 5025a8bda28dSNaoya Horiguchi 5026a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5027a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5028e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5029e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5030a8bda28dSNaoya Horiguchi pages++; 5031a8bda28dSNaoya Horiguchi } 5032a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5033a8bda28dSNaoya Horiguchi continue; 5034a8bda28dSNaoya Horiguchi } 5035a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5036023bdd00SAneesh Kumar K.V pte_t old_pte; 5037023bdd00SAneesh Kumar K.V 5038023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5039023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5040be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5041023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50427da4d641SPeter Zijlstra pages++; 50438f860591SZhang, Yanmin } 5044cb900f41SKirill A. Shutemov spin_unlock(ptl); 50458f860591SZhang, Yanmin } 5046d833352aSMel Gorman /* 5047c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5048d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5049c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5050dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5051dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5052d833352aSMel Gorman */ 5053dff11abeSMike Kravetz if (shared_pmd) 5054ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5055dff11abeSMike Kravetz else 50565491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50570f10851eSJérôme Glisse /* 50580f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 50590f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 50600f10851eSJérôme Glisse * 5061ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50620f10851eSJérôme Glisse */ 506383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5064ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50657da4d641SPeter Zijlstra 50667da4d641SPeter Zijlstra return pages << h->order; 50678f860591SZhang, Yanmin } 50688f860591SZhang, Yanmin 5069a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 5070a1e78772SMel Gorman long from, long to, 50715a6fe125SMel Gorman struct vm_area_struct *vma, 5072ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5073e4e574b7SAdam Litke { 50740db9d74eSMina Almasry long ret, chg, add = -1; 5075a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 507690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 50779119a41eSJoonsoo Kim struct resv_map *resv_map; 5078075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 50790db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5080e4e574b7SAdam Litke 508163489f8eSMike Kravetz /* This should never happen */ 508263489f8eSMike Kravetz if (from > to) { 508363489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 508463489f8eSMike Kravetz return -EINVAL; 508563489f8eSMike Kravetz } 508663489f8eSMike Kravetz 5087a1e78772SMel Gorman /* 508817c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 508917c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 509090481622SDavid Gibson * without using reserves 509117c9d12eSMel Gorman */ 5092ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 509317c9d12eSMel Gorman return 0; 509417c9d12eSMel Gorman 509517c9d12eSMel Gorman /* 5096a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5097a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5098a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5099a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5100a1e78772SMel Gorman */ 51019119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5102f27a5136SMike Kravetz /* 5103f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5104f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5105f27a5136SMike Kravetz * hugetlbfs_get_inode). 5106f27a5136SMike Kravetz */ 51074e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 51089119a41eSJoonsoo Kim 51090db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 51109119a41eSJoonsoo Kim 51119119a41eSJoonsoo Kim } else { 5112e9fe92aeSMina Almasry /* Private mapping. */ 51139119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51145a6fe125SMel Gorman if (!resv_map) 51155a6fe125SMel Gorman return -ENOMEM; 51165a6fe125SMel Gorman 511717c9d12eSMel Gorman chg = to - from; 511817c9d12eSMel Gorman 51195a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51205a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51215a6fe125SMel Gorman } 51225a6fe125SMel Gorman 5123c50ac050SDave Hansen if (chg < 0) { 5124c50ac050SDave Hansen ret = chg; 5125c50ac050SDave Hansen goto out_err; 5126c50ac050SDave Hansen } 512717c9d12eSMel Gorman 5128075a61d0SMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 5129075a61d0SMina Almasry hstate_index(h), chg * pages_per_huge_page(h), &h_cg); 5130075a61d0SMina Almasry 5131075a61d0SMina Almasry if (ret < 0) { 5132075a61d0SMina Almasry ret = -ENOMEM; 5133075a61d0SMina Almasry goto out_err; 5134075a61d0SMina Almasry } 5135075a61d0SMina Almasry 5136075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5137075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5138075a61d0SMina Almasry * of the resv_map. 5139075a61d0SMina Almasry */ 5140075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5141075a61d0SMina Almasry } 5142075a61d0SMina Almasry 51431c5ecae3SMike Kravetz /* 51441c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51451c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51461c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51471c5ecae3SMike Kravetz */ 51481c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 51491c5ecae3SMike Kravetz if (gbl_reserve < 0) { 5150c50ac050SDave Hansen ret = -ENOSPC; 5151075a61d0SMina Almasry goto out_uncharge_cgroup; 5152c50ac050SDave Hansen } 515317c9d12eSMel Gorman 515417c9d12eSMel Gorman /* 515517c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 515690481622SDavid Gibson * Hand the pages back to the subpool if there are not 515717c9d12eSMel Gorman */ 51581c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 515917c9d12eSMel Gorman if (ret < 0) { 5160075a61d0SMina Almasry goto out_put_pages; 516117c9d12eSMel Gorman } 516217c9d12eSMel Gorman 516317c9d12eSMel Gorman /* 516417c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 516517c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 516617c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 516717c9d12eSMel Gorman * the reservation was and the page cache tells how much of 516817c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 516917c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 517017c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 517117c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 517217c9d12eSMel Gorman * else has to be done for private mappings here 517317c9d12eSMel Gorman */ 517433039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5175075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 517633039678SMike Kravetz 51770db9d74eSMina Almasry if (unlikely(add < 0)) { 51780db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5179075a61d0SMina Almasry goto out_put_pages; 51800db9d74eSMina Almasry } else if (unlikely(chg > add)) { 518133039678SMike Kravetz /* 518233039678SMike Kravetz * pages in this range were added to the reserve 518333039678SMike Kravetz * map between region_chg and region_add. This 518433039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 518533039678SMike Kravetz * the subpool and reserve counts modified above 518633039678SMike Kravetz * based on the difference. 518733039678SMike Kravetz */ 518833039678SMike Kravetz long rsv_adjust; 518933039678SMike Kravetz 5190075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5191075a61d0SMina Almasry hstate_index(h), 5192075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5193075a61d0SMina Almasry 519433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 519533039678SMike Kravetz chg - add); 519633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 519733039678SMike Kravetz } 519833039678SMike Kravetz } 5199a43a8c39SChen, Kenneth W return 0; 5200075a61d0SMina Almasry out_put_pages: 5201075a61d0SMina Almasry /* put back original number of pages, chg */ 5202075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5203075a61d0SMina Almasry out_uncharge_cgroup: 5204075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5205075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5206c50ac050SDave Hansen out_err: 52075e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 52080db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 52090db9d74eSMina Almasry * region_add failed or didn't run. 52100db9d74eSMina Almasry */ 52110db9d74eSMina Almasry if (chg >= 0 && add < 0) 52120db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5213f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5214f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 5215c50ac050SDave Hansen return ret; 5216a43a8c39SChen, Kenneth W } 5217a43a8c39SChen, Kenneth W 5218b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5219b5cec28dSMike Kravetz long freed) 5220a43a8c39SChen, Kenneth W { 5221a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52224e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52239119a41eSJoonsoo Kim long chg = 0; 522490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52251c5ecae3SMike Kravetz long gbl_reserve; 522645c682a6SKen Chen 5227f27a5136SMike Kravetz /* 5228f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5229f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5230f27a5136SMike Kravetz */ 5231b5cec28dSMike Kravetz if (resv_map) { 5232b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5233b5cec28dSMike Kravetz /* 5234b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5235b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5236b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5237b5cec28dSMike Kravetz */ 5238b5cec28dSMike Kravetz if (chg < 0) 5239b5cec28dSMike Kravetz return chg; 5240b5cec28dSMike Kravetz } 5241b5cec28dSMike Kravetz 524245c682a6SKen Chen spin_lock(&inode->i_lock); 5243e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 524445c682a6SKen Chen spin_unlock(&inode->i_lock); 524545c682a6SKen Chen 52461c5ecae3SMike Kravetz /* 52471c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52481c5ecae3SMike Kravetz * reservations to be released may be adjusted. 52491c5ecae3SMike Kravetz */ 52501c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52511c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5252b5cec28dSMike Kravetz 5253b5cec28dSMike Kravetz return 0; 5254a43a8c39SChen, Kenneth W } 525593f70f90SNaoya Horiguchi 52563212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 52573212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 52583212b535SSteve Capper struct vm_area_struct *vma, 52593212b535SSteve Capper unsigned long addr, pgoff_t idx) 52603212b535SSteve Capper { 52613212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52623212b535SSteve Capper svma->vm_start; 52633212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52643212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 52653212b535SSteve Capper 52663212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5267de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5268de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 52693212b535SSteve Capper 52703212b535SSteve Capper /* 52713212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 52723212b535SSteve Capper * page table page. 52733212b535SSteve Capper */ 52743212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 52753212b535SSteve Capper vm_flags != svm_flags || 52763212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 52773212b535SSteve Capper return 0; 52783212b535SSteve Capper 52793212b535SSteve Capper return saddr; 52803212b535SSteve Capper } 52813212b535SSteve Capper 528231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 52833212b535SSteve Capper { 52843212b535SSteve Capper unsigned long base = addr & PUD_MASK; 52853212b535SSteve Capper unsigned long end = base + PUD_SIZE; 52863212b535SSteve Capper 52873212b535SSteve Capper /* 52883212b535SSteve Capper * check on proper vm_flags and page table alignment 52893212b535SSteve Capper */ 5290017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 529131aafb45SNicholas Krause return true; 529231aafb45SNicholas Krause return false; 52933212b535SSteve Capper } 52943212b535SSteve Capper 52953212b535SSteve Capper /* 5296017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5297017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5298017b1660SMike Kravetz * shared pmd mappings. 5299017b1660SMike Kravetz */ 5300017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5301017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5302017b1660SMike Kravetz { 530375802ca6SPeter Xu unsigned long a_start, a_end; 5304017b1660SMike Kravetz 5305017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 5306017b1660SMike Kravetz return; 5307017b1660SMike Kravetz 530875802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 530975802ca6SPeter Xu a_start = ALIGN_DOWN(*start, PUD_SIZE); 531075802ca6SPeter Xu a_end = ALIGN(*end, PUD_SIZE); 5311017b1660SMike Kravetz 5312017b1660SMike Kravetz /* 531375802ca6SPeter Xu * Intersect the range with the vma range, since pmd sharing won't be 531475802ca6SPeter Xu * across vma after all 5315017b1660SMike Kravetz */ 531675802ca6SPeter Xu *start = max(vma->vm_start, a_start); 531775802ca6SPeter Xu *end = min(vma->vm_end, a_end); 5318017b1660SMike Kravetz } 5319017b1660SMike Kravetz 5320017b1660SMike Kravetz /* 53213212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53223212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53233212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5324c0d0381aSMike Kravetz * code much cleaner. 5325c0d0381aSMike Kravetz * 5326c0d0381aSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode. 5327c0d0381aSMike Kravetz * For hugetlbfs, this prevents removal of any page table entries associated 5328c0d0381aSMike Kravetz * with the address space. This is important as we are setting up sharing 5329c0d0381aSMike Kravetz * based on existing page table entries (mappings). 53303212b535SSteve Capper */ 53313212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 53323212b535SSteve Capper { 53333212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 53343212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 53353212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 53363212b535SSteve Capper vma->vm_pgoff; 53373212b535SSteve Capper struct vm_area_struct *svma; 53383212b535SSteve Capper unsigned long saddr; 53393212b535SSteve Capper pte_t *spte = NULL; 53403212b535SSteve Capper pte_t *pte; 5341cb900f41SKirill A. Shutemov spinlock_t *ptl; 53423212b535SSteve Capper 53433212b535SSteve Capper if (!vma_shareable(vma, addr)) 53443212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 53453212b535SSteve Capper 53463212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 53473212b535SSteve Capper if (svma == vma) 53483212b535SSteve Capper continue; 53493212b535SSteve Capper 53503212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 53513212b535SSteve Capper if (saddr) { 53527868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 53537868a208SPunit Agrawal vma_mmu_pagesize(svma)); 53543212b535SSteve Capper if (spte) { 53553212b535SSteve Capper get_page(virt_to_page(spte)); 53563212b535SSteve Capper break; 53573212b535SSteve Capper } 53583212b535SSteve Capper } 53593212b535SSteve Capper } 53603212b535SSteve Capper 53613212b535SSteve Capper if (!spte) 53623212b535SSteve Capper goto out; 53633212b535SSteve Capper 53648bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5365dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 53663212b535SSteve Capper pud_populate(mm, pud, 53673212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5368c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5369dc6c9a35SKirill A. Shutemov } else { 53703212b535SSteve Capper put_page(virt_to_page(spte)); 5371dc6c9a35SKirill A. Shutemov } 5372cb900f41SKirill A. Shutemov spin_unlock(ptl); 53733212b535SSteve Capper out: 53743212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 53753212b535SSteve Capper return pte; 53763212b535SSteve Capper } 53773212b535SSteve Capper 53783212b535SSteve Capper /* 53793212b535SSteve Capper * unmap huge page backed by shared pte. 53803212b535SSteve Capper * 53813212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 53823212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 53833212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 53843212b535SSteve Capper * 5385c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 53863212b535SSteve Capper * 53873212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 53883212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 53893212b535SSteve Capper */ 539034ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 539134ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 53923212b535SSteve Capper { 53933212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5394c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5395c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 53963212b535SSteve Capper 539734ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 53983212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 53993212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 54003212b535SSteve Capper return 0; 54013212b535SSteve Capper 54023212b535SSteve Capper pud_clear(pud); 54033212b535SSteve Capper put_page(virt_to_page(ptep)); 5404dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54053212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54063212b535SSteve Capper return 1; 54073212b535SSteve Capper } 54089e5fc74cSSteve Capper #define want_pmd_share() (1) 54099e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54109e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 54119e5fc74cSSteve Capper { 54129e5fc74cSSteve Capper return NULL; 54139e5fc74cSSteve Capper } 5414e81f2d22SZhang Zhen 541534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 541634ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5417e81f2d22SZhang Zhen { 5418e81f2d22SZhang Zhen return 0; 5419e81f2d22SZhang Zhen } 5420017b1660SMike Kravetz 5421017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5422017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5423017b1660SMike Kravetz { 5424017b1660SMike Kravetz } 54259e5fc74cSSteve Capper #define want_pmd_share() (0) 54263212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54273212b535SSteve Capper 54289e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 54299e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 54309e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54319e5fc74cSSteve Capper { 54329e5fc74cSSteve Capper pgd_t *pgd; 5433c2febafcSKirill A. Shutemov p4d_t *p4d; 54349e5fc74cSSteve Capper pud_t *pud; 54359e5fc74cSSteve Capper pte_t *pte = NULL; 54369e5fc74cSSteve Capper 54379e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5438f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5439f4f0a3d8SKirill A. Shutemov if (!p4d) 5440f4f0a3d8SKirill A. Shutemov return NULL; 5441c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 54429e5fc74cSSteve Capper if (pud) { 54439e5fc74cSSteve Capper if (sz == PUD_SIZE) { 54449e5fc74cSSteve Capper pte = (pte_t *)pud; 54459e5fc74cSSteve Capper } else { 54469e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 54479e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 54489e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 54499e5fc74cSSteve Capper else 54509e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54519e5fc74cSSteve Capper } 54529e5fc74cSSteve Capper } 54534e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 54549e5fc74cSSteve Capper 54559e5fc74cSSteve Capper return pte; 54569e5fc74cSSteve Capper } 54579e5fc74cSSteve Capper 54589b19df29SPunit Agrawal /* 54599b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 54609b19df29SPunit Agrawal * entry at address @addr 54619b19df29SPunit Agrawal * 54628ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 54638ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 54649b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 54659b19df29SPunit Agrawal * table. 54669b19df29SPunit Agrawal */ 54677868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 54687868a208SPunit Agrawal unsigned long addr, unsigned long sz) 54699e5fc74cSSteve Capper { 54709e5fc74cSSteve Capper pgd_t *pgd; 5471c2febafcSKirill A. Shutemov p4d_t *p4d; 54728ac0b81aSLi Xinhai pud_t *pud; 54738ac0b81aSLi Xinhai pmd_t *pmd; 54749e5fc74cSSteve Capper 54759e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5476c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5477c2febafcSKirill A. Shutemov return NULL; 5478c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5479c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5480c2febafcSKirill A. Shutemov return NULL; 54819b19df29SPunit Agrawal 5482c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 54838ac0b81aSLi Xinhai if (sz == PUD_SIZE) 54848ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 54859e5fc74cSSteve Capper return (pte_t *)pud; 54868ac0b81aSLi Xinhai if (!pud_present(*pud)) 54878ac0b81aSLi Xinhai return NULL; 54888ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 54899b19df29SPunit Agrawal 54909e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 54918ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 54929e5fc74cSSteve Capper return (pte_t *)pmd; 54939e5fc74cSSteve Capper } 54949e5fc74cSSteve Capper 549561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 549661f77edaSNaoya Horiguchi 549761f77edaSNaoya Horiguchi /* 549861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 549961f77edaSNaoya Horiguchi * behavior. 550061f77edaSNaoya Horiguchi */ 550161f77edaSNaoya Horiguchi struct page * __weak 550261f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 550361f77edaSNaoya Horiguchi int write) 550461f77edaSNaoya Horiguchi { 550561f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 550661f77edaSNaoya Horiguchi } 550761f77edaSNaoya Horiguchi 550861f77edaSNaoya Horiguchi struct page * __weak 55094dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55104dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55114dc71451SAneesh Kumar K.V { 55124dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55134dc71451SAneesh Kumar K.V return NULL; 55144dc71451SAneesh Kumar K.V } 55154dc71451SAneesh Kumar K.V 55164dc71451SAneesh Kumar K.V struct page * __weak 55179e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5518e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55199e5fc74cSSteve Capper { 5520e66f17ffSNaoya Horiguchi struct page *page = NULL; 5521e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5522c9d398faSNaoya Horiguchi pte_t pte; 55233faa52c0SJohn Hubbard 55243faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55253faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55263faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55273faa52c0SJohn Hubbard return NULL; 55283faa52c0SJohn Hubbard 5529e66f17ffSNaoya Horiguchi retry: 5530e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5531e66f17ffSNaoya Horiguchi spin_lock(ptl); 5532e66f17ffSNaoya Horiguchi /* 5533e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5534e66f17ffSNaoya Horiguchi * unmapped from other threads. 5535e66f17ffSNaoya Horiguchi */ 5536e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5537e66f17ffSNaoya Horiguchi goto out; 5538c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5539c9d398faSNaoya Horiguchi if (pte_present(pte)) { 554097534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 55413faa52c0SJohn Hubbard /* 55423faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 55433faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 55443faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 55453faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 55463faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 55473faa52c0SJohn Hubbard * unless the page refcount overflowed: 55483faa52c0SJohn Hubbard */ 55493faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 55503faa52c0SJohn Hubbard page = NULL; 55513faa52c0SJohn Hubbard goto out; 55523faa52c0SJohn Hubbard } 5553e66f17ffSNaoya Horiguchi } else { 5554c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5555e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5556e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5557e66f17ffSNaoya Horiguchi goto retry; 5558e66f17ffSNaoya Horiguchi } 5559e66f17ffSNaoya Horiguchi /* 5560e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5561e66f17ffSNaoya Horiguchi * follow_page_mask(). 5562e66f17ffSNaoya Horiguchi */ 5563e66f17ffSNaoya Horiguchi } 5564e66f17ffSNaoya Horiguchi out: 5565e66f17ffSNaoya Horiguchi spin_unlock(ptl); 55669e5fc74cSSteve Capper return page; 55679e5fc74cSSteve Capper } 55689e5fc74cSSteve Capper 556961f77edaSNaoya Horiguchi struct page * __weak 55709e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5571e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 55729e5fc74cSSteve Capper { 55733faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5574e66f17ffSNaoya Horiguchi return NULL; 55759e5fc74cSSteve Capper 5576e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 55779e5fc74cSSteve Capper } 55789e5fc74cSSteve Capper 5579faaa5b62SAnshuman Khandual struct page * __weak 5580faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5581faaa5b62SAnshuman Khandual { 55823faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5583faaa5b62SAnshuman Khandual return NULL; 5584faaa5b62SAnshuman Khandual 5585faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5586faaa5b62SAnshuman Khandual } 5587faaa5b62SAnshuman Khandual 558831caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 558931caf665SNaoya Horiguchi { 5590bcc54222SNaoya Horiguchi bool ret = true; 5591bcc54222SNaoya Horiguchi 5592309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 559331caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5594bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 5595bcc54222SNaoya Horiguchi ret = false; 5596bcc54222SNaoya Horiguchi goto unlock; 5597bcc54222SNaoya Horiguchi } 5598bcc54222SNaoya Horiguchi clear_page_huge_active(page); 559931caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5600bcc54222SNaoya Horiguchi unlock: 560131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5602bcc54222SNaoya Horiguchi return ret; 560331caf665SNaoya Horiguchi } 560431caf665SNaoya Horiguchi 560531caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 560631caf665SNaoya Horiguchi { 5607309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 560831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5609bcc54222SNaoya Horiguchi set_page_huge_active(page); 561031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 561131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 561231caf665SNaoya Horiguchi put_page(page); 561331caf665SNaoya Horiguchi } 5614ab5ac90aSMichal Hocko 5615ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5616ab5ac90aSMichal Hocko { 5617ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5618ab5ac90aSMichal Hocko 5619ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5620ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5621ab5ac90aSMichal Hocko 5622ab5ac90aSMichal Hocko /* 5623ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5624ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5625ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5626ab5ac90aSMichal Hocko * the temporary status. 5627ab5ac90aSMichal Hocko * 5628ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5629ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5630ab5ac90aSMichal Hocko * the per-node's. 5631ab5ac90aSMichal Hocko */ 5632ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5633ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5634ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5635ab5ac90aSMichal Hocko 5636ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5637ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5638ab5ac90aSMichal Hocko 5639ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5640ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5641ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5642ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5643ab5ac90aSMichal Hocko } 5644ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5645ab5ac90aSMichal Hocko } 5646ab5ac90aSMichal Hocko } 5647cf11e85fSRoman Gushchin 5648cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5649cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5650cf11e85fSRoman Gushchin 5651cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5652cf11e85fSRoman Gushchin { 5653cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5654cf11e85fSRoman Gushchin return 0; 5655cf11e85fSRoman Gushchin } 5656cf11e85fSRoman Gushchin 5657cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5658cf11e85fSRoman Gushchin 5659cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5660cf11e85fSRoman Gushchin { 5661cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5662cf11e85fSRoman Gushchin int nid; 5663cf11e85fSRoman Gushchin 5664cf11e85fSRoman Gushchin cma_reserve_called = true; 5665cf11e85fSRoman Gushchin 5666cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5667cf11e85fSRoman Gushchin return; 5668cf11e85fSRoman Gushchin 5669cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5670cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5671cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5672cf11e85fSRoman Gushchin return; 5673cf11e85fSRoman Gushchin } 5674cf11e85fSRoman Gushchin 5675cf11e85fSRoman Gushchin /* 5676cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5677cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5678cf11e85fSRoman Gushchin */ 5679cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5680cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5681cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5682cf11e85fSRoman Gushchin 5683cf11e85fSRoman Gushchin reserved = 0; 5684cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5685cf11e85fSRoman Gushchin int res; 568629d0f41dSBarry Song char name[20]; 5687cf11e85fSRoman Gushchin 5688cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5689cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5690cf11e85fSRoman Gushchin 569129d0f41dSBarry Song snprintf(name, 20, "hugetlb%d", nid); 5692cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 569329d0f41dSBarry Song 0, false, name, 5694cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5695cf11e85fSRoman Gushchin if (res) { 5696cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5697cf11e85fSRoman Gushchin res, nid); 5698cf11e85fSRoman Gushchin continue; 5699cf11e85fSRoman Gushchin } 5700cf11e85fSRoman Gushchin 5701cf11e85fSRoman Gushchin reserved += size; 5702cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5703cf11e85fSRoman Gushchin size / SZ_1M, nid); 5704cf11e85fSRoman Gushchin 5705cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5706cf11e85fSRoman Gushchin break; 5707cf11e85fSRoman Gushchin } 5708cf11e85fSRoman Gushchin } 5709cf11e85fSRoman Gushchin 5710cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5711cf11e85fSRoman Gushchin { 5712cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5713cf11e85fSRoman Gushchin return; 5714cf11e85fSRoman Gushchin 5715cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5716cf11e85fSRoman Gushchin } 5717cf11e85fSRoman Gushchin 5718cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5719