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 list_del(&nrg->link); 2440db9d74eSMina Almasry 2450db9d74eSMina Almasry nrg->from = from; 2460db9d74eSMina Almasry nrg->to = to; 2470db9d74eSMina Almasry 2480db9d74eSMina Almasry return nrg; 2490db9d74eSMina Almasry } 2500db9d74eSMina Almasry 251075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 252075a61d0SMina Almasry struct file_region *rg) 253075a61d0SMina Almasry { 254075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 255075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 256075a61d0SMina Almasry nrg->css = rg->css; 257075a61d0SMina Almasry if (rg->css) 258075a61d0SMina Almasry css_get(rg->css); 259075a61d0SMina Almasry #endif 260075a61d0SMina Almasry } 261075a61d0SMina Almasry 262075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 263075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 264075a61d0SMina Almasry struct hstate *h, 265075a61d0SMina Almasry struct resv_map *resv, 266075a61d0SMina Almasry struct file_region *nrg) 267075a61d0SMina Almasry { 268075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 269075a61d0SMina Almasry if (h_cg) { 270075a61d0SMina Almasry nrg->reservation_counter = 271075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 272075a61d0SMina Almasry nrg->css = &h_cg->css; 273075a61d0SMina Almasry if (!resv->pages_per_hpage) 274075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 275075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 276075a61d0SMina Almasry * a resv_map. 277075a61d0SMina Almasry */ 278075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 279075a61d0SMina Almasry } else { 280075a61d0SMina Almasry nrg->reservation_counter = NULL; 281075a61d0SMina Almasry nrg->css = NULL; 282075a61d0SMina Almasry } 283075a61d0SMina Almasry #endif 284075a61d0SMina Almasry } 285075a61d0SMina Almasry 286a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 287a9b3f867SMina Almasry struct file_region *org) 288a9b3f867SMina Almasry { 289a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 290a9b3f867SMina Almasry return rg && org && 291a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 292a9b3f867SMina Almasry rg->css == org->css; 293a9b3f867SMina Almasry 294a9b3f867SMina Almasry #else 295a9b3f867SMina Almasry return true; 296a9b3f867SMina Almasry #endif 297a9b3f867SMina Almasry } 298a9b3f867SMina Almasry 299a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 300a9b3f867SMina Almasry { 301a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 302a9b3f867SMina Almasry 303a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 304a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 305a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 306a9b3f867SMina Almasry prg->to = rg->to; 307a9b3f867SMina Almasry 308a9b3f867SMina Almasry list_del(&rg->link); 309a9b3f867SMina Almasry kfree(rg); 310a9b3f867SMina Almasry 3117db5e7b6SWei Yang rg = prg; 312a9b3f867SMina Almasry } 313a9b3f867SMina Almasry 314a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 315a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 316a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 317a9b3f867SMina Almasry nrg->from = rg->from; 318a9b3f867SMina Almasry 319a9b3f867SMina Almasry list_del(&rg->link); 320a9b3f867SMina Almasry kfree(rg); 321a9b3f867SMina Almasry } 322a9b3f867SMina Almasry } 323a9b3f867SMina Almasry 324972a3da3SWei Yang /* 325972a3da3SWei Yang * Must be called with resv->lock held. 326972a3da3SWei Yang * 327972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 328972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 329972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 330972a3da3SWei Yang * the regions for this range. 331d75c6af9SMina Almasry */ 332d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 333075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 334972a3da3SWei Yang struct hstate *h, long *regions_needed) 335d75c6af9SMina Almasry { 3360db9d74eSMina Almasry long add = 0; 337d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3380db9d74eSMina Almasry long last_accounted_offset = f; 339d75c6af9SMina Almasry struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; 340d75c6af9SMina Almasry 3410db9d74eSMina Almasry if (regions_needed) 3420db9d74eSMina Almasry *regions_needed = 0; 343d75c6af9SMina Almasry 3440db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3450db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3460db9d74eSMina Almasry * bounds checking. 3470db9d74eSMina Almasry */ 3480db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 3490db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 3500db9d74eSMina Almasry if (rg->from < f) { 3510db9d74eSMina Almasry /* If this region ends after the last accounted offset, 3520db9d74eSMina Almasry * then we need to update last_accounted_offset. 3530db9d74eSMina Almasry */ 3540db9d74eSMina Almasry if (rg->to > last_accounted_offset) 3550db9d74eSMina Almasry last_accounted_offset = rg->to; 3560db9d74eSMina Almasry continue; 3570db9d74eSMina Almasry } 358d75c6af9SMina Almasry 3590db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 3600db9d74eSMina Almasry * finished. 3610db9d74eSMina Almasry */ 362d75c6af9SMina Almasry if (rg->from > t) 363d75c6af9SMina Almasry break; 364d75c6af9SMina Almasry 3650db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 3660db9d74eSMina Almasry * update last_accounted_offset. 367d75c6af9SMina Almasry */ 3680db9d74eSMina Almasry if (rg->from > last_accounted_offset) { 3690db9d74eSMina Almasry add += rg->from - last_accounted_offset; 370972a3da3SWei Yang if (!regions_needed) { 3710db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3720db9d74eSMina Almasry resv, last_accounted_offset, rg->from); 373075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, 374075a61d0SMina Almasry resv, nrg); 3750db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 376a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 377972a3da3SWei Yang } else 3780db9d74eSMina Almasry *regions_needed += 1; 379d75c6af9SMina Almasry } 380d75c6af9SMina Almasry 3810db9d74eSMina Almasry last_accounted_offset = rg->to; 3820db9d74eSMina Almasry } 3830db9d74eSMina Almasry 3840db9d74eSMina Almasry /* Handle the case where our range extends beyond 3850db9d74eSMina Almasry * last_accounted_offset. 3860db9d74eSMina Almasry */ 3870db9d74eSMina Almasry if (last_accounted_offset < t) { 3880db9d74eSMina Almasry add += t - last_accounted_offset; 389972a3da3SWei Yang if (!regions_needed) { 3900db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3910db9d74eSMina Almasry resv, last_accounted_offset, t); 392075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); 3930db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 394a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 395972a3da3SWei Yang } else 3960db9d74eSMina Almasry *regions_needed += 1; 3970db9d74eSMina Almasry } 3980db9d74eSMina Almasry 3990db9d74eSMina Almasry VM_BUG_ON(add < 0); 4000db9d74eSMina Almasry return add; 4010db9d74eSMina Almasry } 4020db9d74eSMina Almasry 4030db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4040db9d74eSMina Almasry */ 4050db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4060db9d74eSMina Almasry int regions_needed) 4070db9d74eSMina Almasry __must_hold(&resv->lock) 4080db9d74eSMina Almasry { 4090db9d74eSMina Almasry struct list_head allocated_regions; 4100db9d74eSMina Almasry int to_allocate = 0, i = 0; 4110db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4120db9d74eSMina Almasry 4130db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4140db9d74eSMina Almasry 4150db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4160db9d74eSMina Almasry 4170db9d74eSMina Almasry /* 4180db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4190db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4200db9d74eSMina Almasry * 4210db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4220db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4230db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4240db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4250db9d74eSMina Almasry */ 4260db9d74eSMina Almasry while (resv->region_cache_count < 4270db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4280db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4290db9d74eSMina Almasry resv->region_cache_count; 4300db9d74eSMina Almasry 4310db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4320db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4330db9d74eSMina Almasry * needing to allocate for regions_needed. 4340db9d74eSMina Almasry */ 4350db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4360db9d74eSMina Almasry 4370db9d74eSMina Almasry spin_unlock(&resv->lock); 4380db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4390db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4400db9d74eSMina Almasry if (!trg) 4410db9d74eSMina Almasry goto out_of_memory; 4420db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4430db9d74eSMina Almasry } 4440db9d74eSMina Almasry 4450db9d74eSMina Almasry spin_lock(&resv->lock); 4460db9d74eSMina Almasry 447d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 448d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4490db9d74eSMina Almasry } 4500db9d74eSMina Almasry 4510db9d74eSMina Almasry return 0; 4520db9d74eSMina Almasry 4530db9d74eSMina Almasry out_of_memory: 4540db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 455d75c6af9SMina Almasry list_del(&rg->link); 456d75c6af9SMina Almasry kfree(rg); 457d75c6af9SMina Almasry } 4580db9d74eSMina Almasry return -ENOMEM; 459d75c6af9SMina Almasry } 460d75c6af9SMina Almasry 4611dd308a7SMike Kravetz /* 4621dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 4630db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 4640db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 4650db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 4660db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 4670db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 468cf3ad20bSMike Kravetz * 4690db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 4700db9d74eSMina Almasry * 4710db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 4720db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 4737c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 4740db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 4750db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 4760db9d74eSMina Almasry * 1 page will only require at most 1 entry. 4771dd308a7SMike Kravetz */ 4780db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 479075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 480075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 48196822904SAndy Whitcroft { 4820db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 48396822904SAndy Whitcroft 4847b24d861SDavidlohr Bueso spin_lock(&resv->lock); 4850db9d74eSMina Almasry retry: 4860db9d74eSMina Almasry 4870db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 488972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 489972a3da3SWei Yang &actual_regions_needed); 49096822904SAndy Whitcroft 4915e911373SMike Kravetz /* 4920db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4930db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 4940db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 4950db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 4960db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 4970db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 4980db9d74eSMina Almasry * operation. 4995e911373SMike Kravetz */ 5000db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5010db9d74eSMina Almasry resv->region_cache_count < 5020db9d74eSMina Almasry resv->adds_in_progress + 5030db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5040db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5050db9d74eSMina Almasry * allocate file_region entries. 5060db9d74eSMina Almasry */ 5070db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5085e911373SMike Kravetz 5090db9d74eSMina Almasry if (allocate_file_region_entries( 5100db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5110db9d74eSMina Almasry return -ENOMEM; 5125e911373SMike Kravetz } 5135e911373SMike Kravetz 5140db9d74eSMina Almasry goto retry; 5150db9d74eSMina Almasry } 516cf3ad20bSMike Kravetz 517972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5180db9d74eSMina Almasry 5190db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5200db9d74eSMina Almasry 5217b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 522cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 523cf3ad20bSMike Kravetz return add; 52496822904SAndy Whitcroft } 52596822904SAndy Whitcroft 5261dd308a7SMike Kravetz /* 5271dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5281dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5291dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5301dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5311dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5321dd308a7SMike Kravetz * not change the number of huge pages represented by the 5330db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5340db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5350db9d74eSMina Almasry * file_region structure is added. 5360db9d74eSMina Almasry * 5370db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5380db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5390db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5405e911373SMike Kravetz * 5415e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5425e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5435e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5445e911373SMike Kravetz * is needed and can not be allocated. 5451dd308a7SMike Kravetz */ 5460db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5470db9d74eSMina Almasry long *out_regions_needed) 54896822904SAndy Whitcroft { 54996822904SAndy Whitcroft long chg = 0; 55096822904SAndy Whitcroft 5517b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5525e911373SMike Kravetz 553972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 554075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 555972a3da3SWei Yang out_regions_needed); 5565e911373SMike Kravetz 5570db9d74eSMina Almasry if (*out_regions_needed == 0) 5580db9d74eSMina Almasry *out_regions_needed = 1; 5595e911373SMike Kravetz 5600db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5615e911373SMike Kravetz return -ENOMEM; 5625e911373SMike Kravetz 5630db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 56496822904SAndy Whitcroft 5657b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 56696822904SAndy Whitcroft return chg; 56796822904SAndy Whitcroft } 56896822904SAndy Whitcroft 5691dd308a7SMike Kravetz /* 5705e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 5715e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 5725e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 5735e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 5740db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 5750db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 5760db9d74eSMina Almasry * the adds_in_progress counter. 5775e911373SMike Kravetz * 5785e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 5795e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 5805e911373SMike Kravetz * arguments will match the associated region_chg call. 5815e911373SMike Kravetz */ 5820db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 5830db9d74eSMina Almasry long regions_needed) 5845e911373SMike Kravetz { 5855e911373SMike Kravetz spin_lock(&resv->lock); 5865e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 5870db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 5885e911373SMike Kravetz spin_unlock(&resv->lock); 5895e911373SMike Kravetz } 5905e911373SMike Kravetz 5915e911373SMike Kravetz /* 592feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 593feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 594feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 595feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 596feba16e2SMike Kravetz * 597feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 598feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 599feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 600feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 601feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 602feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 603feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6041dd308a7SMike Kravetz */ 605feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 60696822904SAndy Whitcroft { 6071406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 60896822904SAndy Whitcroft struct file_region *rg, *trg; 609feba16e2SMike Kravetz struct file_region *nrg = NULL; 610feba16e2SMike Kravetz long del = 0; 61196822904SAndy Whitcroft 612feba16e2SMike Kravetz retry: 6137b24d861SDavidlohr Bueso spin_lock(&resv->lock); 614feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 615dbe409e4SMike Kravetz /* 616dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 617dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 618dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 619dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 620dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 621dbe409e4SMike Kravetz */ 622dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 623feba16e2SMike Kravetz continue; 624dbe409e4SMike Kravetz 625feba16e2SMike Kravetz if (rg->from >= t) 62696822904SAndy Whitcroft break; 62796822904SAndy Whitcroft 628feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 629feba16e2SMike Kravetz /* 630feba16e2SMike Kravetz * Check for an entry in the cache before dropping 631feba16e2SMike Kravetz * lock and attempting allocation. 632feba16e2SMike Kravetz */ 633feba16e2SMike Kravetz if (!nrg && 634feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 635feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 636feba16e2SMike Kravetz struct file_region, 637feba16e2SMike Kravetz link); 638feba16e2SMike Kravetz list_del(&nrg->link); 639feba16e2SMike Kravetz resv->region_cache_count--; 64096822904SAndy Whitcroft } 64196822904SAndy Whitcroft 642feba16e2SMike Kravetz if (!nrg) { 643feba16e2SMike Kravetz spin_unlock(&resv->lock); 644feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 645feba16e2SMike Kravetz if (!nrg) 646feba16e2SMike Kravetz return -ENOMEM; 647feba16e2SMike Kravetz goto retry; 648feba16e2SMike Kravetz } 649feba16e2SMike Kravetz 650feba16e2SMike Kravetz del += t - f; 65179aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 65279aa925bSMike Kravetz resv, rg, t - f); 653feba16e2SMike Kravetz 654feba16e2SMike Kravetz /* New entry for end of split region */ 655feba16e2SMike Kravetz nrg->from = t; 656feba16e2SMike Kravetz nrg->to = rg->to; 657075a61d0SMina Almasry 658075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 659075a61d0SMina Almasry 660feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 661feba16e2SMike Kravetz 662feba16e2SMike Kravetz /* Original entry is trimmed */ 663feba16e2SMike Kravetz rg->to = f; 664feba16e2SMike Kravetz 665feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 666feba16e2SMike Kravetz nrg = NULL; 66796822904SAndy Whitcroft break; 668feba16e2SMike Kravetz } 669feba16e2SMike Kravetz 670feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 671feba16e2SMike Kravetz del += rg->to - rg->from; 672075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 673075a61d0SMina Almasry rg->to - rg->from); 67496822904SAndy Whitcroft list_del(&rg->link); 67596822904SAndy Whitcroft kfree(rg); 676feba16e2SMike Kravetz continue; 67796822904SAndy Whitcroft } 6787b24d861SDavidlohr Bueso 679feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 680075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 681075a61d0SMina Almasry t - rg->from); 682075a61d0SMina Almasry 68379aa925bSMike Kravetz del += t - rg->from; 68479aa925bSMike Kravetz rg->from = t; 68579aa925bSMike Kravetz } else { /* Trim end of region */ 686075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 687075a61d0SMina Almasry rg->to - f); 68879aa925bSMike Kravetz 68979aa925bSMike Kravetz del += rg->to - f; 69079aa925bSMike Kravetz rg->to = f; 691feba16e2SMike Kravetz } 692feba16e2SMike Kravetz } 693feba16e2SMike Kravetz 6947b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 695feba16e2SMike Kravetz kfree(nrg); 696feba16e2SMike Kravetz return del; 69796822904SAndy Whitcroft } 69896822904SAndy Whitcroft 6991dd308a7SMike Kravetz /* 700b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 701b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 702b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 703b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 704b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 705b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 706b5cec28dSMike Kravetz * counts. 707b5cec28dSMike Kravetz */ 70872e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 709b5cec28dSMike Kravetz { 710b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 711b5cec28dSMike Kravetz long rsv_adjust; 712b5cec28dSMike Kravetz 713b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 71472e2936cSzhong jiang if (rsv_adjust) { 715b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 716b5cec28dSMike Kravetz 717b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 718b5cec28dSMike Kravetz } 719b5cec28dSMike Kravetz } 720b5cec28dSMike Kravetz 721b5cec28dSMike Kravetz /* 7221dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7231dd308a7SMike Kravetz * that intersect with the range [f, t). 7241dd308a7SMike Kravetz */ 7251406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 72684afd99bSAndy Whitcroft { 7271406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 72884afd99bSAndy Whitcroft struct file_region *rg; 72984afd99bSAndy Whitcroft long chg = 0; 73084afd99bSAndy Whitcroft 7317b24d861SDavidlohr Bueso spin_lock(&resv->lock); 73284afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 73384afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 734f2135a4aSWang Sheng-Hui long seg_from; 735f2135a4aSWang Sheng-Hui long seg_to; 73684afd99bSAndy Whitcroft 73784afd99bSAndy Whitcroft if (rg->to <= f) 73884afd99bSAndy Whitcroft continue; 73984afd99bSAndy Whitcroft if (rg->from >= t) 74084afd99bSAndy Whitcroft break; 74184afd99bSAndy Whitcroft 74284afd99bSAndy Whitcroft seg_from = max(rg->from, f); 74384afd99bSAndy Whitcroft seg_to = min(rg->to, t); 74484afd99bSAndy Whitcroft 74584afd99bSAndy Whitcroft chg += seg_to - seg_from; 74684afd99bSAndy Whitcroft } 7477b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 74884afd99bSAndy Whitcroft 74984afd99bSAndy Whitcroft return chg; 75084afd99bSAndy Whitcroft } 75184afd99bSAndy Whitcroft 75296822904SAndy Whitcroft /* 753e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 754e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 755e7c4b0bfSAndy Whitcroft */ 756a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 757a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 758e7c4b0bfSAndy Whitcroft { 759a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 760a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 761e7c4b0bfSAndy Whitcroft } 762e7c4b0bfSAndy Whitcroft 7630fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 7640fe6e20bSNaoya Horiguchi unsigned long address) 7650fe6e20bSNaoya Horiguchi { 7660fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 7670fe6e20bSNaoya Horiguchi } 768dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 7690fe6e20bSNaoya Horiguchi 77084afd99bSAndy Whitcroft /* 77108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 77208fba699SMel Gorman * cases this will be same size as used by the page table entries. 77308fba699SMel Gorman */ 77408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 77508fba699SMel Gorman { 77605ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 77705ea8860SDan Williams return vma->vm_ops->pagesize(vma); 77808fba699SMel Gorman return PAGE_SIZE; 77908fba699SMel Gorman } 780f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 78108fba699SMel Gorman 78208fba699SMel Gorman /* 7833340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 7843340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 78509135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 78609135cc5SDan Williams * version of this symbol is required. 7873340289dSMel Gorman */ 78809135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 7893340289dSMel Gorman { 7903340289dSMel Gorman return vma_kernel_pagesize(vma); 7913340289dSMel Gorman } 7923340289dSMel Gorman 7933340289dSMel Gorman /* 79484afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 79584afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 79684afd99bSAndy Whitcroft * alignment. 79784afd99bSAndy Whitcroft */ 79884afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 79984afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 80004f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 80184afd99bSAndy Whitcroft 802a1e78772SMel Gorman /* 803a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 804a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 805a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 806a1e78772SMel Gorman * 807a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 808a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 809a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 810a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 81184afd99bSAndy Whitcroft * 81284afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 81384afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 81484afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 81584afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 81684afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 81784afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 81884afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 81984afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 820a1e78772SMel Gorman */ 821e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 822e7c4b0bfSAndy Whitcroft { 823e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 824e7c4b0bfSAndy Whitcroft } 825e7c4b0bfSAndy Whitcroft 826e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 827e7c4b0bfSAndy Whitcroft unsigned long value) 828e7c4b0bfSAndy Whitcroft { 829e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 830e7c4b0bfSAndy Whitcroft } 831e7c4b0bfSAndy Whitcroft 832e9fe92aeSMina Almasry static void 833e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 834e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 835e9fe92aeSMina Almasry struct hstate *h) 836e9fe92aeSMina Almasry { 837e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 838e9fe92aeSMina Almasry if (!h_cg || !h) { 839e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 840e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 841e9fe92aeSMina Almasry resv_map->css = NULL; 842e9fe92aeSMina Almasry } else { 843e9fe92aeSMina Almasry resv_map->reservation_counter = 844e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 845e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 846e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 847e9fe92aeSMina Almasry } 848e9fe92aeSMina Almasry #endif 849e9fe92aeSMina Almasry } 850e9fe92aeSMina Almasry 8519119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 85284afd99bSAndy Whitcroft { 85384afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8545e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8555e911373SMike Kravetz 8565e911373SMike Kravetz if (!resv_map || !rg) { 8575e911373SMike Kravetz kfree(resv_map); 8585e911373SMike Kravetz kfree(rg); 85984afd99bSAndy Whitcroft return NULL; 8605e911373SMike Kravetz } 86184afd99bSAndy Whitcroft 86284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 8637b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 86484afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 86584afd99bSAndy Whitcroft 8665e911373SMike Kravetz resv_map->adds_in_progress = 0; 867e9fe92aeSMina Almasry /* 868e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 869e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 870e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 871e9fe92aeSMina Almasry * reservations are to be un-charged from here. 872e9fe92aeSMina Almasry */ 873e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 8745e911373SMike Kravetz 8755e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 8765e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 8775e911373SMike Kravetz resv_map->region_cache_count = 1; 8785e911373SMike Kravetz 87984afd99bSAndy Whitcroft return resv_map; 88084afd99bSAndy Whitcroft } 88184afd99bSAndy Whitcroft 8829119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 88384afd99bSAndy Whitcroft { 88484afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 8855e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 8865e911373SMike Kravetz struct file_region *rg, *trg; 88784afd99bSAndy Whitcroft 88884afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 889feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 8905e911373SMike Kravetz 8915e911373SMike Kravetz /* ... and any entries left in the cache */ 8925e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 8935e911373SMike Kravetz list_del(&rg->link); 8945e911373SMike Kravetz kfree(rg); 8955e911373SMike Kravetz } 8965e911373SMike Kravetz 8975e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 8985e911373SMike Kravetz 89984afd99bSAndy Whitcroft kfree(resv_map); 90084afd99bSAndy Whitcroft } 90184afd99bSAndy Whitcroft 9024e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9034e35f483SJoonsoo Kim { 904f27a5136SMike Kravetz /* 905f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 906f27a5136SMike Kravetz * address space within the inode. This original address space 907f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 908f27a5136SMike Kravetz * address space embedded within the inode. 909f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 910f27a5136SMike Kravetz * this may not be true for device special inodes. 911f27a5136SMike Kravetz */ 912f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9134e35f483SJoonsoo Kim } 9144e35f483SJoonsoo Kim 91584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 916a1e78772SMel Gorman { 91781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9184e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9194e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9204e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9214e35f483SJoonsoo Kim 9224e35f483SJoonsoo Kim return inode_resv_map(inode); 9234e35f483SJoonsoo Kim 9244e35f483SJoonsoo Kim } else { 92584afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 92684afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9274e35f483SJoonsoo Kim } 928a1e78772SMel Gorman } 929a1e78772SMel Gorman 93084afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 931a1e78772SMel Gorman { 93281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 93381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 934a1e78772SMel Gorman 93584afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 93684afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 93704f2cbe3SMel Gorman } 93804f2cbe3SMel Gorman 93904f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 94004f2cbe3SMel Gorman { 94181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 94281d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 943e7c4b0bfSAndy Whitcroft 944e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 94504f2cbe3SMel Gorman } 94604f2cbe3SMel Gorman 94704f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 94804f2cbe3SMel Gorman { 94981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 950e7c4b0bfSAndy Whitcroft 951e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 952a1e78772SMel Gorman } 953a1e78772SMel Gorman 95404f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 955a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 956a1e78772SMel Gorman { 95781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 958f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 959a1e78772SMel Gorman vma->vm_private_data = (void *)0; 960a1e78772SMel Gorman } 961a1e78772SMel Gorman 962a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 963559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 964a1e78772SMel Gorman { 965af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 966af0ed73eSJoonsoo Kim /* 967af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 968af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 969af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 970af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 971af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 972af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 973af0ed73eSJoonsoo Kim * properly, so add work-around here. 974af0ed73eSJoonsoo Kim */ 975af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 976559ec2f8SNicholas Krause return true; 977af0ed73eSJoonsoo Kim else 978559ec2f8SNicholas Krause return false; 979af0ed73eSJoonsoo Kim } 980a63884e9SJoonsoo Kim 981a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 9821fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 9831fb1b0e9SMike Kravetz /* 9841fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 9851fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 9861fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 9877c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 9881fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 9891fb1b0e9SMike Kravetz */ 9901fb1b0e9SMike Kravetz if (chg) 9911fb1b0e9SMike Kravetz return false; 9921fb1b0e9SMike Kravetz else 993559ec2f8SNicholas Krause return true; 9941fb1b0e9SMike Kravetz } 995a63884e9SJoonsoo Kim 996a63884e9SJoonsoo Kim /* 997a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 998a63884e9SJoonsoo Kim * private mappings. 999a63884e9SJoonsoo Kim */ 100067961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 100167961f9dSMike Kravetz /* 100267961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 100367961f9dSMike Kravetz * could have been performed on the private mapping. 100467961f9dSMike Kravetz * Examine the value of chg to determine if reserves 100567961f9dSMike Kravetz * actually exist or were previously consumed. 100667961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 100767961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 100867961f9dSMike Kravetz * private mappings has different (opposite) semantics 100967961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 101067961f9dSMike Kravetz * has already taken this difference in semantics into 101167961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 101267961f9dSMike Kravetz * as in the shared case above. Code could easily be 101367961f9dSMike Kravetz * combined, but keeping it separate draws attention to 101467961f9dSMike Kravetz * subtle differences. 101567961f9dSMike Kravetz */ 101667961f9dSMike Kravetz if (chg) 101767961f9dSMike Kravetz return false; 101867961f9dSMike Kravetz else 1019559ec2f8SNicholas Krause return true; 102067961f9dSMike Kravetz } 1021a63884e9SJoonsoo Kim 1022559ec2f8SNicholas Krause return false; 1023a1e78772SMel Gorman } 1024a1e78772SMel Gorman 1025a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10261da177e4SLinus Torvalds { 10271da177e4SLinus Torvalds int nid = page_to_nid(page); 10280edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1029a5516438SAndi Kleen h->free_huge_pages++; 1030a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10311da177e4SLinus Torvalds } 10321da177e4SLinus Torvalds 103394310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1034bf50bab2SNaoya Horiguchi { 1035bf50bab2SNaoya Horiguchi struct page *page; 1036bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1037bf50bab2SNaoya Horiguchi 1038bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1039bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1040bbe88753SJoonsoo Kim continue; 1041bbe88753SJoonsoo Kim 10426664bfc8SWei Yang if (PageHWPoison(page)) 10436664bfc8SWei Yang continue; 1044bbe88753SJoonsoo Kim 10450edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1046a9869b83SNaoya Horiguchi set_page_refcounted(page); 1047bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1048bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1049bf50bab2SNaoya Horiguchi return page; 1050bf50bab2SNaoya Horiguchi } 1051bf50bab2SNaoya Horiguchi 10526664bfc8SWei Yang return NULL; 10536664bfc8SWei Yang } 10546664bfc8SWei Yang 10553e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10563e59fcb0SMichal Hocko nodemask_t *nmask) 105794310cbcSAnshuman Khandual { 10583e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 10593e59fcb0SMichal Hocko struct zonelist *zonelist; 10603e59fcb0SMichal Hocko struct zone *zone; 10613e59fcb0SMichal Hocko struct zoneref *z; 106298fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 10633e59fcb0SMichal Hocko 10643e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 10653e59fcb0SMichal Hocko 10663e59fcb0SMichal Hocko retry_cpuset: 10673e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 10683e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 106994310cbcSAnshuman Khandual struct page *page; 107094310cbcSAnshuman Khandual 10713e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 10723e59fcb0SMichal Hocko continue; 10733e59fcb0SMichal Hocko /* 10743e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 10753e59fcb0SMichal Hocko * zone aware 10763e59fcb0SMichal Hocko */ 10773e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 10783e59fcb0SMichal Hocko continue; 10793e59fcb0SMichal Hocko node = zone_to_nid(zone); 108094310cbcSAnshuman Khandual 108194310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 108294310cbcSAnshuman Khandual if (page) 108394310cbcSAnshuman Khandual return page; 108494310cbcSAnshuman Khandual } 10853e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 10863e59fcb0SMichal Hocko goto retry_cpuset; 10873e59fcb0SMichal Hocko 108894310cbcSAnshuman Khandual return NULL; 108994310cbcSAnshuman Khandual } 109094310cbcSAnshuman Khandual 1091a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1092a5516438SAndi Kleen struct vm_area_struct *vma, 1093af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1094af0ed73eSJoonsoo Kim long chg) 10951da177e4SLinus Torvalds { 10963e59fcb0SMichal Hocko struct page *page; 1097480eccf9SLee Schermerhorn struct mempolicy *mpol; 109804ec6264SVlastimil Babka gfp_t gfp_mask; 10993e59fcb0SMichal Hocko nodemask_t *nodemask; 110004ec6264SVlastimil Babka int nid; 11011da177e4SLinus Torvalds 1102a1e78772SMel Gorman /* 1103a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1104a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1105a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1106a1e78772SMel Gorman */ 1107af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1108a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1109c0ff7453SMiao Xie goto err; 1110a1e78772SMel Gorman 111104f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1112a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11136eab04a8SJustin P. Mattock goto err; 111404f2cbe3SMel Gorman 111504ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 111604ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11173e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11183e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 111907443a85SJoonsoo Kim SetPagePrivate(page); 1120a63884e9SJoonsoo Kim h->resv_huge_pages--; 1121bf50bab2SNaoya Horiguchi } 1122cc9a6c87SMel Gorman 1123cc9a6c87SMel Gorman mpol_cond_put(mpol); 1124cc9a6c87SMel Gorman return page; 1125cc9a6c87SMel Gorman 1126c0ff7453SMiao Xie err: 1127cc9a6c87SMel Gorman return NULL; 11281da177e4SLinus Torvalds } 11291da177e4SLinus Torvalds 11301cac6f2cSLuiz Capitulino /* 11311cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11321cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11331cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11341cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11351cac6f2cSLuiz Capitulino * node for alloc or free. 11361cac6f2cSLuiz Capitulino */ 11371cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11381cac6f2cSLuiz Capitulino { 11390edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11401cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11411cac6f2cSLuiz Capitulino 11421cac6f2cSLuiz Capitulino return nid; 11431cac6f2cSLuiz Capitulino } 11441cac6f2cSLuiz Capitulino 11451cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11461cac6f2cSLuiz Capitulino { 11471cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11481cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11491cac6f2cSLuiz Capitulino return nid; 11501cac6f2cSLuiz Capitulino } 11511cac6f2cSLuiz Capitulino 11521cac6f2cSLuiz Capitulino /* 11531cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11541cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11551cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11561cac6f2cSLuiz Capitulino * mask. 11571cac6f2cSLuiz Capitulino */ 11581cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 11591cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 11601cac6f2cSLuiz Capitulino { 11611cac6f2cSLuiz Capitulino int nid; 11621cac6f2cSLuiz Capitulino 11631cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11641cac6f2cSLuiz Capitulino 11651cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 11661cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 11671cac6f2cSLuiz Capitulino 11681cac6f2cSLuiz Capitulino return nid; 11691cac6f2cSLuiz Capitulino } 11701cac6f2cSLuiz Capitulino 11711cac6f2cSLuiz Capitulino /* 11721cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 11731cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 11741cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 11751cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 11761cac6f2cSLuiz Capitulino */ 11771cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 11781cac6f2cSLuiz Capitulino { 11791cac6f2cSLuiz Capitulino int nid; 11801cac6f2cSLuiz Capitulino 11811cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11821cac6f2cSLuiz Capitulino 11831cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 11841cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 11851cac6f2cSLuiz Capitulino 11861cac6f2cSLuiz Capitulino return nid; 11871cac6f2cSLuiz Capitulino } 11881cac6f2cSLuiz Capitulino 11891cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 11901cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 11911cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 11921cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 11931cac6f2cSLuiz Capitulino nr_nodes--) 11941cac6f2cSLuiz Capitulino 11951cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 11961cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 11971cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 11981cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 11991cac6f2cSLuiz Capitulino nr_nodes--) 12001cac6f2cSLuiz Capitulino 1201e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1202944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1203d00181b9SKirill A. Shutemov unsigned int order) 1204944d9fecSLuiz Capitulino { 1205944d9fecSLuiz Capitulino int i; 1206944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1207944d9fecSLuiz Capitulino struct page *p = page + 1; 1208944d9fecSLuiz Capitulino 1209c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 121047e29d32SJohn Hubbard if (hpage_pincount_available(page)) 121147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 121247e29d32SJohn Hubbard 1213944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12141d798ca3SKirill A. Shutemov clear_compound_head(p); 1215944d9fecSLuiz Capitulino set_page_refcounted(p); 1216944d9fecSLuiz Capitulino } 1217944d9fecSLuiz Capitulino 1218944d9fecSLuiz Capitulino set_compound_order(page, 0); 1219ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1220944d9fecSLuiz Capitulino __ClearPageHead(page); 1221944d9fecSLuiz Capitulino } 1222944d9fecSLuiz Capitulino 1223d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1224944d9fecSLuiz Capitulino { 1225cf11e85fSRoman Gushchin /* 1226cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1227cf11e85fSRoman Gushchin * cma_release() returns false. 1228cf11e85fSRoman Gushchin */ 1229dbda8feaSBarry Song #ifdef CONFIG_CMA 1230dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1231cf11e85fSRoman Gushchin return; 1232dbda8feaSBarry Song #endif 1233cf11e85fSRoman Gushchin 1234944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1235944d9fecSLuiz Capitulino } 1236944d9fecSLuiz Capitulino 12374eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1238d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1239d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1240944d9fecSLuiz Capitulino { 12415e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1242953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1243953f064aSLi Xinhai nid = numa_mem_id(); 1244944d9fecSLuiz Capitulino 1245dbda8feaSBarry Song #ifdef CONFIG_CMA 1246dbda8feaSBarry Song { 1247cf11e85fSRoman Gushchin struct page *page; 1248cf11e85fSRoman Gushchin int node; 1249cf11e85fSRoman Gushchin 1250953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1251953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1252953f064aSLi Xinhai huge_page_order(h), true); 1253953f064aSLi Xinhai if (page) 1254953f064aSLi Xinhai return page; 1255953f064aSLi Xinhai } 1256953f064aSLi Xinhai 1257953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1258cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1259953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1260cf11e85fSRoman Gushchin continue; 1261cf11e85fSRoman Gushchin 1262cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1263cf11e85fSRoman Gushchin huge_page_order(h), true); 1264cf11e85fSRoman Gushchin if (page) 1265cf11e85fSRoman Gushchin return page; 1266cf11e85fSRoman Gushchin } 1267cf11e85fSRoman Gushchin } 1268953f064aSLi Xinhai } 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); 15039dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15041adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 15052f37511cSWei Yang spin_lock(&hugetlb_lock); 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 and lock address space (mapping) in write mode. 1573c0d0381aSMike Kravetz * 1574336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1575336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1576336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1577c0d0381aSMike Kravetz */ 1578c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1579c0d0381aSMike Kravetz { 1580336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1581c0d0381aSMike Kravetz 1582c0d0381aSMike Kravetz if (!mapping) 1583c0d0381aSMike Kravetz return mapping; 1584c0d0381aSMike Kravetz 1585c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1586c0d0381aSMike Kravetz return mapping; 1587c0d0381aSMike Kravetz 1588c0d0381aSMike Kravetz return NULL; 1589c0d0381aSMike Kravetz } 1590c0d0381aSMike Kravetz 159113d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 159213d60f4bSZhang Yi { 159313d60f4bSZhang Yi struct page *page_head = compound_head(page); 159413d60f4bSZhang Yi pgoff_t index = page_index(page_head); 159513d60f4bSZhang Yi unsigned long compound_idx; 159613d60f4bSZhang Yi 159713d60f4bSZhang Yi if (!PageHuge(page_head)) 159813d60f4bSZhang Yi return page_index(page); 159913d60f4bSZhang Yi 160013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 160113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 160213d60f4bSZhang Yi else 160313d60f4bSZhang Yi compound_idx = page - page_head; 160413d60f4bSZhang Yi 160513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 160613d60f4bSZhang Yi } 160713d60f4bSZhang Yi 16080c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1609f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1610f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16111da177e4SLinus Torvalds { 1612af0fb9dfSMichal Hocko int order = huge_page_order(h); 16131da177e4SLinus Torvalds struct page *page; 1614f60858f9SMike Kravetz bool alloc_try_hard = true; 1615f96efd58SJoe Jin 1616f60858f9SMike Kravetz /* 1617f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1618f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1619f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1620f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1621f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1622f60858f9SMike Kravetz */ 1623f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1624f60858f9SMike Kravetz alloc_try_hard = false; 1625f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1626f60858f9SMike Kravetz if (alloc_try_hard) 1627f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1628af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1629af0fb9dfSMichal Hocko nid = numa_mem_id(); 1630af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1631af0fb9dfSMichal Hocko if (page) 1632af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1633af0fb9dfSMichal Hocko else 1634af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 163563b4613cSNishanth Aravamudan 1636f60858f9SMike Kravetz /* 1637f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1638f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1639f60858f9SMike Kravetz * normal 'try hard' allocations. 1640f60858f9SMike Kravetz */ 1641f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1642f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1643f60858f9SMike Kravetz 1644f60858f9SMike Kravetz /* 1645f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1646f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1647f60858f9SMike Kravetz * overall state change. 1648f60858f9SMike Kravetz */ 1649f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1650f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1651f60858f9SMike Kravetz 165263b4613cSNishanth Aravamudan return page; 165363b4613cSNishanth Aravamudan } 165463b4613cSNishanth Aravamudan 1655af0fb9dfSMichal Hocko /* 16560c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16570c397daeSMichal Hocko * should use this function to get new hugetlb pages 16580c397daeSMichal Hocko */ 16590c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1660f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1661f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16620c397daeSMichal Hocko { 16630c397daeSMichal Hocko struct page *page; 16640c397daeSMichal Hocko 16650c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16660c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16670c397daeSMichal Hocko else 16680c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1669f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16700c397daeSMichal Hocko if (!page) 16710c397daeSMichal Hocko return NULL; 16720c397daeSMichal Hocko 16730c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16740c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16750c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16760c397daeSMichal Hocko 16770c397daeSMichal Hocko return page; 16780c397daeSMichal Hocko } 16790c397daeSMichal Hocko 16800c397daeSMichal Hocko /* 1681af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1682af0fb9dfSMichal Hocko * manner. 1683af0fb9dfSMichal Hocko */ 1684f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1685f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1686b2261026SJoonsoo Kim { 1687b2261026SJoonsoo Kim struct page *page; 1688b2261026SJoonsoo Kim int nr_nodes, node; 1689af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1690b2261026SJoonsoo Kim 1691b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1692f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1693f60858f9SMike Kravetz node_alloc_noretry); 1694af0fb9dfSMichal Hocko if (page) 1695b2261026SJoonsoo Kim break; 1696b2261026SJoonsoo Kim } 1697b2261026SJoonsoo Kim 1698af0fb9dfSMichal Hocko if (!page) 1699af0fb9dfSMichal Hocko return 0; 1700b2261026SJoonsoo Kim 1701af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1702af0fb9dfSMichal Hocko 1703af0fb9dfSMichal Hocko return 1; 1704b2261026SJoonsoo Kim } 1705b2261026SJoonsoo Kim 1706e8c5c824SLee Schermerhorn /* 1707e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1708e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1709e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1710e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1711e8c5c824SLee Schermerhorn */ 17126ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17136ae11b27SLee Schermerhorn bool acct_surplus) 1714e8c5c824SLee Schermerhorn { 1715b2261026SJoonsoo Kim int nr_nodes, node; 1716e8c5c824SLee Schermerhorn int ret = 0; 1717e8c5c824SLee Schermerhorn 1718b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1719685f3457SLee Schermerhorn /* 1720685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1721685f3457SLee Schermerhorn * nodes with surplus pages. 1722685f3457SLee Schermerhorn */ 1723b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1724b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1725e8c5c824SLee Schermerhorn struct page *page = 1726b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1727e8c5c824SLee Schermerhorn struct page, lru); 1728e8c5c824SLee Schermerhorn list_del(&page->lru); 1729e8c5c824SLee Schermerhorn h->free_huge_pages--; 1730b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1731685f3457SLee Schermerhorn if (acct_surplus) { 1732685f3457SLee Schermerhorn h->surplus_huge_pages--; 1733b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1734685f3457SLee Schermerhorn } 1735e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1736e8c5c824SLee Schermerhorn ret = 1; 17379a76db09SLee Schermerhorn break; 1738e8c5c824SLee Schermerhorn } 1739b2261026SJoonsoo Kim } 1740e8c5c824SLee Schermerhorn 1741e8c5c824SLee Schermerhorn return ret; 1742e8c5c824SLee Schermerhorn } 1743e8c5c824SLee Schermerhorn 1744c8721bbbSNaoya Horiguchi /* 1745c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1746faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1747faf53defSNaoya Horiguchi * This function returns values like below: 1748faf53defSNaoya Horiguchi * 1749faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1750faf53defSNaoya Horiguchi * (allocated or reserved.) 1751faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1752faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1753c8721bbbSNaoya Horiguchi */ 1754c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1755c8721bbbSNaoya Horiguchi { 17566bc9b564SNaoya Horiguchi int rc = -EBUSY; 1757082d5b6bSGerald Schaefer 1758faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1759faf53defSNaoya Horiguchi if (!PageHuge(page)) 1760faf53defSNaoya Horiguchi return 0; 1761faf53defSNaoya Horiguchi 1762c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1763faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1764faf53defSNaoya Horiguchi rc = 0; 1765faf53defSNaoya Horiguchi goto out; 1766faf53defSNaoya Horiguchi } 1767faf53defSNaoya Horiguchi 1768faf53defSNaoya Horiguchi if (!page_count(page)) { 17692247bb33SGerald Schaefer struct page *head = compound_head(page); 17702247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17712247bb33SGerald Schaefer int nid = page_to_nid(head); 17726bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1773082d5b6bSGerald Schaefer goto out; 1774c3114a84SAnshuman Khandual /* 1775c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1776c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1777c3114a84SAnshuman Khandual */ 1778c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1779c3114a84SAnshuman Khandual SetPageHWPoison(page); 1780c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1781c3114a84SAnshuman Khandual } 17822247bb33SGerald Schaefer list_del(&head->lru); 1783c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1784c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1785c1470b33Szhong jiang h->max_huge_pages--; 17862247bb33SGerald Schaefer update_and_free_page(h, head); 17876bc9b564SNaoya Horiguchi rc = 0; 1788c8721bbbSNaoya Horiguchi } 1789082d5b6bSGerald Schaefer out: 1790c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1791082d5b6bSGerald Schaefer return rc; 1792c8721bbbSNaoya Horiguchi } 1793c8721bbbSNaoya Horiguchi 1794c8721bbbSNaoya Horiguchi /* 1795c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1796c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 17972247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 17982247bb33SGerald Schaefer * part of it lies within the given range. 1799082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1800082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1801c8721bbbSNaoya Horiguchi */ 1802082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1803c8721bbbSNaoya Horiguchi { 1804c8721bbbSNaoya Horiguchi unsigned long pfn; 1805eb03aa00SGerald Schaefer struct page *page; 1806082d5b6bSGerald Schaefer int rc = 0; 1807c8721bbbSNaoya Horiguchi 1808d0177639SLi Zhong if (!hugepages_supported()) 1809082d5b6bSGerald Schaefer return rc; 1810d0177639SLi Zhong 1811eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1812eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1813eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1814eb03aa00SGerald Schaefer if (rc) 1815082d5b6bSGerald Schaefer break; 1816eb03aa00SGerald Schaefer } 1817082d5b6bSGerald Schaefer 1818082d5b6bSGerald Schaefer return rc; 1819c8721bbbSNaoya Horiguchi } 1820c8721bbbSNaoya Horiguchi 1821ab5ac90aSMichal Hocko /* 1822ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1823ab5ac90aSMichal Hocko */ 18240c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1825aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18267893d1d5SAdam Litke { 18279980d744SMichal Hocko struct page *page = NULL; 18287893d1d5SAdam Litke 1829bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1830aa888a74SAndi Kleen return NULL; 1831aa888a74SAndi Kleen 1832d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 18339980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18349980d744SMichal Hocko goto out_unlock; 1835d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1836d1c3fb1fSNishanth Aravamudan 1837f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18389980d744SMichal Hocko if (!page) 18390c397daeSMichal Hocko return NULL; 1840d1c3fb1fSNishanth Aravamudan 18417893d1d5SAdam Litke spin_lock(&hugetlb_lock); 18429980d744SMichal Hocko /* 18439980d744SMichal Hocko * We could have raced with the pool size change. 18449980d744SMichal Hocko * Double check that and simply deallocate the new page 18459980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18469980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18479980d744SMichal Hocko * codeflow 18489980d744SMichal Hocko */ 18499980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18509980d744SMichal Hocko SetPageHugeTemporary(page); 18512bf753e6SKai Shen spin_unlock(&hugetlb_lock); 18529980d744SMichal Hocko put_page(page); 18532bf753e6SKai Shen return NULL; 18549980d744SMichal Hocko } else { 18559980d744SMichal Hocko h->surplus_huge_pages++; 18564704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18577893d1d5SAdam Litke } 18589980d744SMichal Hocko 18599980d744SMichal Hocko out_unlock: 1860d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 18617893d1d5SAdam Litke 18627893d1d5SAdam Litke return page; 18637893d1d5SAdam Litke } 18647893d1d5SAdam Litke 1865bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1866ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1867ab5ac90aSMichal Hocko { 1868ab5ac90aSMichal Hocko struct page *page; 1869ab5ac90aSMichal Hocko 1870ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1871ab5ac90aSMichal Hocko return NULL; 1872ab5ac90aSMichal Hocko 1873f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1874ab5ac90aSMichal Hocko if (!page) 1875ab5ac90aSMichal Hocko return NULL; 1876ab5ac90aSMichal Hocko 1877ab5ac90aSMichal Hocko /* 1878ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1879ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1880ab5ac90aSMichal Hocko */ 1881ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1882ab5ac90aSMichal Hocko 1883ab5ac90aSMichal Hocko return page; 1884ab5ac90aSMichal Hocko } 1885ab5ac90aSMichal Hocko 1886e4e574b7SAdam Litke /* 1887099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1888099730d6SDave Hansen */ 1889e0ec90eeSDave Hansen static 18900c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1891099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1892099730d6SDave Hansen { 1893aaf14e40SMichal Hocko struct page *page; 1894aaf14e40SMichal Hocko struct mempolicy *mpol; 1895aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1896aaf14e40SMichal Hocko int nid; 1897aaf14e40SMichal Hocko nodemask_t *nodemask; 1898aaf14e40SMichal Hocko 1899aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19000c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1901aaf14e40SMichal Hocko mpol_cond_put(mpol); 1902aaf14e40SMichal Hocko 1903aaf14e40SMichal Hocko return page; 1904099730d6SDave Hansen } 1905099730d6SDave Hansen 1906ab5ac90aSMichal Hocko /* page migration callback function */ 19073e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1908d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19094db9b2efSMichal Hocko { 19104db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 19114db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19123e59fcb0SMichal Hocko struct page *page; 19133e59fcb0SMichal Hocko 19143e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19153e59fcb0SMichal Hocko if (page) { 19163e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 19173e59fcb0SMichal Hocko return page; 19184db9b2efSMichal Hocko } 19194db9b2efSMichal Hocko } 19204db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 19214db9b2efSMichal Hocko 19220c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19234db9b2efSMichal Hocko } 19244db9b2efSMichal Hocko 1925ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1926389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1927389c8178SMichal Hocko unsigned long address) 1928ebd63723SMichal Hocko { 1929ebd63723SMichal Hocko struct mempolicy *mpol; 1930ebd63723SMichal Hocko nodemask_t *nodemask; 1931ebd63723SMichal Hocko struct page *page; 1932ebd63723SMichal Hocko gfp_t gfp_mask; 1933ebd63723SMichal Hocko int node; 1934ebd63723SMichal Hocko 1935ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1936ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1937d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1938ebd63723SMichal Hocko mpol_cond_put(mpol); 1939ebd63723SMichal Hocko 1940ebd63723SMichal Hocko return page; 1941ebd63723SMichal Hocko } 1942ebd63723SMichal Hocko 1943bf50bab2SNaoya Horiguchi /* 194425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1945e4e574b7SAdam Litke * of size 'delta'. 1946e4e574b7SAdam Litke */ 19470a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19481b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1949e4e574b7SAdam Litke { 1950e4e574b7SAdam Litke struct list_head surplus_list; 1951e4e574b7SAdam Litke struct page *page, *tmp; 19520a4f3d1bSLiu Xiang int ret; 19530a4f3d1bSLiu Xiang long i; 19540a4f3d1bSLiu Xiang long needed, allocated; 195528073b02SHillf Danton bool alloc_ok = true; 1956e4e574b7SAdam Litke 1957a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1958ac09b3a1SAdam Litke if (needed <= 0) { 1959a5516438SAndi Kleen h->resv_huge_pages += delta; 1960e4e574b7SAdam Litke return 0; 1961ac09b3a1SAdam Litke } 1962e4e574b7SAdam Litke 1963e4e574b7SAdam Litke allocated = 0; 1964e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1965e4e574b7SAdam Litke 1966e4e574b7SAdam Litke ret = -ENOMEM; 1967e4e574b7SAdam Litke retry: 1968e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1969e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19700c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1971aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 197228073b02SHillf Danton if (!page) { 197328073b02SHillf Danton alloc_ok = false; 197428073b02SHillf Danton break; 197528073b02SHillf Danton } 1976e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 197769ed779aSDavid Rientjes cond_resched(); 1978e4e574b7SAdam Litke } 197928073b02SHillf Danton allocated += i; 1980e4e574b7SAdam Litke 1981e4e574b7SAdam Litke /* 1982e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1983e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1984e4e574b7SAdam Litke */ 1985e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1986a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1987a5516438SAndi Kleen (h->free_huge_pages + allocated); 198828073b02SHillf Danton if (needed > 0) { 198928073b02SHillf Danton if (alloc_ok) 1990e4e574b7SAdam Litke goto retry; 199128073b02SHillf Danton /* 199228073b02SHillf Danton * We were not able to allocate enough pages to 199328073b02SHillf Danton * satisfy the entire reservation so we free what 199428073b02SHillf Danton * we've allocated so far. 199528073b02SHillf Danton */ 199628073b02SHillf Danton goto free; 199728073b02SHillf Danton } 1998e4e574b7SAdam Litke /* 1999e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 200025985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2001e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2002ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2003ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2004ac09b3a1SAdam Litke * before they are reserved. 2005e4e574b7SAdam Litke */ 2006e4e574b7SAdam Litke needed += allocated; 2007a5516438SAndi Kleen h->resv_huge_pages += delta; 2008e4e574b7SAdam Litke ret = 0; 2009a9869b83SNaoya Horiguchi 201019fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 201119fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 201219fc3f0aSAdam Litke if ((--needed) < 0) 201319fc3f0aSAdam Litke break; 2014a9869b83SNaoya Horiguchi /* 2015a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2016a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2017a9869b83SNaoya Horiguchi */ 2018e5dfacebSHui Su VM_BUG_ON_PAGE(!put_page_testzero(page), page); 2019a5516438SAndi Kleen enqueue_huge_page(h, page); 202019fc3f0aSAdam Litke } 202128073b02SHillf Danton free: 2022b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 202319fc3f0aSAdam Litke 202419fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2025c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2026a9869b83SNaoya Horiguchi put_page(page); 202719fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2028e4e574b7SAdam Litke 2029e4e574b7SAdam Litke return ret; 2030e4e574b7SAdam Litke } 2031e4e574b7SAdam Litke 2032e4e574b7SAdam Litke /* 2033e5bbc8a6SMike Kravetz * This routine has two main purposes: 2034e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2035e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2036e5bbc8a6SMike Kravetz * to the associated reservation map. 2037e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2038e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2039e5bbc8a6SMike Kravetz * 2040e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2041e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2042e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2043e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2044e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2045e4e574b7SAdam Litke */ 2046a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2047a5516438SAndi Kleen unsigned long unused_resv_pages) 2048e4e574b7SAdam Litke { 2049e4e574b7SAdam Litke unsigned long nr_pages; 2050e4e574b7SAdam Litke 2051aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2052bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2053e5bbc8a6SMike Kravetz goto out; 2054aa888a74SAndi Kleen 2055e5bbc8a6SMike Kravetz /* 2056e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2057e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2058e5bbc8a6SMike Kravetz */ 2059a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2060e4e574b7SAdam Litke 2061685f3457SLee Schermerhorn /* 2062685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20639b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20649b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20659b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 20669e7ee400SRandy Dunlap * free_pool_huge_page() will balance the freed pages across the 20679b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2068e5bbc8a6SMike Kravetz * 2069e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2070e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2071e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2072685f3457SLee Schermerhorn */ 2073685f3457SLee Schermerhorn while (nr_pages--) { 2074e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2075e5bbc8a6SMike Kravetz unused_resv_pages--; 20768cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2077e5bbc8a6SMike Kravetz goto out; 20787848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2079e4e574b7SAdam Litke } 2080e5bbc8a6SMike Kravetz 2081e5bbc8a6SMike Kravetz out: 2082e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2083e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2084e4e574b7SAdam Litke } 2085e4e574b7SAdam Litke 20865e911373SMike Kravetz 2087c37f9fb1SAndy Whitcroft /* 2088feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 20895e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2090cf3ad20bSMike Kravetz * 2091cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2092cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2093cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2094cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2095cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2096feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2097feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2098feba16e2SMike Kravetz * is called in such cases. 2099cf3ad20bSMike Kravetz * 2100cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2101cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2102cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2103cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2104cf3ad20bSMike Kravetz * take appropriate action. 210596b96a96SMike Kravetz * 210696b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 210796b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 210896b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 210996b96a96SMike Kravetz * reservation exists. 2110c37f9fb1SAndy Whitcroft */ 21115e911373SMike Kravetz enum vma_resv_mode { 21125e911373SMike Kravetz VMA_NEEDS_RESV, 21135e911373SMike Kravetz VMA_COMMIT_RESV, 2114feba16e2SMike Kravetz VMA_END_RESV, 211596b96a96SMike Kravetz VMA_ADD_RESV, 21165e911373SMike Kravetz }; 2117cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2118cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21195e911373SMike Kravetz enum vma_resv_mode mode) 2120c37f9fb1SAndy Whitcroft { 21214e35f483SJoonsoo Kim struct resv_map *resv; 21224e35f483SJoonsoo Kim pgoff_t idx; 2123cf3ad20bSMike Kravetz long ret; 21240db9d74eSMina Almasry long dummy_out_regions_needed; 2125c37f9fb1SAndy Whitcroft 21264e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21274e35f483SJoonsoo Kim if (!resv) 2128c37f9fb1SAndy Whitcroft return 1; 2129c37f9fb1SAndy Whitcroft 21304e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21315e911373SMike Kravetz switch (mode) { 21325e911373SMike Kravetz case VMA_NEEDS_RESV: 21330db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21340db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21350db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21360db9d74eSMina Almasry * ever require 1 region. 21370db9d74eSMina Almasry */ 21380db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21395e911373SMike Kravetz break; 21405e911373SMike Kravetz case VMA_COMMIT_RESV: 2141075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21420db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21430db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21445e911373SMike Kravetz break; 2145feba16e2SMike Kravetz case VMA_END_RESV: 21460db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21475e911373SMike Kravetz ret = 0; 21485e911373SMike Kravetz break; 214996b96a96SMike Kravetz case VMA_ADD_RESV: 21500db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2151075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21520db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21530db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21540db9d74eSMina Almasry } else { 21550db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 215696b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 215796b96a96SMike Kravetz } 215896b96a96SMike Kravetz break; 21595e911373SMike Kravetz default: 21605e911373SMike Kravetz BUG(); 21615e911373SMike Kravetz } 216284afd99bSAndy Whitcroft 21634e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2164cf3ad20bSMike Kravetz return ret; 216567961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 216667961f9dSMike Kravetz /* 216767961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 216867961f9dSMike Kravetz * However, a file associated with mapping could have been 216967961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 217067961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 217167961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 217267961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 217367961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 217467961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 217567961f9dSMike Kravetz * reservation has already been consumed. As a result, the 217667961f9dSMike Kravetz * return value of this routine is the opposite of the 217767961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 217867961f9dSMike Kravetz */ 217967961f9dSMike Kravetz if (ret) 218067961f9dSMike Kravetz return 0; 218167961f9dSMike Kravetz else 218267961f9dSMike Kravetz return 1; 218367961f9dSMike Kravetz } 21844e35f483SJoonsoo Kim else 2185cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 218684afd99bSAndy Whitcroft } 2187cf3ad20bSMike Kravetz 2188cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2189a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2190c37f9fb1SAndy Whitcroft { 21915e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2192cf3ad20bSMike Kravetz } 2193c37f9fb1SAndy Whitcroft 2194cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2195cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2196cf3ad20bSMike Kravetz { 21975e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 21985e911373SMike Kravetz } 21995e911373SMike Kravetz 2200feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22015e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22025e911373SMike Kravetz { 2203feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2204c37f9fb1SAndy Whitcroft } 2205c37f9fb1SAndy Whitcroft 220696b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 220796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 220896b96a96SMike Kravetz { 220996b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 221096b96a96SMike Kravetz } 221196b96a96SMike Kravetz 221296b96a96SMike Kravetz /* 221396b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 221496b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 221596b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 221696b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 221796b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 221896b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 221996b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 222096b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 222196b96a96SMike Kravetz * to be made by free_huge_page. 222296b96a96SMike Kravetz */ 222396b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 222496b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 222596b96a96SMike Kravetz struct page *page) 222696b96a96SMike Kravetz { 222796b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 222896b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 222996b96a96SMike Kravetz 223096b96a96SMike Kravetz if (unlikely(rc < 0)) { 223196b96a96SMike Kravetz /* 223296b96a96SMike Kravetz * Rare out of memory condition in reserve map 223396b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 223496b96a96SMike Kravetz * global reserve count will not be incremented 223596b96a96SMike Kravetz * by free_huge_page. This will make it appear 223696b96a96SMike Kravetz * as though the reservation for this page was 223796b96a96SMike Kravetz * consumed. This may prevent the task from 223896b96a96SMike Kravetz * faulting in the page at a later time. This 223996b96a96SMike Kravetz * is better than inconsistent global huge page 224096b96a96SMike Kravetz * accounting of reserve counts. 224196b96a96SMike Kravetz */ 224296b96a96SMike Kravetz ClearPagePrivate(page); 224396b96a96SMike Kravetz } else if (rc) { 224496b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 224596b96a96SMike Kravetz if (unlikely(rc < 0)) 224696b96a96SMike Kravetz /* 224796b96a96SMike Kravetz * See above comment about rare out of 224896b96a96SMike Kravetz * memory condition. 224996b96a96SMike Kravetz */ 225096b96a96SMike Kravetz ClearPagePrivate(page); 225196b96a96SMike Kravetz } else 225296b96a96SMike Kravetz vma_end_reservation(h, vma, address); 225396b96a96SMike Kravetz } 225496b96a96SMike Kravetz } 225596b96a96SMike Kravetz 225670c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 225704f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2258348ea204SAdam Litke { 225990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2260a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2261348ea204SAdam Litke struct page *page; 2262d85f69b0SMike Kravetz long map_chg, map_commit; 2263d85f69b0SMike Kravetz long gbl_chg; 22646d76dcf4SAneesh Kumar K.V int ret, idx; 22656d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 226608cf9fafSMina Almasry bool deferred_reserve; 22672fc39cecSAdam Litke 22686d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2269a1e78772SMel Gorman /* 2270d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2271d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2272d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2273a1e78772SMel Gorman */ 2274d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2275d85f69b0SMike Kravetz if (map_chg < 0) 227676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2277d85f69b0SMike Kravetz 2278d85f69b0SMike Kravetz /* 2279d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2280d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2281d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2282d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2283d85f69b0SMike Kravetz * checked against any subpool limit. 2284d85f69b0SMike Kravetz */ 2285d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2286d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2287d85f69b0SMike Kravetz if (gbl_chg < 0) { 2288feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 228976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 22905e911373SMike Kravetz } 229190d8b7e6SAdam Litke 2292d85f69b0SMike Kravetz /* 2293d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2294d85f69b0SMike Kravetz * map, there could be reservations associated with the 2295d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2296d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2297d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2298d85f69b0SMike Kravetz * the subpool reservations. 2299d85f69b0SMike Kravetz */ 2300d85f69b0SMike Kravetz if (avoid_reserve) 2301d85f69b0SMike Kravetz gbl_chg = 1; 2302d85f69b0SMike Kravetz } 2303d85f69b0SMike Kravetz 230408cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 230508cf9fafSMina Almasry */ 230608cf9fafSMina Almasry deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); 230708cf9fafSMina Almasry if (deferred_reserve) { 230808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 230908cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 23108f34af6fSJianyu Zhan if (ret) 23118f34af6fSJianyu Zhan goto out_subpool_put; 231208cf9fafSMina Almasry } 231308cf9fafSMina Almasry 231408cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 231508cf9fafSMina Almasry if (ret) 231608cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 23178f34af6fSJianyu Zhan 2318a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2319d85f69b0SMike Kravetz /* 2320d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2321d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2322d85f69b0SMike Kravetz * a reservation exists for the allocation. 2323d85f69b0SMike Kravetz */ 2324d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 232581a6fcaeSJoonsoo Kim if (!page) { 232694ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 23270c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 23288f34af6fSJianyu Zhan if (!page) 23298f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2330a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2331a88c7695SNaoya Horiguchi SetPagePrivate(page); 2332a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2333a88c7695SNaoya Horiguchi } 233479dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 233515a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 233681a6fcaeSJoonsoo Kim /* Fall through */ 2337a1e78772SMel Gorman } 233881a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 233908cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 234008cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 234108cf9fafSMina Almasry */ 234208cf9fafSMina Almasry if (deferred_reserve) { 234308cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 234408cf9fafSMina Almasry h_cg, page); 234508cf9fafSMina Almasry } 234608cf9fafSMina Almasry 234781a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2348a1e78772SMel Gorman 234990481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2350a1e78772SMel Gorman 2351d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2352d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 235333039678SMike Kravetz /* 235433039678SMike Kravetz * The page was added to the reservation map between 235533039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 235633039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 235733039678SMike Kravetz * Adjust for the subpool count incremented above AND 235833039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 235933039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 236033039678SMike Kravetz * no longer applies. 236133039678SMike Kravetz */ 236233039678SMike Kravetz long rsv_adjust; 236333039678SMike Kravetz 236433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 236533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 236679aa925bSMike Kravetz if (deferred_reserve) 236779aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 236879aa925bSMike Kravetz pages_per_huge_page(h), page); 236933039678SMike Kravetz } 23707893d1d5SAdam Litke return page; 23718f34af6fSJianyu Zhan 23728f34af6fSJianyu Zhan out_uncharge_cgroup: 23738f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 237408cf9fafSMina Almasry out_uncharge_cgroup_reservation: 237508cf9fafSMina Almasry if (deferred_reserve) 237608cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 237708cf9fafSMina Almasry h_cg); 23788f34af6fSJianyu Zhan out_subpool_put: 2379d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 23808f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2381feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 23828f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2383b45b5bd6SDavid Gibson } 2384b45b5bd6SDavid Gibson 2385e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2386e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2387e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2388aa888a74SAndi Kleen { 2389aa888a74SAndi Kleen struct huge_bootmem_page *m; 2390b2261026SJoonsoo Kim int nr_nodes, node; 2391aa888a74SAndi Kleen 2392b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2393aa888a74SAndi Kleen void *addr; 2394aa888a74SAndi Kleen 2395eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 23968b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 239797ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2398aa888a74SAndi Kleen if (addr) { 2399aa888a74SAndi Kleen /* 2400aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2401aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2402aa888a74SAndi Kleen * puts them into the mem_map). 2403aa888a74SAndi Kleen */ 2404aa888a74SAndi Kleen m = addr; 2405aa888a74SAndi Kleen goto found; 2406aa888a74SAndi Kleen } 2407aa888a74SAndi Kleen } 2408aa888a74SAndi Kleen return 0; 2409aa888a74SAndi Kleen 2410aa888a74SAndi Kleen found: 2411df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2412aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2413330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2414aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2415aa888a74SAndi Kleen m->hstate = h; 2416aa888a74SAndi Kleen return 1; 2417aa888a74SAndi Kleen } 2418aa888a74SAndi Kleen 2419d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2420d00181b9SKirill A. Shutemov unsigned int order) 242118229df5SAndy Whitcroft { 242218229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 242318229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 242418229df5SAndy Whitcroft else 242518229df5SAndy Whitcroft prep_compound_page(page, order); 242618229df5SAndy Whitcroft } 242718229df5SAndy Whitcroft 2428aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2429aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2430aa888a74SAndi Kleen { 2431aa888a74SAndi Kleen struct huge_bootmem_page *m; 2432aa888a74SAndi Kleen 2433aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 243440d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2435aa888a74SAndi Kleen struct hstate *h = m->hstate; 2436ee8f248dSBecky Bruce 2437aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 243818229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2439ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2440aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2441af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2442af0fb9dfSMichal Hocko 2443b0320c7bSRafael Aquini /* 2444b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2445b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2446b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2447b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2448b0320c7bSRafael Aquini */ 2449bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 24503dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2451520495feSCannon Matthews cond_resched(); 2452aa888a74SAndi Kleen } 2453aa888a74SAndi Kleen } 2454aa888a74SAndi Kleen 24558faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 24561da177e4SLinus Torvalds { 24571da177e4SLinus Torvalds unsigned long i; 2458f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2459f60858f9SMike Kravetz 2460f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2461f60858f9SMike Kravetz /* 2462f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2463f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2464f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2465f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2466f60858f9SMike Kravetz */ 2467f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2468f60858f9SMike Kravetz GFP_KERNEL); 2469f60858f9SMike Kravetz } else { 2470f60858f9SMike Kravetz /* allocations done at boot time */ 2471f60858f9SMike Kravetz node_alloc_noretry = NULL; 2472f60858f9SMike Kravetz } 2473f60858f9SMike Kravetz 2474f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2475f60858f9SMike Kravetz if (node_alloc_noretry) 2476f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 24771da177e4SLinus Torvalds 2478e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2479bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2480dbda8feaSBarry Song if (hugetlb_cma_size) { 2481cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 2482cf11e85fSRoman Gushchin break; 2483cf11e85fSRoman Gushchin } 2484aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2485aa888a74SAndi Kleen break; 24860c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2487f60858f9SMike Kravetz &node_states[N_MEMORY], 2488f60858f9SMike Kravetz node_alloc_noretry)) 24891da177e4SLinus Torvalds break; 249069ed779aSDavid Rientjes cond_resched(); 24911da177e4SLinus Torvalds } 2492d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2493d715cf80SLiam R. Howlett char buf[32]; 2494d715cf80SLiam R. Howlett 2495c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2496d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2497d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 24988faa8b07SAndi Kleen h->max_huge_pages = i; 2499e5ff2159SAndi Kleen } 2500f60858f9SMike Kravetz 2501f60858f9SMike Kravetz kfree(node_alloc_noretry); 2502d715cf80SLiam R. Howlett } 2503e5ff2159SAndi Kleen 2504e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2505e5ff2159SAndi Kleen { 2506e5ff2159SAndi Kleen struct hstate *h; 2507e5ff2159SAndi Kleen 2508e5ff2159SAndi Kleen for_each_hstate(h) { 2509641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2510641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2511641844f5SNaoya Horiguchi 25128faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2513bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 25148faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2515e5ff2159SAndi Kleen } 2516641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2517e5ff2159SAndi Kleen } 2518e5ff2159SAndi Kleen 2519e5ff2159SAndi Kleen static void __init report_hugepages(void) 2520e5ff2159SAndi Kleen { 2521e5ff2159SAndi Kleen struct hstate *h; 2522e5ff2159SAndi Kleen 2523e5ff2159SAndi Kleen for_each_hstate(h) { 25244abd32dbSAndi Kleen char buf[32]; 2525c6247f72SMatthew Wilcox 2526c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2527ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2528c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2529e5ff2159SAndi Kleen } 2530e5ff2159SAndi Kleen } 2531e5ff2159SAndi Kleen 25321da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 25336ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 25346ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25351da177e4SLinus Torvalds { 25364415cc8dSChristoph Lameter int i; 25374415cc8dSChristoph Lameter 2538bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2539aa888a74SAndi Kleen return; 2540aa888a74SAndi Kleen 25416ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 25421da177e4SLinus Torvalds struct page *page, *next; 2543a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2544a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2545a5516438SAndi Kleen if (count >= h->nr_huge_pages) 25466b0c880dSAdam Litke return; 25471da177e4SLinus Torvalds if (PageHighMem(page)) 25481da177e4SLinus Torvalds continue; 25491da177e4SLinus Torvalds list_del(&page->lru); 2550e5ff2159SAndi Kleen update_and_free_page(h, page); 2551a5516438SAndi Kleen h->free_huge_pages--; 2552a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 25531da177e4SLinus Torvalds } 25541da177e4SLinus Torvalds } 25551da177e4SLinus Torvalds } 25561da177e4SLinus Torvalds #else 25576ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 25586ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25591da177e4SLinus Torvalds { 25601da177e4SLinus Torvalds } 25611da177e4SLinus Torvalds #endif 25621da177e4SLinus Torvalds 256320a0307cSWu Fengguang /* 256420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 256520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 256620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 256720a0307cSWu Fengguang */ 25686ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 25696ae11b27SLee Schermerhorn int delta) 257020a0307cSWu Fengguang { 2571b2261026SJoonsoo Kim int nr_nodes, node; 257220a0307cSWu Fengguang 257320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 257420a0307cSWu Fengguang 2575e8c5c824SLee Schermerhorn if (delta < 0) { 2576b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2577b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2578b2261026SJoonsoo Kim goto found; 2579b2261026SJoonsoo Kim } 2580b2261026SJoonsoo Kim } else { 2581b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2582b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2583b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2584b2261026SJoonsoo Kim goto found; 2585e8c5c824SLee Schermerhorn } 25869a76db09SLee Schermerhorn } 2587b2261026SJoonsoo Kim return 0; 258820a0307cSWu Fengguang 2589b2261026SJoonsoo Kim found: 259020a0307cSWu Fengguang h->surplus_huge_pages += delta; 2591b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2592b2261026SJoonsoo Kim return 1; 259320a0307cSWu Fengguang } 259420a0307cSWu Fengguang 2595a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2596fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 25976ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25981da177e4SLinus Torvalds { 25997893d1d5SAdam Litke unsigned long min_count, ret; 2600f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2601f60858f9SMike Kravetz 2602f60858f9SMike Kravetz /* 2603f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2604f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2605f60858f9SMike Kravetz * the requested huge pages. 2606f60858f9SMike Kravetz */ 2607f60858f9SMike Kravetz if (node_alloc_noretry) 2608f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2609f60858f9SMike Kravetz else 2610f60858f9SMike Kravetz return -ENOMEM; 26111da177e4SLinus Torvalds 26124eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 26134eb0716eSAlexandre Ghiti 26144eb0716eSAlexandre Ghiti /* 2615fd875dcaSMike Kravetz * Check for a node specific request. 2616fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2617fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2618fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2619fd875dcaSMike Kravetz */ 2620fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2621fd875dcaSMike Kravetz unsigned long old_count = count; 2622fd875dcaSMike Kravetz 2623fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2624fd875dcaSMike Kravetz /* 2625fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2626fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2627fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2628fd875dcaSMike Kravetz * largest possible value to align with their intention. 2629fd875dcaSMike Kravetz */ 2630fd875dcaSMike Kravetz if (count < old_count) 2631fd875dcaSMike Kravetz count = ULONG_MAX; 2632fd875dcaSMike Kravetz } 2633fd875dcaSMike Kravetz 2634fd875dcaSMike Kravetz /* 26354eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 26364eb0716eSAlexandre Ghiti * page range allocation. 26374eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 26384eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 26394eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 26404eb0716eSAlexandre Ghiti */ 26414eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 26424eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 26434eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2644f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 26454eb0716eSAlexandre Ghiti return -EINVAL; 26464eb0716eSAlexandre Ghiti } 26474eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 26484eb0716eSAlexandre Ghiti } 2649aa888a74SAndi Kleen 26507893d1d5SAdam Litke /* 26517893d1d5SAdam Litke * Increase the pool size 26527893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 26537893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2654d1c3fb1fSNishanth Aravamudan * 26550c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2656d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2657d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2658d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2659d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 26607893d1d5SAdam Litke */ 2661a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 26626ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 26637893d1d5SAdam Litke break; 26647893d1d5SAdam Litke } 26657893d1d5SAdam Litke 2666a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 26677893d1d5SAdam Litke /* 26687893d1d5SAdam Litke * If this allocation races such that we no longer need the 26697893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 26707893d1d5SAdam Litke * and reducing the surplus. 26717893d1d5SAdam Litke */ 26727893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2673649920c6SJia He 2674649920c6SJia He /* yield cpu to avoid soft lockup */ 2675649920c6SJia He cond_resched(); 2676649920c6SJia He 2677f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2678f60858f9SMike Kravetz node_alloc_noretry); 26797893d1d5SAdam Litke spin_lock(&hugetlb_lock); 26807893d1d5SAdam Litke if (!ret) 26817893d1d5SAdam Litke goto out; 26827893d1d5SAdam Litke 2683536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2684536240f2SMel Gorman if (signal_pending(current)) 2685536240f2SMel Gorman goto out; 26867893d1d5SAdam Litke } 26877893d1d5SAdam Litke 26887893d1d5SAdam Litke /* 26897893d1d5SAdam Litke * Decrease the pool size 26907893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 26917893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 26927893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 26937893d1d5SAdam Litke * to the desired size as pages become free. 2694d1c3fb1fSNishanth Aravamudan * 2695d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2696d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2697d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 26980c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2699d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2700d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2701d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27027893d1d5SAdam Litke */ 2703a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27046b0c880dSAdam Litke min_count = max(count, min_count); 27056ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2706a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 27076ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 27081da177e4SLinus Torvalds break; 270955f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 27101da177e4SLinus Torvalds } 2711a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 27126ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 27137893d1d5SAdam Litke break; 27147893d1d5SAdam Litke } 27157893d1d5SAdam Litke out: 27164eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 27171da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 27184eb0716eSAlexandre Ghiti 2719f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2720f60858f9SMike Kravetz 27214eb0716eSAlexandre Ghiti return 0; 27221da177e4SLinus Torvalds } 27231da177e4SLinus Torvalds 2724a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2725a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2726a3437870SNishanth Aravamudan 2727a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2728a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2729a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2730a3437870SNishanth Aravamudan 2731a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2732a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2733a3437870SNishanth Aravamudan 27349a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 27359a305230SLee Schermerhorn 27369a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2737a3437870SNishanth Aravamudan { 2738a3437870SNishanth Aravamudan int i; 27399a305230SLee Schermerhorn 2740a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 27419a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 27429a305230SLee Schermerhorn if (nidp) 27439a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2744a3437870SNishanth Aravamudan return &hstates[i]; 27459a305230SLee Schermerhorn } 27469a305230SLee Schermerhorn 27479a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2748a3437870SNishanth Aravamudan } 2749a3437870SNishanth Aravamudan 275006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2751a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2752a3437870SNishanth Aravamudan { 27539a305230SLee Schermerhorn struct hstate *h; 27549a305230SLee Schermerhorn unsigned long nr_huge_pages; 27559a305230SLee Schermerhorn int nid; 27569a305230SLee Schermerhorn 27579a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27589a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27599a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 27609a305230SLee Schermerhorn else 27619a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 27629a305230SLee Schermerhorn 2763*ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2764a3437870SNishanth Aravamudan } 2765adbe8726SEric B Munson 2766238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2767238d3c13SDavid Rientjes struct hstate *h, int nid, 2768238d3c13SDavid Rientjes unsigned long count, size_t len) 2769a3437870SNishanth Aravamudan { 2770a3437870SNishanth Aravamudan int err; 27712d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2772a3437870SNishanth Aravamudan 27732d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 27742d0adf7eSOscar Salvador return -EINVAL; 2775adbe8726SEric B Munson 27769a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 27779a305230SLee Schermerhorn /* 27789a305230SLee Schermerhorn * global hstate attribute 27799a305230SLee Schermerhorn */ 27809a305230SLee Schermerhorn if (!(obey_mempolicy && 27812d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 27822d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 27832d0adf7eSOscar Salvador else 27842d0adf7eSOscar Salvador n_mask = &nodes_allowed; 27852d0adf7eSOscar Salvador } else { 27869a305230SLee Schermerhorn /* 2787fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2788fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 27899a305230SLee Schermerhorn */ 27902d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 27912d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2792fd875dcaSMike Kravetz } 27939a305230SLee Schermerhorn 27942d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 279506808b08SLee Schermerhorn 27964eb0716eSAlexandre Ghiti return err ? err : len; 279706808b08SLee Schermerhorn } 279806808b08SLee Schermerhorn 2799238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2800238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2801238d3c13SDavid Rientjes size_t len) 2802238d3c13SDavid Rientjes { 2803238d3c13SDavid Rientjes struct hstate *h; 2804238d3c13SDavid Rientjes unsigned long count; 2805238d3c13SDavid Rientjes int nid; 2806238d3c13SDavid Rientjes int err; 2807238d3c13SDavid Rientjes 2808238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2809238d3c13SDavid Rientjes if (err) 2810238d3c13SDavid Rientjes return err; 2811238d3c13SDavid Rientjes 2812238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2813238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2814238d3c13SDavid Rientjes } 2815238d3c13SDavid Rientjes 281606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 281706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 281806808b08SLee Schermerhorn { 281906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 282006808b08SLee Schermerhorn } 282106808b08SLee Schermerhorn 282206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 282306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 282406808b08SLee Schermerhorn { 2825238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2826a3437870SNishanth Aravamudan } 2827a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2828a3437870SNishanth Aravamudan 282906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 283006808b08SLee Schermerhorn 283106808b08SLee Schermerhorn /* 283206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 283306808b08SLee Schermerhorn * huge page alloc/free. 283406808b08SLee Schermerhorn */ 283506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 2836*ae7a927dSJoe Perches struct kobj_attribute *attr, 2837*ae7a927dSJoe Perches char *buf) 283806808b08SLee Schermerhorn { 283906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 284006808b08SLee Schermerhorn } 284106808b08SLee Schermerhorn 284206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 284306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 284406808b08SLee Schermerhorn { 2845238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 284606808b08SLee Schermerhorn } 284706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 284806808b08SLee Schermerhorn #endif 284906808b08SLee Schermerhorn 285006808b08SLee Schermerhorn 2851a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2852a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2853a3437870SNishanth Aravamudan { 28549a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2855*ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 2856a3437870SNishanth Aravamudan } 2857adbe8726SEric B Munson 2858a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2859a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2860a3437870SNishanth Aravamudan { 2861a3437870SNishanth Aravamudan int err; 2862a3437870SNishanth Aravamudan unsigned long input; 28639a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2864a3437870SNishanth Aravamudan 2865bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2866adbe8726SEric B Munson return -EINVAL; 2867adbe8726SEric B Munson 28683dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2869a3437870SNishanth Aravamudan if (err) 287073ae31e5SEric B Munson return err; 2871a3437870SNishanth Aravamudan 2872a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2873a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2874a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2875a3437870SNishanth Aravamudan 2876a3437870SNishanth Aravamudan return count; 2877a3437870SNishanth Aravamudan } 2878a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2879a3437870SNishanth Aravamudan 2880a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2881a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2882a3437870SNishanth Aravamudan { 28839a305230SLee Schermerhorn struct hstate *h; 28849a305230SLee Schermerhorn unsigned long free_huge_pages; 28859a305230SLee Schermerhorn int nid; 28869a305230SLee Schermerhorn 28879a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 28889a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 28899a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 28909a305230SLee Schermerhorn else 28919a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 28929a305230SLee Schermerhorn 2893*ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 2894a3437870SNishanth Aravamudan } 2895a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2896a3437870SNishanth Aravamudan 2897a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2898a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2899a3437870SNishanth Aravamudan { 29009a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2901*ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 2902a3437870SNishanth Aravamudan } 2903a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2904a3437870SNishanth Aravamudan 2905a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2906a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2907a3437870SNishanth Aravamudan { 29089a305230SLee Schermerhorn struct hstate *h; 29099a305230SLee Schermerhorn unsigned long surplus_huge_pages; 29109a305230SLee Schermerhorn int nid; 29119a305230SLee Schermerhorn 29129a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29139a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29149a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 29159a305230SLee Schermerhorn else 29169a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 29179a305230SLee Schermerhorn 2918*ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 2919a3437870SNishanth Aravamudan } 2920a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2921a3437870SNishanth Aravamudan 2922a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2923a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2924a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2925a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2926a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2927a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 292806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 292906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 293006808b08SLee Schermerhorn #endif 2931a3437870SNishanth Aravamudan NULL, 2932a3437870SNishanth Aravamudan }; 2933a3437870SNishanth Aravamudan 293467e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2935a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2936a3437870SNishanth Aravamudan }; 2937a3437870SNishanth Aravamudan 2938094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 29399a305230SLee Schermerhorn struct kobject **hstate_kobjs, 294067e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2941a3437870SNishanth Aravamudan { 2942a3437870SNishanth Aravamudan int retval; 2943972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2944a3437870SNishanth Aravamudan 29459a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 29469a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2947a3437870SNishanth Aravamudan return -ENOMEM; 2948a3437870SNishanth Aravamudan 29499a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2950a3437870SNishanth Aravamudan if (retval) 29519a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2952a3437870SNishanth Aravamudan 2953a3437870SNishanth Aravamudan return retval; 2954a3437870SNishanth Aravamudan } 2955a3437870SNishanth Aravamudan 2956a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2957a3437870SNishanth Aravamudan { 2958a3437870SNishanth Aravamudan struct hstate *h; 2959a3437870SNishanth Aravamudan int err; 2960a3437870SNishanth Aravamudan 2961a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2962a3437870SNishanth Aravamudan if (!hugepages_kobj) 2963a3437870SNishanth Aravamudan return; 2964a3437870SNishanth Aravamudan 2965a3437870SNishanth Aravamudan for_each_hstate(h) { 29669a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 29679a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2968a3437870SNishanth Aravamudan if (err) 2969282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 2970a3437870SNishanth Aravamudan } 2971a3437870SNishanth Aravamudan } 2972a3437870SNishanth Aravamudan 29739a305230SLee Schermerhorn #ifdef CONFIG_NUMA 29749a305230SLee Schermerhorn 29759a305230SLee Schermerhorn /* 29769a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 297710fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 297810fbcf4cSKay Sievers * index of a node device or _hstate == node id. 297910fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 29809a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 29819a305230SLee Schermerhorn */ 29829a305230SLee Schermerhorn struct node_hstate { 29839a305230SLee Schermerhorn struct kobject *hugepages_kobj; 29849a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 29859a305230SLee Schermerhorn }; 2986b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 29879a305230SLee Schermerhorn 29889a305230SLee Schermerhorn /* 298910fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 29909a305230SLee Schermerhorn */ 29919a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 29929a305230SLee Schermerhorn &nr_hugepages_attr.attr, 29939a305230SLee Schermerhorn &free_hugepages_attr.attr, 29949a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 29959a305230SLee Schermerhorn NULL, 29969a305230SLee Schermerhorn }; 29979a305230SLee Schermerhorn 299867e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 29999a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30009a305230SLee Schermerhorn }; 30019a305230SLee Schermerhorn 30029a305230SLee Schermerhorn /* 300310fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30049a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30059a305230SLee Schermerhorn */ 30069a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 30079a305230SLee Schermerhorn { 30089a305230SLee Schermerhorn int nid; 30099a305230SLee Schermerhorn 30109a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 30119a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 30129a305230SLee Schermerhorn int i; 30139a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 30149a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 30159a305230SLee Schermerhorn if (nidp) 30169a305230SLee Schermerhorn *nidp = nid; 30179a305230SLee Schermerhorn return &hstates[i]; 30189a305230SLee Schermerhorn } 30199a305230SLee Schermerhorn } 30209a305230SLee Schermerhorn 30219a305230SLee Schermerhorn BUG(); 30229a305230SLee Schermerhorn return NULL; 30239a305230SLee Schermerhorn } 30249a305230SLee Schermerhorn 30259a305230SLee Schermerhorn /* 302610fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 30279a305230SLee Schermerhorn * No-op if no hstate attributes attached. 30289a305230SLee Schermerhorn */ 30293cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 30309a305230SLee Schermerhorn { 30319a305230SLee Schermerhorn struct hstate *h; 303210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30339a305230SLee Schermerhorn 30349a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30359b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 30369a305230SLee Schermerhorn 3037972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3038972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3039972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3040972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3041972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3042972dc4deSAneesh Kumar K.V } 30439a305230SLee Schermerhorn } 30449a305230SLee Schermerhorn 30459a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 30469a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 30479a305230SLee Schermerhorn } 30489a305230SLee Schermerhorn 30499a305230SLee Schermerhorn 30509a305230SLee Schermerhorn /* 305110fbcf4cSKay Sievers * Register hstate attributes for a single node device. 30529a305230SLee Schermerhorn * No-op if attributes already registered. 30539a305230SLee Schermerhorn */ 30543cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 30559a305230SLee Schermerhorn { 30569a305230SLee Schermerhorn struct hstate *h; 305710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30589a305230SLee Schermerhorn int err; 30599a305230SLee Schermerhorn 30609a305230SLee Schermerhorn if (nhs->hugepages_kobj) 30619a305230SLee Schermerhorn return; /* already allocated */ 30629a305230SLee Schermerhorn 30639a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 306410fbcf4cSKay Sievers &node->dev.kobj); 30659a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30669a305230SLee Schermerhorn return; 30679a305230SLee Schermerhorn 30689a305230SLee Schermerhorn for_each_hstate(h) { 30699a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 30709a305230SLee Schermerhorn nhs->hstate_kobjs, 30719a305230SLee Schermerhorn &per_node_hstate_attr_group); 30729a305230SLee Schermerhorn if (err) { 3073282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 307410fbcf4cSKay Sievers h->name, node->dev.id); 30759a305230SLee Schermerhorn hugetlb_unregister_node(node); 30769a305230SLee Schermerhorn break; 30779a305230SLee Schermerhorn } 30789a305230SLee Schermerhorn } 30799a305230SLee Schermerhorn } 30809a305230SLee Schermerhorn 30819a305230SLee Schermerhorn /* 30829b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 308310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 308410fbcf4cSKay Sievers * registered their associated device by this time. 30859a305230SLee Schermerhorn */ 30867d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 30879a305230SLee Schermerhorn { 30889a305230SLee Schermerhorn int nid; 30899a305230SLee Schermerhorn 30908cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 30918732794bSWen Congyang struct node *node = node_devices[nid]; 309210fbcf4cSKay Sievers if (node->dev.id == nid) 30939a305230SLee Schermerhorn hugetlb_register_node(node); 30949a305230SLee Schermerhorn } 30959a305230SLee Schermerhorn 30969a305230SLee Schermerhorn /* 309710fbcf4cSKay Sievers * Let the node device driver know we're here so it can 30989a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 30999a305230SLee Schermerhorn */ 31009a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31019a305230SLee Schermerhorn hugetlb_unregister_node); 31029a305230SLee Schermerhorn } 31039a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31049a305230SLee Schermerhorn 31059a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31069a305230SLee Schermerhorn { 31079a305230SLee Schermerhorn BUG(); 31089a305230SLee Schermerhorn if (nidp) 31099a305230SLee Schermerhorn *nidp = -1; 31109a305230SLee Schermerhorn return NULL; 31119a305230SLee Schermerhorn } 31129a305230SLee Schermerhorn 31139a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 31149a305230SLee Schermerhorn 31159a305230SLee Schermerhorn #endif 31169a305230SLee Schermerhorn 3117a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3118a3437870SNishanth Aravamudan { 31198382d914SDavidlohr Bueso int i; 31208382d914SDavidlohr Bueso 3121c2833a5bSMike Kravetz if (!hugepages_supported()) { 3122c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3123c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 31240ef89d25SBenjamin Herrenschmidt return 0; 3125d715cf80SLiam R. Howlett } 3126d715cf80SLiam R. Howlett 3127282f4214SMike Kravetz /* 3128282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3129282f4214SMike Kravetz * architectures depend on setup being done here. 3130282f4214SMike Kravetz */ 3131a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3132282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3133282f4214SMike Kravetz /* 3134282f4214SMike Kravetz * If we did not parse a default huge page size, set 3135282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3136282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3137282f4214SMike Kravetz * specified, set that here as well. 3138282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3139282f4214SMike Kravetz * setting. A warning will be printed in this case. 3140282f4214SMike Kravetz */ 3141282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3142f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3143282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3144282f4214SMike Kravetz char buf[32]; 3145282f4214SMike Kravetz 3146282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3147282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3148282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3149282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3150282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3151282f4214SMike Kravetz default_hstate_max_huge_pages); 3152282f4214SMike Kravetz } 3153282f4214SMike Kravetz default_hstate.max_huge_pages = 3154282f4214SMike Kravetz default_hstate_max_huge_pages; 3155282f4214SMike Kravetz } 3156f8b74815SVaishali Thakkar } 3157a3437870SNishanth Aravamudan 3158cf11e85fSRoman Gushchin hugetlb_cma_check(); 3159a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3160aa888a74SAndi Kleen gather_bootmem_prealloc(); 3161a3437870SNishanth Aravamudan report_hugepages(); 3162a3437870SNishanth Aravamudan 3163a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 31649a305230SLee Schermerhorn hugetlb_register_all_nodes(); 31657179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 31669a305230SLee Schermerhorn 31678382d914SDavidlohr Bueso #ifdef CONFIG_SMP 31688382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 31698382d914SDavidlohr Bueso #else 31708382d914SDavidlohr Bueso num_fault_mutexes = 1; 31718382d914SDavidlohr Bueso #endif 3172c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 31736da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 31746da2ec56SKees Cook GFP_KERNEL); 3175c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 31768382d914SDavidlohr Bueso 31778382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3178c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3179a3437870SNishanth Aravamudan return 0; 3180a3437870SNishanth Aravamudan } 31813e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3182a3437870SNishanth Aravamudan 3183ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3184ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 31859fee021dSVaishali Thakkar { 3186ae94da89SMike Kravetz return size == HPAGE_SIZE; 31879fee021dSVaishali Thakkar } 31889fee021dSVaishali Thakkar 3189d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3190a3437870SNishanth Aravamudan { 3191a3437870SNishanth Aravamudan struct hstate *h; 31928faa8b07SAndi Kleen unsigned long i; 31938faa8b07SAndi Kleen 3194a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3195a3437870SNishanth Aravamudan return; 3196a3437870SNishanth Aravamudan } 319747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3198a3437870SNishanth Aravamudan BUG_ON(order == 0); 319947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3200a3437870SNishanth Aravamudan h->order = order; 3201a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 32028faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32038faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 32040edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 320554f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 320654f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3207a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3208a3437870SNishanth Aravamudan huge_page_size(h)/1024); 32098faa8b07SAndi Kleen 3210a3437870SNishanth Aravamudan parsed_hstate = h; 3211a3437870SNishanth Aravamudan } 3212a3437870SNishanth Aravamudan 3213282f4214SMike Kravetz /* 3214282f4214SMike Kravetz * hugepages command line processing 3215282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3216282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3217282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3218282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3219282f4214SMike Kravetz */ 3220282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3221a3437870SNishanth Aravamudan { 3222a3437870SNishanth Aravamudan unsigned long *mhp; 32238faa8b07SAndi Kleen static unsigned long *last_mhp; 3224a3437870SNishanth Aravamudan 32259fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3226282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 32279fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3228282f4214SMike Kravetz return 0; 32299fee021dSVaishali Thakkar } 3230282f4214SMike Kravetz 3231a3437870SNishanth Aravamudan /* 3232282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3233282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3234282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3235282f4214SMike Kravetz * default_hugepagesz. 3236a3437870SNishanth Aravamudan */ 32379fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3238a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3239a3437870SNishanth Aravamudan else 3240a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3241a3437870SNishanth Aravamudan 32428faa8b07SAndi Kleen if (mhp == last_mhp) { 3243282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3244282f4214SMike Kravetz return 0; 32458faa8b07SAndi Kleen } 32468faa8b07SAndi Kleen 3247a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3248a3437870SNishanth Aravamudan *mhp = 0; 3249a3437870SNishanth Aravamudan 32508faa8b07SAndi Kleen /* 32518faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 32528faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 32538faa8b07SAndi Kleen * use the bootmem allocator. 32548faa8b07SAndi Kleen */ 325547d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 32568faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 32578faa8b07SAndi Kleen 32588faa8b07SAndi Kleen last_mhp = mhp; 32598faa8b07SAndi Kleen 3260a3437870SNishanth Aravamudan return 1; 3261a3437870SNishanth Aravamudan } 3262282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3263e11bfbfcSNick Piggin 3264282f4214SMike Kravetz /* 3265282f4214SMike Kravetz * hugepagesz command line processing 3266282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3267282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3268282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3269282f4214SMike Kravetz * hugepagesz argument was valid. 3270282f4214SMike Kravetz */ 3271359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3272e11bfbfcSNick Piggin { 3273359f2544SMike Kravetz unsigned long size; 3274282f4214SMike Kravetz struct hstate *h; 3275282f4214SMike Kravetz 3276282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3277359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3278359f2544SMike Kravetz 3279359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3280282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3281359f2544SMike Kravetz return 0; 3282359f2544SMike Kravetz } 3283359f2544SMike Kravetz 3284282f4214SMike Kravetz h = size_to_hstate(size); 3285282f4214SMike Kravetz if (h) { 3286282f4214SMike Kravetz /* 3287282f4214SMike Kravetz * hstate for this size already exists. This is normally 3288282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3289282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3290282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3291282f4214SMike Kravetz * previously specified. 3292282f4214SMike Kravetz */ 3293282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3294282f4214SMike Kravetz default_hstate.max_huge_pages) { 3295282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 329638237830SMike Kravetz return 0; 329738237830SMike Kravetz } 329838237830SMike Kravetz 3299282f4214SMike Kravetz /* 3300282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3301282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3302282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3303282f4214SMike Kravetz */ 3304282f4214SMike Kravetz parsed_hstate = h; 3305282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3306e11bfbfcSNick Piggin return 1; 3307e11bfbfcSNick Piggin } 3308282f4214SMike Kravetz 3309359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3310282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3311359f2544SMike Kravetz return 1; 3312359f2544SMike Kravetz } 3313359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3314359f2544SMike Kravetz 3315282f4214SMike Kravetz /* 3316282f4214SMike Kravetz * default_hugepagesz command line input 3317282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3318282f4214SMike Kravetz */ 3319ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3320e11bfbfcSNick Piggin { 3321ae94da89SMike Kravetz unsigned long size; 3322ae94da89SMike Kravetz 3323282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3324282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3325282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3326282f4214SMike Kravetz return 0; 3327282f4214SMike Kravetz } 3328282f4214SMike Kravetz 3329282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3330282f4214SMike Kravetz 3331282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3332282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3333282f4214SMike Kravetz return 0; 3334282f4214SMike Kravetz } 3335282f4214SMike Kravetz 3336282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3337282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3338282f4214SMike Kravetz parsed_default_hugepagesz = true; 3339282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3340282f4214SMike Kravetz 3341282f4214SMike Kravetz /* 3342282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3343282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3344282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3345282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3346282f4214SMike Kravetz * allocated here from bootmem allocator. 3347282f4214SMike Kravetz */ 3348282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3349282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3350282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3351282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3352282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3353282f4214SMike Kravetz } 3354282f4214SMike Kravetz 3355e11bfbfcSNick Piggin return 1; 3356e11bfbfcSNick Piggin } 3357ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3358a3437870SNishanth Aravamudan 33598ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 33608a213460SNishanth Aravamudan { 33618a213460SNishanth Aravamudan int node; 33628a213460SNishanth Aravamudan unsigned int nr = 0; 33638ca39e68SMuchun Song nodemask_t *mpol_allowed; 33648ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 33658ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 33668a213460SNishanth Aravamudan 33678ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 33688ca39e68SMuchun Song 33698ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 33708ca39e68SMuchun Song if (!mpol_allowed || 33718ca39e68SMuchun Song (mpol_allowed && node_isset(node, *mpol_allowed))) 33728a213460SNishanth Aravamudan nr += array[node]; 33738ca39e68SMuchun Song } 33748a213460SNishanth Aravamudan 33758a213460SNishanth Aravamudan return nr; 33768a213460SNishanth Aravamudan } 33778a213460SNishanth Aravamudan 33788a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 337917743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 338017743798SMuchun Song void *buffer, size_t *length, 338117743798SMuchun Song loff_t *ppos, unsigned long *out) 338217743798SMuchun Song { 338317743798SMuchun Song struct ctl_table dup_table; 338417743798SMuchun Song 338517743798SMuchun Song /* 338617743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 338717743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 338817743798SMuchun Song */ 338917743798SMuchun Song dup_table = *table; 339017743798SMuchun Song dup_table.data = out; 339117743798SMuchun Song 339217743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 339317743798SMuchun Song } 339417743798SMuchun Song 339506808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 339606808b08SLee Schermerhorn struct ctl_table *table, int write, 339732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 33981da177e4SLinus Torvalds { 3399e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3400238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 340108d4a246SMichal Hocko int ret; 3402e5ff2159SAndi Kleen 3403457c1b27SNishanth Aravamudan if (!hugepages_supported()) 340486613628SJan Stancek return -EOPNOTSUPP; 3405457c1b27SNishanth Aravamudan 340617743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 340717743798SMuchun Song &tmp); 340808d4a246SMichal Hocko if (ret) 340908d4a246SMichal Hocko goto out; 3410e5ff2159SAndi Kleen 3411238d3c13SDavid Rientjes if (write) 3412238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3413238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 341408d4a246SMichal Hocko out: 341508d4a246SMichal Hocko return ret; 34161da177e4SLinus Torvalds } 3417396faf03SMel Gorman 341806808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 341932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 342006808b08SLee Schermerhorn { 342106808b08SLee Schermerhorn 342206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 342306808b08SLee Schermerhorn buffer, length, ppos); 342406808b08SLee Schermerhorn } 342506808b08SLee Schermerhorn 342606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 342706808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 342832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 342906808b08SLee Schermerhorn { 343006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 343106808b08SLee Schermerhorn buffer, length, ppos); 343206808b08SLee Schermerhorn } 343306808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 343406808b08SLee Schermerhorn 3435a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 343632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3437a3d0c6aaSNishanth Aravamudan { 3438a5516438SAndi Kleen struct hstate *h = &default_hstate; 3439e5ff2159SAndi Kleen unsigned long tmp; 344008d4a246SMichal Hocko int ret; 3441e5ff2159SAndi Kleen 3442457c1b27SNishanth Aravamudan if (!hugepages_supported()) 344386613628SJan Stancek return -EOPNOTSUPP; 3444457c1b27SNishanth Aravamudan 3445e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3446e5ff2159SAndi Kleen 3447bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3448adbe8726SEric B Munson return -EINVAL; 3449adbe8726SEric B Munson 345017743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 345117743798SMuchun Song &tmp); 345208d4a246SMichal Hocko if (ret) 345308d4a246SMichal Hocko goto out; 3454e5ff2159SAndi Kleen 3455e5ff2159SAndi Kleen if (write) { 3456064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3457e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3458a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3459e5ff2159SAndi Kleen } 346008d4a246SMichal Hocko out: 346108d4a246SMichal Hocko return ret; 3462a3d0c6aaSNishanth Aravamudan } 3463a3d0c6aaSNishanth Aravamudan 34641da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 34651da177e4SLinus Torvalds 3466e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 34671da177e4SLinus Torvalds { 3468fcb2b0c5SRoman Gushchin struct hstate *h; 3469fcb2b0c5SRoman Gushchin unsigned long total = 0; 3470fcb2b0c5SRoman Gushchin 3471457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3472457c1b27SNishanth Aravamudan return; 3473fcb2b0c5SRoman Gushchin 3474fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3475fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3476fcb2b0c5SRoman Gushchin 3477fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3478fcb2b0c5SRoman Gushchin 3479fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3480e1759c21SAlexey Dobriyan seq_printf(m, 34811da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 34821da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3483b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 34847893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 34854f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3486fcb2b0c5SRoman Gushchin count, 3487a5516438SAndi Kleen h->free_huge_pages, 3488a5516438SAndi Kleen h->resv_huge_pages, 3489a5516438SAndi Kleen h->surplus_huge_pages, 3490fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3491fcb2b0c5SRoman Gushchin } 3492fcb2b0c5SRoman Gushchin 3493fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 34941da177e4SLinus Torvalds } 34951da177e4SLinus Torvalds 34967981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 34971da177e4SLinus Torvalds { 3498a5516438SAndi Kleen struct hstate *h = &default_hstate; 34997981593bSJoe Perches 3500457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3501457c1b27SNishanth Aravamudan return 0; 35027981593bSJoe Perches 35037981593bSJoe Perches return sysfs_emit_at(buf, len, 35041da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3505a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3506a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3507a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3508a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3509a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35101da177e4SLinus Torvalds } 35111da177e4SLinus Torvalds 3512949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3513949f7ec5SDavid Rientjes { 3514949f7ec5SDavid Rientjes struct hstate *h; 3515949f7ec5SDavid Rientjes int nid; 3516949f7ec5SDavid Rientjes 3517457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3518457c1b27SNishanth Aravamudan return; 3519457c1b27SNishanth Aravamudan 3520949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3521949f7ec5SDavid Rientjes for_each_hstate(h) 3522949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3523949f7ec5SDavid Rientjes nid, 3524949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3525949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3526949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3527949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3528949f7ec5SDavid Rientjes } 3529949f7ec5SDavid Rientjes 35305d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 35315d317b2bSNaoya Horiguchi { 35325d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 35335d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 35345d317b2bSNaoya Horiguchi } 35355d317b2bSNaoya Horiguchi 35361da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 35371da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 35381da177e4SLinus Torvalds { 3539d0028588SWanpeng Li struct hstate *h; 3540d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3541d0028588SWanpeng Li 3542d0028588SWanpeng Li for_each_hstate(h) 3543d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3544d0028588SWanpeng Li return nr_total_pages; 35451da177e4SLinus Torvalds } 35461da177e4SLinus Torvalds 3547a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3548fc1b8a73SMel Gorman { 3549fc1b8a73SMel Gorman int ret = -ENOMEM; 3550fc1b8a73SMel Gorman 3551fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3552fc1b8a73SMel Gorman /* 3553fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3554fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3555fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3556fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3557fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3558fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3559fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3560fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3561fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3562fc1b8a73SMel Gorman * 3563fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3564fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3565fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3566fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3567fc1b8a73SMel Gorman * semantics that cpuset has. 35688ca39e68SMuchun Song * 35698ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 35708ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 35718ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 35728ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 35738ca39e68SMuchun Song * above. 3574fc1b8a73SMel Gorman */ 3575fc1b8a73SMel Gorman if (delta > 0) { 3576a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3577fc1b8a73SMel Gorman goto out; 3578fc1b8a73SMel Gorman 35798ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3580a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3581fc1b8a73SMel Gorman goto out; 3582fc1b8a73SMel Gorman } 3583fc1b8a73SMel Gorman } 3584fc1b8a73SMel Gorman 3585fc1b8a73SMel Gorman ret = 0; 3586fc1b8a73SMel Gorman if (delta < 0) 3587a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3588fc1b8a73SMel Gorman 3589fc1b8a73SMel Gorman out: 3590fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3591fc1b8a73SMel Gorman return ret; 3592fc1b8a73SMel Gorman } 3593fc1b8a73SMel Gorman 359484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 359584afd99bSAndy Whitcroft { 3596f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 359784afd99bSAndy Whitcroft 359884afd99bSAndy Whitcroft /* 359984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 360084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 360184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 360225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 360384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 360484afd99bSAndy Whitcroft * new reference here without additional locking. 360584afd99bSAndy Whitcroft */ 36064e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3607f522c3acSJoonsoo Kim kref_get(&resv->refs); 360884afd99bSAndy Whitcroft } 360984afd99bSAndy Whitcroft 3610a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3611a1e78772SMel Gorman { 3612a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3613f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 361490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36154e35f483SJoonsoo Kim unsigned long reserve, start, end; 36161c5ecae3SMike Kravetz long gbl_reserve; 361784afd99bSAndy Whitcroft 36184e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36194e35f483SJoonsoo Kim return; 36204e35f483SJoonsoo Kim 3621a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3622a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 362384afd99bSAndy Whitcroft 36244e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3625e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36267251ff78SAdam Litke if (reserve) { 36271c5ecae3SMike Kravetz /* 36281c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36291c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36301c5ecae3SMike Kravetz */ 36311c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36321c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36337251ff78SAdam Litke } 3634e9fe92aeSMina Almasry 3635e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3636a1e78772SMel Gorman } 3637a1e78772SMel Gorman 363831383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 363931383c68SDan Williams { 364031383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 364131383c68SDan Williams return -EINVAL; 364231383c68SDan Williams return 0; 364331383c68SDan Williams } 364431383c68SDan Williams 364505ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 364605ea8860SDan Williams { 364705ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 364805ea8860SDan Williams 364905ea8860SDan Williams return 1UL << huge_page_shift(hstate); 365005ea8860SDan Williams } 365105ea8860SDan Williams 36521da177e4SLinus Torvalds /* 36531da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 36541da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 36551da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 36561da177e4SLinus Torvalds * this far. 36571da177e4SLinus Torvalds */ 3658b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 36591da177e4SLinus Torvalds { 36601da177e4SLinus Torvalds BUG(); 3661d0217ac0SNick Piggin return 0; 36621da177e4SLinus Torvalds } 36631da177e4SLinus Torvalds 3664eec3636aSJane Chu /* 3665eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3666eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3667eec3636aSJane Chu * This is because under System V memory model, mappings created via 3668eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3669eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3670eec3636aSJane Chu */ 3671f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3672d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 367384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3674a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3675dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 367605ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 36771da177e4SLinus Torvalds }; 36781da177e4SLinus Torvalds 36791e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 36801e8f889bSDavid Gibson int writable) 368163551ae0SDavid Gibson { 368263551ae0SDavid Gibson pte_t entry; 368363551ae0SDavid Gibson 36841e8f889bSDavid Gibson if (writable) { 3685106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3686106c992aSGerald Schaefer vma->vm_page_prot))); 368763551ae0SDavid Gibson } else { 3688106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3689106c992aSGerald Schaefer vma->vm_page_prot)); 369063551ae0SDavid Gibson } 369163551ae0SDavid Gibson entry = pte_mkyoung(entry); 369263551ae0SDavid Gibson entry = pte_mkhuge(entry); 3693d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 369463551ae0SDavid Gibson 369563551ae0SDavid Gibson return entry; 369663551ae0SDavid Gibson } 369763551ae0SDavid Gibson 36981e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 36991e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37001e8f889bSDavid Gibson { 37011e8f889bSDavid Gibson pte_t entry; 37021e8f889bSDavid Gibson 3703106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 370432f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37054b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37061e8f889bSDavid Gibson } 37071e8f889bSDavid Gibson 3708d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37094a705fefSNaoya Horiguchi { 37104a705fefSNaoya Horiguchi swp_entry_t swp; 37114a705fefSNaoya Horiguchi 37124a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3713d5ed7444SAneesh Kumar K.V return false; 37144a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3715d79d176aSBaoquan He if (is_migration_entry(swp)) 3716d5ed7444SAneesh Kumar K.V return true; 37174a705fefSNaoya Horiguchi else 3718d5ed7444SAneesh Kumar K.V return false; 37194a705fefSNaoya Horiguchi } 37204a705fefSNaoya Horiguchi 37213e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 37224a705fefSNaoya Horiguchi { 37234a705fefSNaoya Horiguchi swp_entry_t swp; 37244a705fefSNaoya Horiguchi 37254a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37263e5c3600SBaoquan He return false; 37274a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3728d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 37293e5c3600SBaoquan He return true; 37304a705fefSNaoya Horiguchi else 37313e5c3600SBaoquan He return false; 37324a705fefSNaoya Horiguchi } 37331e8f889bSDavid Gibson 373463551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 373563551ae0SDavid Gibson struct vm_area_struct *vma) 373663551ae0SDavid Gibson { 37375e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 373863551ae0SDavid Gibson struct page *ptepage; 37391c59827dSHugh Dickins unsigned long addr; 37401e8f889bSDavid Gibson int cow; 3741a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3742a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3743c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3744ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3745e8569dd2SAndreas Sandberg int ret = 0; 37461e8f889bSDavid Gibson 37471e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 374863551ae0SDavid Gibson 3749ac46d4f3SJérôme Glisse if (cow) { 37507269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 37516f4f13e8SJérôme Glisse vma->vm_start, 3752ac46d4f3SJérôme Glisse vma->vm_end); 3753ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3754c0d0381aSMike Kravetz } else { 3755c0d0381aSMike Kravetz /* 3756c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3757c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3758c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3759c0d0381aSMike Kravetz * huge_pmd_unshare. 3760c0d0381aSMike Kravetz */ 3761c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3762ac46d4f3SJérôme Glisse } 3763e8569dd2SAndreas Sandberg 3764a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3765cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 37667868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3767c74df32cSHugh Dickins if (!src_pte) 3768c74df32cSHugh Dickins continue; 3769a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3770e8569dd2SAndreas Sandberg if (!dst_pte) { 3771e8569dd2SAndreas Sandberg ret = -ENOMEM; 3772e8569dd2SAndreas Sandberg break; 3773e8569dd2SAndreas Sandberg } 3774c5c99429SLarry Woodman 37755e41540cSMike Kravetz /* 37765e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 37775e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 37785e41540cSMike Kravetz * 37795e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 37805e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 37815e41540cSMike Kravetz * Check here before taking page table lock, and once again 37825e41540cSMike Kravetz * after taking the lock below. 37835e41540cSMike Kravetz */ 37845e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 37855e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3786c5c99429SLarry Woodman continue; 3787c5c99429SLarry Woodman 3788cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3789cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3790cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 37914a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 37925e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 37935e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 37945e41540cSMike Kravetz /* 37955e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 37965e41540cSMike Kravetz * unlikely case dst entry !none as this implies 37975e41540cSMike Kravetz * sharing with another vma. 37985e41540cSMike Kravetz */ 37994a705fefSNaoya Horiguchi ; 38004a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38014a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38024a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38034a705fefSNaoya Horiguchi 38044a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38054a705fefSNaoya Horiguchi /* 38064a705fefSNaoya Horiguchi * COW mappings require pages in both 38074a705fefSNaoya Horiguchi * parent and child to be set to read. 38084a705fefSNaoya Horiguchi */ 38094a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38104a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3811e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3812e5251fd4SPunit Agrawal entry, sz); 38134a705fefSNaoya Horiguchi } 3814e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38154a705fefSNaoya Horiguchi } else { 381634ee645eSJoerg Roedel if (cow) { 38170f10851eSJérôme Glisse /* 38180f10851eSJérôme Glisse * No need to notify as we are downgrading page 38190f10851eSJérôme Glisse * table protection not changing it to point 38200f10851eSJérôme Glisse * to a new page. 38210f10851eSJérôme Glisse * 3822ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 38230f10851eSJérôme Glisse */ 38247f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 382534ee645eSJoerg Roedel } 38260253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 382763551ae0SDavid Gibson ptepage = pte_page(entry); 382863551ae0SDavid Gibson get_page(ptepage); 382953f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 383063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 38315d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 38321c59827dSHugh Dickins } 3833cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3834cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 383563551ae0SDavid Gibson } 383663551ae0SDavid Gibson 3837e8569dd2SAndreas Sandberg if (cow) 3838ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3839c0d0381aSMike Kravetz else 3840c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3841e8569dd2SAndreas Sandberg 3842e8569dd2SAndreas Sandberg return ret; 384363551ae0SDavid Gibson } 384463551ae0SDavid Gibson 384524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 384624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 384724669e58SAneesh Kumar K.V struct page *ref_page) 384863551ae0SDavid Gibson { 384963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 385063551ae0SDavid Gibson unsigned long address; 3851c7546f8fSDavid Gibson pte_t *ptep; 385263551ae0SDavid Gibson pte_t pte; 3853cb900f41SKirill A. Shutemov spinlock_t *ptl; 385463551ae0SDavid Gibson struct page *page; 3855a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3856a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3857ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3858a5516438SAndi Kleen 385963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3860a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3861a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 386263551ae0SDavid Gibson 386307e32661SAneesh Kumar K.V /* 386407e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 386507e32661SAneesh Kumar K.V * to huge page. 386607e32661SAneesh Kumar K.V */ 3867ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 386824669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3869dff11abeSMike Kravetz 3870dff11abeSMike Kravetz /* 3871dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3872dff11abeSMike Kravetz */ 38736f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 38746f4f13e8SJérôme Glisse end); 3875ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3876ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3877569f48b8SHillf Danton address = start; 3878569f48b8SHillf Danton for (; address < end; address += sz) { 38797868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3880c7546f8fSDavid Gibson if (!ptep) 3881c7546f8fSDavid Gibson continue; 3882c7546f8fSDavid Gibson 3883cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 388434ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 388531d49da5SAneesh Kumar K.V spin_unlock(ptl); 3886dff11abeSMike Kravetz /* 3887dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3888dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3889dff11abeSMike Kravetz */ 389031d49da5SAneesh Kumar K.V continue; 389131d49da5SAneesh Kumar K.V } 389239dde65cSChen, Kenneth W 38936629326bSHillf Danton pte = huge_ptep_get(ptep); 389431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 389531d49da5SAneesh Kumar K.V spin_unlock(ptl); 389631d49da5SAneesh Kumar K.V continue; 389731d49da5SAneesh Kumar K.V } 38986629326bSHillf Danton 38996629326bSHillf Danton /* 39009fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39019fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39026629326bSHillf Danton */ 39039fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39049386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 390531d49da5SAneesh Kumar K.V spin_unlock(ptl); 390631d49da5SAneesh Kumar K.V continue; 39078c4894c6SNaoya Horiguchi } 39086629326bSHillf Danton 39096629326bSHillf Danton page = pte_page(pte); 391004f2cbe3SMel Gorman /* 391104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 391204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 391304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 391404f2cbe3SMel Gorman */ 391504f2cbe3SMel Gorman if (ref_page) { 391631d49da5SAneesh Kumar K.V if (page != ref_page) { 391731d49da5SAneesh Kumar K.V spin_unlock(ptl); 391831d49da5SAneesh Kumar K.V continue; 391931d49da5SAneesh Kumar K.V } 392004f2cbe3SMel Gorman /* 392104f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 392204f2cbe3SMel Gorman * future faults in this VMA will fail rather than 392304f2cbe3SMel Gorman * looking like data was lost 392404f2cbe3SMel Gorman */ 392504f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 392604f2cbe3SMel Gorman } 392704f2cbe3SMel Gorman 3928c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3929b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3930106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 39316649a386SKen Chen set_page_dirty(page); 39329e81130bSHillf Danton 39335d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3934d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 393531d49da5SAneesh Kumar K.V 3936cb900f41SKirill A. Shutemov spin_unlock(ptl); 3937e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 393824669e58SAneesh Kumar K.V /* 393931d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 394024669e58SAneesh Kumar K.V */ 394131d49da5SAneesh Kumar K.V if (ref_page) 394231d49da5SAneesh Kumar K.V break; 3943fe1668aeSChen, Kenneth W } 3944ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 394524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 39461da177e4SLinus Torvalds } 394763551ae0SDavid Gibson 3948d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3949d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3950d833352aSMel Gorman unsigned long end, struct page *ref_page) 3951d833352aSMel Gorman { 3952d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3953d833352aSMel Gorman 3954d833352aSMel Gorman /* 3955d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3956d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3957d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3958d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3959d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3960c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3961d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3962c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3963d833352aSMel Gorman */ 3964d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3965d833352aSMel Gorman } 3966d833352aSMel Gorman 3967502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 396804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3969502717f4SChen, Kenneth W { 397024669e58SAneesh Kumar K.V struct mm_struct *mm; 397124669e58SAneesh Kumar K.V struct mmu_gather tlb; 3972dff11abeSMike Kravetz unsigned long tlb_start = start; 3973dff11abeSMike Kravetz unsigned long tlb_end = end; 3974dff11abeSMike Kravetz 3975dff11abeSMike Kravetz /* 3976dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 3977dff11abeSMike Kravetz * start/end for worst case tlb flushing. 3978dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 3979dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 3980dff11abeSMike Kravetz * the largest possible range. 3981dff11abeSMike Kravetz */ 3982dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 398324669e58SAneesh Kumar K.V 398424669e58SAneesh Kumar K.V mm = vma->vm_mm; 398524669e58SAneesh Kumar K.V 3986dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 398724669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 3988dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 3989502717f4SChen, Kenneth W } 3990502717f4SChen, Kenneth W 399104f2cbe3SMel Gorman /* 399204f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 399304f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 399404f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 399504f2cbe3SMel Gorman * same region. 399604f2cbe3SMel Gorman */ 39972f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 39982a4b3dedSHarvey Harrison struct page *page, unsigned long address) 399904f2cbe3SMel Gorman { 40007526674dSAdam Litke struct hstate *h = hstate_vma(vma); 400104f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 400204f2cbe3SMel Gorman struct address_space *mapping; 400304f2cbe3SMel Gorman pgoff_t pgoff; 400404f2cbe3SMel Gorman 400504f2cbe3SMel Gorman /* 400604f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 400704f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 400804f2cbe3SMel Gorman */ 40097526674dSAdam Litke address = address & huge_page_mask(h); 401036e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 401136e4f20aSMichal Hocko vma->vm_pgoff; 401293c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 401304f2cbe3SMel Gorman 40144eb2b1dcSMel Gorman /* 40154eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40164eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40174eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40184eb2b1dcSMel Gorman */ 401983cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40206b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 402104f2cbe3SMel Gorman /* Do not unmap the current VMA */ 402204f2cbe3SMel Gorman if (iter_vma == vma) 402304f2cbe3SMel Gorman continue; 402404f2cbe3SMel Gorman 402504f2cbe3SMel Gorman /* 40262f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40272f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 40282f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 40292f84a899SMel Gorman */ 40302f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 40312f84a899SMel Gorman continue; 40322f84a899SMel Gorman 40332f84a899SMel Gorman /* 403404f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 403504f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 403604f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 403704f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 403804f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 403904f2cbe3SMel Gorman */ 404004f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 404124669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 404224669e58SAneesh Kumar K.V address + huge_page_size(h), page); 404304f2cbe3SMel Gorman } 404483cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 404504f2cbe3SMel Gorman } 404604f2cbe3SMel Gorman 40470fe6e20bSNaoya Horiguchi /* 40480fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4049ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4050ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4051ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 40520fe6e20bSNaoya Horiguchi */ 40532b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4054974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4055cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 40561e8f889bSDavid Gibson { 40573999f52eSAneesh Kumar K.V pte_t pte; 4058a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 40591e8f889bSDavid Gibson struct page *old_page, *new_page; 40602b740303SSouptick Joarder int outside_reserve = 0; 40612b740303SSouptick Joarder vm_fault_t ret = 0; 4062974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4063ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 40641e8f889bSDavid Gibson 40653999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 40661e8f889bSDavid Gibson old_page = pte_page(pte); 40671e8f889bSDavid Gibson 406804f2cbe3SMel Gorman retry_avoidcopy: 40691e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 40701e8f889bSDavid Gibson * and just make the page writable */ 407137a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 40725a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 40735b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 407483c54070SNick Piggin return 0; 40751e8f889bSDavid Gibson } 40761e8f889bSDavid Gibson 407704f2cbe3SMel Gorman /* 407804f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 407904f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 408004f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 408104f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 408204f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 408304f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 408404f2cbe3SMel Gorman * of the full address range. 408504f2cbe3SMel Gorman */ 40865944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 408704f2cbe3SMel Gorman old_page != pagecache_page) 408804f2cbe3SMel Gorman outside_reserve = 1; 408904f2cbe3SMel Gorman 409009cbfeafSKirill A. Shutemov get_page(old_page); 4091b76c8cfbSLarry Woodman 4092ad4404a2SDavidlohr Bueso /* 4093ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4094ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4095ad4404a2SDavidlohr Bueso */ 4096cb900f41SKirill A. Shutemov spin_unlock(ptl); 40975b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 40981e8f889bSDavid Gibson 40992fc39cecSAdam Litke if (IS_ERR(new_page)) { 410004f2cbe3SMel Gorman /* 410104f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 410204f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 410304f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 410404f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 410504f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 410604f2cbe3SMel Gorman */ 410704f2cbe3SMel Gorman if (outside_reserve) { 410809cbfeafSKirill A. Shutemov put_page(old_page); 410904f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 41105b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 411104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4112cb900f41SKirill A. Shutemov spin_lock(ptl); 41135b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4114a9af0c5dSNaoya Horiguchi if (likely(ptep && 4115a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 411604f2cbe3SMel Gorman goto retry_avoidcopy; 4117a734bcc8SHillf Danton /* 4118cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4119cb900f41SKirill A. Shutemov * lock, and our job is done. 4120a734bcc8SHillf Danton */ 4121a734bcc8SHillf Danton return 0; 412204f2cbe3SMel Gorman } 412304f2cbe3SMel Gorman 41242b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4125ad4404a2SDavidlohr Bueso goto out_release_old; 41261e8f889bSDavid Gibson } 41271e8f889bSDavid Gibson 41280fe6e20bSNaoya Horiguchi /* 41290fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 41300fe6e20bSNaoya Horiguchi * anon_vma prepared. 41310fe6e20bSNaoya Horiguchi */ 413244e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4133ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4134ad4404a2SDavidlohr Bueso goto out_release_all; 413544e2aa93SDean Nelson } 41360fe6e20bSNaoya Horiguchi 4137974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 413847ad8475SAndrea Arcangeli pages_per_huge_page(h)); 41390ed361deSNick Piggin __SetPageUptodate(new_page); 41401e8f889bSDavid Gibson 41417269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 41426f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4143ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4144ad4404a2SDavidlohr Bueso 4145b76c8cfbSLarry Woodman /* 4146cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4147b76c8cfbSLarry Woodman * before the page tables are altered 4148b76c8cfbSLarry Woodman */ 4149cb900f41SKirill A. Shutemov spin_lock(ptl); 41505b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4151a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 415207443a85SJoonsoo Kim ClearPagePrivate(new_page); 415307443a85SJoonsoo Kim 41541e8f889bSDavid Gibson /* Break COW */ 41555b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4156ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 41575b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 41581e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4159d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 41605b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 4161cb6acd01SMike Kravetz set_page_huge_active(new_page); 41621e8f889bSDavid Gibson /* Make the old page be freed below */ 41631e8f889bSDavid Gibson new_page = old_page; 41641e8f889bSDavid Gibson } 4165cb900f41SKirill A. Shutemov spin_unlock(ptl); 4166ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4167ad4404a2SDavidlohr Bueso out_release_all: 41685b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 416909cbfeafSKirill A. Shutemov put_page(new_page); 4170ad4404a2SDavidlohr Bueso out_release_old: 417109cbfeafSKirill A. Shutemov put_page(old_page); 41728312034fSJoonsoo Kim 4173ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4174ad4404a2SDavidlohr Bueso return ret; 41751e8f889bSDavid Gibson } 41761e8f889bSDavid Gibson 417704f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4178a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4179a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 418004f2cbe3SMel Gorman { 418104f2cbe3SMel Gorman struct address_space *mapping; 4182e7c4b0bfSAndy Whitcroft pgoff_t idx; 418304f2cbe3SMel Gorman 418404f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4185a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 418604f2cbe3SMel Gorman 418704f2cbe3SMel Gorman return find_lock_page(mapping, idx); 418804f2cbe3SMel Gorman } 418904f2cbe3SMel Gorman 41903ae77f43SHugh Dickins /* 41913ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 41923ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 41933ae77f43SHugh Dickins */ 41943ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 41952a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 41962a15efc9SHugh Dickins { 41972a15efc9SHugh Dickins struct address_space *mapping; 41982a15efc9SHugh Dickins pgoff_t idx; 41992a15efc9SHugh Dickins struct page *page; 42002a15efc9SHugh Dickins 42012a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42022a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42032a15efc9SHugh Dickins 42042a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42052a15efc9SHugh Dickins if (page) 42062a15efc9SHugh Dickins put_page(page); 42072a15efc9SHugh Dickins return page != NULL; 42082a15efc9SHugh Dickins } 42092a15efc9SHugh Dickins 4210ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4211ab76ad54SMike Kravetz pgoff_t idx) 4212ab76ad54SMike Kravetz { 4213ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4214ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4215ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4216ab76ad54SMike Kravetz 4217ab76ad54SMike Kravetz if (err) 4218ab76ad54SMike Kravetz return err; 4219ab76ad54SMike Kravetz ClearPagePrivate(page); 4220ab76ad54SMike Kravetz 422122146c3cSMike Kravetz /* 422222146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 422322146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 422422146c3cSMike Kravetz */ 422522146c3cSMike Kravetz set_page_dirty(page); 422622146c3cSMike Kravetz 4227ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4228ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4229ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4230ab76ad54SMike Kravetz return 0; 4231ab76ad54SMike Kravetz } 4232ab76ad54SMike Kravetz 42332b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 42342b740303SSouptick Joarder struct vm_area_struct *vma, 42358382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4236788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4237ac9b9c66SHugh Dickins { 4238a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42392b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4240409eb8c2SHillf Danton int anon_rmap = 0; 42414c887265SAdam Litke unsigned long size; 42424c887265SAdam Litke struct page *page; 42431e8f889bSDavid Gibson pte_t new_pte; 4244cb900f41SKirill A. Shutemov spinlock_t *ptl; 4245285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4246cb6acd01SMike Kravetz bool new_page = false; 42474c887265SAdam Litke 424804f2cbe3SMel Gorman /* 424904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 425004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 425125985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 425204f2cbe3SMel Gorman */ 425304f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4254910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 425504f2cbe3SMel Gorman current->pid); 425604f2cbe3SMel Gorman return ret; 425704f2cbe3SMel Gorman } 425804f2cbe3SMel Gorman 42594c887265SAdam Litke /* 426087bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 426187bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 426287bf91d3SMike Kravetz * once for faults beyond end of file. 42634c887265SAdam Litke */ 4264a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4265ebed4bfcSHugh Dickins if (idx >= size) 4266ebed4bfcSHugh Dickins goto out; 42671a1aad8aSMike Kravetz 426887bf91d3SMike Kravetz retry: 426987bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 427087bf91d3SMike Kravetz if (!page) { 42711a1aad8aSMike Kravetz /* 42721a1aad8aSMike Kravetz * Check for page in userfault range 42731a1aad8aSMike Kravetz */ 42741a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 42751a1aad8aSMike Kravetz u32 hash; 42761a1aad8aSMike Kravetz struct vm_fault vmf = { 42771a1aad8aSMike Kravetz .vma = vma, 4278285b8dcaSHuang Ying .address = haddr, 42791a1aad8aSMike Kravetz .flags = flags, 42801a1aad8aSMike Kravetz /* 42811a1aad8aSMike Kravetz * Hard to debug if it ends up being 42821a1aad8aSMike Kravetz * used by a callee that assumes 42831a1aad8aSMike Kravetz * something about the other 42841a1aad8aSMike Kravetz * uninitialized fields... same as in 42851a1aad8aSMike Kravetz * memory.c 42861a1aad8aSMike Kravetz */ 42871a1aad8aSMike Kravetz }; 42881a1aad8aSMike Kravetz 42891a1aad8aSMike Kravetz /* 4290c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4291c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4292c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 42931a1aad8aSMike Kravetz */ 4294188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 42951a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4296c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 42971a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4298c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 42991a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43001a1aad8aSMike Kravetz goto out; 43011a1aad8aSMike Kravetz } 43021a1aad8aSMike Kravetz 4303285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43042fc39cecSAdam Litke if (IS_ERR(page)) { 43054643d67eSMike Kravetz /* 43064643d67eSMike Kravetz * Returning error will result in faulting task being 43074643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43084643d67eSMike Kravetz * tasks from racing to fault in the same page which 43094643d67eSMike Kravetz * could result in false unable to allocate errors. 43104643d67eSMike Kravetz * Page migration does not take the fault mutex, but 43114643d67eSMike Kravetz * does a clear then write of pte's under page table 43124643d67eSMike Kravetz * lock. Page fault code could race with migration, 43134643d67eSMike Kravetz * notice the clear pte and try to allocate a page 43144643d67eSMike Kravetz * here. Before returning error, get ptl and make 43154643d67eSMike Kravetz * sure there really is no pte entry. 43164643d67eSMike Kravetz */ 43174643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 43184643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 43194643d67eSMike Kravetz ret = 0; 43204643d67eSMike Kravetz spin_unlock(ptl); 43214643d67eSMike Kravetz goto out; 43224643d67eSMike Kravetz } 43234643d67eSMike Kravetz spin_unlock(ptl); 43242b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 43256bda666aSChristoph Lameter goto out; 43266bda666aSChristoph Lameter } 432747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 43280ed361deSNick Piggin __SetPageUptodate(page); 4329cb6acd01SMike Kravetz new_page = true; 4330ac9b9c66SHugh Dickins 4331f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4332ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 43336bda666aSChristoph Lameter if (err) { 43346bda666aSChristoph Lameter put_page(page); 43356bda666aSChristoph Lameter if (err == -EEXIST) 43366bda666aSChristoph Lameter goto retry; 43376bda666aSChristoph Lameter goto out; 43386bda666aSChristoph Lameter } 433923be7468SMel Gorman } else { 43406bda666aSChristoph Lameter lock_page(page); 43410fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 43420fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 43430fe6e20bSNaoya Horiguchi goto backout_unlocked; 434423be7468SMel Gorman } 4345409eb8c2SHillf Danton anon_rmap = 1; 43460fe6e20bSNaoya Horiguchi } 43470fe6e20bSNaoya Horiguchi } else { 434857303d80SAndy Whitcroft /* 4349998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4350998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4351998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4352fd6a03edSNaoya Horiguchi */ 4353fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 4354aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 4355972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4356fd6a03edSNaoya Horiguchi goto backout_unlocked; 43576bda666aSChristoph Lameter } 4358998b4382SNaoya Horiguchi } 43591e8f889bSDavid Gibson 436057303d80SAndy Whitcroft /* 436157303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 436257303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 436357303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 436457303d80SAndy Whitcroft * the spinlock. 436557303d80SAndy Whitcroft */ 43665e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4367285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 43682b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 43692b26736cSAndy Whitcroft goto backout_unlocked; 43702b26736cSAndy Whitcroft } 43715e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4372285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 43735e911373SMike Kravetz } 437457303d80SAndy Whitcroft 43758bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 437683c54070SNick Piggin ret = 0; 43777f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 43784c887265SAdam Litke goto backout; 43794c887265SAdam Litke 438007443a85SJoonsoo Kim if (anon_rmap) { 438107443a85SJoonsoo Kim ClearPagePrivate(page); 4382285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4383ac714904SChoi Gi-yong } else 438453f9263bSKirill A. Shutemov page_dup_rmap(page, true); 43851e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 43861e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4387285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 43881e8f889bSDavid Gibson 43895d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4390788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 43911e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4392974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 43931e8f889bSDavid Gibson } 43941e8f889bSDavid Gibson 4395cb900f41SKirill A. Shutemov spin_unlock(ptl); 4396cb6acd01SMike Kravetz 4397cb6acd01SMike Kravetz /* 4398cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 4399cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 4400cb6acd01SMike Kravetz * isolated for migration. 4401cb6acd01SMike Kravetz */ 4402cb6acd01SMike Kravetz if (new_page) 4403cb6acd01SMike Kravetz set_page_huge_active(page); 4404cb6acd01SMike Kravetz 44054c887265SAdam Litke unlock_page(page); 44064c887265SAdam Litke out: 4407ac9b9c66SHugh Dickins return ret; 44084c887265SAdam Litke 44094c887265SAdam Litke backout: 4410cb900f41SKirill A. Shutemov spin_unlock(ptl); 44112b26736cSAndy Whitcroft backout_unlocked: 44124c887265SAdam Litke unlock_page(page); 4413285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 44144c887265SAdam Litke put_page(page); 44154c887265SAdam Litke goto out; 4416ac9b9c66SHugh Dickins } 4417ac9b9c66SHugh Dickins 44188382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4419188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44208382d914SDavidlohr Bueso { 44218382d914SDavidlohr Bueso unsigned long key[2]; 44228382d914SDavidlohr Bueso u32 hash; 44238382d914SDavidlohr Bueso 44248382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 44258382d914SDavidlohr Bueso key[1] = idx; 44268382d914SDavidlohr Bueso 442755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 44288382d914SDavidlohr Bueso 44298382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 44308382d914SDavidlohr Bueso } 44318382d914SDavidlohr Bueso #else 44328382d914SDavidlohr Bueso /* 44338382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 44348382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 44358382d914SDavidlohr Bueso */ 4436188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44378382d914SDavidlohr Bueso { 44388382d914SDavidlohr Bueso return 0; 44398382d914SDavidlohr Bueso } 44408382d914SDavidlohr Bueso #endif 44418382d914SDavidlohr Bueso 44422b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4443788c7df4SHugh Dickins unsigned long address, unsigned int flags) 444486e5216fSAdam Litke { 44458382d914SDavidlohr Bueso pte_t *ptep, entry; 4446cb900f41SKirill A. Shutemov spinlock_t *ptl; 44472b740303SSouptick Joarder vm_fault_t ret; 44488382d914SDavidlohr Bueso u32 hash; 44498382d914SDavidlohr Bueso pgoff_t idx; 44500fe6e20bSNaoya Horiguchi struct page *page = NULL; 445157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4452a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44538382d914SDavidlohr Bueso struct address_space *mapping; 44540f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4455285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 445686e5216fSAdam Litke 4457285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4458fd6a03edSNaoya Horiguchi if (ptep) { 4459c0d0381aSMike Kravetz /* 4460c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4461c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4462c0d0381aSMike Kravetz * not actually modifying content here. 4463c0d0381aSMike Kravetz */ 4464fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4465290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4466cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4467290408d4SNaoya Horiguchi return 0; 4468290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4469aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4470972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4471b43a9990SMike Kravetz } 44728382d914SDavidlohr Bueso 4473c0d0381aSMike Kravetz /* 4474c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 447587bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 447687bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 447787bf91d3SMike Kravetz * and making the ptep no longer valid. 447887bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4479c0d0381aSMike Kravetz * 4480c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4481c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4482c0d0381aSMike Kravetz * something has changed. 4483c0d0381aSMike Kravetz */ 4484ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4485c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4486c0d0381aSMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4487c0d0381aSMike Kravetz if (!ptep) { 4488c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4489c0d0381aSMike Kravetz return VM_FAULT_OOM; 4490c0d0381aSMike Kravetz } 4491ddeaab32SMike Kravetz 44923935baa9SDavid Gibson /* 44933935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 44943935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 44953935baa9SDavid Gibson * the same page in the page cache. 44963935baa9SDavid Gibson */ 4497c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4498188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4499c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45008382d914SDavidlohr Bueso 45017f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45027f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45038382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4504b4d1d99fSDavid Gibson goto out_mutex; 45053935baa9SDavid Gibson } 450686e5216fSAdam Litke 450783c54070SNick Piggin ret = 0; 45081e8f889bSDavid Gibson 450957303d80SAndy Whitcroft /* 45100f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45110f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45127c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45137c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 45147c8de358SEthon Paul * properly handle it. 45150f792cf9SNaoya Horiguchi */ 45160f792cf9SNaoya Horiguchi if (!pte_present(entry)) 45170f792cf9SNaoya Horiguchi goto out_mutex; 45180f792cf9SNaoya Horiguchi 45190f792cf9SNaoya Horiguchi /* 452057303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 452157303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 452257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 452357303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 452457303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 452557303d80SAndy Whitcroft * consumed. 452657303d80SAndy Whitcroft */ 4527106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4528285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 45292b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4530b4d1d99fSDavid Gibson goto out_mutex; 45312b26736cSAndy Whitcroft } 45325e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4533285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 453457303d80SAndy Whitcroft 4535f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 453657303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4537285b8dcaSHuang Ying vma, haddr); 453857303d80SAndy Whitcroft } 453957303d80SAndy Whitcroft 45400f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 45410fe6e20bSNaoya Horiguchi 45421e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4543b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4544cb900f41SKirill A. Shutemov goto out_ptl; 4545b4d1d99fSDavid Gibson 45460f792cf9SNaoya Horiguchi /* 45470f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 45480f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 45490f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 45500f792cf9SNaoya Horiguchi */ 45510f792cf9SNaoya Horiguchi page = pte_page(entry); 45520f792cf9SNaoya Horiguchi if (page != pagecache_page) 45530f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 45540f792cf9SNaoya Horiguchi need_wait_lock = 1; 45550f792cf9SNaoya Horiguchi goto out_ptl; 45560f792cf9SNaoya Horiguchi } 45570f792cf9SNaoya Horiguchi 45580f792cf9SNaoya Horiguchi get_page(page); 4559b4d1d99fSDavid Gibson 4560788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4561106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4562974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4563cb900f41SKirill A. Shutemov pagecache_page, ptl); 45640f792cf9SNaoya Horiguchi goto out_put_page; 4565b4d1d99fSDavid Gibson } 4566106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4567b4d1d99fSDavid Gibson } 4568b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4569285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4570788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4571285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 45720f792cf9SNaoya Horiguchi out_put_page: 45730f792cf9SNaoya Horiguchi if (page != pagecache_page) 45740f792cf9SNaoya Horiguchi unlock_page(page); 45750f792cf9SNaoya Horiguchi put_page(page); 4576cb900f41SKirill A. Shutemov out_ptl: 4577cb900f41SKirill A. Shutemov spin_unlock(ptl); 457857303d80SAndy Whitcroft 457957303d80SAndy Whitcroft if (pagecache_page) { 458057303d80SAndy Whitcroft unlock_page(pagecache_page); 458157303d80SAndy Whitcroft put_page(pagecache_page); 458257303d80SAndy Whitcroft } 4583b4d1d99fSDavid Gibson out_mutex: 4584c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4585c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 45860f792cf9SNaoya Horiguchi /* 45870f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 45880f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 45890f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 45900f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 45910f792cf9SNaoya Horiguchi * here without taking refcount. 45920f792cf9SNaoya Horiguchi */ 45930f792cf9SNaoya Horiguchi if (need_wait_lock) 45940f792cf9SNaoya Horiguchi wait_on_page_locked(page); 45951e8f889bSDavid Gibson return ret; 459686e5216fSAdam Litke } 459786e5216fSAdam Litke 45988fb5debcSMike Kravetz /* 45998fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46008fb5debcSMike Kravetz * modifications for huge pages. 46018fb5debcSMike Kravetz */ 46028fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46038fb5debcSMike Kravetz pte_t *dst_pte, 46048fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46058fb5debcSMike Kravetz unsigned long dst_addr, 46068fb5debcSMike Kravetz unsigned long src_addr, 46078fb5debcSMike Kravetz struct page **pagep) 46088fb5debcSMike Kravetz { 46091e392147SAndrea Arcangeli struct address_space *mapping; 46101e392147SAndrea Arcangeli pgoff_t idx; 46111e392147SAndrea Arcangeli unsigned long size; 46121c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46138fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 46148fb5debcSMike Kravetz pte_t _dst_pte; 46158fb5debcSMike Kravetz spinlock_t *ptl; 46168fb5debcSMike Kravetz int ret; 46178fb5debcSMike Kravetz struct page *page; 46188fb5debcSMike Kravetz 46198fb5debcSMike Kravetz if (!*pagep) { 46208fb5debcSMike Kravetz ret = -ENOMEM; 46218fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 46228fb5debcSMike Kravetz if (IS_ERR(page)) 46238fb5debcSMike Kravetz goto out; 46248fb5debcSMike Kravetz 46258fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 46268fb5debcSMike Kravetz (const void __user *) src_addr, 4627810a56b9SMike Kravetz pages_per_huge_page(h), false); 46288fb5debcSMike Kravetz 4629c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 46308fb5debcSMike Kravetz if (unlikely(ret)) { 46319e368259SAndrea Arcangeli ret = -ENOENT; 46328fb5debcSMike Kravetz *pagep = page; 46338fb5debcSMike Kravetz /* don't free the page */ 46348fb5debcSMike Kravetz goto out; 46358fb5debcSMike Kravetz } 46368fb5debcSMike Kravetz } else { 46378fb5debcSMike Kravetz page = *pagep; 46388fb5debcSMike Kravetz *pagep = NULL; 46398fb5debcSMike Kravetz } 46408fb5debcSMike Kravetz 46418fb5debcSMike Kravetz /* 46428fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 46438fb5debcSMike Kravetz * preceding stores to the page contents become visible before 46448fb5debcSMike Kravetz * the set_pte_at() write. 46458fb5debcSMike Kravetz */ 46468fb5debcSMike Kravetz __SetPageUptodate(page); 46478fb5debcSMike Kravetz 46481e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 46491e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 46501e392147SAndrea Arcangeli 46511c9e8defSMike Kravetz /* 46521c9e8defSMike Kravetz * If shared, add to page cache 46531c9e8defSMike Kravetz */ 46541c9e8defSMike Kravetz if (vm_shared) { 46551e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 46561e392147SAndrea Arcangeli ret = -EFAULT; 46571e392147SAndrea Arcangeli if (idx >= size) 46581e392147SAndrea Arcangeli goto out_release_nounlock; 46591c9e8defSMike Kravetz 46601e392147SAndrea Arcangeli /* 46611e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 46621e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 46631e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 46641e392147SAndrea Arcangeli * the caller. 46651e392147SAndrea Arcangeli */ 46661c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 46671c9e8defSMike Kravetz if (ret) 46681c9e8defSMike Kravetz goto out_release_nounlock; 46691c9e8defSMike Kravetz } 46701c9e8defSMike Kravetz 46718fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 46728fb5debcSMike Kravetz spin_lock(ptl); 46738fb5debcSMike Kravetz 46741e392147SAndrea Arcangeli /* 46751e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 46761e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 46771e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 46781e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 46791e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 46801e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 46811e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 46821e392147SAndrea Arcangeli */ 46831e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 46841e392147SAndrea Arcangeli ret = -EFAULT; 46851e392147SAndrea Arcangeli if (idx >= size) 46861e392147SAndrea Arcangeli goto out_release_unlock; 46871e392147SAndrea Arcangeli 46888fb5debcSMike Kravetz ret = -EEXIST; 46898fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 46908fb5debcSMike Kravetz goto out_release_unlock; 46918fb5debcSMike Kravetz 46921c9e8defSMike Kravetz if (vm_shared) { 46931c9e8defSMike Kravetz page_dup_rmap(page, true); 46941c9e8defSMike Kravetz } else { 46958fb5debcSMike Kravetz ClearPagePrivate(page); 46968fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 46971c9e8defSMike Kravetz } 46988fb5debcSMike Kravetz 46998fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47008fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47018fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47028fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47038fb5debcSMike Kravetz 47048fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47058fb5debcSMike Kravetz 47068fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47078fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47088fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47098fb5debcSMike Kravetz 47108fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47118fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47128fb5debcSMike Kravetz 47138fb5debcSMike Kravetz spin_unlock(ptl); 4714cb6acd01SMike Kravetz set_page_huge_active(page); 47151c9e8defSMike Kravetz if (vm_shared) 47161c9e8defSMike Kravetz unlock_page(page); 47178fb5debcSMike Kravetz ret = 0; 47188fb5debcSMike Kravetz out: 47198fb5debcSMike Kravetz return ret; 47208fb5debcSMike Kravetz out_release_unlock: 47218fb5debcSMike Kravetz spin_unlock(ptl); 47221c9e8defSMike Kravetz if (vm_shared) 47231c9e8defSMike Kravetz unlock_page(page); 47245af10dfdSAndrea Arcangeli out_release_nounlock: 47258fb5debcSMike Kravetz put_page(page); 47268fb5debcSMike Kravetz goto out; 47278fb5debcSMike Kravetz } 47288fb5debcSMike Kravetz 472928a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 473063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 473128a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 47324f6da934SPeter Xu long i, unsigned int flags, int *locked) 473363551ae0SDavid Gibson { 4734d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4735d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 473628a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4737a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 47382be7cfedSDaniel Jordan int err = -EFAULT; 473963551ae0SDavid Gibson 474063551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 474163551ae0SDavid Gibson pte_t *pte; 4742cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 47432a15efc9SHugh Dickins int absent; 474463551ae0SDavid Gibson struct page *page; 474563551ae0SDavid Gibson 47464c887265SAdam Litke /* 474702057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 474802057967SDavid Rientjes * potentially allocating memory. 474902057967SDavid Rientjes */ 4750fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 475102057967SDavid Rientjes remainder = 0; 475202057967SDavid Rientjes break; 475302057967SDavid Rientjes } 475402057967SDavid Rientjes 475502057967SDavid Rientjes /* 47564c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 47572a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 47584c887265SAdam Litke * first, for the page indexing below to work. 4759cb900f41SKirill A. Shutemov * 4760cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 47614c887265SAdam Litke */ 47627868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 47637868a208SPunit Agrawal huge_page_size(h)); 4764cb900f41SKirill A. Shutemov if (pte) 4765cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 47662a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 476763551ae0SDavid Gibson 47682a15efc9SHugh Dickins /* 47692a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 47703ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 47713ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 47723ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 47733ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 47742a15efc9SHugh Dickins */ 47753ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 47763ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4777cb900f41SKirill A. Shutemov if (pte) 4778cb900f41SKirill A. Shutemov spin_unlock(ptl); 47792a15efc9SHugh Dickins remainder = 0; 47802a15efc9SHugh Dickins break; 47812a15efc9SHugh Dickins } 47822a15efc9SHugh Dickins 47839cc3a5bdSNaoya Horiguchi /* 47849cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 47859cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 47869cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 47879cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 47889cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 47899cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 47909cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 47919cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 47929cc3a5bdSNaoya Horiguchi */ 47939cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4794106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4795106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 47962b740303SSouptick Joarder vm_fault_t ret; 479787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 47984c887265SAdam Litke 4799cb900f41SKirill A. Shutemov if (pte) 4800cb900f41SKirill A. Shutemov spin_unlock(ptl); 480187ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 480287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 48034f6da934SPeter Xu if (locked) 480471335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 480571335f37SPeter Xu FAULT_FLAG_KILLABLE; 480687ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 480787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 480887ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 480987ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 48104426e945SPeter Xu /* 48114426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 48124426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 48134426e945SPeter Xu */ 481487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 481587ffc118SAndrea Arcangeli } 481687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 481787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 48182be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 48191c59827dSHugh Dickins remainder = 0; 48201c59827dSHugh Dickins break; 48211c59827dSHugh Dickins } 482287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 48234f6da934SPeter Xu if (locked && 48241ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 48254f6da934SPeter Xu *locked = 0; 482687ffc118SAndrea Arcangeli *nr_pages = 0; 482787ffc118SAndrea Arcangeli /* 482887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 482987ffc118SAndrea Arcangeli * error, it will return zero 483087ffc118SAndrea Arcangeli * instead. 483187ffc118SAndrea Arcangeli * 483287ffc118SAndrea Arcangeli * No need to update "position" as the 483387ffc118SAndrea Arcangeli * caller will not check it after 483487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 483587ffc118SAndrea Arcangeli */ 483687ffc118SAndrea Arcangeli return i; 483787ffc118SAndrea Arcangeli } 483887ffc118SAndrea Arcangeli continue; 483987ffc118SAndrea Arcangeli } 484063551ae0SDavid Gibson 4841a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 48427f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 48438fde12caSLinus Torvalds 48448fde12caSLinus Torvalds /* 4845acbfb087SZhigang Lu * If subpage information not requested, update counters 4846acbfb087SZhigang Lu * and skip the same_page loop below. 4847acbfb087SZhigang Lu */ 4848acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4849acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4850acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4851acbfb087SZhigang Lu vaddr += huge_page_size(h); 4852acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4853acbfb087SZhigang Lu i += pages_per_huge_page(h); 4854acbfb087SZhigang Lu spin_unlock(ptl); 4855acbfb087SZhigang Lu continue; 4856acbfb087SZhigang Lu } 4857acbfb087SZhigang Lu 4858d5d4b0aaSChen, Kenneth W same_page: 4859d6692183SChen, Kenneth W if (pages) { 486069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 48613faa52c0SJohn Hubbard /* 48623faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: 48633faa52c0SJohn Hubbard * a) we hold the ptl lock, and b) we've just checked 48643faa52c0SJohn Hubbard * that the huge page is present in the page tables. If 48653faa52c0SJohn Hubbard * the huge page is present, then the tail pages must 48663faa52c0SJohn Hubbard * also be present. The ptl prevents the head page and 48673faa52c0SJohn Hubbard * tail pages from being rearranged in any way. So this 48683faa52c0SJohn Hubbard * page must be available at this point, unless the page 48693faa52c0SJohn Hubbard * refcount overflowed: 48703faa52c0SJohn Hubbard */ 48713faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { 48723faa52c0SJohn Hubbard spin_unlock(ptl); 48733faa52c0SJohn Hubbard remainder = 0; 48743faa52c0SJohn Hubbard err = -ENOMEM; 48753faa52c0SJohn Hubbard break; 48763faa52c0SJohn Hubbard } 4877d6692183SChen, Kenneth W } 487863551ae0SDavid Gibson 487963551ae0SDavid Gibson if (vmas) 488063551ae0SDavid Gibson vmas[i] = vma; 488163551ae0SDavid Gibson 488263551ae0SDavid Gibson vaddr += PAGE_SIZE; 4883d5d4b0aaSChen, Kenneth W ++pfn_offset; 488463551ae0SDavid Gibson --remainder; 488563551ae0SDavid Gibson ++i; 4886d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4887a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4888d5d4b0aaSChen, Kenneth W /* 4889d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4890d5d4b0aaSChen, Kenneth W * of this compound page. 4891d5d4b0aaSChen, Kenneth W */ 4892d5d4b0aaSChen, Kenneth W goto same_page; 4893d5d4b0aaSChen, Kenneth W } 4894cb900f41SKirill A. Shutemov spin_unlock(ptl); 489563551ae0SDavid Gibson } 489628a35716SMichel Lespinasse *nr_pages = remainder; 489787ffc118SAndrea Arcangeli /* 489887ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 489987ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 490087ffc118SAndrea Arcangeli * branch. 490187ffc118SAndrea Arcangeli */ 490263551ae0SDavid Gibson *position = vaddr; 490363551ae0SDavid Gibson 49042be7cfedSDaniel Jordan return i ? i : err; 490563551ae0SDavid Gibson } 49068f860591SZhang, Yanmin 49075491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 49085491ae7bSAneesh Kumar K.V /* 49095491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 49105491ae7bSAneesh Kumar K.V * implement this. 49115491ae7bSAneesh Kumar K.V */ 49125491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 49135491ae7bSAneesh Kumar K.V #endif 49145491ae7bSAneesh Kumar K.V 49157da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 49168f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 49178f860591SZhang, Yanmin { 49188f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 49198f860591SZhang, Yanmin unsigned long start = address; 49208f860591SZhang, Yanmin pte_t *ptep; 49218f860591SZhang, Yanmin pte_t pte; 4922a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49237da4d641SPeter Zijlstra unsigned long pages = 0; 4924dff11abeSMike Kravetz bool shared_pmd = false; 4925ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4926dff11abeSMike Kravetz 4927dff11abeSMike Kravetz /* 4928dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4929ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4930dff11abeSMike Kravetz * range if PMD sharing is possible. 4931dff11abeSMike Kravetz */ 49327269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 49337269f999SJérôme Glisse 0, vma, mm, start, end); 4934ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 49358f860591SZhang, Yanmin 49368f860591SZhang, Yanmin BUG_ON(address >= end); 4937ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 49388f860591SZhang, Yanmin 4939ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 494083cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4941a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4942cb900f41SKirill A. Shutemov spinlock_t *ptl; 49437868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 49448f860591SZhang, Yanmin if (!ptep) 49458f860591SZhang, Yanmin continue; 4946cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 494734ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 49487da4d641SPeter Zijlstra pages++; 4949cb900f41SKirill A. Shutemov spin_unlock(ptl); 4950dff11abeSMike Kravetz shared_pmd = true; 495139dde65cSChen, Kenneth W continue; 49527da4d641SPeter Zijlstra } 4953a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4954a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4955a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4956a8bda28dSNaoya Horiguchi continue; 4957a8bda28dSNaoya Horiguchi } 4958a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4959a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4960a8bda28dSNaoya Horiguchi 4961a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4962a8bda28dSNaoya Horiguchi pte_t newpte; 4963a8bda28dSNaoya Horiguchi 4964a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4965a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4966e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4967e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4968a8bda28dSNaoya Horiguchi pages++; 4969a8bda28dSNaoya Horiguchi } 4970a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4971a8bda28dSNaoya Horiguchi continue; 4972a8bda28dSNaoya Horiguchi } 4973a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 4974023bdd00SAneesh Kumar K.V pte_t old_pte; 4975023bdd00SAneesh Kumar K.V 4976023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 4977023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 4978be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 4979023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 49807da4d641SPeter Zijlstra pages++; 49818f860591SZhang, Yanmin } 4982cb900f41SKirill A. Shutemov spin_unlock(ptl); 49838f860591SZhang, Yanmin } 4984d833352aSMel Gorman /* 4985c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4986d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4987c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4988dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 4989dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 4990d833352aSMel Gorman */ 4991dff11abeSMike Kravetz if (shared_pmd) 4992ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 4993dff11abeSMike Kravetz else 49945491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 49950f10851eSJérôme Glisse /* 49960f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 49970f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 49980f10851eSJérôme Glisse * 4999ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50000f10851eSJérôme Glisse */ 500183cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5002ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50037da4d641SPeter Zijlstra 50047da4d641SPeter Zijlstra return pages << h->order; 50058f860591SZhang, Yanmin } 50068f860591SZhang, Yanmin 5007a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 5008a1e78772SMel Gorman long from, long to, 50095a6fe125SMel Gorman struct vm_area_struct *vma, 5010ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5011e4e574b7SAdam Litke { 50120db9d74eSMina Almasry long ret, chg, add = -1; 5013a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 501490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 50159119a41eSJoonsoo Kim struct resv_map *resv_map; 5016075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 50170db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5018e4e574b7SAdam Litke 501963489f8eSMike Kravetz /* This should never happen */ 502063489f8eSMike Kravetz if (from > to) { 502163489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 502263489f8eSMike Kravetz return -EINVAL; 502363489f8eSMike Kravetz } 502463489f8eSMike Kravetz 5025a1e78772SMel Gorman /* 502617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 502717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 502890481622SDavid Gibson * without using reserves 502917c9d12eSMel Gorman */ 5030ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 503117c9d12eSMel Gorman return 0; 503217c9d12eSMel Gorman 503317c9d12eSMel Gorman /* 5034a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5035a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5036a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5037a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5038a1e78772SMel Gorman */ 50399119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5040f27a5136SMike Kravetz /* 5041f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5042f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5043f27a5136SMike Kravetz * hugetlbfs_get_inode). 5044f27a5136SMike Kravetz */ 50454e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 50469119a41eSJoonsoo Kim 50470db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 50489119a41eSJoonsoo Kim 50499119a41eSJoonsoo Kim } else { 5050e9fe92aeSMina Almasry /* Private mapping. */ 50519119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 50525a6fe125SMel Gorman if (!resv_map) 50535a6fe125SMel Gorman return -ENOMEM; 50545a6fe125SMel Gorman 505517c9d12eSMel Gorman chg = to - from; 505617c9d12eSMel Gorman 50575a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 50585a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 50595a6fe125SMel Gorman } 50605a6fe125SMel Gorman 5061c50ac050SDave Hansen if (chg < 0) { 5062c50ac050SDave Hansen ret = chg; 5063c50ac050SDave Hansen goto out_err; 5064c50ac050SDave Hansen } 506517c9d12eSMel Gorman 5066075a61d0SMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 5067075a61d0SMina Almasry hstate_index(h), chg * pages_per_huge_page(h), &h_cg); 5068075a61d0SMina Almasry 5069075a61d0SMina Almasry if (ret < 0) { 5070075a61d0SMina Almasry ret = -ENOMEM; 5071075a61d0SMina Almasry goto out_err; 5072075a61d0SMina Almasry } 5073075a61d0SMina Almasry 5074075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5075075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5076075a61d0SMina Almasry * of the resv_map. 5077075a61d0SMina Almasry */ 5078075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5079075a61d0SMina Almasry } 5080075a61d0SMina Almasry 50811c5ecae3SMike Kravetz /* 50821c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 50831c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 50841c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 50851c5ecae3SMike Kravetz */ 50861c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 50871c5ecae3SMike Kravetz if (gbl_reserve < 0) { 5088c50ac050SDave Hansen ret = -ENOSPC; 5089075a61d0SMina Almasry goto out_uncharge_cgroup; 5090c50ac050SDave Hansen } 509117c9d12eSMel Gorman 509217c9d12eSMel Gorman /* 509317c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 509490481622SDavid Gibson * Hand the pages back to the subpool if there are not 509517c9d12eSMel Gorman */ 50961c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 509717c9d12eSMel Gorman if (ret < 0) { 5098075a61d0SMina Almasry goto out_put_pages; 509917c9d12eSMel Gorman } 510017c9d12eSMel Gorman 510117c9d12eSMel Gorman /* 510217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 510317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 510417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 510517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 510617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 510717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 510817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 510917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 511017c9d12eSMel Gorman * else has to be done for private mappings here 511117c9d12eSMel Gorman */ 511233039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5113075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 511433039678SMike Kravetz 51150db9d74eSMina Almasry if (unlikely(add < 0)) { 51160db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 51177fc2513aSDan Carpenter ret = add; 5118075a61d0SMina Almasry goto out_put_pages; 51190db9d74eSMina Almasry } else if (unlikely(chg > add)) { 512033039678SMike Kravetz /* 512133039678SMike Kravetz * pages in this range were added to the reserve 512233039678SMike Kravetz * map between region_chg and region_add. This 512333039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 512433039678SMike Kravetz * the subpool and reserve counts modified above 512533039678SMike Kravetz * based on the difference. 512633039678SMike Kravetz */ 512733039678SMike Kravetz long rsv_adjust; 512833039678SMike Kravetz 5129075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5130075a61d0SMina Almasry hstate_index(h), 5131075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5132075a61d0SMina Almasry 513333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 513433039678SMike Kravetz chg - add); 513533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 513633039678SMike Kravetz } 513733039678SMike Kravetz } 5138a43a8c39SChen, Kenneth W return 0; 5139075a61d0SMina Almasry out_put_pages: 5140075a61d0SMina Almasry /* put back original number of pages, chg */ 5141075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5142075a61d0SMina Almasry out_uncharge_cgroup: 5143075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5144075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5145c50ac050SDave Hansen out_err: 51465e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 51470db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 51480db9d74eSMina Almasry * region_add failed or didn't run. 51490db9d74eSMina Almasry */ 51500db9d74eSMina Almasry if (chg >= 0 && add < 0) 51510db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5152f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5153f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 5154c50ac050SDave Hansen return ret; 5155a43a8c39SChen, Kenneth W } 5156a43a8c39SChen, Kenneth W 5157b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5158b5cec28dSMike Kravetz long freed) 5159a43a8c39SChen, Kenneth W { 5160a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 51614e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 51629119a41eSJoonsoo Kim long chg = 0; 516390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 51641c5ecae3SMike Kravetz long gbl_reserve; 516545c682a6SKen Chen 5166f27a5136SMike Kravetz /* 5167f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5168f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5169f27a5136SMike Kravetz */ 5170b5cec28dSMike Kravetz if (resv_map) { 5171b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5172b5cec28dSMike Kravetz /* 5173b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5174b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5175b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5176b5cec28dSMike Kravetz */ 5177b5cec28dSMike Kravetz if (chg < 0) 5178b5cec28dSMike Kravetz return chg; 5179b5cec28dSMike Kravetz } 5180b5cec28dSMike Kravetz 518145c682a6SKen Chen spin_lock(&inode->i_lock); 5182e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 518345c682a6SKen Chen spin_unlock(&inode->i_lock); 518445c682a6SKen Chen 51851c5ecae3SMike Kravetz /* 51861c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 51871c5ecae3SMike Kravetz * reservations to be released may be adjusted. 51881c5ecae3SMike Kravetz */ 51891c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 51901c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5191b5cec28dSMike Kravetz 5192b5cec28dSMike Kravetz return 0; 5193a43a8c39SChen, Kenneth W } 519493f70f90SNaoya Horiguchi 51953212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 51963212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 51973212b535SSteve Capper struct vm_area_struct *vma, 51983212b535SSteve Capper unsigned long addr, pgoff_t idx) 51993212b535SSteve Capper { 52003212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52013212b535SSteve Capper svma->vm_start; 52023212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52033212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 52043212b535SSteve Capper 52053212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5206de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5207de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 52083212b535SSteve Capper 52093212b535SSteve Capper /* 52103212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 52113212b535SSteve Capper * page table page. 52123212b535SSteve Capper */ 52133212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 52143212b535SSteve Capper vm_flags != svm_flags || 52153212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 52163212b535SSteve Capper return 0; 52173212b535SSteve Capper 52183212b535SSteve Capper return saddr; 52193212b535SSteve Capper } 52203212b535SSteve Capper 522131aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 52223212b535SSteve Capper { 52233212b535SSteve Capper unsigned long base = addr & PUD_MASK; 52243212b535SSteve Capper unsigned long end = base + PUD_SIZE; 52253212b535SSteve Capper 52263212b535SSteve Capper /* 52273212b535SSteve Capper * check on proper vm_flags and page table alignment 52283212b535SSteve Capper */ 5229017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 523031aafb45SNicholas Krause return true; 523131aafb45SNicholas Krause return false; 52323212b535SSteve Capper } 52333212b535SSteve Capper 52343212b535SSteve Capper /* 5235017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5236017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5237017b1660SMike Kravetz * shared pmd mappings. 5238017b1660SMike Kravetz */ 5239017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5240017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5241017b1660SMike Kravetz { 524275802ca6SPeter Xu unsigned long a_start, a_end; 5243017b1660SMike Kravetz 5244017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 5245017b1660SMike Kravetz return; 5246017b1660SMike Kravetz 524775802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 524875802ca6SPeter Xu a_start = ALIGN_DOWN(*start, PUD_SIZE); 524975802ca6SPeter Xu a_end = ALIGN(*end, PUD_SIZE); 5250017b1660SMike Kravetz 5251017b1660SMike Kravetz /* 525275802ca6SPeter Xu * Intersect the range with the vma range, since pmd sharing won't be 525375802ca6SPeter Xu * across vma after all 5254017b1660SMike Kravetz */ 525575802ca6SPeter Xu *start = max(vma->vm_start, a_start); 525675802ca6SPeter Xu *end = min(vma->vm_end, a_end); 5257017b1660SMike Kravetz } 5258017b1660SMike Kravetz 5259017b1660SMike Kravetz /* 52603212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 52613212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 52623212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5263c0d0381aSMike Kravetz * code much cleaner. 5264c0d0381aSMike Kravetz * 52650bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 52660bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 52670bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 52680bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 52690bf7b64eSMike Kravetz * 52700bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 52710bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 52720bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 52730bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 52740bf7b64eSMike Kravetz * only required for subsequent processing. 52753212b535SSteve Capper */ 52763212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 52773212b535SSteve Capper { 52783212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 52793212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 52803212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 52813212b535SSteve Capper vma->vm_pgoff; 52823212b535SSteve Capper struct vm_area_struct *svma; 52833212b535SSteve Capper unsigned long saddr; 52843212b535SSteve Capper pte_t *spte = NULL; 52853212b535SSteve Capper pte_t *pte; 5286cb900f41SKirill A. Shutemov spinlock_t *ptl; 52873212b535SSteve Capper 52883212b535SSteve Capper if (!vma_shareable(vma, addr)) 52893212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 52903212b535SSteve Capper 52910bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 52923212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 52933212b535SSteve Capper if (svma == vma) 52943212b535SSteve Capper continue; 52953212b535SSteve Capper 52963212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 52973212b535SSteve Capper if (saddr) { 52987868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 52997868a208SPunit Agrawal vma_mmu_pagesize(svma)); 53003212b535SSteve Capper if (spte) { 53013212b535SSteve Capper get_page(virt_to_page(spte)); 53023212b535SSteve Capper break; 53033212b535SSteve Capper } 53043212b535SSteve Capper } 53053212b535SSteve Capper } 53063212b535SSteve Capper 53073212b535SSteve Capper if (!spte) 53083212b535SSteve Capper goto out; 53093212b535SSteve Capper 53108bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5311dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 53123212b535SSteve Capper pud_populate(mm, pud, 53133212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5314c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5315dc6c9a35SKirill A. Shutemov } else { 53163212b535SSteve Capper put_page(virt_to_page(spte)); 5317dc6c9a35SKirill A. Shutemov } 5318cb900f41SKirill A. Shutemov spin_unlock(ptl); 53193212b535SSteve Capper out: 53203212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 53213212b535SSteve Capper return pte; 53223212b535SSteve Capper } 53233212b535SSteve Capper 53243212b535SSteve Capper /* 53253212b535SSteve Capper * unmap huge page backed by shared pte. 53263212b535SSteve Capper * 53273212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 53283212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 53293212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 53303212b535SSteve Capper * 5331c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 53323212b535SSteve Capper * 53333212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 53343212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 53353212b535SSteve Capper */ 533634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 533734ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 53383212b535SSteve Capper { 53393212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5340c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5341c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 53423212b535SSteve Capper 534334ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 53443212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 53453212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 53463212b535SSteve Capper return 0; 53473212b535SSteve Capper 53483212b535SSteve Capper pud_clear(pud); 53493212b535SSteve Capper put_page(virt_to_page(ptep)); 5350dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 53513212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 53523212b535SSteve Capper return 1; 53533212b535SSteve Capper } 53549e5fc74cSSteve Capper #define want_pmd_share() (1) 53559e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 53569e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 53579e5fc74cSSteve Capper { 53589e5fc74cSSteve Capper return NULL; 53599e5fc74cSSteve Capper } 5360e81f2d22SZhang Zhen 536134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 536234ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5363e81f2d22SZhang Zhen { 5364e81f2d22SZhang Zhen return 0; 5365e81f2d22SZhang Zhen } 5366017b1660SMike Kravetz 5367017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5368017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5369017b1660SMike Kravetz { 5370017b1660SMike Kravetz } 53719e5fc74cSSteve Capper #define want_pmd_share() (0) 53723212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 53733212b535SSteve Capper 53749e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 53759e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 53769e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 53779e5fc74cSSteve Capper { 53789e5fc74cSSteve Capper pgd_t *pgd; 5379c2febafcSKirill A. Shutemov p4d_t *p4d; 53809e5fc74cSSteve Capper pud_t *pud; 53819e5fc74cSSteve Capper pte_t *pte = NULL; 53829e5fc74cSSteve Capper 53839e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5384f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5385f4f0a3d8SKirill A. Shutemov if (!p4d) 5386f4f0a3d8SKirill A. Shutemov return NULL; 5387c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 53889e5fc74cSSteve Capper if (pud) { 53899e5fc74cSSteve Capper if (sz == PUD_SIZE) { 53909e5fc74cSSteve Capper pte = (pte_t *)pud; 53919e5fc74cSSteve Capper } else { 53929e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 53939e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 53949e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 53959e5fc74cSSteve Capper else 53969e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 53979e5fc74cSSteve Capper } 53989e5fc74cSSteve Capper } 53994e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 54009e5fc74cSSteve Capper 54019e5fc74cSSteve Capper return pte; 54029e5fc74cSSteve Capper } 54039e5fc74cSSteve Capper 54049b19df29SPunit Agrawal /* 54059b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 54069b19df29SPunit Agrawal * entry at address @addr 54079b19df29SPunit Agrawal * 54088ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 54098ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 54109b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 54119b19df29SPunit Agrawal * table. 54129b19df29SPunit Agrawal */ 54137868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 54147868a208SPunit Agrawal unsigned long addr, unsigned long sz) 54159e5fc74cSSteve Capper { 54169e5fc74cSSteve Capper pgd_t *pgd; 5417c2febafcSKirill A. Shutemov p4d_t *p4d; 54188ac0b81aSLi Xinhai pud_t *pud; 54198ac0b81aSLi Xinhai pmd_t *pmd; 54209e5fc74cSSteve Capper 54219e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5422c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5423c2febafcSKirill A. Shutemov return NULL; 5424c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5425c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5426c2febafcSKirill A. Shutemov return NULL; 54279b19df29SPunit Agrawal 5428c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 54298ac0b81aSLi Xinhai if (sz == PUD_SIZE) 54308ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 54319e5fc74cSSteve Capper return (pte_t *)pud; 54328ac0b81aSLi Xinhai if (!pud_present(*pud)) 54338ac0b81aSLi Xinhai return NULL; 54348ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 54359b19df29SPunit Agrawal 54369e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 54378ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 54389e5fc74cSSteve Capper return (pte_t *)pmd; 54399e5fc74cSSteve Capper } 54409e5fc74cSSteve Capper 544161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 544261f77edaSNaoya Horiguchi 544361f77edaSNaoya Horiguchi /* 544461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 544561f77edaSNaoya Horiguchi * behavior. 544661f77edaSNaoya Horiguchi */ 544761f77edaSNaoya Horiguchi struct page * __weak 544861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 544961f77edaSNaoya Horiguchi int write) 545061f77edaSNaoya Horiguchi { 545161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 545261f77edaSNaoya Horiguchi } 545361f77edaSNaoya Horiguchi 545461f77edaSNaoya Horiguchi struct page * __weak 54554dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 54564dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 54574dc71451SAneesh Kumar K.V { 54584dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 54594dc71451SAneesh Kumar K.V return NULL; 54604dc71451SAneesh Kumar K.V } 54614dc71451SAneesh Kumar K.V 54624dc71451SAneesh Kumar K.V struct page * __weak 54639e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5464e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 54659e5fc74cSSteve Capper { 5466e66f17ffSNaoya Horiguchi struct page *page = NULL; 5467e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5468c9d398faSNaoya Horiguchi pte_t pte; 54693faa52c0SJohn Hubbard 54703faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 54713faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 54723faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 54733faa52c0SJohn Hubbard return NULL; 54743faa52c0SJohn Hubbard 5475e66f17ffSNaoya Horiguchi retry: 5476e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5477e66f17ffSNaoya Horiguchi spin_lock(ptl); 5478e66f17ffSNaoya Horiguchi /* 5479e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5480e66f17ffSNaoya Horiguchi * unmapped from other threads. 5481e66f17ffSNaoya Horiguchi */ 5482e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5483e66f17ffSNaoya Horiguchi goto out; 5484c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5485c9d398faSNaoya Horiguchi if (pte_present(pte)) { 548697534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 54873faa52c0SJohn Hubbard /* 54883faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 54893faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 54903faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 54913faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 54923faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 54933faa52c0SJohn Hubbard * unless the page refcount overflowed: 54943faa52c0SJohn Hubbard */ 54953faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 54963faa52c0SJohn Hubbard page = NULL; 54973faa52c0SJohn Hubbard goto out; 54983faa52c0SJohn Hubbard } 5499e66f17ffSNaoya Horiguchi } else { 5500c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5501e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5502e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5503e66f17ffSNaoya Horiguchi goto retry; 5504e66f17ffSNaoya Horiguchi } 5505e66f17ffSNaoya Horiguchi /* 5506e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5507e66f17ffSNaoya Horiguchi * follow_page_mask(). 5508e66f17ffSNaoya Horiguchi */ 5509e66f17ffSNaoya Horiguchi } 5510e66f17ffSNaoya Horiguchi out: 5511e66f17ffSNaoya Horiguchi spin_unlock(ptl); 55129e5fc74cSSteve Capper return page; 55139e5fc74cSSteve Capper } 55149e5fc74cSSteve Capper 551561f77edaSNaoya Horiguchi struct page * __weak 55169e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5517e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 55189e5fc74cSSteve Capper { 55193faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5520e66f17ffSNaoya Horiguchi return NULL; 55219e5fc74cSSteve Capper 5522e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 55239e5fc74cSSteve Capper } 55249e5fc74cSSteve Capper 5525faaa5b62SAnshuman Khandual struct page * __weak 5526faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5527faaa5b62SAnshuman Khandual { 55283faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5529faaa5b62SAnshuman Khandual return NULL; 5530faaa5b62SAnshuman Khandual 5531faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5532faaa5b62SAnshuman Khandual } 5533faaa5b62SAnshuman Khandual 553431caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 553531caf665SNaoya Horiguchi { 5536bcc54222SNaoya Horiguchi bool ret = true; 5537bcc54222SNaoya Horiguchi 5538309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 553931caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5540bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 5541bcc54222SNaoya Horiguchi ret = false; 5542bcc54222SNaoya Horiguchi goto unlock; 5543bcc54222SNaoya Horiguchi } 5544bcc54222SNaoya Horiguchi clear_page_huge_active(page); 554531caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5546bcc54222SNaoya Horiguchi unlock: 554731caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5548bcc54222SNaoya Horiguchi return ret; 554931caf665SNaoya Horiguchi } 555031caf665SNaoya Horiguchi 555131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 555231caf665SNaoya Horiguchi { 5553309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 555431caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5555bcc54222SNaoya Horiguchi set_page_huge_active(page); 555631caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 555731caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 555831caf665SNaoya Horiguchi put_page(page); 555931caf665SNaoya Horiguchi } 5560ab5ac90aSMichal Hocko 5561ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5562ab5ac90aSMichal Hocko { 5563ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5564ab5ac90aSMichal Hocko 5565ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5566ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5567ab5ac90aSMichal Hocko 5568ab5ac90aSMichal Hocko /* 5569ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5570ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5571ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5572ab5ac90aSMichal Hocko * the temporary status. 5573ab5ac90aSMichal Hocko * 5574ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5575ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5576ab5ac90aSMichal Hocko * the per-node's. 5577ab5ac90aSMichal Hocko */ 5578ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5579ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5580ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5581ab5ac90aSMichal Hocko 5582ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5583ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5584ab5ac90aSMichal Hocko 5585ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5586ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5587ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5588ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5589ab5ac90aSMichal Hocko } 5590ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5591ab5ac90aSMichal Hocko } 5592ab5ac90aSMichal Hocko } 5593cf11e85fSRoman Gushchin 5594cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5595cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5596cf11e85fSRoman Gushchin 5597cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5598cf11e85fSRoman Gushchin { 5599cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5600cf11e85fSRoman Gushchin return 0; 5601cf11e85fSRoman Gushchin } 5602cf11e85fSRoman Gushchin 5603cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5604cf11e85fSRoman Gushchin 5605cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5606cf11e85fSRoman Gushchin { 5607cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5608cf11e85fSRoman Gushchin int nid; 5609cf11e85fSRoman Gushchin 5610cf11e85fSRoman Gushchin cma_reserve_called = true; 5611cf11e85fSRoman Gushchin 5612cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5613cf11e85fSRoman Gushchin return; 5614cf11e85fSRoman Gushchin 5615cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5616cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5617cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5618cf11e85fSRoman Gushchin return; 5619cf11e85fSRoman Gushchin } 5620cf11e85fSRoman Gushchin 5621cf11e85fSRoman Gushchin /* 5622cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5623cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5624cf11e85fSRoman Gushchin */ 5625cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5626cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5627cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5628cf11e85fSRoman Gushchin 5629cf11e85fSRoman Gushchin reserved = 0; 5630cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5631cf11e85fSRoman Gushchin int res; 56322281f797SBarry Song char name[CMA_MAX_NAME]; 5633cf11e85fSRoman Gushchin 5634cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5635cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5636cf11e85fSRoman Gushchin 56372281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5638cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 563929d0f41dSBarry Song 0, false, name, 5640cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5641cf11e85fSRoman Gushchin if (res) { 5642cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5643cf11e85fSRoman Gushchin res, nid); 5644cf11e85fSRoman Gushchin continue; 5645cf11e85fSRoman Gushchin } 5646cf11e85fSRoman Gushchin 5647cf11e85fSRoman Gushchin reserved += size; 5648cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5649cf11e85fSRoman Gushchin size / SZ_1M, nid); 5650cf11e85fSRoman Gushchin 5651cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5652cf11e85fSRoman Gushchin break; 5653cf11e85fSRoman Gushchin } 5654cf11e85fSRoman Gushchin } 5655cf11e85fSRoman Gushchin 5656cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5657cf11e85fSRoman Gushchin { 5658cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5659cf11e85fSRoman Gushchin return; 5660cf11e85fSRoman Gushchin 5661cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5662cf11e85fSRoman Gushchin } 5663cf11e85fSRoman Gushchin 5664cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5665