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> 2263489f8eSMike Kravetz #include <linux/mmdebug.h> 23174cd4b1SIngo Molnar #include <linux/sched/signal.h> 240fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 25c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 26fd6a03edSNaoya Horiguchi #include <linux/swap.h> 27fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 288382d914SDavidlohr Bueso #include <linux/jhash.h> 2998fa15f3SAnshuman Khandual #include <linux/numa.h> 30c77c0a8aSWaiman Long #include <linux/llist.h> 31cf11e85fSRoman Gushchin #include <linux/cma.h> 32d6606683SLinus Torvalds 3363551ae0SDavid Gibson #include <asm/page.h> 3463551ae0SDavid Gibson #include <asm/pgtable.h> 3524669e58SAneesh Kumar K.V #include <asm/tlb.h> 3663551ae0SDavid Gibson 3724669e58SAneesh Kumar K.V #include <linux/io.h> 3863551ae0SDavid Gibson #include <linux/hugetlb.h> 399dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 409a305230SLee Schermerhorn #include <linux/node.h> 411a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 42ab5ac90aSMichal Hocko #include <linux/page_owner.h> 437835e98bSNick Piggin #include "internal.h" 441da177e4SLinus Torvalds 45c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 46e5ff2159SAndi Kleen unsigned int default_hstate_idx; 47e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 48cf11e85fSRoman Gushchin 49cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 50cf11e85fSRoman Gushchin 51641844f5SNaoya Horiguchi /* 52641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 53641844f5SNaoya Horiguchi * at boot time. 54641844f5SNaoya Horiguchi */ 55641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 56e5ff2159SAndi Kleen 5753ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5853ba51d2SJon Tollefson 59e5ff2159SAndi Kleen /* for command line parsing */ 60e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 61e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 629fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 63282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 64e5ff2159SAndi Kleen 653935baa9SDavid Gibson /* 6631caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 6731caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 683935baa9SDavid Gibson */ 69c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 700bd0f9fbSEric Paris 718382d914SDavidlohr Bueso /* 728382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 738382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 748382d914SDavidlohr Bueso */ 758382d914SDavidlohr Bueso static int num_fault_mutexes; 76c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 778382d914SDavidlohr Bueso 787ca02d0aSMike Kravetz /* Forward declaration */ 797ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 807ca02d0aSMike Kravetz 8190481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 8290481622SDavid Gibson { 8390481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 8490481622SDavid Gibson 8590481622SDavid Gibson spin_unlock(&spool->lock); 8690481622SDavid Gibson 8790481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 887c8de358SEthon Paul * remain, give up any reservations based on minimum size and 897ca02d0aSMike Kravetz * free the subpool */ 907ca02d0aSMike Kravetz if (free) { 917ca02d0aSMike Kravetz if (spool->min_hpages != -1) 927ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 937ca02d0aSMike Kravetz -spool->min_hpages); 9490481622SDavid Gibson kfree(spool); 9590481622SDavid Gibson } 967ca02d0aSMike Kravetz } 9790481622SDavid Gibson 987ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 997ca02d0aSMike Kravetz long min_hpages) 10090481622SDavid Gibson { 10190481622SDavid Gibson struct hugepage_subpool *spool; 10290481622SDavid Gibson 103c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 10490481622SDavid Gibson if (!spool) 10590481622SDavid Gibson return NULL; 10690481622SDavid Gibson 10790481622SDavid Gibson spin_lock_init(&spool->lock); 10890481622SDavid Gibson spool->count = 1; 1097ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1107ca02d0aSMike Kravetz spool->hstate = h; 1117ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1127ca02d0aSMike Kravetz 1137ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1147ca02d0aSMike Kravetz kfree(spool); 1157ca02d0aSMike Kravetz return NULL; 1167ca02d0aSMike Kravetz } 1177ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 11890481622SDavid Gibson 11990481622SDavid Gibson return spool; 12090481622SDavid Gibson } 12190481622SDavid Gibson 12290481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 12390481622SDavid Gibson { 12490481622SDavid Gibson spin_lock(&spool->lock); 12590481622SDavid Gibson BUG_ON(!spool->count); 12690481622SDavid Gibson spool->count--; 12790481622SDavid Gibson unlock_or_release_subpool(spool); 12890481622SDavid Gibson } 12990481622SDavid Gibson 1301c5ecae3SMike Kravetz /* 1311c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1321c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1331c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1341c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1351c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1367c8de358SEthon Paul * a subpool minimum size must be maintained. 1371c5ecae3SMike Kravetz */ 1381c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13990481622SDavid Gibson long delta) 14090481622SDavid Gibson { 1411c5ecae3SMike Kravetz long ret = delta; 14290481622SDavid Gibson 14390481622SDavid Gibson if (!spool) 1441c5ecae3SMike Kravetz return ret; 14590481622SDavid Gibson 14690481622SDavid Gibson spin_lock(&spool->lock); 14790481622SDavid Gibson 1481c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1491c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1501c5ecae3SMike Kravetz spool->used_hpages += delta; 1511c5ecae3SMike Kravetz else { 1521c5ecae3SMike Kravetz ret = -ENOMEM; 1531c5ecae3SMike Kravetz goto unlock_ret; 1541c5ecae3SMike Kravetz } 1551c5ecae3SMike Kravetz } 1561c5ecae3SMike Kravetz 15709a95e29SMike Kravetz /* minimum size accounting */ 15809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1591c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1601c5ecae3SMike Kravetz /* 1611c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1621c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1631c5ecae3SMike Kravetz */ 1641c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1651c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1661c5ecae3SMike Kravetz } else { 1671c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1681c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1691c5ecae3SMike Kravetz } 1701c5ecae3SMike Kravetz } 1711c5ecae3SMike Kravetz 1721c5ecae3SMike Kravetz unlock_ret: 1731c5ecae3SMike Kravetz spin_unlock(&spool->lock); 17490481622SDavid Gibson return ret; 17590481622SDavid Gibson } 17690481622SDavid Gibson 1771c5ecae3SMike Kravetz /* 1781c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1791c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1801c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1811c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1821c5ecae3SMike Kravetz */ 1831c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 18490481622SDavid Gibson long delta) 18590481622SDavid Gibson { 1861c5ecae3SMike Kravetz long ret = delta; 1871c5ecae3SMike Kravetz 18890481622SDavid Gibson if (!spool) 1891c5ecae3SMike Kravetz return delta; 19090481622SDavid Gibson 19190481622SDavid Gibson spin_lock(&spool->lock); 1921c5ecae3SMike Kravetz 1931c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 19490481622SDavid Gibson spool->used_hpages -= delta; 1951c5ecae3SMike Kravetz 19609a95e29SMike Kravetz /* minimum size accounting */ 19709a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1981c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1991c5ecae3SMike Kravetz ret = 0; 2001c5ecae3SMike Kravetz else 2011c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2021c5ecae3SMike Kravetz 2031c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2041c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2051c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2061c5ecae3SMike Kravetz } 2071c5ecae3SMike Kravetz 2081c5ecae3SMike Kravetz /* 2091c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2101c5ecae3SMike Kravetz * quota reference, free it now. 2111c5ecae3SMike Kravetz */ 21290481622SDavid Gibson unlock_or_release_subpool(spool); 2131c5ecae3SMike Kravetz 2141c5ecae3SMike Kravetz return ret; 21590481622SDavid Gibson } 21690481622SDavid Gibson 21790481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 21890481622SDavid Gibson { 21990481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 22090481622SDavid Gibson } 22190481622SDavid Gibson 22290481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 22390481622SDavid Gibson { 224496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 22590481622SDavid Gibson } 22690481622SDavid Gibson 2270db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2280db9d74eSMina Almasry * it for use. 2290db9d74eSMina Almasry */ 2300db9d74eSMina Almasry static struct file_region * 2310db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2320db9d74eSMina Almasry { 2330db9d74eSMina Almasry struct file_region *nrg = NULL; 2340db9d74eSMina Almasry 2350db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2360db9d74eSMina Almasry 2370db9d74eSMina Almasry resv->region_cache_count--; 2380db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2390db9d74eSMina Almasry VM_BUG_ON(!nrg); 2400db9d74eSMina Almasry list_del(&nrg->link); 2410db9d74eSMina Almasry 2420db9d74eSMina Almasry nrg->from = from; 2430db9d74eSMina Almasry nrg->to = to; 2440db9d74eSMina Almasry 2450db9d74eSMina Almasry return nrg; 2460db9d74eSMina Almasry } 2470db9d74eSMina Almasry 248075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 249075a61d0SMina Almasry struct file_region *rg) 250075a61d0SMina Almasry { 251075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 252075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 253075a61d0SMina Almasry nrg->css = rg->css; 254075a61d0SMina Almasry if (rg->css) 255075a61d0SMina Almasry css_get(rg->css); 256075a61d0SMina Almasry #endif 257075a61d0SMina Almasry } 258075a61d0SMina Almasry 259075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 260075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 261075a61d0SMina Almasry struct hstate *h, 262075a61d0SMina Almasry struct resv_map *resv, 263075a61d0SMina Almasry struct file_region *nrg) 264075a61d0SMina Almasry { 265075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 266075a61d0SMina Almasry if (h_cg) { 267075a61d0SMina Almasry nrg->reservation_counter = 268075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 269075a61d0SMina Almasry nrg->css = &h_cg->css; 270075a61d0SMina Almasry if (!resv->pages_per_hpage) 271075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 272075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 273075a61d0SMina Almasry * a resv_map. 274075a61d0SMina Almasry */ 275075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 276075a61d0SMina Almasry } else { 277075a61d0SMina Almasry nrg->reservation_counter = NULL; 278075a61d0SMina Almasry nrg->css = NULL; 279075a61d0SMina Almasry } 280075a61d0SMina Almasry #endif 281075a61d0SMina Almasry } 282075a61d0SMina Almasry 283a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 284a9b3f867SMina Almasry struct file_region *org) 285a9b3f867SMina Almasry { 286a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 287a9b3f867SMina Almasry return rg && org && 288a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 289a9b3f867SMina Almasry rg->css == org->css; 290a9b3f867SMina Almasry 291a9b3f867SMina Almasry #else 292a9b3f867SMina Almasry return true; 293a9b3f867SMina Almasry #endif 294a9b3f867SMina Almasry } 295a9b3f867SMina Almasry 296a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 297a9b3f867SMina Almasry { 298a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 299a9b3f867SMina Almasry 300a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 301a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 302a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 303a9b3f867SMina Almasry prg->to = rg->to; 304a9b3f867SMina Almasry 305a9b3f867SMina Almasry list_del(&rg->link); 306a9b3f867SMina Almasry kfree(rg); 307a9b3f867SMina Almasry 308a9b3f867SMina Almasry coalesce_file_region(resv, prg); 309a9b3f867SMina Almasry return; 310a9b3f867SMina Almasry } 311a9b3f867SMina Almasry 312a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 313a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 314a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 315a9b3f867SMina Almasry nrg->from = rg->from; 316a9b3f867SMina Almasry 317a9b3f867SMina Almasry list_del(&rg->link); 318a9b3f867SMina Almasry kfree(rg); 319a9b3f867SMina Almasry 320a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 321a9b3f867SMina Almasry return; 322a9b3f867SMina Almasry } 323a9b3f867SMina Almasry } 324a9b3f867SMina Almasry 325d75c6af9SMina Almasry /* Must be called with resv->lock held. Calling this with count_only == true 326d75c6af9SMina Almasry * will count the number of pages to be added but will not modify the linked 3270db9d74eSMina Almasry * list. If regions_needed != NULL and count_only == true, then regions_needed 3280db9d74eSMina Almasry * will indicate the number of file_regions needed in the cache to carry out to 3290db9d74eSMina Almasry * add the regions for this range. 330d75c6af9SMina Almasry */ 331d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 332075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 333075a61d0SMina Almasry struct hstate *h, long *regions_needed, 334075a61d0SMina Almasry bool count_only) 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; 3700db9d74eSMina Almasry if (!count_only) { 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); 3770db9d74eSMina Almasry } else if (regions_needed) 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; 3890db9d74eSMina Almasry if (!count_only) { 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); 3950db9d74eSMina Almasry } else if (regions_needed) 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 4470db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 4480db9d74eSMina Almasry list_del(&rg->link); 4490db9d74eSMina Almasry list_add(&rg->link, &resv->region_cache); 4500db9d74eSMina Almasry resv->region_cache_count++; 4510db9d74eSMina Almasry } 4520db9d74eSMina Almasry } 4530db9d74eSMina Almasry 4540db9d74eSMina Almasry return 0; 4550db9d74eSMina Almasry 4560db9d74eSMina Almasry out_of_memory: 4570db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 458d75c6af9SMina Almasry list_del(&rg->link); 459d75c6af9SMina Almasry kfree(rg); 460d75c6af9SMina Almasry } 4610db9d74eSMina Almasry return -ENOMEM; 462d75c6af9SMina Almasry } 463d75c6af9SMina Almasry 4641dd308a7SMike Kravetz /* 4651dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 4660db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 4670db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 4680db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 4690db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 4700db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 471cf3ad20bSMike Kravetz * 4720db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 4730db9d74eSMina Almasry * 4740db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 4750db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 4767c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 4770db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 4780db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 4790db9d74eSMina Almasry * 1 page will only require at most 1 entry. 4801dd308a7SMike Kravetz */ 4810db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 482075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 483075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 48496822904SAndy Whitcroft { 4850db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 48696822904SAndy Whitcroft 4877b24d861SDavidlohr Bueso spin_lock(&resv->lock); 4880db9d74eSMina Almasry retry: 4890db9d74eSMina Almasry 4900db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 491075a61d0SMina Almasry add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed, 492075a61d0SMina Almasry true); 49396822904SAndy Whitcroft 4945e911373SMike Kravetz /* 4950db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4960db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 4970db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 4980db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 4990db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5000db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5010db9d74eSMina Almasry * operation. 5025e911373SMike Kravetz */ 5030db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5040db9d74eSMina Almasry resv->region_cache_count < 5050db9d74eSMina Almasry resv->adds_in_progress + 5060db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5070db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5080db9d74eSMina Almasry * allocate file_region entries. 5090db9d74eSMina Almasry */ 5100db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5115e911373SMike Kravetz 5120db9d74eSMina Almasry if (allocate_file_region_entries( 5130db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5140db9d74eSMina Almasry return -ENOMEM; 5155e911373SMike Kravetz } 5165e911373SMike Kravetz 5170db9d74eSMina Almasry goto retry; 5180db9d74eSMina Almasry } 519cf3ad20bSMike Kravetz 520075a61d0SMina Almasry add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false); 5210db9d74eSMina Almasry 5220db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5230db9d74eSMina Almasry 5247b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 525cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 526cf3ad20bSMike Kravetz return add; 52796822904SAndy Whitcroft } 52896822904SAndy Whitcroft 5291dd308a7SMike Kravetz /* 5301dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5311dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5321dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5331dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5341dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5351dd308a7SMike Kravetz * not change the number of huge pages represented by the 5360db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5370db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5380db9d74eSMina Almasry * file_region structure is added. 5390db9d74eSMina Almasry * 5400db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5410db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5420db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5435e911373SMike Kravetz * 5445e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5455e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5465e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5475e911373SMike Kravetz * is needed and can not be allocated. 5481dd308a7SMike Kravetz */ 5490db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5500db9d74eSMina Almasry long *out_regions_needed) 55196822904SAndy Whitcroft { 55296822904SAndy Whitcroft long chg = 0; 55396822904SAndy Whitcroft 5547b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5555e911373SMike Kravetz 5560db9d74eSMina Almasry /* Count how many hugepages in this range are NOT respresented. */ 557075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 558075a61d0SMina Almasry out_regions_needed, true); 5595e911373SMike Kravetz 5600db9d74eSMina Almasry if (*out_regions_needed == 0) 5610db9d74eSMina Almasry *out_regions_needed = 1; 5625e911373SMike Kravetz 5630db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5645e911373SMike Kravetz return -ENOMEM; 5655e911373SMike Kravetz 5660db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 56796822904SAndy Whitcroft 5687b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 56996822904SAndy Whitcroft return chg; 57096822904SAndy Whitcroft } 57196822904SAndy Whitcroft 5721dd308a7SMike Kravetz /* 5735e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 5745e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 5755e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 5765e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 5770db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 5780db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 5790db9d74eSMina Almasry * the adds_in_progress counter. 5805e911373SMike Kravetz * 5815e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 5825e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 5835e911373SMike Kravetz * arguments will match the associated region_chg call. 5845e911373SMike Kravetz */ 5850db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 5860db9d74eSMina Almasry long regions_needed) 5875e911373SMike Kravetz { 5885e911373SMike Kravetz spin_lock(&resv->lock); 5895e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 5900db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 5915e911373SMike Kravetz spin_unlock(&resv->lock); 5925e911373SMike Kravetz } 5935e911373SMike Kravetz 5945e911373SMike Kravetz /* 595feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 596feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 597feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 598feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 599feba16e2SMike Kravetz * 600feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 601feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 602feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 603feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 604feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 605feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 606feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6071dd308a7SMike Kravetz */ 608feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 60996822904SAndy Whitcroft { 6101406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 61196822904SAndy Whitcroft struct file_region *rg, *trg; 612feba16e2SMike Kravetz struct file_region *nrg = NULL; 613feba16e2SMike Kravetz long del = 0; 61496822904SAndy Whitcroft 615feba16e2SMike Kravetz retry: 6167b24d861SDavidlohr Bueso spin_lock(&resv->lock); 617feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 618dbe409e4SMike Kravetz /* 619dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 620dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 621dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 622dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 623dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 624dbe409e4SMike Kravetz */ 625dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 626feba16e2SMike Kravetz continue; 627dbe409e4SMike Kravetz 628feba16e2SMike Kravetz if (rg->from >= t) 62996822904SAndy Whitcroft break; 63096822904SAndy Whitcroft 631feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 632feba16e2SMike Kravetz /* 633feba16e2SMike Kravetz * Check for an entry in the cache before dropping 634feba16e2SMike Kravetz * lock and attempting allocation. 635feba16e2SMike Kravetz */ 636feba16e2SMike Kravetz if (!nrg && 637feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 638feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 639feba16e2SMike Kravetz struct file_region, 640feba16e2SMike Kravetz link); 641feba16e2SMike Kravetz list_del(&nrg->link); 642feba16e2SMike Kravetz resv->region_cache_count--; 64396822904SAndy Whitcroft } 64496822904SAndy Whitcroft 645feba16e2SMike Kravetz if (!nrg) { 646feba16e2SMike Kravetz spin_unlock(&resv->lock); 647feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 648feba16e2SMike Kravetz if (!nrg) 649feba16e2SMike Kravetz return -ENOMEM; 650feba16e2SMike Kravetz goto retry; 651feba16e2SMike Kravetz } 652feba16e2SMike Kravetz 653feba16e2SMike Kravetz del += t - f; 654feba16e2SMike Kravetz 655feba16e2SMike Kravetz /* New entry for end of split region */ 656feba16e2SMike Kravetz nrg->from = t; 657feba16e2SMike Kravetz nrg->to = rg->to; 658075a61d0SMina Almasry 659075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 660075a61d0SMina Almasry 661feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 662feba16e2SMike Kravetz 663feba16e2SMike Kravetz /* Original entry is trimmed */ 664feba16e2SMike Kravetz rg->to = f; 665feba16e2SMike Kravetz 666075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region( 667075a61d0SMina Almasry resv, rg, nrg->to - nrg->from); 668075a61d0SMina Almasry 669feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 670feba16e2SMike Kravetz nrg = NULL; 67196822904SAndy Whitcroft break; 672feba16e2SMike Kravetz } 673feba16e2SMike Kravetz 674feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 675feba16e2SMike Kravetz del += rg->to - rg->from; 676075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 677075a61d0SMina Almasry rg->to - rg->from); 67896822904SAndy Whitcroft list_del(&rg->link); 67996822904SAndy Whitcroft kfree(rg); 680feba16e2SMike Kravetz continue; 68196822904SAndy Whitcroft } 6827b24d861SDavidlohr Bueso 683feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 684feba16e2SMike Kravetz del += t - rg->from; 685feba16e2SMike Kravetz rg->from = t; 686075a61d0SMina Almasry 687075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 688075a61d0SMina Almasry t - rg->from); 689feba16e2SMike Kravetz } else { /* Trim end of region */ 690feba16e2SMike Kravetz del += rg->to - f; 691feba16e2SMike Kravetz rg->to = f; 692075a61d0SMina Almasry 693075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 694075a61d0SMina Almasry rg->to - f); 695feba16e2SMike Kravetz } 696feba16e2SMike Kravetz } 697feba16e2SMike Kravetz 6987b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 699feba16e2SMike Kravetz kfree(nrg); 700feba16e2SMike Kravetz return del; 70196822904SAndy Whitcroft } 70296822904SAndy Whitcroft 7031dd308a7SMike Kravetz /* 704b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 705b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 706b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 707b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 708b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 709b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 710b5cec28dSMike Kravetz * counts. 711b5cec28dSMike Kravetz */ 71272e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 713b5cec28dSMike Kravetz { 714b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 715b5cec28dSMike Kravetz long rsv_adjust; 716b5cec28dSMike Kravetz 717b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 71872e2936cSzhong jiang if (rsv_adjust) { 719b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 720b5cec28dSMike Kravetz 721b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 722b5cec28dSMike Kravetz } 723b5cec28dSMike Kravetz } 724b5cec28dSMike Kravetz 725b5cec28dSMike Kravetz /* 7261dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7271dd308a7SMike Kravetz * that intersect with the range [f, t). 7281dd308a7SMike Kravetz */ 7291406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 73084afd99bSAndy Whitcroft { 7311406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 73284afd99bSAndy Whitcroft struct file_region *rg; 73384afd99bSAndy Whitcroft long chg = 0; 73484afd99bSAndy Whitcroft 7357b24d861SDavidlohr Bueso spin_lock(&resv->lock); 73684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 73784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 738f2135a4aSWang Sheng-Hui long seg_from; 739f2135a4aSWang Sheng-Hui long seg_to; 74084afd99bSAndy Whitcroft 74184afd99bSAndy Whitcroft if (rg->to <= f) 74284afd99bSAndy Whitcroft continue; 74384afd99bSAndy Whitcroft if (rg->from >= t) 74484afd99bSAndy Whitcroft break; 74584afd99bSAndy Whitcroft 74684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 74784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 74884afd99bSAndy Whitcroft 74984afd99bSAndy Whitcroft chg += seg_to - seg_from; 75084afd99bSAndy Whitcroft } 7517b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 75284afd99bSAndy Whitcroft 75384afd99bSAndy Whitcroft return chg; 75484afd99bSAndy Whitcroft } 75584afd99bSAndy Whitcroft 75696822904SAndy Whitcroft /* 757e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 758e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 759e7c4b0bfSAndy Whitcroft */ 760a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 761a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 762e7c4b0bfSAndy Whitcroft { 763a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 764a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 765e7c4b0bfSAndy Whitcroft } 766e7c4b0bfSAndy Whitcroft 7670fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 7680fe6e20bSNaoya Horiguchi unsigned long address) 7690fe6e20bSNaoya Horiguchi { 7700fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 7710fe6e20bSNaoya Horiguchi } 772dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 7730fe6e20bSNaoya Horiguchi 77484afd99bSAndy Whitcroft /* 77508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 77608fba699SMel Gorman * cases this will be same size as used by the page table entries. 77708fba699SMel Gorman */ 77808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 77908fba699SMel Gorman { 78005ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 78105ea8860SDan Williams return vma->vm_ops->pagesize(vma); 78208fba699SMel Gorman return PAGE_SIZE; 78308fba699SMel Gorman } 784f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 78508fba699SMel Gorman 78608fba699SMel Gorman /* 7873340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 7883340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 78909135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 79009135cc5SDan Williams * version of this symbol is required. 7913340289dSMel Gorman */ 79209135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 7933340289dSMel Gorman { 7943340289dSMel Gorman return vma_kernel_pagesize(vma); 7953340289dSMel Gorman } 7963340289dSMel Gorman 7973340289dSMel Gorman /* 79884afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 79984afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 80084afd99bSAndy Whitcroft * alignment. 80184afd99bSAndy Whitcroft */ 80284afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 80384afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 80404f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 80584afd99bSAndy Whitcroft 806a1e78772SMel Gorman /* 807a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 808a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 809a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 810a1e78772SMel Gorman * 811a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 812a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 813a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 814a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 81584afd99bSAndy Whitcroft * 81684afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 81784afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 81884afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 81984afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 82084afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 82184afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 82284afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 82384afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 824a1e78772SMel Gorman */ 825e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 826e7c4b0bfSAndy Whitcroft { 827e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 828e7c4b0bfSAndy Whitcroft } 829e7c4b0bfSAndy Whitcroft 830e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 831e7c4b0bfSAndy Whitcroft unsigned long value) 832e7c4b0bfSAndy Whitcroft { 833e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 834e7c4b0bfSAndy Whitcroft } 835e7c4b0bfSAndy Whitcroft 836e9fe92aeSMina Almasry static void 837e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 838e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 839e9fe92aeSMina Almasry struct hstate *h) 840e9fe92aeSMina Almasry { 841e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 842e9fe92aeSMina Almasry if (!h_cg || !h) { 843e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 844e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 845e9fe92aeSMina Almasry resv_map->css = NULL; 846e9fe92aeSMina Almasry } else { 847e9fe92aeSMina Almasry resv_map->reservation_counter = 848e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 849e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 850e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 851e9fe92aeSMina Almasry } 852e9fe92aeSMina Almasry #endif 853e9fe92aeSMina Almasry } 854e9fe92aeSMina Almasry 8559119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 85684afd99bSAndy Whitcroft { 85784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8585e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8595e911373SMike Kravetz 8605e911373SMike Kravetz if (!resv_map || !rg) { 8615e911373SMike Kravetz kfree(resv_map); 8625e911373SMike Kravetz kfree(rg); 86384afd99bSAndy Whitcroft return NULL; 8645e911373SMike Kravetz } 86584afd99bSAndy Whitcroft 86684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 8677b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 86884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 86984afd99bSAndy Whitcroft 8705e911373SMike Kravetz resv_map->adds_in_progress = 0; 871e9fe92aeSMina Almasry /* 872e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 873e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 874e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 875e9fe92aeSMina Almasry * reservations are to be un-charged from here. 876e9fe92aeSMina Almasry */ 877e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 8785e911373SMike Kravetz 8795e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 8805e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 8815e911373SMike Kravetz resv_map->region_cache_count = 1; 8825e911373SMike Kravetz 88384afd99bSAndy Whitcroft return resv_map; 88484afd99bSAndy Whitcroft } 88584afd99bSAndy Whitcroft 8869119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 88784afd99bSAndy Whitcroft { 88884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 8895e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 8905e911373SMike Kravetz struct file_region *rg, *trg; 89184afd99bSAndy Whitcroft 89284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 893feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 8945e911373SMike Kravetz 8955e911373SMike Kravetz /* ... and any entries left in the cache */ 8965e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 8975e911373SMike Kravetz list_del(&rg->link); 8985e911373SMike Kravetz kfree(rg); 8995e911373SMike Kravetz } 9005e911373SMike Kravetz 9015e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9025e911373SMike Kravetz 90384afd99bSAndy Whitcroft kfree(resv_map); 90484afd99bSAndy Whitcroft } 90584afd99bSAndy Whitcroft 9064e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9074e35f483SJoonsoo Kim { 908f27a5136SMike Kravetz /* 909f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 910f27a5136SMike Kravetz * address space within the inode. This original address space 911f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 912f27a5136SMike Kravetz * address space embedded within the inode. 913f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 914f27a5136SMike Kravetz * this may not be true for device special inodes. 915f27a5136SMike Kravetz */ 916f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9174e35f483SJoonsoo Kim } 9184e35f483SJoonsoo Kim 91984afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 920a1e78772SMel Gorman { 92181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9224e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9234e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9244e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9254e35f483SJoonsoo Kim 9264e35f483SJoonsoo Kim return inode_resv_map(inode); 9274e35f483SJoonsoo Kim 9284e35f483SJoonsoo Kim } else { 92984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 93084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9314e35f483SJoonsoo Kim } 932a1e78772SMel Gorman } 933a1e78772SMel Gorman 93484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 935a1e78772SMel Gorman { 93681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 93781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 938a1e78772SMel Gorman 93984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 94084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 94104f2cbe3SMel Gorman } 94204f2cbe3SMel Gorman 94304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 94404f2cbe3SMel Gorman { 94581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 94681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 947e7c4b0bfSAndy Whitcroft 948e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 94904f2cbe3SMel Gorman } 95004f2cbe3SMel Gorman 95104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 95204f2cbe3SMel Gorman { 95381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 954e7c4b0bfSAndy Whitcroft 955e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 956a1e78772SMel Gorman } 957a1e78772SMel Gorman 95804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 959a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 960a1e78772SMel Gorman { 96181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 962f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 963a1e78772SMel Gorman vma->vm_private_data = (void *)0; 964a1e78772SMel Gorman } 965a1e78772SMel Gorman 966a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 967559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 968a1e78772SMel Gorman { 969af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 970af0ed73eSJoonsoo Kim /* 971af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 972af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 973af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 974af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 975af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 976af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 977af0ed73eSJoonsoo Kim * properly, so add work-around here. 978af0ed73eSJoonsoo Kim */ 979af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 980559ec2f8SNicholas Krause return true; 981af0ed73eSJoonsoo Kim else 982559ec2f8SNicholas Krause return false; 983af0ed73eSJoonsoo Kim } 984a63884e9SJoonsoo Kim 985a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 9861fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 9871fb1b0e9SMike Kravetz /* 9881fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 9891fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 9901fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 9917c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 9921fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 9931fb1b0e9SMike Kravetz */ 9941fb1b0e9SMike Kravetz if (chg) 9951fb1b0e9SMike Kravetz return false; 9961fb1b0e9SMike Kravetz else 997559ec2f8SNicholas Krause return true; 9981fb1b0e9SMike Kravetz } 999a63884e9SJoonsoo Kim 1000a63884e9SJoonsoo Kim /* 1001a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1002a63884e9SJoonsoo Kim * private mappings. 1003a63884e9SJoonsoo Kim */ 100467961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 100567961f9dSMike Kravetz /* 100667961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 100767961f9dSMike Kravetz * could have been performed on the private mapping. 100867961f9dSMike Kravetz * Examine the value of chg to determine if reserves 100967961f9dSMike Kravetz * actually exist or were previously consumed. 101067961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 101167961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 101267961f9dSMike Kravetz * private mappings has different (opposite) semantics 101367961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 101467961f9dSMike Kravetz * has already taken this difference in semantics into 101567961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 101667961f9dSMike Kravetz * as in the shared case above. Code could easily be 101767961f9dSMike Kravetz * combined, but keeping it separate draws attention to 101867961f9dSMike Kravetz * subtle differences. 101967961f9dSMike Kravetz */ 102067961f9dSMike Kravetz if (chg) 102167961f9dSMike Kravetz return false; 102267961f9dSMike Kravetz else 1023559ec2f8SNicholas Krause return true; 102467961f9dSMike Kravetz } 1025a63884e9SJoonsoo Kim 1026559ec2f8SNicholas Krause return false; 1027a1e78772SMel Gorman } 1028a1e78772SMel Gorman 1029a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10301da177e4SLinus Torvalds { 10311da177e4SLinus Torvalds int nid = page_to_nid(page); 10320edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1033a5516438SAndi Kleen h->free_huge_pages++; 1034a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10351da177e4SLinus Torvalds } 10361da177e4SLinus Torvalds 103794310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1038bf50bab2SNaoya Horiguchi { 1039bf50bab2SNaoya Horiguchi struct page *page; 1040bf50bab2SNaoya Horiguchi 1041c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 1042243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 1043c8721bbbSNaoya Horiguchi break; 1044c8721bbbSNaoya Horiguchi /* 1045c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 1046c8721bbbSNaoya Horiguchi * the allocation fails. 1047c8721bbbSNaoya Horiguchi */ 1048c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 1049bf50bab2SNaoya Horiguchi return NULL; 10500edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1051a9869b83SNaoya Horiguchi set_page_refcounted(page); 1052bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1053bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1054bf50bab2SNaoya Horiguchi return page; 1055bf50bab2SNaoya Horiguchi } 1056bf50bab2SNaoya Horiguchi 10573e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10583e59fcb0SMichal Hocko nodemask_t *nmask) 105994310cbcSAnshuman Khandual { 10603e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 10613e59fcb0SMichal Hocko struct zonelist *zonelist; 10623e59fcb0SMichal Hocko struct zone *zone; 10633e59fcb0SMichal Hocko struct zoneref *z; 106498fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 10653e59fcb0SMichal Hocko 10663e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 10673e59fcb0SMichal Hocko 10683e59fcb0SMichal Hocko retry_cpuset: 10693e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 10703e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 107194310cbcSAnshuman Khandual struct page *page; 107294310cbcSAnshuman Khandual 10733e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 10743e59fcb0SMichal Hocko continue; 10753e59fcb0SMichal Hocko /* 10763e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 10773e59fcb0SMichal Hocko * zone aware 10783e59fcb0SMichal Hocko */ 10793e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 10803e59fcb0SMichal Hocko continue; 10813e59fcb0SMichal Hocko node = zone_to_nid(zone); 108294310cbcSAnshuman Khandual 108394310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 108494310cbcSAnshuman Khandual if (page) 108594310cbcSAnshuman Khandual return page; 108694310cbcSAnshuman Khandual } 10873e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 10883e59fcb0SMichal Hocko goto retry_cpuset; 10893e59fcb0SMichal Hocko 109094310cbcSAnshuman Khandual return NULL; 109194310cbcSAnshuman Khandual } 109294310cbcSAnshuman Khandual 109386cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 109486cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 109586cdb465SNaoya Horiguchi { 10967ed2c31dSAnshuman Khandual if (hugepage_movable_supported(h)) 109786cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 109886cdb465SNaoya Horiguchi else 109986cdb465SNaoya Horiguchi return GFP_HIGHUSER; 110086cdb465SNaoya Horiguchi } 110186cdb465SNaoya Horiguchi 1102a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1103a5516438SAndi Kleen struct vm_area_struct *vma, 1104af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1105af0ed73eSJoonsoo Kim long chg) 11061da177e4SLinus Torvalds { 11073e59fcb0SMichal Hocko struct page *page; 1108480eccf9SLee Schermerhorn struct mempolicy *mpol; 110904ec6264SVlastimil Babka gfp_t gfp_mask; 11103e59fcb0SMichal Hocko nodemask_t *nodemask; 111104ec6264SVlastimil Babka int nid; 11121da177e4SLinus Torvalds 1113a1e78772SMel Gorman /* 1114a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1115a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1116a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1117a1e78772SMel Gorman */ 1118af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1119a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1120c0ff7453SMiao Xie goto err; 1121a1e78772SMel Gorman 112204f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1123a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11246eab04a8SJustin P. Mattock goto err; 112504f2cbe3SMel Gorman 112604ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 112704ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11283e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11293e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 113007443a85SJoonsoo Kim SetPagePrivate(page); 1131a63884e9SJoonsoo Kim h->resv_huge_pages--; 1132bf50bab2SNaoya Horiguchi } 1133cc9a6c87SMel Gorman 1134cc9a6c87SMel Gorman mpol_cond_put(mpol); 1135cc9a6c87SMel Gorman return page; 1136cc9a6c87SMel Gorman 1137c0ff7453SMiao Xie err: 1138cc9a6c87SMel Gorman return NULL; 11391da177e4SLinus Torvalds } 11401da177e4SLinus Torvalds 11411cac6f2cSLuiz Capitulino /* 11421cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11431cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11441cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11451cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11461cac6f2cSLuiz Capitulino * node for alloc or free. 11471cac6f2cSLuiz Capitulino */ 11481cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11491cac6f2cSLuiz Capitulino { 11500edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11511cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11521cac6f2cSLuiz Capitulino 11531cac6f2cSLuiz Capitulino return nid; 11541cac6f2cSLuiz Capitulino } 11551cac6f2cSLuiz Capitulino 11561cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11571cac6f2cSLuiz Capitulino { 11581cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11591cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11601cac6f2cSLuiz Capitulino return nid; 11611cac6f2cSLuiz Capitulino } 11621cac6f2cSLuiz Capitulino 11631cac6f2cSLuiz Capitulino /* 11641cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11651cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11661cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11671cac6f2cSLuiz Capitulino * mask. 11681cac6f2cSLuiz Capitulino */ 11691cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 11701cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 11711cac6f2cSLuiz Capitulino { 11721cac6f2cSLuiz Capitulino int nid; 11731cac6f2cSLuiz Capitulino 11741cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11751cac6f2cSLuiz Capitulino 11761cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 11771cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 11781cac6f2cSLuiz Capitulino 11791cac6f2cSLuiz Capitulino return nid; 11801cac6f2cSLuiz Capitulino } 11811cac6f2cSLuiz Capitulino 11821cac6f2cSLuiz Capitulino /* 11831cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 11841cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 11851cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 11861cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 11871cac6f2cSLuiz Capitulino */ 11881cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 11891cac6f2cSLuiz Capitulino { 11901cac6f2cSLuiz Capitulino int nid; 11911cac6f2cSLuiz Capitulino 11921cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11931cac6f2cSLuiz Capitulino 11941cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 11951cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 11961cac6f2cSLuiz Capitulino 11971cac6f2cSLuiz Capitulino return nid; 11981cac6f2cSLuiz Capitulino } 11991cac6f2cSLuiz Capitulino 12001cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12011cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12021cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12031cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12041cac6f2cSLuiz Capitulino nr_nodes--) 12051cac6f2cSLuiz Capitulino 12061cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12071cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12081cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12091cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12101cac6f2cSLuiz Capitulino nr_nodes--) 12111cac6f2cSLuiz Capitulino 1212e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1213944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1214d00181b9SKirill A. Shutemov unsigned int order) 1215944d9fecSLuiz Capitulino { 1216944d9fecSLuiz Capitulino int i; 1217944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1218944d9fecSLuiz Capitulino struct page *p = page + 1; 1219944d9fecSLuiz Capitulino 1220c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 122147e29d32SJohn Hubbard if (hpage_pincount_available(page)) 122247e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 122347e29d32SJohn Hubbard 1224944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12251d798ca3SKirill A. Shutemov clear_compound_head(p); 1226944d9fecSLuiz Capitulino set_page_refcounted(p); 1227944d9fecSLuiz Capitulino } 1228944d9fecSLuiz Capitulino 1229944d9fecSLuiz Capitulino set_compound_order(page, 0); 1230944d9fecSLuiz Capitulino __ClearPageHead(page); 1231944d9fecSLuiz Capitulino } 1232944d9fecSLuiz Capitulino 1233d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1234944d9fecSLuiz Capitulino { 1235cf11e85fSRoman Gushchin /* 1236cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1237cf11e85fSRoman Gushchin * cma_release() returns false. 1238cf11e85fSRoman Gushchin */ 1239cf11e85fSRoman Gushchin if (IS_ENABLED(CONFIG_CMA) && 1240cf11e85fSRoman Gushchin cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1241cf11e85fSRoman Gushchin return; 1242cf11e85fSRoman Gushchin 1243944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1244944d9fecSLuiz Capitulino } 1245944d9fecSLuiz Capitulino 12464eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1247d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1248d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1249944d9fecSLuiz Capitulino { 12505e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1251944d9fecSLuiz Capitulino 1252cf11e85fSRoman Gushchin if (IS_ENABLED(CONFIG_CMA)) { 1253cf11e85fSRoman Gushchin struct page *page; 1254cf11e85fSRoman Gushchin int node; 1255cf11e85fSRoman Gushchin 1256cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1257cf11e85fSRoman Gushchin if (!hugetlb_cma[node]) 1258cf11e85fSRoman Gushchin continue; 1259cf11e85fSRoman Gushchin 1260cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1261cf11e85fSRoman Gushchin huge_page_order(h), true); 1262cf11e85fSRoman Gushchin if (page) 1263cf11e85fSRoman Gushchin return page; 1264cf11e85fSRoman Gushchin } 1265cf11e85fSRoman Gushchin } 1266cf11e85fSRoman Gushchin 12675e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1268944d9fecSLuiz Capitulino } 1269944d9fecSLuiz Capitulino 1270944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1271d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 12724eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 12734eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 12744eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 12754eb0716eSAlexandre Ghiti { 12764eb0716eSAlexandre Ghiti return NULL; 12774eb0716eSAlexandre Ghiti } 12784eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1279944d9fecSLuiz Capitulino 1280e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1281d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 12824eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 12834eb0716eSAlexandre Ghiti { 12844eb0716eSAlexandre Ghiti return NULL; 12854eb0716eSAlexandre Ghiti } 1286d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1287944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1288d00181b9SKirill A. Shutemov unsigned int order) { } 1289944d9fecSLuiz Capitulino #endif 1290944d9fecSLuiz Capitulino 1291a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 12926af2acb6SAdam Litke { 12936af2acb6SAdam Litke int i; 1294a5516438SAndi Kleen 12954eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1296944d9fecSLuiz Capitulino return; 129718229df5SAndy Whitcroft 1298a5516438SAndi Kleen h->nr_huge_pages--; 1299a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1300a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 130132f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 130232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1303a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1304a7407a27SLuiz Capitulino 1 << PG_writeback); 13056af2acb6SAdam Litke } 1306309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13071adc4d41SMina Almasry VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1308f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13096af2acb6SAdam Litke set_page_refcounted(page); 1310944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1311cf11e85fSRoman Gushchin /* 1312cf11e85fSRoman Gushchin * Temporarily drop the hugetlb_lock, because 1313cf11e85fSRoman Gushchin * we might block in free_gigantic_page(). 1314cf11e85fSRoman Gushchin */ 1315cf11e85fSRoman Gushchin spin_unlock(&hugetlb_lock); 1316944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1317944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1318cf11e85fSRoman Gushchin spin_lock(&hugetlb_lock); 1319944d9fecSLuiz Capitulino } else { 1320a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13216af2acb6SAdam Litke } 1322944d9fecSLuiz Capitulino } 13236af2acb6SAdam Litke 1324e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1325e5ff2159SAndi Kleen { 1326e5ff2159SAndi Kleen struct hstate *h; 1327e5ff2159SAndi Kleen 1328e5ff2159SAndi Kleen for_each_hstate(h) { 1329e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1330e5ff2159SAndi Kleen return h; 1331e5ff2159SAndi Kleen } 1332e5ff2159SAndi Kleen return NULL; 1333e5ff2159SAndi Kleen } 1334e5ff2159SAndi Kleen 1335bcc54222SNaoya Horiguchi /* 1336bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1337bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1338bcc54222SNaoya Horiguchi * 1339bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1340bcc54222SNaoya Horiguchi */ 1341bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1342bcc54222SNaoya Horiguchi { 1343bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1344bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1345bcc54222SNaoya Horiguchi } 1346bcc54222SNaoya Horiguchi 1347bcc54222SNaoya Horiguchi /* never called for tail page */ 1348bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1349bcc54222SNaoya Horiguchi { 1350bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1351bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1352bcc54222SNaoya Horiguchi } 1353bcc54222SNaoya Horiguchi 1354bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1355bcc54222SNaoya Horiguchi { 1356bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1357bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1358bcc54222SNaoya Horiguchi } 1359bcc54222SNaoya Horiguchi 1360ab5ac90aSMichal Hocko /* 1361ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1362ab5ac90aSMichal Hocko * code 1363ab5ac90aSMichal Hocko */ 1364ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1365ab5ac90aSMichal Hocko { 1366ab5ac90aSMichal Hocko if (!PageHuge(page)) 1367ab5ac90aSMichal Hocko return false; 1368ab5ac90aSMichal Hocko 1369ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1370ab5ac90aSMichal Hocko } 1371ab5ac90aSMichal Hocko 1372ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1373ab5ac90aSMichal Hocko { 1374ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1375ab5ac90aSMichal Hocko } 1376ab5ac90aSMichal Hocko 1377ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1378ab5ac90aSMichal Hocko { 1379ab5ac90aSMichal Hocko page[2].mapping = NULL; 1380ab5ac90aSMichal Hocko } 1381ab5ac90aSMichal Hocko 1382c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 138327a85ef1SDavid Gibson { 1384a5516438SAndi Kleen /* 1385a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1386a5516438SAndi Kleen * compound page destructor. 1387a5516438SAndi Kleen */ 1388e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 13897893d1d5SAdam Litke int nid = page_to_nid(page); 139090481622SDavid Gibson struct hugepage_subpool *spool = 139190481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 139207443a85SJoonsoo Kim bool restore_reserve; 139327a85ef1SDavid Gibson 1394b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1395b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 13968ace22bcSYongkai Wu 13978ace22bcSYongkai Wu set_page_private(page, 0); 13988ace22bcSYongkai Wu page->mapping = NULL; 139907443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 140016c794b4SJoonsoo Kim ClearPagePrivate(page); 140127a85ef1SDavid Gibson 14021c5ecae3SMike Kravetz /* 14030919e1b6SMike Kravetz * If PagePrivate() was set on page, page allocation consumed a 14040919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14050919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14060919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14070919e1b6SMike Kravetz * reservtion, do not call hugepage_subpool_put_pages() as this will 14080919e1b6SMike Kravetz * remove the reserved page from the subpool. 14090919e1b6SMike Kravetz */ 14100919e1b6SMike Kravetz if (!restore_reserve) { 14110919e1b6SMike Kravetz /* 14120919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14130919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14140919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14150919e1b6SMike Kravetz * operation. 14161c5ecae3SMike Kravetz */ 14171c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14181c5ecae3SMike Kravetz restore_reserve = true; 14190919e1b6SMike Kravetz } 14201c5ecae3SMike Kravetz 142127a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1422bcc54222SNaoya Horiguchi clear_page_huge_active(page); 14236d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14246d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 142508cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 142608cf9fafSMina Almasry pages_per_huge_page(h), page); 142707443a85SJoonsoo Kim if (restore_reserve) 142807443a85SJoonsoo Kim h->resv_huge_pages++; 142907443a85SJoonsoo Kim 1430ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1431ab5ac90aSMichal Hocko list_del(&page->lru); 1432ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1433ab5ac90aSMichal Hocko update_and_free_page(h, page); 1434ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14350edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14360edaecfaSAneesh Kumar K.V list_del(&page->lru); 1437a5516438SAndi Kleen update_and_free_page(h, page); 1438a5516438SAndi Kleen h->surplus_huge_pages--; 1439a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 14407893d1d5SAdam Litke } else { 14415d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1442a5516438SAndi Kleen enqueue_huge_page(h, page); 14437893d1d5SAdam Litke } 144427a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 144527a85ef1SDavid Gibson } 144627a85ef1SDavid Gibson 1447c77c0a8aSWaiman Long /* 1448c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1449c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1450c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1451c77c0a8aSWaiman Long * 1452c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1453c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1454c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1455c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1456c77c0a8aSWaiman Long */ 1457c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1458c77c0a8aSWaiman Long 1459c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1460c77c0a8aSWaiman Long { 1461c77c0a8aSWaiman Long struct llist_node *node; 1462c77c0a8aSWaiman Long struct page *page; 1463c77c0a8aSWaiman Long 1464c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1465c77c0a8aSWaiman Long 1466c77c0a8aSWaiman Long while (node) { 1467c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1468c77c0a8aSWaiman Long struct page, mapping); 1469c77c0a8aSWaiman Long node = node->next; 1470c77c0a8aSWaiman Long __free_huge_page(page); 1471c77c0a8aSWaiman Long } 1472c77c0a8aSWaiman Long } 1473c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1474c77c0a8aSWaiman Long 1475c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1476c77c0a8aSWaiman Long { 1477c77c0a8aSWaiman Long /* 1478c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1479c77c0a8aSWaiman Long */ 1480c77c0a8aSWaiman Long if (!in_task()) { 1481c77c0a8aSWaiman Long /* 1482c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1483c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1484c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1485c77c0a8aSWaiman Long */ 1486c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1487c77c0a8aSWaiman Long &hpage_freelist)) 1488c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1489c77c0a8aSWaiman Long return; 1490c77c0a8aSWaiman Long } 1491c77c0a8aSWaiman Long 1492c77c0a8aSWaiman Long __free_huge_page(page); 1493c77c0a8aSWaiman Long } 1494c77c0a8aSWaiman Long 1495a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1496b7ba30c6SAndi Kleen { 14970edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1498f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1499b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 15009dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15011adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1502a5516438SAndi Kleen h->nr_huge_pages++; 1503a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1504b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1505b7ba30c6SAndi Kleen } 1506b7ba30c6SAndi Kleen 1507d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 150820a0307cSWu Fengguang { 150920a0307cSWu Fengguang int i; 151020a0307cSWu Fengguang int nr_pages = 1 << order; 151120a0307cSWu Fengguang struct page *p = page + 1; 151220a0307cSWu Fengguang 151320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 151420a0307cSWu Fengguang set_compound_order(page, order); 1515ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1516de09d31dSKirill A. Shutemov __SetPageHead(page); 151720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1518ef5a22beSAndrea Arcangeli /* 1519ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1520ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1521ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15227c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1523ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1524ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1525ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1526ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1527ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1528ef5a22beSAndrea Arcangeli * after get_user_pages(). 1529ef5a22beSAndrea Arcangeli */ 1530ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 153158a84aa9SYouquan Song set_page_count(p, 0); 15321d798ca3SKirill A. Shutemov set_compound_head(p, page); 153320a0307cSWu Fengguang } 1534b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 153547e29d32SJohn Hubbard 153647e29d32SJohn Hubbard if (hpage_pincount_available(page)) 153747e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 153820a0307cSWu Fengguang } 153920a0307cSWu Fengguang 15407795912cSAndrew Morton /* 15417795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15427795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15437795912cSAndrew Morton * details. 15447795912cSAndrew Morton */ 154520a0307cSWu Fengguang int PageHuge(struct page *page) 154620a0307cSWu Fengguang { 154720a0307cSWu Fengguang if (!PageCompound(page)) 154820a0307cSWu Fengguang return 0; 154920a0307cSWu Fengguang 155020a0307cSWu Fengguang page = compound_head(page); 1551f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 155220a0307cSWu Fengguang } 155343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 155443131e14SNaoya Horiguchi 155527c73ae7SAndrea Arcangeli /* 155627c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 155727c73ae7SAndrea Arcangeli * normal or transparent huge pages. 155827c73ae7SAndrea Arcangeli */ 155927c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 156027c73ae7SAndrea Arcangeli { 156127c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 156227c73ae7SAndrea Arcangeli return 0; 156327c73ae7SAndrea Arcangeli 1564d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 156527c73ae7SAndrea Arcangeli } 156627c73ae7SAndrea Arcangeli 1567c0d0381aSMike Kravetz /* 1568c0d0381aSMike Kravetz * Find address_space associated with hugetlbfs page. 1569c0d0381aSMike Kravetz * Upon entry page is locked and page 'was' mapped although mapped state 1570c0d0381aSMike Kravetz * could change. If necessary, use anon_vma to find vma and associated 1571c0d0381aSMike Kravetz * address space. The returned mapping may be stale, but it can not be 1572c0d0381aSMike Kravetz * invalid as page lock (which is held) is required to destroy mapping. 1573c0d0381aSMike Kravetz */ 1574c0d0381aSMike Kravetz static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) 1575c0d0381aSMike Kravetz { 1576c0d0381aSMike Kravetz struct anon_vma *anon_vma; 1577c0d0381aSMike Kravetz pgoff_t pgoff_start, pgoff_end; 1578c0d0381aSMike Kravetz struct anon_vma_chain *avc; 1579c0d0381aSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1580c0d0381aSMike Kravetz 1581c0d0381aSMike Kravetz /* Simple file based mapping */ 1582c0d0381aSMike Kravetz if (mapping) 1583c0d0381aSMike Kravetz return mapping; 1584c0d0381aSMike Kravetz 1585c0d0381aSMike Kravetz /* 1586c0d0381aSMike Kravetz * Even anonymous hugetlbfs mappings are associated with an 1587c0d0381aSMike Kravetz * underlying hugetlbfs file (see hugetlb_file_setup in mmap 1588c0d0381aSMike Kravetz * code). Find a vma associated with the anonymous vma, and 1589c0d0381aSMike Kravetz * use the file pointer to get address_space. 1590c0d0381aSMike Kravetz */ 1591c0d0381aSMike Kravetz anon_vma = page_lock_anon_vma_read(hpage); 1592c0d0381aSMike Kravetz if (!anon_vma) 1593c0d0381aSMike Kravetz return mapping; /* NULL */ 1594c0d0381aSMike Kravetz 1595c0d0381aSMike Kravetz /* Use first found vma */ 1596c0d0381aSMike Kravetz pgoff_start = page_to_pgoff(hpage); 1597c0d0381aSMike Kravetz pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1; 1598c0d0381aSMike Kravetz anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 1599c0d0381aSMike Kravetz pgoff_start, pgoff_end) { 1600c0d0381aSMike Kravetz struct vm_area_struct *vma = avc->vma; 1601c0d0381aSMike Kravetz 1602c0d0381aSMike Kravetz mapping = vma->vm_file->f_mapping; 1603c0d0381aSMike Kravetz break; 1604c0d0381aSMike Kravetz } 1605c0d0381aSMike Kravetz 1606c0d0381aSMike Kravetz anon_vma_unlock_read(anon_vma); 1607c0d0381aSMike Kravetz return mapping; 1608c0d0381aSMike Kravetz } 1609c0d0381aSMike Kravetz 1610c0d0381aSMike Kravetz /* 1611c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1612c0d0381aSMike Kravetz * 1613c0d0381aSMike Kravetz * Upon entry, the page is locked which allows us to find the mapping 1614c0d0381aSMike Kravetz * even in the case of an anon page. However, locking order dictates 1615c0d0381aSMike Kravetz * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs 1616c0d0381aSMike Kravetz * specific. So, we first try to lock the sema while still holding the 1617c0d0381aSMike Kravetz * page lock. If this works, great! If not, then we need to drop the 1618c0d0381aSMike Kravetz * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of 1619c0d0381aSMike Kravetz * course, need to revalidate state along the way. 1620c0d0381aSMike Kravetz */ 1621c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1622c0d0381aSMike Kravetz { 1623c0d0381aSMike Kravetz struct address_space *mapping, *mapping2; 1624c0d0381aSMike Kravetz 1625c0d0381aSMike Kravetz mapping = _get_hugetlb_page_mapping(hpage); 1626c0d0381aSMike Kravetz retry: 1627c0d0381aSMike Kravetz if (!mapping) 1628c0d0381aSMike Kravetz return mapping; 1629c0d0381aSMike Kravetz 1630c0d0381aSMike Kravetz /* 1631c0d0381aSMike Kravetz * If no contention, take lock and return 1632c0d0381aSMike Kravetz */ 1633c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1634c0d0381aSMike Kravetz return mapping; 1635c0d0381aSMike Kravetz 1636c0d0381aSMike Kravetz /* 1637c0d0381aSMike Kravetz * Must drop page lock and wait on mapping sema. 1638c0d0381aSMike Kravetz * Note: Once page lock is dropped, mapping could become invalid. 1639c0d0381aSMike Kravetz * As a hack, increase map count until we lock page again. 1640c0d0381aSMike Kravetz */ 1641c0d0381aSMike Kravetz atomic_inc(&hpage->_mapcount); 1642c0d0381aSMike Kravetz unlock_page(hpage); 1643c0d0381aSMike Kravetz i_mmap_lock_write(mapping); 1644c0d0381aSMike Kravetz lock_page(hpage); 1645c0d0381aSMike Kravetz atomic_add_negative(-1, &hpage->_mapcount); 1646c0d0381aSMike Kravetz 1647c0d0381aSMike Kravetz /* verify page is still mapped */ 1648c0d0381aSMike Kravetz if (!page_mapped(hpage)) { 1649c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1650c0d0381aSMike Kravetz return NULL; 1651c0d0381aSMike Kravetz } 1652c0d0381aSMike Kravetz 1653c0d0381aSMike Kravetz /* 1654c0d0381aSMike Kravetz * Get address space again and verify it is the same one 1655c0d0381aSMike Kravetz * we locked. If not, drop lock and retry. 1656c0d0381aSMike Kravetz */ 1657c0d0381aSMike Kravetz mapping2 = _get_hugetlb_page_mapping(hpage); 1658c0d0381aSMike Kravetz if (mapping2 != mapping) { 1659c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1660c0d0381aSMike Kravetz mapping = mapping2; 1661c0d0381aSMike Kravetz goto retry; 1662c0d0381aSMike Kravetz } 1663c0d0381aSMike Kravetz 1664c0d0381aSMike Kravetz return mapping; 1665c0d0381aSMike Kravetz } 1666c0d0381aSMike Kravetz 166713d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 166813d60f4bSZhang Yi { 166913d60f4bSZhang Yi struct page *page_head = compound_head(page); 167013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 167113d60f4bSZhang Yi unsigned long compound_idx; 167213d60f4bSZhang Yi 167313d60f4bSZhang Yi if (!PageHuge(page_head)) 167413d60f4bSZhang Yi return page_index(page); 167513d60f4bSZhang Yi 167613d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 167713d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 167813d60f4bSZhang Yi else 167913d60f4bSZhang Yi compound_idx = page - page_head; 168013d60f4bSZhang Yi 168113d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 168213d60f4bSZhang Yi } 168313d60f4bSZhang Yi 16840c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1685f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1686f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16871da177e4SLinus Torvalds { 1688af0fb9dfSMichal Hocko int order = huge_page_order(h); 16891da177e4SLinus Torvalds struct page *page; 1690f60858f9SMike Kravetz bool alloc_try_hard = true; 1691f96efd58SJoe Jin 1692f60858f9SMike Kravetz /* 1693f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1694f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1695f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1696f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1697f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1698f60858f9SMike Kravetz */ 1699f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1700f60858f9SMike Kravetz alloc_try_hard = false; 1701f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1702f60858f9SMike Kravetz if (alloc_try_hard) 1703f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1704af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1705af0fb9dfSMichal Hocko nid = numa_mem_id(); 1706af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1707af0fb9dfSMichal Hocko if (page) 1708af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1709af0fb9dfSMichal Hocko else 1710af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 171163b4613cSNishanth Aravamudan 1712f60858f9SMike Kravetz /* 1713f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1714f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1715f60858f9SMike Kravetz * normal 'try hard' allocations. 1716f60858f9SMike Kravetz */ 1717f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1718f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1719f60858f9SMike Kravetz 1720f60858f9SMike Kravetz /* 1721f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1722f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1723f60858f9SMike Kravetz * overall state change. 1724f60858f9SMike Kravetz */ 1725f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1726f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1727f60858f9SMike Kravetz 172863b4613cSNishanth Aravamudan return page; 172963b4613cSNishanth Aravamudan } 173063b4613cSNishanth Aravamudan 1731af0fb9dfSMichal Hocko /* 17320c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 17330c397daeSMichal Hocko * should use this function to get new hugetlb pages 17340c397daeSMichal Hocko */ 17350c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1736f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1737f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17380c397daeSMichal Hocko { 17390c397daeSMichal Hocko struct page *page; 17400c397daeSMichal Hocko 17410c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17420c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 17430c397daeSMichal Hocko else 17440c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1745f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 17460c397daeSMichal Hocko if (!page) 17470c397daeSMichal Hocko return NULL; 17480c397daeSMichal Hocko 17490c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17500c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 17510c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 17520c397daeSMichal Hocko 17530c397daeSMichal Hocko return page; 17540c397daeSMichal Hocko } 17550c397daeSMichal Hocko 17560c397daeSMichal Hocko /* 1757af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1758af0fb9dfSMichal Hocko * manner. 1759af0fb9dfSMichal Hocko */ 1760f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1761f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1762b2261026SJoonsoo Kim { 1763b2261026SJoonsoo Kim struct page *page; 1764b2261026SJoonsoo Kim int nr_nodes, node; 1765af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1766b2261026SJoonsoo Kim 1767b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1768f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1769f60858f9SMike Kravetz node_alloc_noretry); 1770af0fb9dfSMichal Hocko if (page) 1771b2261026SJoonsoo Kim break; 1772b2261026SJoonsoo Kim } 1773b2261026SJoonsoo Kim 1774af0fb9dfSMichal Hocko if (!page) 1775af0fb9dfSMichal Hocko return 0; 1776b2261026SJoonsoo Kim 1777af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1778af0fb9dfSMichal Hocko 1779af0fb9dfSMichal Hocko return 1; 1780b2261026SJoonsoo Kim } 1781b2261026SJoonsoo Kim 1782e8c5c824SLee Schermerhorn /* 1783e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1784e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1785e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1786e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1787e8c5c824SLee Schermerhorn */ 17886ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17896ae11b27SLee Schermerhorn bool acct_surplus) 1790e8c5c824SLee Schermerhorn { 1791b2261026SJoonsoo Kim int nr_nodes, node; 1792e8c5c824SLee Schermerhorn int ret = 0; 1793e8c5c824SLee Schermerhorn 1794b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1795685f3457SLee Schermerhorn /* 1796685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1797685f3457SLee Schermerhorn * nodes with surplus pages. 1798685f3457SLee Schermerhorn */ 1799b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1800b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1801e8c5c824SLee Schermerhorn struct page *page = 1802b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1803e8c5c824SLee Schermerhorn struct page, lru); 1804e8c5c824SLee Schermerhorn list_del(&page->lru); 1805e8c5c824SLee Schermerhorn h->free_huge_pages--; 1806b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1807685f3457SLee Schermerhorn if (acct_surplus) { 1808685f3457SLee Schermerhorn h->surplus_huge_pages--; 1809b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1810685f3457SLee Schermerhorn } 1811e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1812e8c5c824SLee Schermerhorn ret = 1; 18139a76db09SLee Schermerhorn break; 1814e8c5c824SLee Schermerhorn } 1815b2261026SJoonsoo Kim } 1816e8c5c824SLee Schermerhorn 1817e8c5c824SLee Schermerhorn return ret; 1818e8c5c824SLee Schermerhorn } 1819e8c5c824SLee Schermerhorn 1820c8721bbbSNaoya Horiguchi /* 1821c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1822faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1823faf53defSNaoya Horiguchi * This function returns values like below: 1824faf53defSNaoya Horiguchi * 1825faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1826faf53defSNaoya Horiguchi * (allocated or reserved.) 1827faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1828faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1829c8721bbbSNaoya Horiguchi */ 1830c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1831c8721bbbSNaoya Horiguchi { 18326bc9b564SNaoya Horiguchi int rc = -EBUSY; 1833082d5b6bSGerald Schaefer 1834faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1835faf53defSNaoya Horiguchi if (!PageHuge(page)) 1836faf53defSNaoya Horiguchi return 0; 1837faf53defSNaoya Horiguchi 1838c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1839faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1840faf53defSNaoya Horiguchi rc = 0; 1841faf53defSNaoya Horiguchi goto out; 1842faf53defSNaoya Horiguchi } 1843faf53defSNaoya Horiguchi 1844faf53defSNaoya Horiguchi if (!page_count(page)) { 18452247bb33SGerald Schaefer struct page *head = compound_head(page); 18462247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 18472247bb33SGerald Schaefer int nid = page_to_nid(head); 18486bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1849082d5b6bSGerald Schaefer goto out; 1850c3114a84SAnshuman Khandual /* 1851c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1852c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1853c3114a84SAnshuman Khandual */ 1854c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1855c3114a84SAnshuman Khandual SetPageHWPoison(page); 1856c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1857c3114a84SAnshuman Khandual } 18582247bb33SGerald Schaefer list_del(&head->lru); 1859c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1860c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1861c1470b33Szhong jiang h->max_huge_pages--; 18622247bb33SGerald Schaefer update_and_free_page(h, head); 18636bc9b564SNaoya Horiguchi rc = 0; 1864c8721bbbSNaoya Horiguchi } 1865082d5b6bSGerald Schaefer out: 1866c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1867082d5b6bSGerald Schaefer return rc; 1868c8721bbbSNaoya Horiguchi } 1869c8721bbbSNaoya Horiguchi 1870c8721bbbSNaoya Horiguchi /* 1871c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1872c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18732247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18742247bb33SGerald Schaefer * part of it lies within the given range. 1875082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1876082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1877c8721bbbSNaoya Horiguchi */ 1878082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1879c8721bbbSNaoya Horiguchi { 1880c8721bbbSNaoya Horiguchi unsigned long pfn; 1881eb03aa00SGerald Schaefer struct page *page; 1882082d5b6bSGerald Schaefer int rc = 0; 1883c8721bbbSNaoya Horiguchi 1884d0177639SLi Zhong if (!hugepages_supported()) 1885082d5b6bSGerald Schaefer return rc; 1886d0177639SLi Zhong 1887eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1888eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1889eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1890eb03aa00SGerald Schaefer if (rc) 1891082d5b6bSGerald Schaefer break; 1892eb03aa00SGerald Schaefer } 1893082d5b6bSGerald Schaefer 1894082d5b6bSGerald Schaefer return rc; 1895c8721bbbSNaoya Horiguchi } 1896c8721bbbSNaoya Horiguchi 1897ab5ac90aSMichal Hocko /* 1898ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1899ab5ac90aSMichal Hocko */ 19000c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1901aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 19027893d1d5SAdam Litke { 19039980d744SMichal Hocko struct page *page = NULL; 19047893d1d5SAdam Litke 1905bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1906aa888a74SAndi Kleen return NULL; 1907aa888a74SAndi Kleen 1908d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 19099980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 19109980d744SMichal Hocko goto out_unlock; 1911d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1912d1c3fb1fSNishanth Aravamudan 1913f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 19149980d744SMichal Hocko if (!page) 19150c397daeSMichal Hocko return NULL; 1916d1c3fb1fSNishanth Aravamudan 19177893d1d5SAdam Litke spin_lock(&hugetlb_lock); 19189980d744SMichal Hocko /* 19199980d744SMichal Hocko * We could have raced with the pool size change. 19209980d744SMichal Hocko * Double check that and simply deallocate the new page 19219980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 19229980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 19239980d744SMichal Hocko * codeflow 19249980d744SMichal Hocko */ 19259980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 19269980d744SMichal Hocko SetPageHugeTemporary(page); 19272bf753e6SKai Shen spin_unlock(&hugetlb_lock); 19289980d744SMichal Hocko put_page(page); 19292bf753e6SKai Shen return NULL; 19309980d744SMichal Hocko } else { 19319980d744SMichal Hocko h->surplus_huge_pages++; 19324704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 19337893d1d5SAdam Litke } 19349980d744SMichal Hocko 19359980d744SMichal Hocko out_unlock: 1936d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 19377893d1d5SAdam Litke 19387893d1d5SAdam Litke return page; 19397893d1d5SAdam Litke } 19407893d1d5SAdam Litke 19419a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1942ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1943ab5ac90aSMichal Hocko { 1944ab5ac90aSMichal Hocko struct page *page; 1945ab5ac90aSMichal Hocko 1946ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1947ab5ac90aSMichal Hocko return NULL; 1948ab5ac90aSMichal Hocko 1949f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1950ab5ac90aSMichal Hocko if (!page) 1951ab5ac90aSMichal Hocko return NULL; 1952ab5ac90aSMichal Hocko 1953ab5ac90aSMichal Hocko /* 1954ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1955ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1956ab5ac90aSMichal Hocko */ 1957ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1958ab5ac90aSMichal Hocko 1959ab5ac90aSMichal Hocko return page; 1960ab5ac90aSMichal Hocko } 1961ab5ac90aSMichal Hocko 1962e4e574b7SAdam Litke /* 1963099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1964099730d6SDave Hansen */ 1965e0ec90eeSDave Hansen static 19660c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1967099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1968099730d6SDave Hansen { 1969aaf14e40SMichal Hocko struct page *page; 1970aaf14e40SMichal Hocko struct mempolicy *mpol; 1971aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1972aaf14e40SMichal Hocko int nid; 1973aaf14e40SMichal Hocko nodemask_t *nodemask; 1974aaf14e40SMichal Hocko 1975aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19760c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1977aaf14e40SMichal Hocko mpol_cond_put(mpol); 1978aaf14e40SMichal Hocko 1979aaf14e40SMichal Hocko return page; 1980099730d6SDave Hansen } 1981099730d6SDave Hansen 1982ab5ac90aSMichal Hocko /* page migration callback function */ 1983bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1984bf50bab2SNaoya Horiguchi { 1985aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 19864ef91848SJoonsoo Kim struct page *page = NULL; 1987bf50bab2SNaoya Horiguchi 1988aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1989aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1990aaf14e40SMichal Hocko 1991bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 19924ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 19933e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1994bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1995bf50bab2SNaoya Horiguchi 199694ae8ba7SAneesh Kumar K.V if (!page) 19970c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1998bf50bab2SNaoya Horiguchi 1999bf50bab2SNaoya Horiguchi return page; 2000bf50bab2SNaoya Horiguchi } 2001bf50bab2SNaoya Horiguchi 2002ab5ac90aSMichal Hocko /* page migration callback function */ 20033e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 20043e59fcb0SMichal Hocko nodemask_t *nmask) 20054db9b2efSMichal Hocko { 2006aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 20074db9b2efSMichal Hocko 20084db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 20094db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 20103e59fcb0SMichal Hocko struct page *page; 20113e59fcb0SMichal Hocko 20123e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 20133e59fcb0SMichal Hocko if (page) { 20143e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 20153e59fcb0SMichal Hocko return page; 20164db9b2efSMichal Hocko } 20174db9b2efSMichal Hocko } 20184db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 20194db9b2efSMichal Hocko 20200c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 20214db9b2efSMichal Hocko } 20224db9b2efSMichal Hocko 2023ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2024389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2025389c8178SMichal Hocko unsigned long address) 2026ebd63723SMichal Hocko { 2027ebd63723SMichal Hocko struct mempolicy *mpol; 2028ebd63723SMichal Hocko nodemask_t *nodemask; 2029ebd63723SMichal Hocko struct page *page; 2030ebd63723SMichal Hocko gfp_t gfp_mask; 2031ebd63723SMichal Hocko int node; 2032ebd63723SMichal Hocko 2033ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2034ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2035ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 2036ebd63723SMichal Hocko mpol_cond_put(mpol); 2037ebd63723SMichal Hocko 2038ebd63723SMichal Hocko return page; 2039ebd63723SMichal Hocko } 2040ebd63723SMichal Hocko 2041bf50bab2SNaoya Horiguchi /* 204225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2043e4e574b7SAdam Litke * of size 'delta'. 2044e4e574b7SAdam Litke */ 2045a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 20461b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2047e4e574b7SAdam Litke { 2048e4e574b7SAdam Litke struct list_head surplus_list; 2049e4e574b7SAdam Litke struct page *page, *tmp; 2050e4e574b7SAdam Litke int ret, i; 2051e4e574b7SAdam Litke int needed, allocated; 205228073b02SHillf Danton bool alloc_ok = true; 2053e4e574b7SAdam Litke 2054a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2055ac09b3a1SAdam Litke if (needed <= 0) { 2056a5516438SAndi Kleen h->resv_huge_pages += delta; 2057e4e574b7SAdam Litke return 0; 2058ac09b3a1SAdam Litke } 2059e4e574b7SAdam Litke 2060e4e574b7SAdam Litke allocated = 0; 2061e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2062e4e574b7SAdam Litke 2063e4e574b7SAdam Litke ret = -ENOMEM; 2064e4e574b7SAdam Litke retry: 2065e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 2066e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 20670c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2068aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 206928073b02SHillf Danton if (!page) { 207028073b02SHillf Danton alloc_ok = false; 207128073b02SHillf Danton break; 207228073b02SHillf Danton } 2073e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 207469ed779aSDavid Rientjes cond_resched(); 2075e4e574b7SAdam Litke } 207628073b02SHillf Danton allocated += i; 2077e4e574b7SAdam Litke 2078e4e574b7SAdam Litke /* 2079e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2080e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2081e4e574b7SAdam Litke */ 2082e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 2083a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2084a5516438SAndi Kleen (h->free_huge_pages + allocated); 208528073b02SHillf Danton if (needed > 0) { 208628073b02SHillf Danton if (alloc_ok) 2087e4e574b7SAdam Litke goto retry; 208828073b02SHillf Danton /* 208928073b02SHillf Danton * We were not able to allocate enough pages to 209028073b02SHillf Danton * satisfy the entire reservation so we free what 209128073b02SHillf Danton * we've allocated so far. 209228073b02SHillf Danton */ 209328073b02SHillf Danton goto free; 209428073b02SHillf Danton } 2095e4e574b7SAdam Litke /* 2096e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 209725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2098e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2099ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2100ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2101ac09b3a1SAdam Litke * before they are reserved. 2102e4e574b7SAdam Litke */ 2103e4e574b7SAdam Litke needed += allocated; 2104a5516438SAndi Kleen h->resv_huge_pages += delta; 2105e4e574b7SAdam Litke ret = 0; 2106a9869b83SNaoya Horiguchi 210719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 210819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 210919fc3f0aSAdam Litke if ((--needed) < 0) 211019fc3f0aSAdam Litke break; 2111a9869b83SNaoya Horiguchi /* 2112a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2113a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2114a9869b83SNaoya Horiguchi */ 2115a9869b83SNaoya Horiguchi put_page_testzero(page); 2116309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 2117a5516438SAndi Kleen enqueue_huge_page(h, page); 211819fc3f0aSAdam Litke } 211928073b02SHillf Danton free: 2120b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 212119fc3f0aSAdam Litke 212219fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2123c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2124a9869b83SNaoya Horiguchi put_page(page); 212519fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2126e4e574b7SAdam Litke 2127e4e574b7SAdam Litke return ret; 2128e4e574b7SAdam Litke } 2129e4e574b7SAdam Litke 2130e4e574b7SAdam Litke /* 2131e5bbc8a6SMike Kravetz * This routine has two main purposes: 2132e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2133e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2134e5bbc8a6SMike Kravetz * to the associated reservation map. 2135e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2136e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2137e5bbc8a6SMike Kravetz * 2138e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2139e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2140e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2141e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2142e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2143e4e574b7SAdam Litke */ 2144a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2145a5516438SAndi Kleen unsigned long unused_resv_pages) 2146e4e574b7SAdam Litke { 2147e4e574b7SAdam Litke unsigned long nr_pages; 2148e4e574b7SAdam Litke 2149aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2150bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2151e5bbc8a6SMike Kravetz goto out; 2152aa888a74SAndi Kleen 2153e5bbc8a6SMike Kravetz /* 2154e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2155e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2156e5bbc8a6SMike Kravetz */ 2157a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2158e4e574b7SAdam Litke 2159685f3457SLee Schermerhorn /* 2160685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 21619b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 21629b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 21639b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 21649b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 21659b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2166e5bbc8a6SMike Kravetz * 2167e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2168e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2169e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2170685f3457SLee Schermerhorn */ 2171685f3457SLee Schermerhorn while (nr_pages--) { 2172e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2173e5bbc8a6SMike Kravetz unused_resv_pages--; 21748cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2175e5bbc8a6SMike Kravetz goto out; 21767848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2177e4e574b7SAdam Litke } 2178e5bbc8a6SMike Kravetz 2179e5bbc8a6SMike Kravetz out: 2180e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2181e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2182e4e574b7SAdam Litke } 2183e4e574b7SAdam Litke 21845e911373SMike Kravetz 2185c37f9fb1SAndy Whitcroft /* 2186feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21875e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2188cf3ad20bSMike Kravetz * 2189cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2190cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2191cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2192cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2193cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2194feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2195feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2196feba16e2SMike Kravetz * is called in such cases. 2197cf3ad20bSMike Kravetz * 2198cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2199cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2200cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2201cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2202cf3ad20bSMike Kravetz * take appropriate action. 220396b96a96SMike Kravetz * 220496b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 220596b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 220696b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 220796b96a96SMike Kravetz * reservation exists. 2208c37f9fb1SAndy Whitcroft */ 22095e911373SMike Kravetz enum vma_resv_mode { 22105e911373SMike Kravetz VMA_NEEDS_RESV, 22115e911373SMike Kravetz VMA_COMMIT_RESV, 2212feba16e2SMike Kravetz VMA_END_RESV, 221396b96a96SMike Kravetz VMA_ADD_RESV, 22145e911373SMike Kravetz }; 2215cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2216cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 22175e911373SMike Kravetz enum vma_resv_mode mode) 2218c37f9fb1SAndy Whitcroft { 22194e35f483SJoonsoo Kim struct resv_map *resv; 22204e35f483SJoonsoo Kim pgoff_t idx; 2221cf3ad20bSMike Kravetz long ret; 22220db9d74eSMina Almasry long dummy_out_regions_needed; 2223c37f9fb1SAndy Whitcroft 22244e35f483SJoonsoo Kim resv = vma_resv_map(vma); 22254e35f483SJoonsoo Kim if (!resv) 2226c37f9fb1SAndy Whitcroft return 1; 2227c37f9fb1SAndy Whitcroft 22284e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 22295e911373SMike Kravetz switch (mode) { 22305e911373SMike Kravetz case VMA_NEEDS_RESV: 22310db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 22320db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 22330db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 22340db9d74eSMina Almasry * ever require 1 region. 22350db9d74eSMina Almasry */ 22360db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 22375e911373SMike Kravetz break; 22385e911373SMike Kravetz case VMA_COMMIT_RESV: 2239075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22400db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22410db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22425e911373SMike Kravetz break; 2243feba16e2SMike Kravetz case VMA_END_RESV: 22440db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 22455e911373SMike Kravetz ret = 0; 22465e911373SMike Kravetz break; 224796b96a96SMike Kravetz case VMA_ADD_RESV: 22480db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2249075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22500db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22510db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22520db9d74eSMina Almasry } else { 22530db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 225496b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 225596b96a96SMike Kravetz } 225696b96a96SMike Kravetz break; 22575e911373SMike Kravetz default: 22585e911373SMike Kravetz BUG(); 22595e911373SMike Kravetz } 226084afd99bSAndy Whitcroft 22614e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2262cf3ad20bSMike Kravetz return ret; 226367961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 226467961f9dSMike Kravetz /* 226567961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 226667961f9dSMike Kravetz * However, a file associated with mapping could have been 226767961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 226867961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 226967961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 227067961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 227167961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 227267961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 227367961f9dSMike Kravetz * reservation has already been consumed. As a result, the 227467961f9dSMike Kravetz * return value of this routine is the opposite of the 227567961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 227667961f9dSMike Kravetz */ 227767961f9dSMike Kravetz if (ret) 227867961f9dSMike Kravetz return 0; 227967961f9dSMike Kravetz else 228067961f9dSMike Kravetz return 1; 228167961f9dSMike Kravetz } 22824e35f483SJoonsoo Kim else 2283cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 228484afd99bSAndy Whitcroft } 2285cf3ad20bSMike Kravetz 2286cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2287a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2288c37f9fb1SAndy Whitcroft { 22895e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2290cf3ad20bSMike Kravetz } 2291c37f9fb1SAndy Whitcroft 2292cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2293cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2294cf3ad20bSMike Kravetz { 22955e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22965e911373SMike Kravetz } 22975e911373SMike Kravetz 2298feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22995e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 23005e911373SMike Kravetz { 2301feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2302c37f9fb1SAndy Whitcroft } 2303c37f9fb1SAndy Whitcroft 230496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 230596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 230696b96a96SMike Kravetz { 230796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 230896b96a96SMike Kravetz } 230996b96a96SMike Kravetz 231096b96a96SMike Kravetz /* 231196b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 231296b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 231396b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 231496b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 231596b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 231696b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 231796b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 231896b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 231996b96a96SMike Kravetz * to be made by free_huge_page. 232096b96a96SMike Kravetz */ 232196b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 232296b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 232396b96a96SMike Kravetz struct page *page) 232496b96a96SMike Kravetz { 232596b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 232696b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 232796b96a96SMike Kravetz 232896b96a96SMike Kravetz if (unlikely(rc < 0)) { 232996b96a96SMike Kravetz /* 233096b96a96SMike Kravetz * Rare out of memory condition in reserve map 233196b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 233296b96a96SMike Kravetz * global reserve count will not be incremented 233396b96a96SMike Kravetz * by free_huge_page. This will make it appear 233496b96a96SMike Kravetz * as though the reservation for this page was 233596b96a96SMike Kravetz * consumed. This may prevent the task from 233696b96a96SMike Kravetz * faulting in the page at a later time. This 233796b96a96SMike Kravetz * is better than inconsistent global huge page 233896b96a96SMike Kravetz * accounting of reserve counts. 233996b96a96SMike Kravetz */ 234096b96a96SMike Kravetz ClearPagePrivate(page); 234196b96a96SMike Kravetz } else if (rc) { 234296b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 234396b96a96SMike Kravetz if (unlikely(rc < 0)) 234496b96a96SMike Kravetz /* 234596b96a96SMike Kravetz * See above comment about rare out of 234696b96a96SMike Kravetz * memory condition. 234796b96a96SMike Kravetz */ 234896b96a96SMike Kravetz ClearPagePrivate(page); 234996b96a96SMike Kravetz } else 235096b96a96SMike Kravetz vma_end_reservation(h, vma, address); 235196b96a96SMike Kravetz } 235296b96a96SMike Kravetz } 235396b96a96SMike Kravetz 235470c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 235504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2356348ea204SAdam Litke { 235790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2358a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2359348ea204SAdam Litke struct page *page; 2360d85f69b0SMike Kravetz long map_chg, map_commit; 2361d85f69b0SMike Kravetz long gbl_chg; 23626d76dcf4SAneesh Kumar K.V int ret, idx; 23636d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 236408cf9fafSMina Almasry bool deferred_reserve; 23652fc39cecSAdam Litke 23666d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2367a1e78772SMel Gorman /* 2368d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2369d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2370d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2371a1e78772SMel Gorman */ 2372d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2373d85f69b0SMike Kravetz if (map_chg < 0) 237476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2375d85f69b0SMike Kravetz 2376d85f69b0SMike Kravetz /* 2377d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2378d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2379d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2380d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2381d85f69b0SMike Kravetz * checked against any subpool limit. 2382d85f69b0SMike Kravetz */ 2383d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2384d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2385d85f69b0SMike Kravetz if (gbl_chg < 0) { 2386feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 238776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 23885e911373SMike Kravetz } 238990d8b7e6SAdam Litke 2390d85f69b0SMike Kravetz /* 2391d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2392d85f69b0SMike Kravetz * map, there could be reservations associated with the 2393d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2394d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2395d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2396d85f69b0SMike Kravetz * the subpool reservations. 2397d85f69b0SMike Kravetz */ 2398d85f69b0SMike Kravetz if (avoid_reserve) 2399d85f69b0SMike Kravetz gbl_chg = 1; 2400d85f69b0SMike Kravetz } 2401d85f69b0SMike Kravetz 240208cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 240308cf9fafSMina Almasry */ 240408cf9fafSMina Almasry deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); 240508cf9fafSMina Almasry if (deferred_reserve) { 240608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 240708cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 24088f34af6fSJianyu Zhan if (ret) 24098f34af6fSJianyu Zhan goto out_subpool_put; 241008cf9fafSMina Almasry } 241108cf9fafSMina Almasry 241208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 241308cf9fafSMina Almasry if (ret) 241408cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 24158f34af6fSJianyu Zhan 2416a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2417d85f69b0SMike Kravetz /* 2418d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2419d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2420d85f69b0SMike Kravetz * a reservation exists for the allocation. 2421d85f69b0SMike Kravetz */ 2422d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 242381a6fcaeSJoonsoo Kim if (!page) { 242494ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 24250c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 24268f34af6fSJianyu Zhan if (!page) 24278f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2428a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2429a88c7695SNaoya Horiguchi SetPagePrivate(page); 2430a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2431a88c7695SNaoya Horiguchi } 243279dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 243379dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 243481a6fcaeSJoonsoo Kim /* Fall through */ 2435a1e78772SMel Gorman } 243681a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 243708cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 243808cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 243908cf9fafSMina Almasry */ 244008cf9fafSMina Almasry if (deferred_reserve) { 244108cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 244208cf9fafSMina Almasry h_cg, page); 244308cf9fafSMina Almasry } 244408cf9fafSMina Almasry 244581a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2446a1e78772SMel Gorman 244790481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2448a1e78772SMel Gorman 2449d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2450d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 245133039678SMike Kravetz /* 245233039678SMike Kravetz * The page was added to the reservation map between 245333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 245433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 245533039678SMike Kravetz * Adjust for the subpool count incremented above AND 245633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 245733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 245833039678SMike Kravetz * no longer applies. 245933039678SMike Kravetz */ 246033039678SMike Kravetz long rsv_adjust; 246133039678SMike Kravetz 246233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 246333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 246433039678SMike Kravetz } 24657893d1d5SAdam Litke return page; 24668f34af6fSJianyu Zhan 24678f34af6fSJianyu Zhan out_uncharge_cgroup: 24688f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 246908cf9fafSMina Almasry out_uncharge_cgroup_reservation: 247008cf9fafSMina Almasry if (deferred_reserve) 247108cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 247208cf9fafSMina Almasry h_cg); 24738f34af6fSJianyu Zhan out_subpool_put: 2474d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 24758f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2476feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 24778f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2478b45b5bd6SDavid Gibson } 2479b45b5bd6SDavid Gibson 2480e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2481e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2482e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2483aa888a74SAndi Kleen { 2484aa888a74SAndi Kleen struct huge_bootmem_page *m; 2485b2261026SJoonsoo Kim int nr_nodes, node; 2486aa888a74SAndi Kleen 2487b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2488aa888a74SAndi Kleen void *addr; 2489aa888a74SAndi Kleen 2490eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 24918b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 249297ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2493aa888a74SAndi Kleen if (addr) { 2494aa888a74SAndi Kleen /* 2495aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2496aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2497aa888a74SAndi Kleen * puts them into the mem_map). 2498aa888a74SAndi Kleen */ 2499aa888a74SAndi Kleen m = addr; 2500aa888a74SAndi Kleen goto found; 2501aa888a74SAndi Kleen } 2502aa888a74SAndi Kleen } 2503aa888a74SAndi Kleen return 0; 2504aa888a74SAndi Kleen 2505aa888a74SAndi Kleen found: 2506df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2507aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2508330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2509aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2510aa888a74SAndi Kleen m->hstate = h; 2511aa888a74SAndi Kleen return 1; 2512aa888a74SAndi Kleen } 2513aa888a74SAndi Kleen 2514d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2515d00181b9SKirill A. Shutemov unsigned int order) 251618229df5SAndy Whitcroft { 251718229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 251818229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 251918229df5SAndy Whitcroft else 252018229df5SAndy Whitcroft prep_compound_page(page, order); 252118229df5SAndy Whitcroft } 252218229df5SAndy Whitcroft 2523aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2524aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2525aa888a74SAndi Kleen { 2526aa888a74SAndi Kleen struct huge_bootmem_page *m; 2527aa888a74SAndi Kleen 2528aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 252940d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2530aa888a74SAndi Kleen struct hstate *h = m->hstate; 2531ee8f248dSBecky Bruce 2532aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 253318229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2534ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2535aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2536af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2537af0fb9dfSMichal Hocko 2538b0320c7bSRafael Aquini /* 2539b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2540b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2541b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2542b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2543b0320c7bSRafael Aquini */ 2544bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 25453dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2546520495feSCannon Matthews cond_resched(); 2547aa888a74SAndi Kleen } 2548aa888a74SAndi Kleen } 2549aa888a74SAndi Kleen 25508faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 25511da177e4SLinus Torvalds { 25521da177e4SLinus Torvalds unsigned long i; 2553f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2554f60858f9SMike Kravetz 2555f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2556f60858f9SMike Kravetz /* 2557f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2558f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2559f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2560f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2561f60858f9SMike Kravetz */ 2562f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2563f60858f9SMike Kravetz GFP_KERNEL); 2564f60858f9SMike Kravetz } else { 2565f60858f9SMike Kravetz /* allocations done at boot time */ 2566f60858f9SMike Kravetz node_alloc_noretry = NULL; 2567f60858f9SMike Kravetz } 2568f60858f9SMike Kravetz 2569f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2570f60858f9SMike Kravetz if (node_alloc_noretry) 2571f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 25721da177e4SLinus Torvalds 2573e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2574bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2575cf11e85fSRoman Gushchin if (IS_ENABLED(CONFIG_CMA) && hugetlb_cma[0]) { 2576cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 2577cf11e85fSRoman Gushchin break; 2578cf11e85fSRoman Gushchin } 2579aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2580aa888a74SAndi Kleen break; 25810c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2582f60858f9SMike Kravetz &node_states[N_MEMORY], 2583f60858f9SMike Kravetz node_alloc_noretry)) 25841da177e4SLinus Torvalds break; 258569ed779aSDavid Rientjes cond_resched(); 25861da177e4SLinus Torvalds } 2587d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2588d715cf80SLiam R. Howlett char buf[32]; 2589d715cf80SLiam R. Howlett 2590c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2591d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2592d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 25938faa8b07SAndi Kleen h->max_huge_pages = i; 2594e5ff2159SAndi Kleen } 2595f60858f9SMike Kravetz 2596f60858f9SMike Kravetz kfree(node_alloc_noretry); 2597d715cf80SLiam R. Howlett } 2598e5ff2159SAndi Kleen 2599e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2600e5ff2159SAndi Kleen { 2601e5ff2159SAndi Kleen struct hstate *h; 2602e5ff2159SAndi Kleen 2603e5ff2159SAndi Kleen for_each_hstate(h) { 2604641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2605641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2606641844f5SNaoya Horiguchi 26078faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2608bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 26098faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2610e5ff2159SAndi Kleen } 2611641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2612e5ff2159SAndi Kleen } 2613e5ff2159SAndi Kleen 2614e5ff2159SAndi Kleen static void __init report_hugepages(void) 2615e5ff2159SAndi Kleen { 2616e5ff2159SAndi Kleen struct hstate *h; 2617e5ff2159SAndi Kleen 2618e5ff2159SAndi Kleen for_each_hstate(h) { 26194abd32dbSAndi Kleen char buf[32]; 2620c6247f72SMatthew Wilcox 2621c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2622ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2623c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2624e5ff2159SAndi Kleen } 2625e5ff2159SAndi Kleen } 2626e5ff2159SAndi Kleen 26271da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 26286ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 26296ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26301da177e4SLinus Torvalds { 26314415cc8dSChristoph Lameter int i; 26324415cc8dSChristoph Lameter 2633bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2634aa888a74SAndi Kleen return; 2635aa888a74SAndi Kleen 26366ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 26371da177e4SLinus Torvalds struct page *page, *next; 2638a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2639a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2640a5516438SAndi Kleen if (count >= h->nr_huge_pages) 26416b0c880dSAdam Litke return; 26421da177e4SLinus Torvalds if (PageHighMem(page)) 26431da177e4SLinus Torvalds continue; 26441da177e4SLinus Torvalds list_del(&page->lru); 2645e5ff2159SAndi Kleen update_and_free_page(h, page); 2646a5516438SAndi Kleen h->free_huge_pages--; 2647a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 26481da177e4SLinus Torvalds } 26491da177e4SLinus Torvalds } 26501da177e4SLinus Torvalds } 26511da177e4SLinus Torvalds #else 26526ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 26536ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26541da177e4SLinus Torvalds { 26551da177e4SLinus Torvalds } 26561da177e4SLinus Torvalds #endif 26571da177e4SLinus Torvalds 265820a0307cSWu Fengguang /* 265920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 266020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 266120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 266220a0307cSWu Fengguang */ 26636ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 26646ae11b27SLee Schermerhorn int delta) 266520a0307cSWu Fengguang { 2666b2261026SJoonsoo Kim int nr_nodes, node; 266720a0307cSWu Fengguang 266820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 266920a0307cSWu Fengguang 2670e8c5c824SLee Schermerhorn if (delta < 0) { 2671b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2672b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2673b2261026SJoonsoo Kim goto found; 2674b2261026SJoonsoo Kim } 2675b2261026SJoonsoo Kim } else { 2676b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2677b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2678b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2679b2261026SJoonsoo Kim goto found; 2680e8c5c824SLee Schermerhorn } 26819a76db09SLee Schermerhorn } 2682b2261026SJoonsoo Kim return 0; 268320a0307cSWu Fengguang 2684b2261026SJoonsoo Kim found: 268520a0307cSWu Fengguang h->surplus_huge_pages += delta; 2686b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2687b2261026SJoonsoo Kim return 1; 268820a0307cSWu Fengguang } 268920a0307cSWu Fengguang 2690a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2691fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 26926ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26931da177e4SLinus Torvalds { 26947893d1d5SAdam Litke unsigned long min_count, ret; 2695f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2696f60858f9SMike Kravetz 2697f60858f9SMike Kravetz /* 2698f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2699f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2700f60858f9SMike Kravetz * the requested huge pages. 2701f60858f9SMike Kravetz */ 2702f60858f9SMike Kravetz if (node_alloc_noretry) 2703f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2704f60858f9SMike Kravetz else 2705f60858f9SMike Kravetz return -ENOMEM; 27061da177e4SLinus Torvalds 27074eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 27084eb0716eSAlexandre Ghiti 27094eb0716eSAlexandre Ghiti /* 2710fd875dcaSMike Kravetz * Check for a node specific request. 2711fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2712fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2713fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2714fd875dcaSMike Kravetz */ 2715fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2716fd875dcaSMike Kravetz unsigned long old_count = count; 2717fd875dcaSMike Kravetz 2718fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2719fd875dcaSMike Kravetz /* 2720fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2721fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2722fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2723fd875dcaSMike Kravetz * largest possible value to align with their intention. 2724fd875dcaSMike Kravetz */ 2725fd875dcaSMike Kravetz if (count < old_count) 2726fd875dcaSMike Kravetz count = ULONG_MAX; 2727fd875dcaSMike Kravetz } 2728fd875dcaSMike Kravetz 2729fd875dcaSMike Kravetz /* 27304eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 27314eb0716eSAlexandre Ghiti * page range allocation. 27324eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 27334eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 27344eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 27354eb0716eSAlexandre Ghiti */ 27364eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 27374eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 27384eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2739f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 27404eb0716eSAlexandre Ghiti return -EINVAL; 27414eb0716eSAlexandre Ghiti } 27424eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 27434eb0716eSAlexandre Ghiti } 2744aa888a74SAndi Kleen 27457893d1d5SAdam Litke /* 27467893d1d5SAdam Litke * Increase the pool size 27477893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 27487893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2749d1c3fb1fSNishanth Aravamudan * 27500c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2751d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2752d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2753d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2754d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 27557893d1d5SAdam Litke */ 2756a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 27576ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 27587893d1d5SAdam Litke break; 27597893d1d5SAdam Litke } 27607893d1d5SAdam Litke 2761a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 27627893d1d5SAdam Litke /* 27637893d1d5SAdam Litke * If this allocation races such that we no longer need the 27647893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 27657893d1d5SAdam Litke * and reducing the surplus. 27667893d1d5SAdam Litke */ 27677893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2768649920c6SJia He 2769649920c6SJia He /* yield cpu to avoid soft lockup */ 2770649920c6SJia He cond_resched(); 2771649920c6SJia He 2772f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2773f60858f9SMike Kravetz node_alloc_noretry); 27747893d1d5SAdam Litke spin_lock(&hugetlb_lock); 27757893d1d5SAdam Litke if (!ret) 27767893d1d5SAdam Litke goto out; 27777893d1d5SAdam Litke 2778536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2779536240f2SMel Gorman if (signal_pending(current)) 2780536240f2SMel Gorman goto out; 27817893d1d5SAdam Litke } 27827893d1d5SAdam Litke 27837893d1d5SAdam Litke /* 27847893d1d5SAdam Litke * Decrease the pool size 27857893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27867893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27877893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27887893d1d5SAdam Litke * to the desired size as pages become free. 2789d1c3fb1fSNishanth Aravamudan * 2790d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2791d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2792d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27930c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2794d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2795d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2796d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27977893d1d5SAdam Litke */ 2798a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27996b0c880dSAdam Litke min_count = max(count, min_count); 28006ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2801a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 28026ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 28031da177e4SLinus Torvalds break; 280455f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 28051da177e4SLinus Torvalds } 2806a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 28076ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 28087893d1d5SAdam Litke break; 28097893d1d5SAdam Litke } 28107893d1d5SAdam Litke out: 28114eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 28121da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 28134eb0716eSAlexandre Ghiti 2814f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2815f60858f9SMike Kravetz 28164eb0716eSAlexandre Ghiti return 0; 28171da177e4SLinus Torvalds } 28181da177e4SLinus Torvalds 2819a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2820a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2821a3437870SNishanth Aravamudan 2822a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2823a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2824a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2825a3437870SNishanth Aravamudan 2826a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2827a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2828a3437870SNishanth Aravamudan 28299a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 28309a305230SLee Schermerhorn 28319a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2832a3437870SNishanth Aravamudan { 2833a3437870SNishanth Aravamudan int i; 28349a305230SLee Schermerhorn 2835a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 28369a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 28379a305230SLee Schermerhorn if (nidp) 28389a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2839a3437870SNishanth Aravamudan return &hstates[i]; 28409a305230SLee Schermerhorn } 28419a305230SLee Schermerhorn 28429a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2843a3437870SNishanth Aravamudan } 2844a3437870SNishanth Aravamudan 284506808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2846a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2847a3437870SNishanth Aravamudan { 28489a305230SLee Schermerhorn struct hstate *h; 28499a305230SLee Schermerhorn unsigned long nr_huge_pages; 28509a305230SLee Schermerhorn int nid; 28519a305230SLee Schermerhorn 28529a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 28539a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 28549a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 28559a305230SLee Schermerhorn else 28569a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 28579a305230SLee Schermerhorn 28589a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2859a3437870SNishanth Aravamudan } 2860adbe8726SEric B Munson 2861238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2862238d3c13SDavid Rientjes struct hstate *h, int nid, 2863238d3c13SDavid Rientjes unsigned long count, size_t len) 2864a3437870SNishanth Aravamudan { 2865a3437870SNishanth Aravamudan int err; 28662d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2867a3437870SNishanth Aravamudan 28682d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 28692d0adf7eSOscar Salvador return -EINVAL; 2870adbe8726SEric B Munson 28719a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 28729a305230SLee Schermerhorn /* 28739a305230SLee Schermerhorn * global hstate attribute 28749a305230SLee Schermerhorn */ 28759a305230SLee Schermerhorn if (!(obey_mempolicy && 28762d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 28772d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 28782d0adf7eSOscar Salvador else 28792d0adf7eSOscar Salvador n_mask = &nodes_allowed; 28802d0adf7eSOscar Salvador } else { 28819a305230SLee Schermerhorn /* 2882fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2883fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28849a305230SLee Schermerhorn */ 28852d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28862d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2887fd875dcaSMike Kravetz } 28889a305230SLee Schermerhorn 28892d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 289006808b08SLee Schermerhorn 28914eb0716eSAlexandre Ghiti return err ? err : len; 289206808b08SLee Schermerhorn } 289306808b08SLee Schermerhorn 2894238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2895238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2896238d3c13SDavid Rientjes size_t len) 2897238d3c13SDavid Rientjes { 2898238d3c13SDavid Rientjes struct hstate *h; 2899238d3c13SDavid Rientjes unsigned long count; 2900238d3c13SDavid Rientjes int nid; 2901238d3c13SDavid Rientjes int err; 2902238d3c13SDavid Rientjes 2903238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2904238d3c13SDavid Rientjes if (err) 2905238d3c13SDavid Rientjes return err; 2906238d3c13SDavid Rientjes 2907238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2908238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2909238d3c13SDavid Rientjes } 2910238d3c13SDavid Rientjes 291106808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 291206808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 291306808b08SLee Schermerhorn { 291406808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 291506808b08SLee Schermerhorn } 291606808b08SLee Schermerhorn 291706808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 291806808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 291906808b08SLee Schermerhorn { 2920238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2921a3437870SNishanth Aravamudan } 2922a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2923a3437870SNishanth Aravamudan 292406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 292506808b08SLee Schermerhorn 292606808b08SLee Schermerhorn /* 292706808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 292806808b08SLee Schermerhorn * huge page alloc/free. 292906808b08SLee Schermerhorn */ 293006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 293106808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 293206808b08SLee Schermerhorn { 293306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 293406808b08SLee Schermerhorn } 293506808b08SLee Schermerhorn 293606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 293706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 293806808b08SLee Schermerhorn { 2939238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 294006808b08SLee Schermerhorn } 294106808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 294206808b08SLee Schermerhorn #endif 294306808b08SLee Schermerhorn 294406808b08SLee Schermerhorn 2945a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2946a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2947a3437870SNishanth Aravamudan { 29489a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2949a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2950a3437870SNishanth Aravamudan } 2951adbe8726SEric B Munson 2952a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2953a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2954a3437870SNishanth Aravamudan { 2955a3437870SNishanth Aravamudan int err; 2956a3437870SNishanth Aravamudan unsigned long input; 29579a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2958a3437870SNishanth Aravamudan 2959bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2960adbe8726SEric B Munson return -EINVAL; 2961adbe8726SEric B Munson 29623dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2963a3437870SNishanth Aravamudan if (err) 296473ae31e5SEric B Munson return err; 2965a3437870SNishanth Aravamudan 2966a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2967a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2968a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2969a3437870SNishanth Aravamudan 2970a3437870SNishanth Aravamudan return count; 2971a3437870SNishanth Aravamudan } 2972a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2973a3437870SNishanth Aravamudan 2974a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2975a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2976a3437870SNishanth Aravamudan { 29779a305230SLee Schermerhorn struct hstate *h; 29789a305230SLee Schermerhorn unsigned long free_huge_pages; 29799a305230SLee Schermerhorn int nid; 29809a305230SLee Schermerhorn 29819a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29829a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29839a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29849a305230SLee Schermerhorn else 29859a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29869a305230SLee Schermerhorn 29879a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2988a3437870SNishanth Aravamudan } 2989a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2990a3437870SNishanth Aravamudan 2991a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2992a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2993a3437870SNishanth Aravamudan { 29949a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2995a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2996a3437870SNishanth Aravamudan } 2997a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2998a3437870SNishanth Aravamudan 2999a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3000a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3001a3437870SNishanth Aravamudan { 30029a305230SLee Schermerhorn struct hstate *h; 30039a305230SLee Schermerhorn unsigned long surplus_huge_pages; 30049a305230SLee Schermerhorn int nid; 30059a305230SLee Schermerhorn 30069a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 30079a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 30089a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 30099a305230SLee Schermerhorn else 30109a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 30119a305230SLee Schermerhorn 30129a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 3013a3437870SNishanth Aravamudan } 3014a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3015a3437870SNishanth Aravamudan 3016a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3017a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3018a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3019a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3020a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3021a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 302206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 302306808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 302406808b08SLee Schermerhorn #endif 3025a3437870SNishanth Aravamudan NULL, 3026a3437870SNishanth Aravamudan }; 3027a3437870SNishanth Aravamudan 302867e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3029a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3030a3437870SNishanth Aravamudan }; 3031a3437870SNishanth Aravamudan 3032094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 30339a305230SLee Schermerhorn struct kobject **hstate_kobjs, 303467e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3035a3437870SNishanth Aravamudan { 3036a3437870SNishanth Aravamudan int retval; 3037972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3038a3437870SNishanth Aravamudan 30399a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 30409a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3041a3437870SNishanth Aravamudan return -ENOMEM; 3042a3437870SNishanth Aravamudan 30439a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3044a3437870SNishanth Aravamudan if (retval) 30459a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3046a3437870SNishanth Aravamudan 3047a3437870SNishanth Aravamudan return retval; 3048a3437870SNishanth Aravamudan } 3049a3437870SNishanth Aravamudan 3050a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3051a3437870SNishanth Aravamudan { 3052a3437870SNishanth Aravamudan struct hstate *h; 3053a3437870SNishanth Aravamudan int err; 3054a3437870SNishanth Aravamudan 3055a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3056a3437870SNishanth Aravamudan if (!hugepages_kobj) 3057a3437870SNishanth Aravamudan return; 3058a3437870SNishanth Aravamudan 3059a3437870SNishanth Aravamudan for_each_hstate(h) { 30609a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 30619a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3062a3437870SNishanth Aravamudan if (err) 3063282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3064a3437870SNishanth Aravamudan } 3065a3437870SNishanth Aravamudan } 3066a3437870SNishanth Aravamudan 30679a305230SLee Schermerhorn #ifdef CONFIG_NUMA 30689a305230SLee Schermerhorn 30699a305230SLee Schermerhorn /* 30709a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 307110fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 307210fbcf4cSKay Sievers * index of a node device or _hstate == node id. 307310fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 30749a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 30759a305230SLee Schermerhorn */ 30769a305230SLee Schermerhorn struct node_hstate { 30779a305230SLee Schermerhorn struct kobject *hugepages_kobj; 30789a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30799a305230SLee Schermerhorn }; 3080b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30819a305230SLee Schermerhorn 30829a305230SLee Schermerhorn /* 308310fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30849a305230SLee Schermerhorn */ 30859a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30869a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30879a305230SLee Schermerhorn &free_hugepages_attr.attr, 30889a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30899a305230SLee Schermerhorn NULL, 30909a305230SLee Schermerhorn }; 30919a305230SLee Schermerhorn 309267e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30939a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30949a305230SLee Schermerhorn }; 30959a305230SLee Schermerhorn 30969a305230SLee Schermerhorn /* 309710fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30989a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30999a305230SLee Schermerhorn */ 31009a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31019a305230SLee Schermerhorn { 31029a305230SLee Schermerhorn int nid; 31039a305230SLee Schermerhorn 31049a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 31059a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 31069a305230SLee Schermerhorn int i; 31079a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 31089a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 31099a305230SLee Schermerhorn if (nidp) 31109a305230SLee Schermerhorn *nidp = nid; 31119a305230SLee Schermerhorn return &hstates[i]; 31129a305230SLee Schermerhorn } 31139a305230SLee Schermerhorn } 31149a305230SLee Schermerhorn 31159a305230SLee Schermerhorn BUG(); 31169a305230SLee Schermerhorn return NULL; 31179a305230SLee Schermerhorn } 31189a305230SLee Schermerhorn 31199a305230SLee Schermerhorn /* 312010fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 31219a305230SLee Schermerhorn * No-op if no hstate attributes attached. 31229a305230SLee Schermerhorn */ 31233cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 31249a305230SLee Schermerhorn { 31259a305230SLee Schermerhorn struct hstate *h; 312610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 31279a305230SLee Schermerhorn 31289a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31299b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 31309a305230SLee Schermerhorn 3131972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3132972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3133972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3134972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3135972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3136972dc4deSAneesh Kumar K.V } 31379a305230SLee Schermerhorn } 31389a305230SLee Schermerhorn 31399a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 31409a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 31419a305230SLee Schermerhorn } 31429a305230SLee Schermerhorn 31439a305230SLee Schermerhorn 31449a305230SLee Schermerhorn /* 314510fbcf4cSKay Sievers * Register hstate attributes for a single node device. 31469a305230SLee Schermerhorn * No-op if attributes already registered. 31479a305230SLee Schermerhorn */ 31483cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 31499a305230SLee Schermerhorn { 31509a305230SLee Schermerhorn struct hstate *h; 315110fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 31529a305230SLee Schermerhorn int err; 31539a305230SLee Schermerhorn 31549a305230SLee Schermerhorn if (nhs->hugepages_kobj) 31559a305230SLee Schermerhorn return; /* already allocated */ 31569a305230SLee Schermerhorn 31579a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 315810fbcf4cSKay Sievers &node->dev.kobj); 31599a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31609a305230SLee Schermerhorn return; 31619a305230SLee Schermerhorn 31629a305230SLee Schermerhorn for_each_hstate(h) { 31639a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 31649a305230SLee Schermerhorn nhs->hstate_kobjs, 31659a305230SLee Schermerhorn &per_node_hstate_attr_group); 31669a305230SLee Schermerhorn if (err) { 3167282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 316810fbcf4cSKay Sievers h->name, node->dev.id); 31699a305230SLee Schermerhorn hugetlb_unregister_node(node); 31709a305230SLee Schermerhorn break; 31719a305230SLee Schermerhorn } 31729a305230SLee Schermerhorn } 31739a305230SLee Schermerhorn } 31749a305230SLee Schermerhorn 31759a305230SLee Schermerhorn /* 31769b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 317710fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 317810fbcf4cSKay Sievers * registered their associated device by this time. 31799a305230SLee Schermerhorn */ 31807d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31819a305230SLee Schermerhorn { 31829a305230SLee Schermerhorn int nid; 31839a305230SLee Schermerhorn 31848cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31858732794bSWen Congyang struct node *node = node_devices[nid]; 318610fbcf4cSKay Sievers if (node->dev.id == nid) 31879a305230SLee Schermerhorn hugetlb_register_node(node); 31889a305230SLee Schermerhorn } 31899a305230SLee Schermerhorn 31909a305230SLee Schermerhorn /* 319110fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31929a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31939a305230SLee Schermerhorn */ 31949a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31959a305230SLee Schermerhorn hugetlb_unregister_node); 31969a305230SLee Schermerhorn } 31979a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31989a305230SLee Schermerhorn 31999a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 32009a305230SLee Schermerhorn { 32019a305230SLee Schermerhorn BUG(); 32029a305230SLee Schermerhorn if (nidp) 32039a305230SLee Schermerhorn *nidp = -1; 32049a305230SLee Schermerhorn return NULL; 32059a305230SLee Schermerhorn } 32069a305230SLee Schermerhorn 32079a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 32089a305230SLee Schermerhorn 32099a305230SLee Schermerhorn #endif 32109a305230SLee Schermerhorn 3211a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3212a3437870SNishanth Aravamudan { 32138382d914SDavidlohr Bueso int i; 32148382d914SDavidlohr Bueso 3215c2833a5bSMike Kravetz if (!hugepages_supported()) { 3216c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3217c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 32180ef89d25SBenjamin Herrenschmidt return 0; 3219d715cf80SLiam R. Howlett } 3220d715cf80SLiam R. Howlett 3221282f4214SMike Kravetz /* 3222282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3223282f4214SMike Kravetz * architectures depend on setup being done here. 3224282f4214SMike Kravetz */ 3225a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3226282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3227282f4214SMike Kravetz /* 3228282f4214SMike Kravetz * If we did not parse a default huge page size, set 3229282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3230282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3231282f4214SMike Kravetz * specified, set that here as well. 3232282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3233282f4214SMike Kravetz * setting. A warning will be printed in this case. 3234282f4214SMike Kravetz */ 3235282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3236f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3237282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3238282f4214SMike Kravetz char buf[32]; 3239282f4214SMike Kravetz 3240282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3241282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3242282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3243282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3244282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3245282f4214SMike Kravetz default_hstate_max_huge_pages); 3246282f4214SMike Kravetz } 3247282f4214SMike Kravetz default_hstate.max_huge_pages = 3248282f4214SMike Kravetz default_hstate_max_huge_pages; 3249282f4214SMike Kravetz } 3250f8b74815SVaishali Thakkar } 3251a3437870SNishanth Aravamudan 3252cf11e85fSRoman Gushchin hugetlb_cma_check(); 3253a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3254aa888a74SAndi Kleen gather_bootmem_prealloc(); 3255a3437870SNishanth Aravamudan report_hugepages(); 3256a3437870SNishanth Aravamudan 3257a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 32589a305230SLee Schermerhorn hugetlb_register_all_nodes(); 32597179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 32609a305230SLee Schermerhorn 32618382d914SDavidlohr Bueso #ifdef CONFIG_SMP 32628382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 32638382d914SDavidlohr Bueso #else 32648382d914SDavidlohr Bueso num_fault_mutexes = 1; 32658382d914SDavidlohr Bueso #endif 3266c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 32676da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 32686da2ec56SKees Cook GFP_KERNEL); 3269c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 32708382d914SDavidlohr Bueso 32718382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3272c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3273a3437870SNishanth Aravamudan return 0; 3274a3437870SNishanth Aravamudan } 32753e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3276a3437870SNishanth Aravamudan 3277ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3278ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32799fee021dSVaishali Thakkar { 3280ae94da89SMike Kravetz return size == HPAGE_SIZE; 32819fee021dSVaishali Thakkar } 32829fee021dSVaishali Thakkar 3283d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3284a3437870SNishanth Aravamudan { 3285a3437870SNishanth Aravamudan struct hstate *h; 32868faa8b07SAndi Kleen unsigned long i; 32878faa8b07SAndi Kleen 3288a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3289a3437870SNishanth Aravamudan return; 3290a3437870SNishanth Aravamudan } 329147d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3292a3437870SNishanth Aravamudan BUG_ON(order == 0); 329347d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3294a3437870SNishanth Aravamudan h->order = order; 3295a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 32968faa8b07SAndi Kleen h->nr_huge_pages = 0; 32978faa8b07SAndi Kleen h->free_huge_pages = 0; 32988faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32998faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 33000edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 330154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 330254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3303a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3304a3437870SNishanth Aravamudan huge_page_size(h)/1024); 33058faa8b07SAndi Kleen 3306a3437870SNishanth Aravamudan parsed_hstate = h; 3307a3437870SNishanth Aravamudan } 3308a3437870SNishanth Aravamudan 3309282f4214SMike Kravetz /* 3310282f4214SMike Kravetz * hugepages command line processing 3311282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3312282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3313282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3314282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3315282f4214SMike Kravetz */ 3316282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3317a3437870SNishanth Aravamudan { 3318a3437870SNishanth Aravamudan unsigned long *mhp; 33198faa8b07SAndi Kleen static unsigned long *last_mhp; 3320a3437870SNishanth Aravamudan 33219fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3322282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 33239fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3324282f4214SMike Kravetz return 0; 33259fee021dSVaishali Thakkar } 3326282f4214SMike Kravetz 3327a3437870SNishanth Aravamudan /* 3328282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3329282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3330282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3331282f4214SMike Kravetz * default_hugepagesz. 3332a3437870SNishanth Aravamudan */ 33339fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3334a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3335a3437870SNishanth Aravamudan else 3336a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3337a3437870SNishanth Aravamudan 33388faa8b07SAndi Kleen if (mhp == last_mhp) { 3339282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3340282f4214SMike Kravetz return 0; 33418faa8b07SAndi Kleen } 33428faa8b07SAndi Kleen 3343a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3344a3437870SNishanth Aravamudan *mhp = 0; 3345a3437870SNishanth Aravamudan 33468faa8b07SAndi Kleen /* 33478faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 33488faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 33498faa8b07SAndi Kleen * use the bootmem allocator. 33508faa8b07SAndi Kleen */ 335147d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 33528faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 33538faa8b07SAndi Kleen 33548faa8b07SAndi Kleen last_mhp = mhp; 33558faa8b07SAndi Kleen 3356a3437870SNishanth Aravamudan return 1; 3357a3437870SNishanth Aravamudan } 3358282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3359e11bfbfcSNick Piggin 3360282f4214SMike Kravetz /* 3361282f4214SMike Kravetz * hugepagesz command line processing 3362282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3363282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3364282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3365282f4214SMike Kravetz * hugepagesz argument was valid. 3366282f4214SMike Kravetz */ 3367359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3368e11bfbfcSNick Piggin { 3369359f2544SMike Kravetz unsigned long size; 3370282f4214SMike Kravetz struct hstate *h; 3371282f4214SMike Kravetz 3372282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3373359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3374359f2544SMike Kravetz 3375359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3376282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3377359f2544SMike Kravetz return 0; 3378359f2544SMike Kravetz } 3379359f2544SMike Kravetz 3380282f4214SMike Kravetz h = size_to_hstate(size); 3381282f4214SMike Kravetz if (h) { 3382282f4214SMike Kravetz /* 3383282f4214SMike Kravetz * hstate for this size already exists. This is normally 3384282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3385282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3386282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3387282f4214SMike Kravetz * previously specified. 3388282f4214SMike Kravetz */ 3389282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3390282f4214SMike Kravetz default_hstate.max_huge_pages) { 3391282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 339238237830SMike Kravetz return 0; 339338237830SMike Kravetz } 339438237830SMike Kravetz 3395282f4214SMike Kravetz /* 3396282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3397282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3398282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3399282f4214SMike Kravetz */ 3400282f4214SMike Kravetz parsed_hstate = h; 3401282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3402e11bfbfcSNick Piggin return 1; 3403e11bfbfcSNick Piggin } 3404282f4214SMike Kravetz 3405359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3406282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3407359f2544SMike Kravetz return 1; 3408359f2544SMike Kravetz } 3409359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3410359f2544SMike Kravetz 3411282f4214SMike Kravetz /* 3412282f4214SMike Kravetz * default_hugepagesz command line input 3413282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3414282f4214SMike Kravetz */ 3415ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3416e11bfbfcSNick Piggin { 3417ae94da89SMike Kravetz unsigned long size; 3418ae94da89SMike Kravetz 3419282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3420282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3421282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3422282f4214SMike Kravetz return 0; 3423282f4214SMike Kravetz } 3424282f4214SMike Kravetz 3425282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3426282f4214SMike Kravetz 3427282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3428282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3429282f4214SMike Kravetz return 0; 3430282f4214SMike Kravetz } 3431282f4214SMike Kravetz 3432282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3433282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3434282f4214SMike Kravetz parsed_default_hugepagesz = true; 3435282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3436282f4214SMike Kravetz 3437282f4214SMike Kravetz /* 3438282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3439282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3440282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3441282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3442282f4214SMike Kravetz * allocated here from bootmem allocator. 3443282f4214SMike Kravetz */ 3444282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3445282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3446282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3447282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3448282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3449282f4214SMike Kravetz } 3450282f4214SMike Kravetz 3451e11bfbfcSNick Piggin return 1; 3452e11bfbfcSNick Piggin } 3453ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3454a3437870SNishanth Aravamudan 34558a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 34568a213460SNishanth Aravamudan { 34578a213460SNishanth Aravamudan int node; 34588a213460SNishanth Aravamudan unsigned int nr = 0; 34598a213460SNishanth Aravamudan 34608a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 34618a213460SNishanth Aravamudan nr += array[node]; 34628a213460SNishanth Aravamudan 34638a213460SNishanth Aravamudan return nr; 34648a213460SNishanth Aravamudan } 34658a213460SNishanth Aravamudan 34668a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 346706808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 346806808b08SLee Schermerhorn struct ctl_table *table, int write, 346932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34701da177e4SLinus Torvalds { 3471e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3472238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 347308d4a246SMichal Hocko int ret; 3474e5ff2159SAndi Kleen 3475457c1b27SNishanth Aravamudan if (!hugepages_supported()) 347686613628SJan Stancek return -EOPNOTSUPP; 3477457c1b27SNishanth Aravamudan 3478e5ff2159SAndi Kleen table->data = &tmp; 3479e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 348008d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 348108d4a246SMichal Hocko if (ret) 348208d4a246SMichal Hocko goto out; 3483e5ff2159SAndi Kleen 3484238d3c13SDavid Rientjes if (write) 3485238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3486238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 348708d4a246SMichal Hocko out: 348808d4a246SMichal Hocko return ret; 34891da177e4SLinus Torvalds } 3490396faf03SMel Gorman 349106808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 349232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 349306808b08SLee Schermerhorn { 349406808b08SLee Schermerhorn 349506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 349606808b08SLee Schermerhorn buffer, length, ppos); 349706808b08SLee Schermerhorn } 349806808b08SLee Schermerhorn 349906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 350006808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 350132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 350206808b08SLee Schermerhorn { 350306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 350406808b08SLee Schermerhorn buffer, length, ppos); 350506808b08SLee Schermerhorn } 350606808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 350706808b08SLee Schermerhorn 3508a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 350932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3510a3d0c6aaSNishanth Aravamudan { 3511a5516438SAndi Kleen struct hstate *h = &default_hstate; 3512e5ff2159SAndi Kleen unsigned long tmp; 351308d4a246SMichal Hocko int ret; 3514e5ff2159SAndi Kleen 3515457c1b27SNishanth Aravamudan if (!hugepages_supported()) 351686613628SJan Stancek return -EOPNOTSUPP; 3517457c1b27SNishanth Aravamudan 3518e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3519e5ff2159SAndi Kleen 3520bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3521adbe8726SEric B Munson return -EINVAL; 3522adbe8726SEric B Munson 3523e5ff2159SAndi Kleen table->data = &tmp; 3524e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 352508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 352608d4a246SMichal Hocko if (ret) 352708d4a246SMichal Hocko goto out; 3528e5ff2159SAndi Kleen 3529e5ff2159SAndi Kleen if (write) { 3530064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3531e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3532a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3533e5ff2159SAndi Kleen } 353408d4a246SMichal Hocko out: 353508d4a246SMichal Hocko return ret; 3536a3d0c6aaSNishanth Aravamudan } 3537a3d0c6aaSNishanth Aravamudan 35381da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 35391da177e4SLinus Torvalds 3540e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 35411da177e4SLinus Torvalds { 3542fcb2b0c5SRoman Gushchin struct hstate *h; 3543fcb2b0c5SRoman Gushchin unsigned long total = 0; 3544fcb2b0c5SRoman Gushchin 3545457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3546457c1b27SNishanth Aravamudan return; 3547fcb2b0c5SRoman Gushchin 3548fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3549fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3550fcb2b0c5SRoman Gushchin 3551fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3552fcb2b0c5SRoman Gushchin 3553fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3554e1759c21SAlexey Dobriyan seq_printf(m, 35551da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 35561da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3557b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35587893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35594f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3560fcb2b0c5SRoman Gushchin count, 3561a5516438SAndi Kleen h->free_huge_pages, 3562a5516438SAndi Kleen h->resv_huge_pages, 3563a5516438SAndi Kleen h->surplus_huge_pages, 3564fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3565fcb2b0c5SRoman Gushchin } 3566fcb2b0c5SRoman Gushchin 3567fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 35681da177e4SLinus Torvalds } 35691da177e4SLinus Torvalds 35701da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 35711da177e4SLinus Torvalds { 3572a5516438SAndi Kleen struct hstate *h = &default_hstate; 3573457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3574457c1b27SNishanth Aravamudan return 0; 35751da177e4SLinus Torvalds return sprintf(buf, 35761da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3577a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3578a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3579a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3580a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3581a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35821da177e4SLinus Torvalds } 35831da177e4SLinus Torvalds 3584949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3585949f7ec5SDavid Rientjes { 3586949f7ec5SDavid Rientjes struct hstate *h; 3587949f7ec5SDavid Rientjes int nid; 3588949f7ec5SDavid Rientjes 3589457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3590457c1b27SNishanth Aravamudan return; 3591457c1b27SNishanth Aravamudan 3592949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3593949f7ec5SDavid Rientjes for_each_hstate(h) 3594949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3595949f7ec5SDavid Rientjes nid, 3596949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3597949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3598949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3599949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3600949f7ec5SDavid Rientjes } 3601949f7ec5SDavid Rientjes 36025d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 36035d317b2bSNaoya Horiguchi { 36045d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 36055d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 36065d317b2bSNaoya Horiguchi } 36075d317b2bSNaoya Horiguchi 36081da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 36091da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 36101da177e4SLinus Torvalds { 3611d0028588SWanpeng Li struct hstate *h; 3612d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3613d0028588SWanpeng Li 3614d0028588SWanpeng Li for_each_hstate(h) 3615d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3616d0028588SWanpeng Li return nr_total_pages; 36171da177e4SLinus Torvalds } 36181da177e4SLinus Torvalds 3619a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3620fc1b8a73SMel Gorman { 3621fc1b8a73SMel Gorman int ret = -ENOMEM; 3622fc1b8a73SMel Gorman 3623fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3624fc1b8a73SMel Gorman /* 3625fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3626fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3627fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3628fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3629fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3630fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3631fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3632fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3633fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3634fc1b8a73SMel Gorman * 3635fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3636fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3637fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3638fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3639fc1b8a73SMel Gorman * semantics that cpuset has. 3640fc1b8a73SMel Gorman */ 3641fc1b8a73SMel Gorman if (delta > 0) { 3642a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3643fc1b8a73SMel Gorman goto out; 3644fc1b8a73SMel Gorman 3645a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3646a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3647fc1b8a73SMel Gorman goto out; 3648fc1b8a73SMel Gorman } 3649fc1b8a73SMel Gorman } 3650fc1b8a73SMel Gorman 3651fc1b8a73SMel Gorman ret = 0; 3652fc1b8a73SMel Gorman if (delta < 0) 3653a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3654fc1b8a73SMel Gorman 3655fc1b8a73SMel Gorman out: 3656fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3657fc1b8a73SMel Gorman return ret; 3658fc1b8a73SMel Gorman } 3659fc1b8a73SMel Gorman 366084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 366184afd99bSAndy Whitcroft { 3662f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 366384afd99bSAndy Whitcroft 366484afd99bSAndy Whitcroft /* 366584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 366684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 366784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 366825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 366984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 367084afd99bSAndy Whitcroft * new reference here without additional locking. 367184afd99bSAndy Whitcroft */ 36724e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3673f522c3acSJoonsoo Kim kref_get(&resv->refs); 367484afd99bSAndy Whitcroft } 367584afd99bSAndy Whitcroft 3676a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3677a1e78772SMel Gorman { 3678a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3679f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 368090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36814e35f483SJoonsoo Kim unsigned long reserve, start, end; 36821c5ecae3SMike Kravetz long gbl_reserve; 368384afd99bSAndy Whitcroft 36844e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36854e35f483SJoonsoo Kim return; 36864e35f483SJoonsoo Kim 3687a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3688a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 368984afd99bSAndy Whitcroft 36904e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3691e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36927251ff78SAdam Litke if (reserve) { 36931c5ecae3SMike Kravetz /* 36941c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36951c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36961c5ecae3SMike Kravetz */ 36971c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36981c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36997251ff78SAdam Litke } 3700e9fe92aeSMina Almasry 3701e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3702a1e78772SMel Gorman } 3703a1e78772SMel Gorman 370431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 370531383c68SDan Williams { 370631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 370731383c68SDan Williams return -EINVAL; 370831383c68SDan Williams return 0; 370931383c68SDan Williams } 371031383c68SDan Williams 371105ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 371205ea8860SDan Williams { 371305ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 371405ea8860SDan Williams 371505ea8860SDan Williams return 1UL << huge_page_shift(hstate); 371605ea8860SDan Williams } 371705ea8860SDan Williams 37181da177e4SLinus Torvalds /* 37191da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 37201da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 37211da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 37221da177e4SLinus Torvalds * this far. 37231da177e4SLinus Torvalds */ 3724b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 37251da177e4SLinus Torvalds { 37261da177e4SLinus Torvalds BUG(); 3727d0217ac0SNick Piggin return 0; 37281da177e4SLinus Torvalds } 37291da177e4SLinus Torvalds 3730eec3636aSJane Chu /* 3731eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3732eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3733eec3636aSJane Chu * This is because under System V memory model, mappings created via 3734eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3735eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3736eec3636aSJane Chu */ 3737f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3738d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 373984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3740a1e78772SMel Gorman .close = hugetlb_vm_op_close, 374131383c68SDan Williams .split = hugetlb_vm_op_split, 374205ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 37431da177e4SLinus Torvalds }; 37441da177e4SLinus Torvalds 37451e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 37461e8f889bSDavid Gibson int writable) 374763551ae0SDavid Gibson { 374863551ae0SDavid Gibson pte_t entry; 374963551ae0SDavid Gibson 37501e8f889bSDavid Gibson if (writable) { 3751106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3752106c992aSGerald Schaefer vma->vm_page_prot))); 375363551ae0SDavid Gibson } else { 3754106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3755106c992aSGerald Schaefer vma->vm_page_prot)); 375663551ae0SDavid Gibson } 375763551ae0SDavid Gibson entry = pte_mkyoung(entry); 375863551ae0SDavid Gibson entry = pte_mkhuge(entry); 3759d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 376063551ae0SDavid Gibson 376163551ae0SDavid Gibson return entry; 376263551ae0SDavid Gibson } 376363551ae0SDavid Gibson 37641e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37651e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37661e8f889bSDavid Gibson { 37671e8f889bSDavid Gibson pte_t entry; 37681e8f889bSDavid Gibson 3769106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 377032f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37714b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37721e8f889bSDavid Gibson } 37731e8f889bSDavid Gibson 3774d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37754a705fefSNaoya Horiguchi { 37764a705fefSNaoya Horiguchi swp_entry_t swp; 37774a705fefSNaoya Horiguchi 37784a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3779d5ed7444SAneesh Kumar K.V return false; 37804a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 37814a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3782d5ed7444SAneesh Kumar K.V return true; 37834a705fefSNaoya Horiguchi else 3784d5ed7444SAneesh Kumar K.V return false; 37854a705fefSNaoya Horiguchi } 37864a705fefSNaoya Horiguchi 37874a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 37884a705fefSNaoya Horiguchi { 37894a705fefSNaoya Horiguchi swp_entry_t swp; 37904a705fefSNaoya Horiguchi 37914a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37924a705fefSNaoya Horiguchi return 0; 37934a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 37944a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 37954a705fefSNaoya Horiguchi return 1; 37964a705fefSNaoya Horiguchi else 37974a705fefSNaoya Horiguchi return 0; 37984a705fefSNaoya Horiguchi } 37991e8f889bSDavid Gibson 380063551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 380163551ae0SDavid Gibson struct vm_area_struct *vma) 380263551ae0SDavid Gibson { 38035e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 380463551ae0SDavid Gibson struct page *ptepage; 38051c59827dSHugh Dickins unsigned long addr; 38061e8f889bSDavid Gibson int cow; 3807a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3808a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3809c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3810ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3811e8569dd2SAndreas Sandberg int ret = 0; 38121e8f889bSDavid Gibson 38131e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 381463551ae0SDavid Gibson 3815ac46d4f3SJérôme Glisse if (cow) { 38167269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 38176f4f13e8SJérôme Glisse vma->vm_start, 3818ac46d4f3SJérôme Glisse vma->vm_end); 3819ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3820c0d0381aSMike Kravetz } else { 3821c0d0381aSMike Kravetz /* 3822c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3823c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3824c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3825c0d0381aSMike Kravetz * huge_pmd_unshare. 3826c0d0381aSMike Kravetz */ 3827c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3828ac46d4f3SJérôme Glisse } 3829e8569dd2SAndreas Sandberg 3830a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3831cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 38327868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3833c74df32cSHugh Dickins if (!src_pte) 3834c74df32cSHugh Dickins continue; 3835a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3836e8569dd2SAndreas Sandberg if (!dst_pte) { 3837e8569dd2SAndreas Sandberg ret = -ENOMEM; 3838e8569dd2SAndreas Sandberg break; 3839e8569dd2SAndreas Sandberg } 3840c5c99429SLarry Woodman 38415e41540cSMike Kravetz /* 38425e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38435e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38445e41540cSMike Kravetz * 38455e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38465e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38475e41540cSMike Kravetz * Check here before taking page table lock, and once again 38485e41540cSMike Kravetz * after taking the lock below. 38495e41540cSMike Kravetz */ 38505e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38515e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3852c5c99429SLarry Woodman continue; 3853c5c99429SLarry Woodman 3854cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3855cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3856cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38574a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38585e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38595e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38605e41540cSMike Kravetz /* 38615e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38625e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38635e41540cSMike Kravetz * sharing with another vma. 38645e41540cSMike Kravetz */ 38654a705fefSNaoya Horiguchi ; 38664a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38674a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38684a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38694a705fefSNaoya Horiguchi 38704a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38714a705fefSNaoya Horiguchi /* 38724a705fefSNaoya Horiguchi * COW mappings require pages in both 38734a705fefSNaoya Horiguchi * parent and child to be set to read. 38744a705fefSNaoya Horiguchi */ 38754a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38764a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3877e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3878e5251fd4SPunit Agrawal entry, sz); 38794a705fefSNaoya Horiguchi } 3880e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38814a705fefSNaoya Horiguchi } else { 388234ee645eSJoerg Roedel if (cow) { 38830f10851eSJérôme Glisse /* 38840f10851eSJérôme Glisse * No need to notify as we are downgrading page 38850f10851eSJérôme Glisse * table protection not changing it to point 38860f10851eSJérôme Glisse * to a new page. 38870f10851eSJérôme Glisse * 3888ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 38890f10851eSJérôme Glisse */ 38907f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 389134ee645eSJoerg Roedel } 38920253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 389363551ae0SDavid Gibson ptepage = pte_page(entry); 389463551ae0SDavid Gibson get_page(ptepage); 389553f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 389663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 38975d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 38981c59827dSHugh Dickins } 3899cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3900cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 390163551ae0SDavid Gibson } 390263551ae0SDavid Gibson 3903e8569dd2SAndreas Sandberg if (cow) 3904ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3905c0d0381aSMike Kravetz else 3906c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3907e8569dd2SAndreas Sandberg 3908e8569dd2SAndreas Sandberg return ret; 390963551ae0SDavid Gibson } 391063551ae0SDavid Gibson 391124669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 391224669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 391324669e58SAneesh Kumar K.V struct page *ref_page) 391463551ae0SDavid Gibson { 391563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 391663551ae0SDavid Gibson unsigned long address; 3917c7546f8fSDavid Gibson pte_t *ptep; 391863551ae0SDavid Gibson pte_t pte; 3919cb900f41SKirill A. Shutemov spinlock_t *ptl; 392063551ae0SDavid Gibson struct page *page; 3921a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3922a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3923ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3924a5516438SAndi Kleen 392563551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3926a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3927a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 392863551ae0SDavid Gibson 392907e32661SAneesh Kumar K.V /* 393007e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 393107e32661SAneesh Kumar K.V * to huge page. 393207e32661SAneesh Kumar K.V */ 3933ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 393424669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3935dff11abeSMike Kravetz 3936dff11abeSMike Kravetz /* 3937dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3938dff11abeSMike Kravetz */ 39396f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39406f4f13e8SJérôme Glisse end); 3941ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3942ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3943569f48b8SHillf Danton address = start; 3944569f48b8SHillf Danton for (; address < end; address += sz) { 39457868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3946c7546f8fSDavid Gibson if (!ptep) 3947c7546f8fSDavid Gibson continue; 3948c7546f8fSDavid Gibson 3949cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 395031d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 395131d49da5SAneesh Kumar K.V spin_unlock(ptl); 3952dff11abeSMike Kravetz /* 3953dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3954dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3955dff11abeSMike Kravetz */ 395631d49da5SAneesh Kumar K.V continue; 395731d49da5SAneesh Kumar K.V } 395839dde65cSChen, Kenneth W 39596629326bSHillf Danton pte = huge_ptep_get(ptep); 396031d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 396131d49da5SAneesh Kumar K.V spin_unlock(ptl); 396231d49da5SAneesh Kumar K.V continue; 396331d49da5SAneesh Kumar K.V } 39646629326bSHillf Danton 39656629326bSHillf Danton /* 39669fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39679fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39686629326bSHillf Danton */ 39699fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39709386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 397131d49da5SAneesh Kumar K.V spin_unlock(ptl); 397231d49da5SAneesh Kumar K.V continue; 39738c4894c6SNaoya Horiguchi } 39746629326bSHillf Danton 39756629326bSHillf Danton page = pte_page(pte); 397604f2cbe3SMel Gorman /* 397704f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 397804f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 397904f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 398004f2cbe3SMel Gorman */ 398104f2cbe3SMel Gorman if (ref_page) { 398231d49da5SAneesh Kumar K.V if (page != ref_page) { 398331d49da5SAneesh Kumar K.V spin_unlock(ptl); 398431d49da5SAneesh Kumar K.V continue; 398531d49da5SAneesh Kumar K.V } 398604f2cbe3SMel Gorman /* 398704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 398804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 398904f2cbe3SMel Gorman * looking like data was lost 399004f2cbe3SMel Gorman */ 399104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 399204f2cbe3SMel Gorman } 399304f2cbe3SMel Gorman 3994c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3995b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3996106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 39976649a386SKen Chen set_page_dirty(page); 39989e81130bSHillf Danton 39995d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4000d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 400131d49da5SAneesh Kumar K.V 4002cb900f41SKirill A. Shutemov spin_unlock(ptl); 4003e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 400424669e58SAneesh Kumar K.V /* 400531d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 400624669e58SAneesh Kumar K.V */ 400731d49da5SAneesh Kumar K.V if (ref_page) 400831d49da5SAneesh Kumar K.V break; 4009fe1668aeSChen, Kenneth W } 4010ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 401124669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 40121da177e4SLinus Torvalds } 401363551ae0SDavid Gibson 4014d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4015d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4016d833352aSMel Gorman unsigned long end, struct page *ref_page) 4017d833352aSMel Gorman { 4018d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4019d833352aSMel Gorman 4020d833352aSMel Gorman /* 4021d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4022d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4023d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4024d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4025d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4026c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4027d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4028c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4029d833352aSMel Gorman */ 4030d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4031d833352aSMel Gorman } 4032d833352aSMel Gorman 4033502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 403404f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4035502717f4SChen, Kenneth W { 403624669e58SAneesh Kumar K.V struct mm_struct *mm; 403724669e58SAneesh Kumar K.V struct mmu_gather tlb; 4038dff11abeSMike Kravetz unsigned long tlb_start = start; 4039dff11abeSMike Kravetz unsigned long tlb_end = end; 4040dff11abeSMike Kravetz 4041dff11abeSMike Kravetz /* 4042dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 4043dff11abeSMike Kravetz * start/end for worst case tlb flushing. 4044dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 4045dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 4046dff11abeSMike Kravetz * the largest possible range. 4047dff11abeSMike Kravetz */ 4048dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 404924669e58SAneesh Kumar K.V 405024669e58SAneesh Kumar K.V mm = vma->vm_mm; 405124669e58SAneesh Kumar K.V 4052dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 405324669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4054dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 4055502717f4SChen, Kenneth W } 4056502717f4SChen, Kenneth W 405704f2cbe3SMel Gorman /* 405804f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 405904f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 406004f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 406104f2cbe3SMel Gorman * same region. 406204f2cbe3SMel Gorman */ 40632f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40642a4b3dedSHarvey Harrison struct page *page, unsigned long address) 406504f2cbe3SMel Gorman { 40667526674dSAdam Litke struct hstate *h = hstate_vma(vma); 406704f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 406804f2cbe3SMel Gorman struct address_space *mapping; 406904f2cbe3SMel Gorman pgoff_t pgoff; 407004f2cbe3SMel Gorman 407104f2cbe3SMel Gorman /* 407204f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 407304f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 407404f2cbe3SMel Gorman */ 40757526674dSAdam Litke address = address & huge_page_mask(h); 407636e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 407736e4f20aSMichal Hocko vma->vm_pgoff; 407893c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 407904f2cbe3SMel Gorman 40804eb2b1dcSMel Gorman /* 40814eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40824eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40834eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40844eb2b1dcSMel Gorman */ 408583cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40866b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 408704f2cbe3SMel Gorman /* Do not unmap the current VMA */ 408804f2cbe3SMel Gorman if (iter_vma == vma) 408904f2cbe3SMel Gorman continue; 409004f2cbe3SMel Gorman 409104f2cbe3SMel Gorman /* 40922f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40932f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 40942f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 40952f84a899SMel Gorman */ 40962f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 40972f84a899SMel Gorman continue; 40982f84a899SMel Gorman 40992f84a899SMel Gorman /* 410004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 410104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 410204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 410304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 410404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 410504f2cbe3SMel Gorman */ 410604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 410724669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 410824669e58SAneesh Kumar K.V address + huge_page_size(h), page); 410904f2cbe3SMel Gorman } 411083cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 411104f2cbe3SMel Gorman } 411204f2cbe3SMel Gorman 41130fe6e20bSNaoya Horiguchi /* 41140fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4115ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4116ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4117ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 41180fe6e20bSNaoya Horiguchi */ 41192b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4120974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4121cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 41221e8f889bSDavid Gibson { 41233999f52eSAneesh Kumar K.V pte_t pte; 4124a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41251e8f889bSDavid Gibson struct page *old_page, *new_page; 41262b740303SSouptick Joarder int outside_reserve = 0; 41272b740303SSouptick Joarder vm_fault_t ret = 0; 4128974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4129ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 41301e8f889bSDavid Gibson 41313999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 41321e8f889bSDavid Gibson old_page = pte_page(pte); 41331e8f889bSDavid Gibson 413404f2cbe3SMel Gorman retry_avoidcopy: 41351e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 41361e8f889bSDavid Gibson * and just make the page writable */ 413737a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 41385a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 41395b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 414083c54070SNick Piggin return 0; 41411e8f889bSDavid Gibson } 41421e8f889bSDavid Gibson 414304f2cbe3SMel Gorman /* 414404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 414504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 414604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 414704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 414804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 414904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 415004f2cbe3SMel Gorman * of the full address range. 415104f2cbe3SMel Gorman */ 41525944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 415304f2cbe3SMel Gorman old_page != pagecache_page) 415404f2cbe3SMel Gorman outside_reserve = 1; 415504f2cbe3SMel Gorman 415609cbfeafSKirill A. Shutemov get_page(old_page); 4157b76c8cfbSLarry Woodman 4158ad4404a2SDavidlohr Bueso /* 4159ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4160ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4161ad4404a2SDavidlohr Bueso */ 4162cb900f41SKirill A. Shutemov spin_unlock(ptl); 41635b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41641e8f889bSDavid Gibson 41652fc39cecSAdam Litke if (IS_ERR(new_page)) { 416604f2cbe3SMel Gorman /* 416704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 416804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 416904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 417004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 417104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 417204f2cbe3SMel Gorman */ 417304f2cbe3SMel Gorman if (outside_reserve) { 417409cbfeafSKirill A. Shutemov put_page(old_page); 417504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 41765b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 417704f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4178cb900f41SKirill A. Shutemov spin_lock(ptl); 41795b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4180a9af0c5dSNaoya Horiguchi if (likely(ptep && 4181a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 418204f2cbe3SMel Gorman goto retry_avoidcopy; 4183a734bcc8SHillf Danton /* 4184cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4185cb900f41SKirill A. Shutemov * lock, and our job is done. 4186a734bcc8SHillf Danton */ 4187a734bcc8SHillf Danton return 0; 418804f2cbe3SMel Gorman } 418904f2cbe3SMel Gorman 41902b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4191ad4404a2SDavidlohr Bueso goto out_release_old; 41921e8f889bSDavid Gibson } 41931e8f889bSDavid Gibson 41940fe6e20bSNaoya Horiguchi /* 41950fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 41960fe6e20bSNaoya Horiguchi * anon_vma prepared. 41970fe6e20bSNaoya Horiguchi */ 419844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4199ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4200ad4404a2SDavidlohr Bueso goto out_release_all; 420144e2aa93SDean Nelson } 42020fe6e20bSNaoya Horiguchi 4203974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 420447ad8475SAndrea Arcangeli pages_per_huge_page(h)); 42050ed361deSNick Piggin __SetPageUptodate(new_page); 42061e8f889bSDavid Gibson 42077269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 42086f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4209ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4210ad4404a2SDavidlohr Bueso 4211b76c8cfbSLarry Woodman /* 4212cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4213b76c8cfbSLarry Woodman * before the page tables are altered 4214b76c8cfbSLarry Woodman */ 4215cb900f41SKirill A. Shutemov spin_lock(ptl); 42165b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4217a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 421807443a85SJoonsoo Kim ClearPagePrivate(new_page); 421907443a85SJoonsoo Kim 42201e8f889bSDavid Gibson /* Break COW */ 42215b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4222ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 42235b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 42241e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4225d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 42265b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 4227cb6acd01SMike Kravetz set_page_huge_active(new_page); 42281e8f889bSDavid Gibson /* Make the old page be freed below */ 42291e8f889bSDavid Gibson new_page = old_page; 42301e8f889bSDavid Gibson } 4231cb900f41SKirill A. Shutemov spin_unlock(ptl); 4232ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4233ad4404a2SDavidlohr Bueso out_release_all: 42345b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 423509cbfeafSKirill A. Shutemov put_page(new_page); 4236ad4404a2SDavidlohr Bueso out_release_old: 423709cbfeafSKirill A. Shutemov put_page(old_page); 42388312034fSJoonsoo Kim 4239ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4240ad4404a2SDavidlohr Bueso return ret; 42411e8f889bSDavid Gibson } 42421e8f889bSDavid Gibson 424304f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4244a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4245a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 424604f2cbe3SMel Gorman { 424704f2cbe3SMel Gorman struct address_space *mapping; 4248e7c4b0bfSAndy Whitcroft pgoff_t idx; 424904f2cbe3SMel Gorman 425004f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4251a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 425204f2cbe3SMel Gorman 425304f2cbe3SMel Gorman return find_lock_page(mapping, idx); 425404f2cbe3SMel Gorman } 425504f2cbe3SMel Gorman 42563ae77f43SHugh Dickins /* 42573ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42583ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42593ae77f43SHugh Dickins */ 42603ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42612a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42622a15efc9SHugh Dickins { 42632a15efc9SHugh Dickins struct address_space *mapping; 42642a15efc9SHugh Dickins pgoff_t idx; 42652a15efc9SHugh Dickins struct page *page; 42662a15efc9SHugh Dickins 42672a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42682a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42692a15efc9SHugh Dickins 42702a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42712a15efc9SHugh Dickins if (page) 42722a15efc9SHugh Dickins put_page(page); 42732a15efc9SHugh Dickins return page != NULL; 42742a15efc9SHugh Dickins } 42752a15efc9SHugh Dickins 4276ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4277ab76ad54SMike Kravetz pgoff_t idx) 4278ab76ad54SMike Kravetz { 4279ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4280ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4281ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4282ab76ad54SMike Kravetz 4283ab76ad54SMike Kravetz if (err) 4284ab76ad54SMike Kravetz return err; 4285ab76ad54SMike Kravetz ClearPagePrivate(page); 4286ab76ad54SMike Kravetz 428722146c3cSMike Kravetz /* 428822146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 428922146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 429022146c3cSMike Kravetz */ 429122146c3cSMike Kravetz set_page_dirty(page); 429222146c3cSMike Kravetz 4293ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4294ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4295ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4296ab76ad54SMike Kravetz return 0; 4297ab76ad54SMike Kravetz } 4298ab76ad54SMike Kravetz 42992b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 43002b740303SSouptick Joarder struct vm_area_struct *vma, 43018382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4302788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4303ac9b9c66SHugh Dickins { 4304a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43052b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4306409eb8c2SHillf Danton int anon_rmap = 0; 43074c887265SAdam Litke unsigned long size; 43084c887265SAdam Litke struct page *page; 43091e8f889bSDavid Gibson pte_t new_pte; 4310cb900f41SKirill A. Shutemov spinlock_t *ptl; 4311285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4312cb6acd01SMike Kravetz bool new_page = false; 43134c887265SAdam Litke 431404f2cbe3SMel Gorman /* 431504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 431604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 431725985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 431804f2cbe3SMel Gorman */ 431904f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4320910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 432104f2cbe3SMel Gorman current->pid); 432204f2cbe3SMel Gorman return ret; 432304f2cbe3SMel Gorman } 432404f2cbe3SMel Gorman 43254c887265SAdam Litke /* 432687bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 432787bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 432887bf91d3SMike Kravetz * once for faults beyond end of file. 43294c887265SAdam Litke */ 4330a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4331ebed4bfcSHugh Dickins if (idx >= size) 4332ebed4bfcSHugh Dickins goto out; 43331a1aad8aSMike Kravetz 433487bf91d3SMike Kravetz retry: 433587bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 433687bf91d3SMike Kravetz if (!page) { 43371a1aad8aSMike Kravetz /* 43381a1aad8aSMike Kravetz * Check for page in userfault range 43391a1aad8aSMike Kravetz */ 43401a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43411a1aad8aSMike Kravetz u32 hash; 43421a1aad8aSMike Kravetz struct vm_fault vmf = { 43431a1aad8aSMike Kravetz .vma = vma, 4344285b8dcaSHuang Ying .address = haddr, 43451a1aad8aSMike Kravetz .flags = flags, 43461a1aad8aSMike Kravetz /* 43471a1aad8aSMike Kravetz * Hard to debug if it ends up being 43481a1aad8aSMike Kravetz * used by a callee that assumes 43491a1aad8aSMike Kravetz * something about the other 43501a1aad8aSMike Kravetz * uninitialized fields... same as in 43511a1aad8aSMike Kravetz * memory.c 43521a1aad8aSMike Kravetz */ 43531a1aad8aSMike Kravetz }; 43541a1aad8aSMike Kravetz 43551a1aad8aSMike Kravetz /* 4356c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4357c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4358c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43591a1aad8aSMike Kravetz */ 4360188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43611a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4362c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43631a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4364c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 43651a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43661a1aad8aSMike Kravetz goto out; 43671a1aad8aSMike Kravetz } 43681a1aad8aSMike Kravetz 4369285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43702fc39cecSAdam Litke if (IS_ERR(page)) { 43714643d67eSMike Kravetz /* 43724643d67eSMike Kravetz * Returning error will result in faulting task being 43734643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43744643d67eSMike Kravetz * tasks from racing to fault in the same page which 43754643d67eSMike Kravetz * could result in false unable to allocate errors. 43764643d67eSMike Kravetz * Page migration does not take the fault mutex, but 43774643d67eSMike Kravetz * does a clear then write of pte's under page table 43784643d67eSMike Kravetz * lock. Page fault code could race with migration, 43794643d67eSMike Kravetz * notice the clear pte and try to allocate a page 43804643d67eSMike Kravetz * here. Before returning error, get ptl and make 43814643d67eSMike Kravetz * sure there really is no pte entry. 43824643d67eSMike Kravetz */ 43834643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 43844643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 43854643d67eSMike Kravetz ret = 0; 43864643d67eSMike Kravetz spin_unlock(ptl); 43874643d67eSMike Kravetz goto out; 43884643d67eSMike Kravetz } 43894643d67eSMike Kravetz spin_unlock(ptl); 43902b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 43916bda666aSChristoph Lameter goto out; 43926bda666aSChristoph Lameter } 439347ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 43940ed361deSNick Piggin __SetPageUptodate(page); 4395cb6acd01SMike Kravetz new_page = true; 4396ac9b9c66SHugh Dickins 4397f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4398ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 43996bda666aSChristoph Lameter if (err) { 44006bda666aSChristoph Lameter put_page(page); 44016bda666aSChristoph Lameter if (err == -EEXIST) 44026bda666aSChristoph Lameter goto retry; 44036bda666aSChristoph Lameter goto out; 44046bda666aSChristoph Lameter } 440523be7468SMel Gorman } else { 44066bda666aSChristoph Lameter lock_page(page); 44070fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 44080fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 44090fe6e20bSNaoya Horiguchi goto backout_unlocked; 441023be7468SMel Gorman } 4411409eb8c2SHillf Danton anon_rmap = 1; 44120fe6e20bSNaoya Horiguchi } 44130fe6e20bSNaoya Horiguchi } else { 441457303d80SAndy Whitcroft /* 4415998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4416998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4417998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4418fd6a03edSNaoya Horiguchi */ 4419fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 4420aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 4421972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4422fd6a03edSNaoya Horiguchi goto backout_unlocked; 44236bda666aSChristoph Lameter } 4424998b4382SNaoya Horiguchi } 44251e8f889bSDavid Gibson 442657303d80SAndy Whitcroft /* 442757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 442857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 442957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 443057303d80SAndy Whitcroft * the spinlock. 443157303d80SAndy Whitcroft */ 44325e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4433285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44342b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44352b26736cSAndy Whitcroft goto backout_unlocked; 44362b26736cSAndy Whitcroft } 44375e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4438285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44395e911373SMike Kravetz } 444057303d80SAndy Whitcroft 44418bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 444283c54070SNick Piggin ret = 0; 44437f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44444c887265SAdam Litke goto backout; 44454c887265SAdam Litke 444607443a85SJoonsoo Kim if (anon_rmap) { 444707443a85SJoonsoo Kim ClearPagePrivate(page); 4448285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4449ac714904SChoi Gi-yong } else 445053f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44511e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44521e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4453285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44541e8f889bSDavid Gibson 44555d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4456788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44571e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4458974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44591e8f889bSDavid Gibson } 44601e8f889bSDavid Gibson 4461cb900f41SKirill A. Shutemov spin_unlock(ptl); 4462cb6acd01SMike Kravetz 4463cb6acd01SMike Kravetz /* 4464cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 4465cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 4466cb6acd01SMike Kravetz * isolated for migration. 4467cb6acd01SMike Kravetz */ 4468cb6acd01SMike Kravetz if (new_page) 4469cb6acd01SMike Kravetz set_page_huge_active(page); 4470cb6acd01SMike Kravetz 44714c887265SAdam Litke unlock_page(page); 44724c887265SAdam Litke out: 4473ac9b9c66SHugh Dickins return ret; 44744c887265SAdam Litke 44754c887265SAdam Litke backout: 4476cb900f41SKirill A. Shutemov spin_unlock(ptl); 44772b26736cSAndy Whitcroft backout_unlocked: 44784c887265SAdam Litke unlock_page(page); 4479285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 44804c887265SAdam Litke put_page(page); 44814c887265SAdam Litke goto out; 4482ac9b9c66SHugh Dickins } 4483ac9b9c66SHugh Dickins 44848382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4485188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44868382d914SDavidlohr Bueso { 44878382d914SDavidlohr Bueso unsigned long key[2]; 44888382d914SDavidlohr Bueso u32 hash; 44898382d914SDavidlohr Bueso 44908382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 44918382d914SDavidlohr Bueso key[1] = idx; 44928382d914SDavidlohr Bueso 449355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 44948382d914SDavidlohr Bueso 44958382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 44968382d914SDavidlohr Bueso } 44978382d914SDavidlohr Bueso #else 44988382d914SDavidlohr Bueso /* 44998382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 45008382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 45018382d914SDavidlohr Bueso */ 4502188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45038382d914SDavidlohr Bueso { 45048382d914SDavidlohr Bueso return 0; 45058382d914SDavidlohr Bueso } 45068382d914SDavidlohr Bueso #endif 45078382d914SDavidlohr Bueso 45082b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4509788c7df4SHugh Dickins unsigned long address, unsigned int flags) 451086e5216fSAdam Litke { 45118382d914SDavidlohr Bueso pte_t *ptep, entry; 4512cb900f41SKirill A. Shutemov spinlock_t *ptl; 45132b740303SSouptick Joarder vm_fault_t ret; 45148382d914SDavidlohr Bueso u32 hash; 45158382d914SDavidlohr Bueso pgoff_t idx; 45160fe6e20bSNaoya Horiguchi struct page *page = NULL; 451757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4518a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45198382d914SDavidlohr Bueso struct address_space *mapping; 45200f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4521285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 452286e5216fSAdam Litke 4523285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4524fd6a03edSNaoya Horiguchi if (ptep) { 4525c0d0381aSMike Kravetz /* 4526c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4527c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4528c0d0381aSMike Kravetz * not actually modifying content here. 4529c0d0381aSMike Kravetz */ 4530fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4531290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4532cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4533290408d4SNaoya Horiguchi return 0; 4534290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4535aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4536972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4537ddeaab32SMike Kravetz } else { 4538b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4539ddeaab32SMike Kravetz if (!ptep) 4540b43a9990SMike Kravetz return VM_FAULT_OOM; 4541b43a9990SMike Kravetz } 45428382d914SDavidlohr Bueso 4543c0d0381aSMike Kravetz /* 4544c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 454587bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 454687bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 454787bf91d3SMike Kravetz * and making the ptep no longer valid. 454887bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4549c0d0381aSMike Kravetz * 4550c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4551c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4552c0d0381aSMike Kravetz * something has changed. 4553c0d0381aSMike Kravetz */ 4554ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4555c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4556c0d0381aSMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4557c0d0381aSMike Kravetz if (!ptep) { 4558c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4559c0d0381aSMike Kravetz return VM_FAULT_OOM; 4560c0d0381aSMike Kravetz } 4561ddeaab32SMike Kravetz 45623935baa9SDavid Gibson /* 45633935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45643935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45653935baa9SDavid Gibson * the same page in the page cache. 45663935baa9SDavid Gibson */ 4567c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4568188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4569c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45708382d914SDavidlohr Bueso 45717f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45727f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45738382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4574b4d1d99fSDavid Gibson goto out_mutex; 45753935baa9SDavid Gibson } 457686e5216fSAdam Litke 457783c54070SNick Piggin ret = 0; 45781e8f889bSDavid Gibson 457957303d80SAndy Whitcroft /* 45800f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45810f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45827c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45837c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 45847c8de358SEthon Paul * properly handle it. 45850f792cf9SNaoya Horiguchi */ 45860f792cf9SNaoya Horiguchi if (!pte_present(entry)) 45870f792cf9SNaoya Horiguchi goto out_mutex; 45880f792cf9SNaoya Horiguchi 45890f792cf9SNaoya Horiguchi /* 459057303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 459157303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 459257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 459357303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 459457303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 459557303d80SAndy Whitcroft * consumed. 459657303d80SAndy Whitcroft */ 4597106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4598285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 45992b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4600b4d1d99fSDavid Gibson goto out_mutex; 46012b26736cSAndy Whitcroft } 46025e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4603285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 460457303d80SAndy Whitcroft 4605f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 460657303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4607285b8dcaSHuang Ying vma, haddr); 460857303d80SAndy Whitcroft } 460957303d80SAndy Whitcroft 46100f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 46110fe6e20bSNaoya Horiguchi 46121e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4613b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4614cb900f41SKirill A. Shutemov goto out_ptl; 4615b4d1d99fSDavid Gibson 46160f792cf9SNaoya Horiguchi /* 46170f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 46180f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 46190f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 46200f792cf9SNaoya Horiguchi */ 46210f792cf9SNaoya Horiguchi page = pte_page(entry); 46220f792cf9SNaoya Horiguchi if (page != pagecache_page) 46230f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 46240f792cf9SNaoya Horiguchi need_wait_lock = 1; 46250f792cf9SNaoya Horiguchi goto out_ptl; 46260f792cf9SNaoya Horiguchi } 46270f792cf9SNaoya Horiguchi 46280f792cf9SNaoya Horiguchi get_page(page); 4629b4d1d99fSDavid Gibson 4630788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4631106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4632974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4633cb900f41SKirill A. Shutemov pagecache_page, ptl); 46340f792cf9SNaoya Horiguchi goto out_put_page; 4635b4d1d99fSDavid Gibson } 4636106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4637b4d1d99fSDavid Gibson } 4638b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4639285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4640788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4641285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46420f792cf9SNaoya Horiguchi out_put_page: 46430f792cf9SNaoya Horiguchi if (page != pagecache_page) 46440f792cf9SNaoya Horiguchi unlock_page(page); 46450f792cf9SNaoya Horiguchi put_page(page); 4646cb900f41SKirill A. Shutemov out_ptl: 4647cb900f41SKirill A. Shutemov spin_unlock(ptl); 464857303d80SAndy Whitcroft 464957303d80SAndy Whitcroft if (pagecache_page) { 465057303d80SAndy Whitcroft unlock_page(pagecache_page); 465157303d80SAndy Whitcroft put_page(pagecache_page); 465257303d80SAndy Whitcroft } 4653b4d1d99fSDavid Gibson out_mutex: 4654c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4655c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46560f792cf9SNaoya Horiguchi /* 46570f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46580f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46590f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46600f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46610f792cf9SNaoya Horiguchi * here without taking refcount. 46620f792cf9SNaoya Horiguchi */ 46630f792cf9SNaoya Horiguchi if (need_wait_lock) 46640f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46651e8f889bSDavid Gibson return ret; 466686e5216fSAdam Litke } 466786e5216fSAdam Litke 46688fb5debcSMike Kravetz /* 46698fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46708fb5debcSMike Kravetz * modifications for huge pages. 46718fb5debcSMike Kravetz */ 46728fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46738fb5debcSMike Kravetz pte_t *dst_pte, 46748fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46758fb5debcSMike Kravetz unsigned long dst_addr, 46768fb5debcSMike Kravetz unsigned long src_addr, 46778fb5debcSMike Kravetz struct page **pagep) 46788fb5debcSMike Kravetz { 46791e392147SAndrea Arcangeli struct address_space *mapping; 46801e392147SAndrea Arcangeli pgoff_t idx; 46811e392147SAndrea Arcangeli unsigned long size; 46821c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46838fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 46848fb5debcSMike Kravetz pte_t _dst_pte; 46858fb5debcSMike Kravetz spinlock_t *ptl; 46868fb5debcSMike Kravetz int ret; 46878fb5debcSMike Kravetz struct page *page; 46888fb5debcSMike Kravetz 46898fb5debcSMike Kravetz if (!*pagep) { 46908fb5debcSMike Kravetz ret = -ENOMEM; 46918fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 46928fb5debcSMike Kravetz if (IS_ERR(page)) 46938fb5debcSMike Kravetz goto out; 46948fb5debcSMike Kravetz 46958fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 46968fb5debcSMike Kravetz (const void __user *) src_addr, 4697810a56b9SMike Kravetz pages_per_huge_page(h), false); 46988fb5debcSMike Kravetz 46998fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 47008fb5debcSMike Kravetz if (unlikely(ret)) { 47019e368259SAndrea Arcangeli ret = -ENOENT; 47028fb5debcSMike Kravetz *pagep = page; 47038fb5debcSMike Kravetz /* don't free the page */ 47048fb5debcSMike Kravetz goto out; 47058fb5debcSMike Kravetz } 47068fb5debcSMike Kravetz } else { 47078fb5debcSMike Kravetz page = *pagep; 47088fb5debcSMike Kravetz *pagep = NULL; 47098fb5debcSMike Kravetz } 47108fb5debcSMike Kravetz 47118fb5debcSMike Kravetz /* 47128fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 47138fb5debcSMike Kravetz * preceding stores to the page contents become visible before 47148fb5debcSMike Kravetz * the set_pte_at() write. 47158fb5debcSMike Kravetz */ 47168fb5debcSMike Kravetz __SetPageUptodate(page); 47178fb5debcSMike Kravetz 47181e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 47191e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 47201e392147SAndrea Arcangeli 47211c9e8defSMike Kravetz /* 47221c9e8defSMike Kravetz * If shared, add to page cache 47231c9e8defSMike Kravetz */ 47241c9e8defSMike Kravetz if (vm_shared) { 47251e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47261e392147SAndrea Arcangeli ret = -EFAULT; 47271e392147SAndrea Arcangeli if (idx >= size) 47281e392147SAndrea Arcangeli goto out_release_nounlock; 47291c9e8defSMike Kravetz 47301e392147SAndrea Arcangeli /* 47311e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 47321e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 47331e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47341e392147SAndrea Arcangeli * the caller. 47351e392147SAndrea Arcangeli */ 47361c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47371c9e8defSMike Kravetz if (ret) 47381c9e8defSMike Kravetz goto out_release_nounlock; 47391c9e8defSMike Kravetz } 47401c9e8defSMike Kravetz 47418fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47428fb5debcSMike Kravetz spin_lock(ptl); 47438fb5debcSMike Kravetz 47441e392147SAndrea Arcangeli /* 47451e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47461e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47471e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47481e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47491e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47501e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47511e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47521e392147SAndrea Arcangeli */ 47531e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47541e392147SAndrea Arcangeli ret = -EFAULT; 47551e392147SAndrea Arcangeli if (idx >= size) 47561e392147SAndrea Arcangeli goto out_release_unlock; 47571e392147SAndrea Arcangeli 47588fb5debcSMike Kravetz ret = -EEXIST; 47598fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47608fb5debcSMike Kravetz goto out_release_unlock; 47618fb5debcSMike Kravetz 47621c9e8defSMike Kravetz if (vm_shared) { 47631c9e8defSMike Kravetz page_dup_rmap(page, true); 47641c9e8defSMike Kravetz } else { 47658fb5debcSMike Kravetz ClearPagePrivate(page); 47668fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47671c9e8defSMike Kravetz } 47688fb5debcSMike Kravetz 47698fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47708fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47718fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47728fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47738fb5debcSMike Kravetz 47748fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47758fb5debcSMike Kravetz 47768fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47778fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47788fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47798fb5debcSMike Kravetz 47808fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47818fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47828fb5debcSMike Kravetz 47838fb5debcSMike Kravetz spin_unlock(ptl); 4784cb6acd01SMike Kravetz set_page_huge_active(page); 47851c9e8defSMike Kravetz if (vm_shared) 47861c9e8defSMike Kravetz unlock_page(page); 47878fb5debcSMike Kravetz ret = 0; 47888fb5debcSMike Kravetz out: 47898fb5debcSMike Kravetz return ret; 47908fb5debcSMike Kravetz out_release_unlock: 47918fb5debcSMike Kravetz spin_unlock(ptl); 47921c9e8defSMike Kravetz if (vm_shared) 47931c9e8defSMike Kravetz unlock_page(page); 47945af10dfdSAndrea Arcangeli out_release_nounlock: 47958fb5debcSMike Kravetz put_page(page); 47968fb5debcSMike Kravetz goto out; 47978fb5debcSMike Kravetz } 47988fb5debcSMike Kravetz 479928a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 480063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 480128a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 48024f6da934SPeter Xu long i, unsigned int flags, int *locked) 480363551ae0SDavid Gibson { 4804d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4805d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 480628a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4807a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48082be7cfedSDaniel Jordan int err = -EFAULT; 480963551ae0SDavid Gibson 481063551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 481163551ae0SDavid Gibson pte_t *pte; 4812cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 48132a15efc9SHugh Dickins int absent; 481463551ae0SDavid Gibson struct page *page; 481563551ae0SDavid Gibson 48164c887265SAdam Litke /* 481702057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 481802057967SDavid Rientjes * potentially allocating memory. 481902057967SDavid Rientjes */ 4820fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 482102057967SDavid Rientjes remainder = 0; 482202057967SDavid Rientjes break; 482302057967SDavid Rientjes } 482402057967SDavid Rientjes 482502057967SDavid Rientjes /* 48264c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48272a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48284c887265SAdam Litke * first, for the page indexing below to work. 4829cb900f41SKirill A. Shutemov * 4830cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48314c887265SAdam Litke */ 48327868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48337868a208SPunit Agrawal huge_page_size(h)); 4834cb900f41SKirill A. Shutemov if (pte) 4835cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48362a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 483763551ae0SDavid Gibson 48382a15efc9SHugh Dickins /* 48392a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48403ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48413ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48423ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48433ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48442a15efc9SHugh Dickins */ 48453ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48463ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4847cb900f41SKirill A. Shutemov if (pte) 4848cb900f41SKirill A. Shutemov spin_unlock(ptl); 48492a15efc9SHugh Dickins remainder = 0; 48502a15efc9SHugh Dickins break; 48512a15efc9SHugh Dickins } 48522a15efc9SHugh Dickins 48539cc3a5bdSNaoya Horiguchi /* 48549cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48559cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48569cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 48579cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 48589cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 48599cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 48609cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 48619cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 48629cc3a5bdSNaoya Horiguchi */ 48639cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4864106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4865106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 48662b740303SSouptick Joarder vm_fault_t ret; 486787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 48684c887265SAdam Litke 4869cb900f41SKirill A. Shutemov if (pte) 4870cb900f41SKirill A. Shutemov spin_unlock(ptl); 487187ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 487287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 48734f6da934SPeter Xu if (locked) 487471335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 487571335f37SPeter Xu FAULT_FLAG_KILLABLE; 487687ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 487787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 487887ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 487987ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 48804426e945SPeter Xu /* 48814426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 48824426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 48834426e945SPeter Xu */ 488487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 488587ffc118SAndrea Arcangeli } 488687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 488787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 48882be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 48891c59827dSHugh Dickins remainder = 0; 48901c59827dSHugh Dickins break; 48911c59827dSHugh Dickins } 489287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 48934f6da934SPeter Xu if (locked && 48941ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 48954f6da934SPeter Xu *locked = 0; 489687ffc118SAndrea Arcangeli *nr_pages = 0; 489787ffc118SAndrea Arcangeli /* 489887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 489987ffc118SAndrea Arcangeli * error, it will return zero 490087ffc118SAndrea Arcangeli * instead. 490187ffc118SAndrea Arcangeli * 490287ffc118SAndrea Arcangeli * No need to update "position" as the 490387ffc118SAndrea Arcangeli * caller will not check it after 490487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 490587ffc118SAndrea Arcangeli */ 490687ffc118SAndrea Arcangeli return i; 490787ffc118SAndrea Arcangeli } 490887ffc118SAndrea Arcangeli continue; 490987ffc118SAndrea Arcangeli } 491063551ae0SDavid Gibson 4911a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 49127f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 49138fde12caSLinus Torvalds 49148fde12caSLinus Torvalds /* 4915acbfb087SZhigang Lu * If subpage information not requested, update counters 4916acbfb087SZhigang Lu * and skip the same_page loop below. 4917acbfb087SZhigang Lu */ 4918acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4919acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4920acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4921acbfb087SZhigang Lu vaddr += huge_page_size(h); 4922acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4923acbfb087SZhigang Lu i += pages_per_huge_page(h); 4924acbfb087SZhigang Lu spin_unlock(ptl); 4925acbfb087SZhigang Lu continue; 4926acbfb087SZhigang Lu } 4927acbfb087SZhigang Lu 4928d5d4b0aaSChen, Kenneth W same_page: 4929d6692183SChen, Kenneth W if (pages) { 493069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 49313faa52c0SJohn Hubbard /* 49323faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: 49333faa52c0SJohn Hubbard * a) we hold the ptl lock, and b) we've just checked 49343faa52c0SJohn Hubbard * that the huge page is present in the page tables. If 49353faa52c0SJohn Hubbard * the huge page is present, then the tail pages must 49363faa52c0SJohn Hubbard * also be present. The ptl prevents the head page and 49373faa52c0SJohn Hubbard * tail pages from being rearranged in any way. So this 49383faa52c0SJohn Hubbard * page must be available at this point, unless the page 49393faa52c0SJohn Hubbard * refcount overflowed: 49403faa52c0SJohn Hubbard */ 49413faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { 49423faa52c0SJohn Hubbard spin_unlock(ptl); 49433faa52c0SJohn Hubbard remainder = 0; 49443faa52c0SJohn Hubbard err = -ENOMEM; 49453faa52c0SJohn Hubbard break; 49463faa52c0SJohn Hubbard } 4947d6692183SChen, Kenneth W } 494863551ae0SDavid Gibson 494963551ae0SDavid Gibson if (vmas) 495063551ae0SDavid Gibson vmas[i] = vma; 495163551ae0SDavid Gibson 495263551ae0SDavid Gibson vaddr += PAGE_SIZE; 4953d5d4b0aaSChen, Kenneth W ++pfn_offset; 495463551ae0SDavid Gibson --remainder; 495563551ae0SDavid Gibson ++i; 4956d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4957a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4958d5d4b0aaSChen, Kenneth W /* 4959d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4960d5d4b0aaSChen, Kenneth W * of this compound page. 4961d5d4b0aaSChen, Kenneth W */ 4962d5d4b0aaSChen, Kenneth W goto same_page; 4963d5d4b0aaSChen, Kenneth W } 4964cb900f41SKirill A. Shutemov spin_unlock(ptl); 496563551ae0SDavid Gibson } 496628a35716SMichel Lespinasse *nr_pages = remainder; 496787ffc118SAndrea Arcangeli /* 496887ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 496987ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 497087ffc118SAndrea Arcangeli * branch. 497187ffc118SAndrea Arcangeli */ 497263551ae0SDavid Gibson *position = vaddr; 497363551ae0SDavid Gibson 49742be7cfedSDaniel Jordan return i ? i : err; 497563551ae0SDavid Gibson } 49768f860591SZhang, Yanmin 49775491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 49785491ae7bSAneesh Kumar K.V /* 49795491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 49805491ae7bSAneesh Kumar K.V * implement this. 49815491ae7bSAneesh Kumar K.V */ 49825491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 49835491ae7bSAneesh Kumar K.V #endif 49845491ae7bSAneesh Kumar K.V 49857da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 49868f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 49878f860591SZhang, Yanmin { 49888f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 49898f860591SZhang, Yanmin unsigned long start = address; 49908f860591SZhang, Yanmin pte_t *ptep; 49918f860591SZhang, Yanmin pte_t pte; 4992a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49937da4d641SPeter Zijlstra unsigned long pages = 0; 4994dff11abeSMike Kravetz bool shared_pmd = false; 4995ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4996dff11abeSMike Kravetz 4997dff11abeSMike Kravetz /* 4998dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4999ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5000dff11abeSMike Kravetz * range if PMD sharing is possible. 5001dff11abeSMike Kravetz */ 50027269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 50037269f999SJérôme Glisse 0, vma, mm, start, end); 5004ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50058f860591SZhang, Yanmin 50068f860591SZhang, Yanmin BUG_ON(address >= end); 5007ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 50088f860591SZhang, Yanmin 5009ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 501083cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5011a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5012cb900f41SKirill A. Shutemov spinlock_t *ptl; 50137868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 50148f860591SZhang, Yanmin if (!ptep) 50158f860591SZhang, Yanmin continue; 5016cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 50177da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 50187da4d641SPeter Zijlstra pages++; 5019cb900f41SKirill A. Shutemov spin_unlock(ptl); 5020dff11abeSMike Kravetz shared_pmd = true; 502139dde65cSChen, Kenneth W continue; 50227da4d641SPeter Zijlstra } 5023a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5024a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5025a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5026a8bda28dSNaoya Horiguchi continue; 5027a8bda28dSNaoya Horiguchi } 5028a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5029a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5030a8bda28dSNaoya Horiguchi 5031a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5032a8bda28dSNaoya Horiguchi pte_t newpte; 5033a8bda28dSNaoya Horiguchi 5034a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5035a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5036e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5037e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5038a8bda28dSNaoya Horiguchi pages++; 5039a8bda28dSNaoya Horiguchi } 5040a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5041a8bda28dSNaoya Horiguchi continue; 5042a8bda28dSNaoya Horiguchi } 5043a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5044023bdd00SAneesh Kumar K.V pte_t old_pte; 5045023bdd00SAneesh Kumar K.V 5046023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5047023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5048be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5049023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50507da4d641SPeter Zijlstra pages++; 50518f860591SZhang, Yanmin } 5052cb900f41SKirill A. Shutemov spin_unlock(ptl); 50538f860591SZhang, Yanmin } 5054d833352aSMel Gorman /* 5055c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5056d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5057c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5058dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5059dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5060d833352aSMel Gorman */ 5061dff11abeSMike Kravetz if (shared_pmd) 5062ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5063dff11abeSMike Kravetz else 50645491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50650f10851eSJérôme Glisse /* 50660f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 50670f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 50680f10851eSJérôme Glisse * 5069ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50700f10851eSJérôme Glisse */ 507183cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5072ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50737da4d641SPeter Zijlstra 50747da4d641SPeter Zijlstra return pages << h->order; 50758f860591SZhang, Yanmin } 50768f860591SZhang, Yanmin 5077a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 5078a1e78772SMel Gorman long from, long to, 50795a6fe125SMel Gorman struct vm_area_struct *vma, 5080ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5081e4e574b7SAdam Litke { 50820db9d74eSMina Almasry long ret, chg, add = -1; 5083a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 508490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 50859119a41eSJoonsoo Kim struct resv_map *resv_map; 5086075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 50870db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5088e4e574b7SAdam Litke 508963489f8eSMike Kravetz /* This should never happen */ 509063489f8eSMike Kravetz if (from > to) { 509163489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 509263489f8eSMike Kravetz return -EINVAL; 509363489f8eSMike Kravetz } 509463489f8eSMike Kravetz 5095a1e78772SMel Gorman /* 509617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 509717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 509890481622SDavid Gibson * without using reserves 509917c9d12eSMel Gorman */ 5100ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 510117c9d12eSMel Gorman return 0; 510217c9d12eSMel Gorman 510317c9d12eSMel Gorman /* 5104a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5105a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5106a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5107a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5108a1e78772SMel Gorman */ 51099119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5110f27a5136SMike Kravetz /* 5111f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5112f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5113f27a5136SMike Kravetz * hugetlbfs_get_inode). 5114f27a5136SMike Kravetz */ 51154e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 51169119a41eSJoonsoo Kim 51170db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 51189119a41eSJoonsoo Kim 51199119a41eSJoonsoo Kim } else { 5120e9fe92aeSMina Almasry /* Private mapping. */ 51219119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51225a6fe125SMel Gorman if (!resv_map) 51235a6fe125SMel Gorman return -ENOMEM; 51245a6fe125SMel Gorman 512517c9d12eSMel Gorman chg = to - from; 512617c9d12eSMel Gorman 51275a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51285a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51295a6fe125SMel Gorman } 51305a6fe125SMel Gorman 5131c50ac050SDave Hansen if (chg < 0) { 5132c50ac050SDave Hansen ret = chg; 5133c50ac050SDave Hansen goto out_err; 5134c50ac050SDave Hansen } 513517c9d12eSMel Gorman 5136075a61d0SMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 5137075a61d0SMina Almasry hstate_index(h), chg * pages_per_huge_page(h), &h_cg); 5138075a61d0SMina Almasry 5139075a61d0SMina Almasry if (ret < 0) { 5140075a61d0SMina Almasry ret = -ENOMEM; 5141075a61d0SMina Almasry goto out_err; 5142075a61d0SMina Almasry } 5143075a61d0SMina Almasry 5144075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5145075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5146075a61d0SMina Almasry * of the resv_map. 5147075a61d0SMina Almasry */ 5148075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5149075a61d0SMina Almasry } 5150075a61d0SMina Almasry 51511c5ecae3SMike Kravetz /* 51521c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51531c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51541c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51551c5ecae3SMike Kravetz */ 51561c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 51571c5ecae3SMike Kravetz if (gbl_reserve < 0) { 5158c50ac050SDave Hansen ret = -ENOSPC; 5159075a61d0SMina Almasry goto out_uncharge_cgroup; 5160c50ac050SDave Hansen } 516117c9d12eSMel Gorman 516217c9d12eSMel Gorman /* 516317c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 516490481622SDavid Gibson * Hand the pages back to the subpool if there are not 516517c9d12eSMel Gorman */ 51661c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 516717c9d12eSMel Gorman if (ret < 0) { 5168075a61d0SMina Almasry goto out_put_pages; 516917c9d12eSMel Gorman } 517017c9d12eSMel Gorman 517117c9d12eSMel Gorman /* 517217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 517317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 517417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 517517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 517617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 517717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 517817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 517917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 518017c9d12eSMel Gorman * else has to be done for private mappings here 518117c9d12eSMel Gorman */ 518233039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5183075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 518433039678SMike Kravetz 51850db9d74eSMina Almasry if (unlikely(add < 0)) { 51860db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5187075a61d0SMina Almasry goto out_put_pages; 51880db9d74eSMina Almasry } else if (unlikely(chg > add)) { 518933039678SMike Kravetz /* 519033039678SMike Kravetz * pages in this range were added to the reserve 519133039678SMike Kravetz * map between region_chg and region_add. This 519233039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 519333039678SMike Kravetz * the subpool and reserve counts modified above 519433039678SMike Kravetz * based on the difference. 519533039678SMike Kravetz */ 519633039678SMike Kravetz long rsv_adjust; 519733039678SMike Kravetz 5198075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5199075a61d0SMina Almasry hstate_index(h), 5200075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5201075a61d0SMina Almasry 520233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 520333039678SMike Kravetz chg - add); 520433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 520533039678SMike Kravetz } 520633039678SMike Kravetz } 5207a43a8c39SChen, Kenneth W return 0; 5208075a61d0SMina Almasry out_put_pages: 5209075a61d0SMina Almasry /* put back original number of pages, chg */ 5210075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5211075a61d0SMina Almasry out_uncharge_cgroup: 5212075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5213075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5214c50ac050SDave Hansen out_err: 52155e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 52160db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 52170db9d74eSMina Almasry * region_add failed or didn't run. 52180db9d74eSMina Almasry */ 52190db9d74eSMina Almasry if (chg >= 0 && add < 0) 52200db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5221f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5222f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 5223c50ac050SDave Hansen return ret; 5224a43a8c39SChen, Kenneth W } 5225a43a8c39SChen, Kenneth W 5226b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5227b5cec28dSMike Kravetz long freed) 5228a43a8c39SChen, Kenneth W { 5229a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52304e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52319119a41eSJoonsoo Kim long chg = 0; 523290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52331c5ecae3SMike Kravetz long gbl_reserve; 523445c682a6SKen Chen 5235f27a5136SMike Kravetz /* 5236f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5237f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5238f27a5136SMike Kravetz */ 5239b5cec28dSMike Kravetz if (resv_map) { 5240b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5241b5cec28dSMike Kravetz /* 5242b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5243b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5244b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5245b5cec28dSMike Kravetz */ 5246b5cec28dSMike Kravetz if (chg < 0) 5247b5cec28dSMike Kravetz return chg; 5248b5cec28dSMike Kravetz } 5249b5cec28dSMike Kravetz 525045c682a6SKen Chen spin_lock(&inode->i_lock); 5251e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 525245c682a6SKen Chen spin_unlock(&inode->i_lock); 525345c682a6SKen Chen 52541c5ecae3SMike Kravetz /* 52551c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52561c5ecae3SMike Kravetz * reservations to be released may be adjusted. 52571c5ecae3SMike Kravetz */ 52581c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52591c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5260b5cec28dSMike Kravetz 5261b5cec28dSMike Kravetz return 0; 5262a43a8c39SChen, Kenneth W } 526393f70f90SNaoya Horiguchi 52643212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 52653212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 52663212b535SSteve Capper struct vm_area_struct *vma, 52673212b535SSteve Capper unsigned long addr, pgoff_t idx) 52683212b535SSteve Capper { 52693212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52703212b535SSteve Capper svma->vm_start; 52713212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52723212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 52733212b535SSteve Capper 52743212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5275de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5276de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 52773212b535SSteve Capper 52783212b535SSteve Capper /* 52793212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 52803212b535SSteve Capper * page table page. 52813212b535SSteve Capper */ 52823212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 52833212b535SSteve Capper vm_flags != svm_flags || 52843212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 52853212b535SSteve Capper return 0; 52863212b535SSteve Capper 52873212b535SSteve Capper return saddr; 52883212b535SSteve Capper } 52893212b535SSteve Capper 529031aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 52913212b535SSteve Capper { 52923212b535SSteve Capper unsigned long base = addr & PUD_MASK; 52933212b535SSteve Capper unsigned long end = base + PUD_SIZE; 52943212b535SSteve Capper 52953212b535SSteve Capper /* 52963212b535SSteve Capper * check on proper vm_flags and page table alignment 52973212b535SSteve Capper */ 5298017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 529931aafb45SNicholas Krause return true; 530031aafb45SNicholas Krause return false; 53013212b535SSteve Capper } 53023212b535SSteve Capper 53033212b535SSteve Capper /* 5304017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5305017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5306017b1660SMike Kravetz * shared pmd mappings. 5307017b1660SMike Kravetz */ 5308017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5309017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5310017b1660SMike Kravetz { 5311353b2de4SMateusz Nosek unsigned long check_addr; 5312017b1660SMike Kravetz 5313017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 5314017b1660SMike Kravetz return; 5315017b1660SMike Kravetz 5316017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 5317017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 5318017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 5319017b1660SMike Kravetz 5320017b1660SMike Kravetz /* 5321017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 5322017b1660SMike Kravetz */ 5323017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 5324017b1660SMike Kravetz if (a_start < *start) 5325017b1660SMike Kravetz *start = a_start; 5326017b1660SMike Kravetz if (a_end > *end) 5327017b1660SMike Kravetz *end = a_end; 5328017b1660SMike Kravetz } 5329017b1660SMike Kravetz } 5330017b1660SMike Kravetz } 5331017b1660SMike Kravetz 5332017b1660SMike Kravetz /* 53333212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53343212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53353212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5336c0d0381aSMike Kravetz * code much cleaner. 5337c0d0381aSMike Kravetz * 5338c0d0381aSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode. 5339c0d0381aSMike Kravetz * For hugetlbfs, this prevents removal of any page table entries associated 5340c0d0381aSMike Kravetz * with the address space. This is important as we are setting up sharing 5341c0d0381aSMike Kravetz * based on existing page table entries (mappings). 53423212b535SSteve Capper */ 53433212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 53443212b535SSteve Capper { 53453212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 53463212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 53473212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 53483212b535SSteve Capper vma->vm_pgoff; 53493212b535SSteve Capper struct vm_area_struct *svma; 53503212b535SSteve Capper unsigned long saddr; 53513212b535SSteve Capper pte_t *spte = NULL; 53523212b535SSteve Capper pte_t *pte; 5353cb900f41SKirill A. Shutemov spinlock_t *ptl; 53543212b535SSteve Capper 53553212b535SSteve Capper if (!vma_shareable(vma, addr)) 53563212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 53573212b535SSteve Capper 53583212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 53593212b535SSteve Capper if (svma == vma) 53603212b535SSteve Capper continue; 53613212b535SSteve Capper 53623212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 53633212b535SSteve Capper if (saddr) { 53647868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 53657868a208SPunit Agrawal vma_mmu_pagesize(svma)); 53663212b535SSteve Capper if (spte) { 53673212b535SSteve Capper get_page(virt_to_page(spte)); 53683212b535SSteve Capper break; 53693212b535SSteve Capper } 53703212b535SSteve Capper } 53713212b535SSteve Capper } 53723212b535SSteve Capper 53733212b535SSteve Capper if (!spte) 53743212b535SSteve Capper goto out; 53753212b535SSteve Capper 53768bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5377dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 53783212b535SSteve Capper pud_populate(mm, pud, 53793212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5380c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5381dc6c9a35SKirill A. Shutemov } else { 53823212b535SSteve Capper put_page(virt_to_page(spte)); 5383dc6c9a35SKirill A. Shutemov } 5384cb900f41SKirill A. Shutemov spin_unlock(ptl); 53853212b535SSteve Capper out: 53863212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 53873212b535SSteve Capper return pte; 53883212b535SSteve Capper } 53893212b535SSteve Capper 53903212b535SSteve Capper /* 53913212b535SSteve Capper * unmap huge page backed by shared pte. 53923212b535SSteve Capper * 53933212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 53943212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 53953212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 53963212b535SSteve Capper * 5397c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 53983212b535SSteve Capper * 53993212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 54003212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 54013212b535SSteve Capper */ 54023212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 54033212b535SSteve Capper { 54043212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5405c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5406c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 54073212b535SSteve Capper 54083212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 54093212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 54103212b535SSteve Capper return 0; 54113212b535SSteve Capper 54123212b535SSteve Capper pud_clear(pud); 54133212b535SSteve Capper put_page(virt_to_page(ptep)); 5414dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54153212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54163212b535SSteve Capper return 1; 54173212b535SSteve Capper } 54189e5fc74cSSteve Capper #define want_pmd_share() (1) 54199e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54209e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 54219e5fc74cSSteve Capper { 54229e5fc74cSSteve Capper return NULL; 54239e5fc74cSSteve Capper } 5424e81f2d22SZhang Zhen 5425e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 5426e81f2d22SZhang Zhen { 5427e81f2d22SZhang Zhen return 0; 5428e81f2d22SZhang Zhen } 5429017b1660SMike Kravetz 5430017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5431017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5432017b1660SMike Kravetz { 5433017b1660SMike Kravetz } 54349e5fc74cSSteve Capper #define want_pmd_share() (0) 54353212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54363212b535SSteve Capper 54379e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 54389e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 54399e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54409e5fc74cSSteve Capper { 54419e5fc74cSSteve Capper pgd_t *pgd; 5442c2febafcSKirill A. Shutemov p4d_t *p4d; 54439e5fc74cSSteve Capper pud_t *pud; 54449e5fc74cSSteve Capper pte_t *pte = NULL; 54459e5fc74cSSteve Capper 54469e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5447f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5448f4f0a3d8SKirill A. Shutemov if (!p4d) 5449f4f0a3d8SKirill A. Shutemov return NULL; 5450c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 54519e5fc74cSSteve Capper if (pud) { 54529e5fc74cSSteve Capper if (sz == PUD_SIZE) { 54539e5fc74cSSteve Capper pte = (pte_t *)pud; 54549e5fc74cSSteve Capper } else { 54559e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 54569e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 54579e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 54589e5fc74cSSteve Capper else 54599e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54609e5fc74cSSteve Capper } 54619e5fc74cSSteve Capper } 54624e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 54639e5fc74cSSteve Capper 54649e5fc74cSSteve Capper return pte; 54659e5fc74cSSteve Capper } 54669e5fc74cSSteve Capper 54679b19df29SPunit Agrawal /* 54689b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 54699b19df29SPunit Agrawal * entry at address @addr 54709b19df29SPunit Agrawal * 54718ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 54728ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 54739b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 54749b19df29SPunit Agrawal * table. 54759b19df29SPunit Agrawal */ 54767868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 54777868a208SPunit Agrawal unsigned long addr, unsigned long sz) 54789e5fc74cSSteve Capper { 54799e5fc74cSSteve Capper pgd_t *pgd; 5480c2febafcSKirill A. Shutemov p4d_t *p4d; 54818ac0b81aSLi Xinhai pud_t *pud; 54828ac0b81aSLi Xinhai pmd_t *pmd; 54839e5fc74cSSteve Capper 54849e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5485c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5486c2febafcSKirill A. Shutemov return NULL; 5487c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5488c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5489c2febafcSKirill A. Shutemov return NULL; 54909b19df29SPunit Agrawal 5491c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 54928ac0b81aSLi Xinhai if (sz == PUD_SIZE) 54938ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 54949e5fc74cSSteve Capper return (pte_t *)pud; 54958ac0b81aSLi Xinhai if (!pud_present(*pud)) 54968ac0b81aSLi Xinhai return NULL; 54978ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 54989b19df29SPunit Agrawal 54999e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 55008ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 55019e5fc74cSSteve Capper return (pte_t *)pmd; 55029e5fc74cSSteve Capper } 55039e5fc74cSSteve Capper 550461f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 550561f77edaSNaoya Horiguchi 550661f77edaSNaoya Horiguchi /* 550761f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 550861f77edaSNaoya Horiguchi * behavior. 550961f77edaSNaoya Horiguchi */ 551061f77edaSNaoya Horiguchi struct page * __weak 551161f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 551261f77edaSNaoya Horiguchi int write) 551361f77edaSNaoya Horiguchi { 551461f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 551561f77edaSNaoya Horiguchi } 551661f77edaSNaoya Horiguchi 551761f77edaSNaoya Horiguchi struct page * __weak 55184dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55194dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55204dc71451SAneesh Kumar K.V { 55214dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55224dc71451SAneesh Kumar K.V return NULL; 55234dc71451SAneesh Kumar K.V } 55244dc71451SAneesh Kumar K.V 55254dc71451SAneesh Kumar K.V struct page * __weak 55269e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5527e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55289e5fc74cSSteve Capper { 5529e66f17ffSNaoya Horiguchi struct page *page = NULL; 5530e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5531c9d398faSNaoya Horiguchi pte_t pte; 55323faa52c0SJohn Hubbard 55333faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55343faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55353faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55363faa52c0SJohn Hubbard return NULL; 55373faa52c0SJohn Hubbard 5538e66f17ffSNaoya Horiguchi retry: 5539e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5540e66f17ffSNaoya Horiguchi spin_lock(ptl); 5541e66f17ffSNaoya Horiguchi /* 5542e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5543e66f17ffSNaoya Horiguchi * unmapped from other threads. 5544e66f17ffSNaoya Horiguchi */ 5545e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5546e66f17ffSNaoya Horiguchi goto out; 5547c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5548c9d398faSNaoya Horiguchi if (pte_present(pte)) { 554997534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 55503faa52c0SJohn Hubbard /* 55513faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 55523faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 55533faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 55543faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 55553faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 55563faa52c0SJohn Hubbard * unless the page refcount overflowed: 55573faa52c0SJohn Hubbard */ 55583faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 55593faa52c0SJohn Hubbard page = NULL; 55603faa52c0SJohn Hubbard goto out; 55613faa52c0SJohn Hubbard } 5562e66f17ffSNaoya Horiguchi } else { 5563c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5564e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5565e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5566e66f17ffSNaoya Horiguchi goto retry; 5567e66f17ffSNaoya Horiguchi } 5568e66f17ffSNaoya Horiguchi /* 5569e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5570e66f17ffSNaoya Horiguchi * follow_page_mask(). 5571e66f17ffSNaoya Horiguchi */ 5572e66f17ffSNaoya Horiguchi } 5573e66f17ffSNaoya Horiguchi out: 5574e66f17ffSNaoya Horiguchi spin_unlock(ptl); 55759e5fc74cSSteve Capper return page; 55769e5fc74cSSteve Capper } 55779e5fc74cSSteve Capper 557861f77edaSNaoya Horiguchi struct page * __weak 55799e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5580e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 55819e5fc74cSSteve Capper { 55823faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5583e66f17ffSNaoya Horiguchi return NULL; 55849e5fc74cSSteve Capper 5585e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 55869e5fc74cSSteve Capper } 55879e5fc74cSSteve Capper 5588faaa5b62SAnshuman Khandual struct page * __weak 5589faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5590faaa5b62SAnshuman Khandual { 55913faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5592faaa5b62SAnshuman Khandual return NULL; 5593faaa5b62SAnshuman Khandual 5594faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5595faaa5b62SAnshuman Khandual } 5596faaa5b62SAnshuman Khandual 559731caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 559831caf665SNaoya Horiguchi { 5599bcc54222SNaoya Horiguchi bool ret = true; 5600bcc54222SNaoya Horiguchi 5601309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 560231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5603bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 5604bcc54222SNaoya Horiguchi ret = false; 5605bcc54222SNaoya Horiguchi goto unlock; 5606bcc54222SNaoya Horiguchi } 5607bcc54222SNaoya Horiguchi clear_page_huge_active(page); 560831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5609bcc54222SNaoya Horiguchi unlock: 561031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5611bcc54222SNaoya Horiguchi return ret; 561231caf665SNaoya Horiguchi } 561331caf665SNaoya Horiguchi 561431caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 561531caf665SNaoya Horiguchi { 5616309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 561731caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5618bcc54222SNaoya Horiguchi set_page_huge_active(page); 561931caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 562031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 562131caf665SNaoya Horiguchi put_page(page); 562231caf665SNaoya Horiguchi } 5623ab5ac90aSMichal Hocko 5624ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5625ab5ac90aSMichal Hocko { 5626ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5627ab5ac90aSMichal Hocko 5628ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5629ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5630ab5ac90aSMichal Hocko 5631ab5ac90aSMichal Hocko /* 5632ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5633ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5634ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5635ab5ac90aSMichal Hocko * the temporary status. 5636ab5ac90aSMichal Hocko * 5637ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5638ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5639ab5ac90aSMichal Hocko * the per-node's. 5640ab5ac90aSMichal Hocko */ 5641ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5642ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5643ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5644ab5ac90aSMichal Hocko 5645ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5646ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5647ab5ac90aSMichal Hocko 5648ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5649ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5650ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5651ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5652ab5ac90aSMichal Hocko } 5653ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5654ab5ac90aSMichal Hocko } 5655ab5ac90aSMichal Hocko } 5656cf11e85fSRoman Gushchin 5657cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5658cf11e85fSRoman Gushchin static unsigned long hugetlb_cma_size __initdata; 5659cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5660cf11e85fSRoman Gushchin 5661cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5662cf11e85fSRoman Gushchin { 5663cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5664cf11e85fSRoman Gushchin return 0; 5665cf11e85fSRoman Gushchin } 5666cf11e85fSRoman Gushchin 5667cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5668cf11e85fSRoman Gushchin 5669cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5670cf11e85fSRoman Gushchin { 5671cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5672cf11e85fSRoman Gushchin int nid; 5673cf11e85fSRoman Gushchin 5674cf11e85fSRoman Gushchin cma_reserve_called = true; 5675cf11e85fSRoman Gushchin 5676cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5677cf11e85fSRoman Gushchin return; 5678cf11e85fSRoman Gushchin 5679cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5680cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5681cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5682cf11e85fSRoman Gushchin return; 5683cf11e85fSRoman Gushchin } 5684cf11e85fSRoman Gushchin 5685cf11e85fSRoman Gushchin /* 5686cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5687cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5688cf11e85fSRoman Gushchin */ 5689cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5690cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5691cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5692cf11e85fSRoman Gushchin 5693cf11e85fSRoman Gushchin reserved = 0; 5694cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5695cf11e85fSRoman Gushchin int res; 5696cf11e85fSRoman Gushchin 5697cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5698cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5699cf11e85fSRoman Gushchin 5700cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 5701cf11e85fSRoman Gushchin 0, false, "hugetlb", 5702cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5703cf11e85fSRoman Gushchin if (res) { 5704cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5705cf11e85fSRoman Gushchin res, nid); 5706cf11e85fSRoman Gushchin continue; 5707cf11e85fSRoman Gushchin } 5708cf11e85fSRoman Gushchin 5709cf11e85fSRoman Gushchin reserved += size; 5710cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5711cf11e85fSRoman Gushchin size / SZ_1M, nid); 5712cf11e85fSRoman Gushchin 5713cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5714cf11e85fSRoman Gushchin break; 5715cf11e85fSRoman Gushchin } 5716cf11e85fSRoman Gushchin } 5717cf11e85fSRoman Gushchin 5718cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5719cf11e85fSRoman Gushchin { 5720cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5721cf11e85fSRoman Gushchin return; 5722cf11e85fSRoman Gushchin 5723cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5724cf11e85fSRoman Gushchin } 5725cf11e85fSRoman Gushchin 5726cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5727