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> 34ca15ca40SMike Rapoport #include <asm/pgalloc.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 49dbda8feaSBarry Song #ifdef CONFIG_CMA 50cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 51dbda8feaSBarry Song #endif 52dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 53cf11e85fSRoman Gushchin 54641844f5SNaoya Horiguchi /* 55641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 56641844f5SNaoya Horiguchi * at boot time. 57641844f5SNaoya Horiguchi */ 58641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 59e5ff2159SAndi Kleen 6053ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6153ba51d2SJon Tollefson 62e5ff2159SAndi Kleen /* for command line parsing */ 63e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 64e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 659fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 66282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 67e5ff2159SAndi Kleen 683935baa9SDavid Gibson /* 6931caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7031caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 713935baa9SDavid Gibson */ 72c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 730bd0f9fbSEric Paris 748382d914SDavidlohr Bueso /* 758382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 768382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 778382d914SDavidlohr Bueso */ 788382d914SDavidlohr Bueso static int num_fault_mutexes; 79c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 808382d914SDavidlohr Bueso 817ca02d0aSMike Kravetz /* Forward declaration */ 827ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 837ca02d0aSMike Kravetz 8490481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 8590481622SDavid Gibson { 8690481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 8790481622SDavid Gibson 8890481622SDavid Gibson spin_unlock(&spool->lock); 8990481622SDavid Gibson 9090481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 917c8de358SEthon Paul * remain, give up any reservations based on minimum size and 927ca02d0aSMike Kravetz * free the subpool */ 937ca02d0aSMike Kravetz if (free) { 947ca02d0aSMike Kravetz if (spool->min_hpages != -1) 957ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 967ca02d0aSMike Kravetz -spool->min_hpages); 9790481622SDavid Gibson kfree(spool); 9890481622SDavid Gibson } 997ca02d0aSMike Kravetz } 10090481622SDavid Gibson 1017ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1027ca02d0aSMike Kravetz long min_hpages) 10390481622SDavid Gibson { 10490481622SDavid Gibson struct hugepage_subpool *spool; 10590481622SDavid Gibson 106c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 10790481622SDavid Gibson if (!spool) 10890481622SDavid Gibson return NULL; 10990481622SDavid Gibson 11090481622SDavid Gibson spin_lock_init(&spool->lock); 11190481622SDavid Gibson spool->count = 1; 1127ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1137ca02d0aSMike Kravetz spool->hstate = h; 1147ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1157ca02d0aSMike Kravetz 1167ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1177ca02d0aSMike Kravetz kfree(spool); 1187ca02d0aSMike Kravetz return NULL; 1197ca02d0aSMike Kravetz } 1207ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 12190481622SDavid Gibson 12290481622SDavid Gibson return spool; 12390481622SDavid Gibson } 12490481622SDavid Gibson 12590481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 12690481622SDavid Gibson { 12790481622SDavid Gibson spin_lock(&spool->lock); 12890481622SDavid Gibson BUG_ON(!spool->count); 12990481622SDavid Gibson spool->count--; 13090481622SDavid Gibson unlock_or_release_subpool(spool); 13190481622SDavid Gibson } 13290481622SDavid Gibson 1331c5ecae3SMike Kravetz /* 1341c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1351c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1361c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1371c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1381c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1397c8de358SEthon Paul * a subpool minimum size must be maintained. 1401c5ecae3SMike Kravetz */ 1411c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 14290481622SDavid Gibson long delta) 14390481622SDavid Gibson { 1441c5ecae3SMike Kravetz long ret = delta; 14590481622SDavid Gibson 14690481622SDavid Gibson if (!spool) 1471c5ecae3SMike Kravetz return ret; 14890481622SDavid Gibson 14990481622SDavid Gibson spin_lock(&spool->lock); 15090481622SDavid Gibson 1511c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1521c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1531c5ecae3SMike Kravetz spool->used_hpages += delta; 1541c5ecae3SMike Kravetz else { 1551c5ecae3SMike Kravetz ret = -ENOMEM; 1561c5ecae3SMike Kravetz goto unlock_ret; 1571c5ecae3SMike Kravetz } 1581c5ecae3SMike Kravetz } 1591c5ecae3SMike Kravetz 16009a95e29SMike Kravetz /* minimum size accounting */ 16109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1621c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1631c5ecae3SMike Kravetz /* 1641c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1651c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1661c5ecae3SMike Kravetz */ 1671c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1681c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1691c5ecae3SMike Kravetz } else { 1701c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1711c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1721c5ecae3SMike Kravetz } 1731c5ecae3SMike Kravetz } 1741c5ecae3SMike Kravetz 1751c5ecae3SMike Kravetz unlock_ret: 1761c5ecae3SMike Kravetz spin_unlock(&spool->lock); 17790481622SDavid Gibson return ret; 17890481622SDavid Gibson } 17990481622SDavid Gibson 1801c5ecae3SMike Kravetz /* 1811c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1821c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1831c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1841c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1851c5ecae3SMike Kravetz */ 1861c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 18790481622SDavid Gibson long delta) 18890481622SDavid Gibson { 1891c5ecae3SMike Kravetz long ret = delta; 1901c5ecae3SMike Kravetz 19190481622SDavid Gibson if (!spool) 1921c5ecae3SMike Kravetz return delta; 19390481622SDavid Gibson 19490481622SDavid Gibson spin_lock(&spool->lock); 1951c5ecae3SMike Kravetz 1961c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 19790481622SDavid Gibson spool->used_hpages -= delta; 1981c5ecae3SMike Kravetz 19909a95e29SMike Kravetz /* minimum size accounting */ 20009a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2011c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2021c5ecae3SMike Kravetz ret = 0; 2031c5ecae3SMike Kravetz else 2041c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2051c5ecae3SMike Kravetz 2061c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2071c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2081c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2091c5ecae3SMike Kravetz } 2101c5ecae3SMike Kravetz 2111c5ecae3SMike Kravetz /* 2121c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2131c5ecae3SMike Kravetz * quota reference, free it now. 2141c5ecae3SMike Kravetz */ 21590481622SDavid Gibson unlock_or_release_subpool(spool); 2161c5ecae3SMike Kravetz 2171c5ecae3SMike Kravetz return ret; 21890481622SDavid Gibson } 21990481622SDavid Gibson 22090481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 22190481622SDavid Gibson { 22290481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 22390481622SDavid Gibson } 22490481622SDavid Gibson 22590481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 22690481622SDavid Gibson { 227496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 22890481622SDavid Gibson } 22990481622SDavid Gibson 2300db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2310db9d74eSMina Almasry * it for use. 2320db9d74eSMina Almasry */ 2330db9d74eSMina Almasry static struct file_region * 2340db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2350db9d74eSMina Almasry { 2360db9d74eSMina Almasry struct file_region *nrg = NULL; 2370db9d74eSMina Almasry 2380db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2390db9d74eSMina Almasry 2400db9d74eSMina Almasry resv->region_cache_count--; 2410db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2420db9d74eSMina Almasry VM_BUG_ON(!nrg); 2430db9d74eSMina Almasry list_del(&nrg->link); 2440db9d74eSMina Almasry 2450db9d74eSMina Almasry nrg->from = from; 2460db9d74eSMina Almasry nrg->to = to; 2470db9d74eSMina Almasry 2480db9d74eSMina Almasry return nrg; 2490db9d74eSMina Almasry } 2500db9d74eSMina Almasry 251075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 252075a61d0SMina Almasry struct file_region *rg) 253075a61d0SMina Almasry { 254075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 255075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 256075a61d0SMina Almasry nrg->css = rg->css; 257075a61d0SMina Almasry if (rg->css) 258075a61d0SMina Almasry css_get(rg->css); 259075a61d0SMina Almasry #endif 260075a61d0SMina Almasry } 261075a61d0SMina Almasry 262075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 263075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 264075a61d0SMina Almasry struct hstate *h, 265075a61d0SMina Almasry struct resv_map *resv, 266075a61d0SMina Almasry struct file_region *nrg) 267075a61d0SMina Almasry { 268075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 269075a61d0SMina Almasry if (h_cg) { 270075a61d0SMina Almasry nrg->reservation_counter = 271075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 272075a61d0SMina Almasry nrg->css = &h_cg->css; 273075a61d0SMina Almasry if (!resv->pages_per_hpage) 274075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 275075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 276075a61d0SMina Almasry * a resv_map. 277075a61d0SMina Almasry */ 278075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 279075a61d0SMina Almasry } else { 280075a61d0SMina Almasry nrg->reservation_counter = NULL; 281075a61d0SMina Almasry nrg->css = NULL; 282075a61d0SMina Almasry } 283075a61d0SMina Almasry #endif 284075a61d0SMina Almasry } 285075a61d0SMina Almasry 286a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 287a9b3f867SMina Almasry struct file_region *org) 288a9b3f867SMina Almasry { 289a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 290a9b3f867SMina Almasry return rg && org && 291a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 292a9b3f867SMina Almasry rg->css == org->css; 293a9b3f867SMina Almasry 294a9b3f867SMina Almasry #else 295a9b3f867SMina Almasry return true; 296a9b3f867SMina Almasry #endif 297a9b3f867SMina Almasry } 298a9b3f867SMina Almasry 299a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 300a9b3f867SMina Almasry { 301a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 302a9b3f867SMina Almasry 303a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 304a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 305a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 306a9b3f867SMina Almasry prg->to = rg->to; 307a9b3f867SMina Almasry 308a9b3f867SMina Almasry list_del(&rg->link); 309a9b3f867SMina Almasry kfree(rg); 310a9b3f867SMina Almasry 311a9b3f867SMina Almasry coalesce_file_region(resv, prg); 312a9b3f867SMina Almasry return; 313a9b3f867SMina Almasry } 314a9b3f867SMina Almasry 315a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 316a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 317a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 318a9b3f867SMina Almasry nrg->from = rg->from; 319a9b3f867SMina Almasry 320a9b3f867SMina Almasry list_del(&rg->link); 321a9b3f867SMina Almasry kfree(rg); 322a9b3f867SMina Almasry 323a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 324a9b3f867SMina Almasry return; 325a9b3f867SMina Almasry } 326a9b3f867SMina Almasry } 327a9b3f867SMina Almasry 328d75c6af9SMina Almasry /* Must be called with resv->lock held. Calling this with count_only == true 329d75c6af9SMina Almasry * will count the number of pages to be added but will not modify the linked 3300db9d74eSMina Almasry * list. If regions_needed != NULL and count_only == true, then regions_needed 3310db9d74eSMina Almasry * will indicate the number of file_regions needed in the cache to carry out to 3320db9d74eSMina Almasry * add the regions for this range. 333d75c6af9SMina Almasry */ 334d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 335075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 336075a61d0SMina Almasry struct hstate *h, long *regions_needed, 337075a61d0SMina Almasry bool count_only) 338d75c6af9SMina Almasry { 3390db9d74eSMina Almasry long add = 0; 340d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3410db9d74eSMina Almasry long last_accounted_offset = f; 342d75c6af9SMina Almasry struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; 343d75c6af9SMina Almasry 3440db9d74eSMina Almasry if (regions_needed) 3450db9d74eSMina Almasry *regions_needed = 0; 346d75c6af9SMina Almasry 3470db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3480db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3490db9d74eSMina Almasry * bounds checking. 3500db9d74eSMina Almasry */ 3510db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 3520db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 3530db9d74eSMina Almasry if (rg->from < f) { 3540db9d74eSMina Almasry /* If this region ends after the last accounted offset, 3550db9d74eSMina Almasry * then we need to update last_accounted_offset. 3560db9d74eSMina Almasry */ 3570db9d74eSMina Almasry if (rg->to > last_accounted_offset) 3580db9d74eSMina Almasry last_accounted_offset = rg->to; 3590db9d74eSMina Almasry continue; 3600db9d74eSMina Almasry } 361d75c6af9SMina Almasry 3620db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 3630db9d74eSMina Almasry * finished. 3640db9d74eSMina Almasry */ 365d75c6af9SMina Almasry if (rg->from > t) 366d75c6af9SMina Almasry break; 367d75c6af9SMina Almasry 3680db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 3690db9d74eSMina Almasry * update last_accounted_offset. 370d75c6af9SMina Almasry */ 3710db9d74eSMina Almasry if (rg->from > last_accounted_offset) { 3720db9d74eSMina Almasry add += rg->from - last_accounted_offset; 3730db9d74eSMina Almasry if (!count_only) { 3740db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3750db9d74eSMina Almasry resv, last_accounted_offset, rg->from); 376075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, 377075a61d0SMina Almasry resv, nrg); 3780db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 379a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 3800db9d74eSMina Almasry } else if (regions_needed) 3810db9d74eSMina Almasry *regions_needed += 1; 382d75c6af9SMina Almasry } 383d75c6af9SMina Almasry 3840db9d74eSMina Almasry last_accounted_offset = rg->to; 3850db9d74eSMina Almasry } 3860db9d74eSMina Almasry 3870db9d74eSMina Almasry /* Handle the case where our range extends beyond 3880db9d74eSMina Almasry * last_accounted_offset. 3890db9d74eSMina Almasry */ 3900db9d74eSMina Almasry if (last_accounted_offset < t) { 3910db9d74eSMina Almasry add += t - last_accounted_offset; 3920db9d74eSMina Almasry if (!count_only) { 3930db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3940db9d74eSMina Almasry resv, last_accounted_offset, t); 395075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); 3960db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 397a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 3980db9d74eSMina Almasry } else if (regions_needed) 3990db9d74eSMina Almasry *regions_needed += 1; 4000db9d74eSMina Almasry } 4010db9d74eSMina Almasry 4020db9d74eSMina Almasry VM_BUG_ON(add < 0); 4030db9d74eSMina Almasry return add; 4040db9d74eSMina Almasry } 4050db9d74eSMina Almasry 4060db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4070db9d74eSMina Almasry */ 4080db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4090db9d74eSMina Almasry int regions_needed) 4100db9d74eSMina Almasry __must_hold(&resv->lock) 4110db9d74eSMina Almasry { 4120db9d74eSMina Almasry struct list_head allocated_regions; 4130db9d74eSMina Almasry int to_allocate = 0, i = 0; 4140db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4150db9d74eSMina Almasry 4160db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4170db9d74eSMina Almasry 4180db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4190db9d74eSMina Almasry 4200db9d74eSMina Almasry /* 4210db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4220db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4230db9d74eSMina Almasry * 4240db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4250db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4260db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4270db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4280db9d74eSMina Almasry */ 4290db9d74eSMina Almasry while (resv->region_cache_count < 4300db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4310db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4320db9d74eSMina Almasry resv->region_cache_count; 4330db9d74eSMina Almasry 4340db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4350db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4360db9d74eSMina Almasry * needing to allocate for regions_needed. 4370db9d74eSMina Almasry */ 4380db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4390db9d74eSMina Almasry 4400db9d74eSMina Almasry spin_unlock(&resv->lock); 4410db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4420db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4430db9d74eSMina Almasry if (!trg) 4440db9d74eSMina Almasry goto out_of_memory; 4450db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4460db9d74eSMina Almasry } 4470db9d74eSMina Almasry 4480db9d74eSMina Almasry spin_lock(&resv->lock); 4490db9d74eSMina Almasry 4500db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 4510db9d74eSMina Almasry list_del(&rg->link); 4520db9d74eSMina Almasry list_add(&rg->link, &resv->region_cache); 4530db9d74eSMina Almasry resv->region_cache_count++; 4540db9d74eSMina Almasry } 4550db9d74eSMina Almasry } 4560db9d74eSMina Almasry 4570db9d74eSMina Almasry return 0; 4580db9d74eSMina Almasry 4590db9d74eSMina Almasry out_of_memory: 4600db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 461d75c6af9SMina Almasry list_del(&rg->link); 462d75c6af9SMina Almasry kfree(rg); 463d75c6af9SMina Almasry } 4640db9d74eSMina Almasry return -ENOMEM; 465d75c6af9SMina Almasry } 466d75c6af9SMina Almasry 4671dd308a7SMike Kravetz /* 4681dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 4690db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 4700db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 4710db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 4720db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 4730db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 474cf3ad20bSMike Kravetz * 4750db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 4760db9d74eSMina Almasry * 4770db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 4780db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 4797c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 4800db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 4810db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 4820db9d74eSMina Almasry * 1 page will only require at most 1 entry. 4831dd308a7SMike Kravetz */ 4840db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 485075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 486075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 48796822904SAndy Whitcroft { 4880db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 48996822904SAndy Whitcroft 4907b24d861SDavidlohr Bueso spin_lock(&resv->lock); 4910db9d74eSMina Almasry retry: 4920db9d74eSMina Almasry 4930db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 494075a61d0SMina Almasry add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed, 495075a61d0SMina Almasry true); 49696822904SAndy Whitcroft 4975e911373SMike Kravetz /* 4980db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4990db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5000db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5010db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5020db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5030db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5040db9d74eSMina Almasry * operation. 5055e911373SMike Kravetz */ 5060db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5070db9d74eSMina Almasry resv->region_cache_count < 5080db9d74eSMina Almasry resv->adds_in_progress + 5090db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5100db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5110db9d74eSMina Almasry * allocate file_region entries. 5120db9d74eSMina Almasry */ 5130db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5145e911373SMike Kravetz 5150db9d74eSMina Almasry if (allocate_file_region_entries( 5160db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5170db9d74eSMina Almasry return -ENOMEM; 5185e911373SMike Kravetz } 5195e911373SMike Kravetz 5200db9d74eSMina Almasry goto retry; 5210db9d74eSMina Almasry } 522cf3ad20bSMike Kravetz 523075a61d0SMina Almasry add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false); 5240db9d74eSMina Almasry 5250db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5260db9d74eSMina Almasry 5277b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 528cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 529cf3ad20bSMike Kravetz return add; 53096822904SAndy Whitcroft } 53196822904SAndy Whitcroft 5321dd308a7SMike Kravetz /* 5331dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5341dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5351dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5361dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5371dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5381dd308a7SMike Kravetz * not change the number of huge pages represented by the 5390db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5400db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5410db9d74eSMina Almasry * file_region structure is added. 5420db9d74eSMina Almasry * 5430db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5440db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5450db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5465e911373SMike Kravetz * 5475e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5485e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5495e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5505e911373SMike Kravetz * is needed and can not be allocated. 5511dd308a7SMike Kravetz */ 5520db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5530db9d74eSMina Almasry long *out_regions_needed) 55496822904SAndy Whitcroft { 55596822904SAndy Whitcroft long chg = 0; 55696822904SAndy Whitcroft 5577b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5585e911373SMike Kravetz 5590db9d74eSMina Almasry /* Count how many hugepages in this range are NOT respresented. */ 560075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 561075a61d0SMina Almasry out_regions_needed, true); 5625e911373SMike Kravetz 5630db9d74eSMina Almasry if (*out_regions_needed == 0) 5640db9d74eSMina Almasry *out_regions_needed = 1; 5655e911373SMike Kravetz 5660db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5675e911373SMike Kravetz return -ENOMEM; 5685e911373SMike Kravetz 5690db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 57096822904SAndy Whitcroft 5717b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 57296822904SAndy Whitcroft return chg; 57396822904SAndy Whitcroft } 57496822904SAndy Whitcroft 5751dd308a7SMike Kravetz /* 5765e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 5775e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 5785e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 5795e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 5800db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 5810db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 5820db9d74eSMina Almasry * the adds_in_progress counter. 5835e911373SMike Kravetz * 5845e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 5855e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 5865e911373SMike Kravetz * arguments will match the associated region_chg call. 5875e911373SMike Kravetz */ 5880db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 5890db9d74eSMina Almasry long regions_needed) 5905e911373SMike Kravetz { 5915e911373SMike Kravetz spin_lock(&resv->lock); 5925e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 5930db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 5945e911373SMike Kravetz spin_unlock(&resv->lock); 5955e911373SMike Kravetz } 5965e911373SMike Kravetz 5975e911373SMike Kravetz /* 598feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 599feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 600feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 601feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 602feba16e2SMike Kravetz * 603feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 604feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 605feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 606feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 607feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 608feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 609feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6101dd308a7SMike Kravetz */ 611feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 61296822904SAndy Whitcroft { 6131406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 61496822904SAndy Whitcroft struct file_region *rg, *trg; 615feba16e2SMike Kravetz struct file_region *nrg = NULL; 616feba16e2SMike Kravetz long del = 0; 61796822904SAndy Whitcroft 618feba16e2SMike Kravetz retry: 6197b24d861SDavidlohr Bueso spin_lock(&resv->lock); 620feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 621dbe409e4SMike Kravetz /* 622dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 623dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 624dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 625dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 626dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 627dbe409e4SMike Kravetz */ 628dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 629feba16e2SMike Kravetz continue; 630dbe409e4SMike Kravetz 631feba16e2SMike Kravetz if (rg->from >= t) 63296822904SAndy Whitcroft break; 63396822904SAndy Whitcroft 634feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 635feba16e2SMike Kravetz /* 636feba16e2SMike Kravetz * Check for an entry in the cache before dropping 637feba16e2SMike Kravetz * lock and attempting allocation. 638feba16e2SMike Kravetz */ 639feba16e2SMike Kravetz if (!nrg && 640feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 641feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 642feba16e2SMike Kravetz struct file_region, 643feba16e2SMike Kravetz link); 644feba16e2SMike Kravetz list_del(&nrg->link); 645feba16e2SMike Kravetz resv->region_cache_count--; 64696822904SAndy Whitcroft } 64796822904SAndy Whitcroft 648feba16e2SMike Kravetz if (!nrg) { 649feba16e2SMike Kravetz spin_unlock(&resv->lock); 650feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 651feba16e2SMike Kravetz if (!nrg) 652feba16e2SMike Kravetz return -ENOMEM; 653feba16e2SMike Kravetz goto retry; 654feba16e2SMike Kravetz } 655feba16e2SMike Kravetz 656feba16e2SMike Kravetz del += t - f; 657feba16e2SMike Kravetz 658feba16e2SMike Kravetz /* New entry for end of split region */ 659feba16e2SMike Kravetz nrg->from = t; 660feba16e2SMike Kravetz nrg->to = rg->to; 661075a61d0SMina Almasry 662075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 663075a61d0SMina Almasry 664feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 665feba16e2SMike Kravetz 666feba16e2SMike Kravetz /* Original entry is trimmed */ 667feba16e2SMike Kravetz rg->to = f; 668feba16e2SMike Kravetz 669075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region( 670075a61d0SMina Almasry resv, rg, nrg->to - nrg->from); 671075a61d0SMina Almasry 672feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 673feba16e2SMike Kravetz nrg = NULL; 67496822904SAndy Whitcroft break; 675feba16e2SMike Kravetz } 676feba16e2SMike Kravetz 677feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 678feba16e2SMike Kravetz del += rg->to - rg->from; 679075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 680075a61d0SMina Almasry rg->to - rg->from); 68196822904SAndy Whitcroft list_del(&rg->link); 68296822904SAndy Whitcroft kfree(rg); 683feba16e2SMike Kravetz continue; 68496822904SAndy Whitcroft } 6857b24d861SDavidlohr Bueso 686feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 687feba16e2SMike Kravetz del += t - rg->from; 688feba16e2SMike Kravetz rg->from = t; 689075a61d0SMina Almasry 690075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 691075a61d0SMina Almasry t - rg->from); 692feba16e2SMike Kravetz } else { /* Trim end of region */ 693feba16e2SMike Kravetz del += rg->to - f; 694feba16e2SMike Kravetz rg->to = f; 695075a61d0SMina Almasry 696075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 697075a61d0SMina Almasry rg->to - f); 698feba16e2SMike Kravetz } 699feba16e2SMike Kravetz } 700feba16e2SMike Kravetz 7017b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 702feba16e2SMike Kravetz kfree(nrg); 703feba16e2SMike Kravetz return del; 70496822904SAndy Whitcroft } 70596822904SAndy Whitcroft 7061dd308a7SMike Kravetz /* 707b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 708b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 709b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 710b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 711b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 712b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 713b5cec28dSMike Kravetz * counts. 714b5cec28dSMike Kravetz */ 71572e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 716b5cec28dSMike Kravetz { 717b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 718b5cec28dSMike Kravetz long rsv_adjust; 719b5cec28dSMike Kravetz 720b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 72172e2936cSzhong jiang if (rsv_adjust) { 722b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 723b5cec28dSMike Kravetz 724b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 725b5cec28dSMike Kravetz } 726b5cec28dSMike Kravetz } 727b5cec28dSMike Kravetz 728b5cec28dSMike Kravetz /* 7291dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7301dd308a7SMike Kravetz * that intersect with the range [f, t). 7311dd308a7SMike Kravetz */ 7321406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 73384afd99bSAndy Whitcroft { 7341406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 73584afd99bSAndy Whitcroft struct file_region *rg; 73684afd99bSAndy Whitcroft long chg = 0; 73784afd99bSAndy Whitcroft 7387b24d861SDavidlohr Bueso spin_lock(&resv->lock); 73984afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 74084afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 741f2135a4aSWang Sheng-Hui long seg_from; 742f2135a4aSWang Sheng-Hui long seg_to; 74384afd99bSAndy Whitcroft 74484afd99bSAndy Whitcroft if (rg->to <= f) 74584afd99bSAndy Whitcroft continue; 74684afd99bSAndy Whitcroft if (rg->from >= t) 74784afd99bSAndy Whitcroft break; 74884afd99bSAndy Whitcroft 74984afd99bSAndy Whitcroft seg_from = max(rg->from, f); 75084afd99bSAndy Whitcroft seg_to = min(rg->to, t); 75184afd99bSAndy Whitcroft 75284afd99bSAndy Whitcroft chg += seg_to - seg_from; 75384afd99bSAndy Whitcroft } 7547b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 75584afd99bSAndy Whitcroft 75684afd99bSAndy Whitcroft return chg; 75784afd99bSAndy Whitcroft } 75884afd99bSAndy Whitcroft 75996822904SAndy Whitcroft /* 760e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 761e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 762e7c4b0bfSAndy Whitcroft */ 763a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 764a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 765e7c4b0bfSAndy Whitcroft { 766a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 767a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 768e7c4b0bfSAndy Whitcroft } 769e7c4b0bfSAndy Whitcroft 7700fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 7710fe6e20bSNaoya Horiguchi unsigned long address) 7720fe6e20bSNaoya Horiguchi { 7730fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 7740fe6e20bSNaoya Horiguchi } 775dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 7760fe6e20bSNaoya Horiguchi 77784afd99bSAndy Whitcroft /* 77808fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 77908fba699SMel Gorman * cases this will be same size as used by the page table entries. 78008fba699SMel Gorman */ 78108fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 78208fba699SMel Gorman { 78305ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 78405ea8860SDan Williams return vma->vm_ops->pagesize(vma); 78508fba699SMel Gorman return PAGE_SIZE; 78608fba699SMel Gorman } 787f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 78808fba699SMel Gorman 78908fba699SMel Gorman /* 7903340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 7913340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 79209135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 79309135cc5SDan Williams * version of this symbol is required. 7943340289dSMel Gorman */ 79509135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 7963340289dSMel Gorman { 7973340289dSMel Gorman return vma_kernel_pagesize(vma); 7983340289dSMel Gorman } 7993340289dSMel Gorman 8003340289dSMel Gorman /* 80184afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 80284afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 80384afd99bSAndy Whitcroft * alignment. 80484afd99bSAndy Whitcroft */ 80584afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 80684afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 80704f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 80884afd99bSAndy Whitcroft 809a1e78772SMel Gorman /* 810a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 811a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 812a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 813a1e78772SMel Gorman * 814a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 815a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 816a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 817a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 81884afd99bSAndy Whitcroft * 81984afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 82084afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 82184afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 82284afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 82384afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 82484afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 82584afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 82684afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 827a1e78772SMel Gorman */ 828e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 829e7c4b0bfSAndy Whitcroft { 830e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 831e7c4b0bfSAndy Whitcroft } 832e7c4b0bfSAndy Whitcroft 833e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 834e7c4b0bfSAndy Whitcroft unsigned long value) 835e7c4b0bfSAndy Whitcroft { 836e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 837e7c4b0bfSAndy Whitcroft } 838e7c4b0bfSAndy Whitcroft 839e9fe92aeSMina Almasry static void 840e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 841e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 842e9fe92aeSMina Almasry struct hstate *h) 843e9fe92aeSMina Almasry { 844e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 845e9fe92aeSMina Almasry if (!h_cg || !h) { 846e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 847e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 848e9fe92aeSMina Almasry resv_map->css = NULL; 849e9fe92aeSMina Almasry } else { 850e9fe92aeSMina Almasry resv_map->reservation_counter = 851e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 852e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 853e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 854e9fe92aeSMina Almasry } 855e9fe92aeSMina Almasry #endif 856e9fe92aeSMina Almasry } 857e9fe92aeSMina Almasry 8589119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 85984afd99bSAndy Whitcroft { 86084afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8615e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8625e911373SMike Kravetz 8635e911373SMike Kravetz if (!resv_map || !rg) { 8645e911373SMike Kravetz kfree(resv_map); 8655e911373SMike Kravetz kfree(rg); 86684afd99bSAndy Whitcroft return NULL; 8675e911373SMike Kravetz } 86884afd99bSAndy Whitcroft 86984afd99bSAndy Whitcroft kref_init(&resv_map->refs); 8707b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 87184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 87284afd99bSAndy Whitcroft 8735e911373SMike Kravetz resv_map->adds_in_progress = 0; 874e9fe92aeSMina Almasry /* 875e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 876e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 877e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 878e9fe92aeSMina Almasry * reservations are to be un-charged from here. 879e9fe92aeSMina Almasry */ 880e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 8815e911373SMike Kravetz 8825e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 8835e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 8845e911373SMike Kravetz resv_map->region_cache_count = 1; 8855e911373SMike Kravetz 88684afd99bSAndy Whitcroft return resv_map; 88784afd99bSAndy Whitcroft } 88884afd99bSAndy Whitcroft 8899119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 89084afd99bSAndy Whitcroft { 89184afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 8925e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 8935e911373SMike Kravetz struct file_region *rg, *trg; 89484afd99bSAndy Whitcroft 89584afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 896feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 8975e911373SMike Kravetz 8985e911373SMike Kravetz /* ... and any entries left in the cache */ 8995e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9005e911373SMike Kravetz list_del(&rg->link); 9015e911373SMike Kravetz kfree(rg); 9025e911373SMike Kravetz } 9035e911373SMike Kravetz 9045e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9055e911373SMike Kravetz 90684afd99bSAndy Whitcroft kfree(resv_map); 90784afd99bSAndy Whitcroft } 90884afd99bSAndy Whitcroft 9094e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9104e35f483SJoonsoo Kim { 911f27a5136SMike Kravetz /* 912f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 913f27a5136SMike Kravetz * address space within the inode. This original address space 914f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 915f27a5136SMike Kravetz * address space embedded within the inode. 916f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 917f27a5136SMike Kravetz * this may not be true for device special inodes. 918f27a5136SMike Kravetz */ 919f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9204e35f483SJoonsoo Kim } 9214e35f483SJoonsoo Kim 92284afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 923a1e78772SMel Gorman { 92481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9254e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9264e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9274e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9284e35f483SJoonsoo Kim 9294e35f483SJoonsoo Kim return inode_resv_map(inode); 9304e35f483SJoonsoo Kim 9314e35f483SJoonsoo Kim } else { 93284afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 93384afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9344e35f483SJoonsoo Kim } 935a1e78772SMel Gorman } 936a1e78772SMel Gorman 93784afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 938a1e78772SMel Gorman { 93981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 94081d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 941a1e78772SMel Gorman 94284afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 94384afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 94404f2cbe3SMel Gorman } 94504f2cbe3SMel Gorman 94604f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 94704f2cbe3SMel Gorman { 94881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 94981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 950e7c4b0bfSAndy Whitcroft 951e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 95204f2cbe3SMel Gorman } 95304f2cbe3SMel Gorman 95404f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 95504f2cbe3SMel Gorman { 95681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 957e7c4b0bfSAndy Whitcroft 958e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 959a1e78772SMel Gorman } 960a1e78772SMel Gorman 96104f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 962a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 963a1e78772SMel Gorman { 96481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 965f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 966a1e78772SMel Gorman vma->vm_private_data = (void *)0; 967a1e78772SMel Gorman } 968a1e78772SMel Gorman 969a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 970559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 971a1e78772SMel Gorman { 972af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 973af0ed73eSJoonsoo Kim /* 974af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 975af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 976af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 977af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 978af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 979af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 980af0ed73eSJoonsoo Kim * properly, so add work-around here. 981af0ed73eSJoonsoo Kim */ 982af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 983559ec2f8SNicholas Krause return true; 984af0ed73eSJoonsoo Kim else 985559ec2f8SNicholas Krause return false; 986af0ed73eSJoonsoo Kim } 987a63884e9SJoonsoo Kim 988a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 9891fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 9901fb1b0e9SMike Kravetz /* 9911fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 9921fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 9931fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 9947c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 9951fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 9961fb1b0e9SMike Kravetz */ 9971fb1b0e9SMike Kravetz if (chg) 9981fb1b0e9SMike Kravetz return false; 9991fb1b0e9SMike Kravetz else 1000559ec2f8SNicholas Krause return true; 10011fb1b0e9SMike Kravetz } 1002a63884e9SJoonsoo Kim 1003a63884e9SJoonsoo Kim /* 1004a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1005a63884e9SJoonsoo Kim * private mappings. 1006a63884e9SJoonsoo Kim */ 100767961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 100867961f9dSMike Kravetz /* 100967961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 101067961f9dSMike Kravetz * could have been performed on the private mapping. 101167961f9dSMike Kravetz * Examine the value of chg to determine if reserves 101267961f9dSMike Kravetz * actually exist or were previously consumed. 101367961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 101467961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 101567961f9dSMike Kravetz * private mappings has different (opposite) semantics 101667961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 101767961f9dSMike Kravetz * has already taken this difference in semantics into 101867961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 101967961f9dSMike Kravetz * as in the shared case above. Code could easily be 102067961f9dSMike Kravetz * combined, but keeping it separate draws attention to 102167961f9dSMike Kravetz * subtle differences. 102267961f9dSMike Kravetz */ 102367961f9dSMike Kravetz if (chg) 102467961f9dSMike Kravetz return false; 102567961f9dSMike Kravetz else 1026559ec2f8SNicholas Krause return true; 102767961f9dSMike Kravetz } 1028a63884e9SJoonsoo Kim 1029559ec2f8SNicholas Krause return false; 1030a1e78772SMel Gorman } 1031a1e78772SMel Gorman 1032a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10331da177e4SLinus Torvalds { 10341da177e4SLinus Torvalds int nid = page_to_nid(page); 10350edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1036a5516438SAndi Kleen h->free_huge_pages++; 1037a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10381da177e4SLinus Torvalds } 10391da177e4SLinus Torvalds 104094310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1041bf50bab2SNaoya Horiguchi { 1042bf50bab2SNaoya Horiguchi struct page *page; 1043bf50bab2SNaoya Horiguchi 1044c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 1045243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 1046c8721bbbSNaoya Horiguchi break; 1047c8721bbbSNaoya Horiguchi /* 1048c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 1049c8721bbbSNaoya Horiguchi * the allocation fails. 1050c8721bbbSNaoya Horiguchi */ 1051c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 1052bf50bab2SNaoya Horiguchi return NULL; 10530edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1054a9869b83SNaoya Horiguchi set_page_refcounted(page); 1055bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1056bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1057bf50bab2SNaoya Horiguchi return page; 1058bf50bab2SNaoya Horiguchi } 1059bf50bab2SNaoya Horiguchi 10603e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10613e59fcb0SMichal Hocko nodemask_t *nmask) 106294310cbcSAnshuman Khandual { 10633e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 10643e59fcb0SMichal Hocko struct zonelist *zonelist; 10653e59fcb0SMichal Hocko struct zone *zone; 10663e59fcb0SMichal Hocko struct zoneref *z; 106798fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 10683e59fcb0SMichal Hocko 10693e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 10703e59fcb0SMichal Hocko 10713e59fcb0SMichal Hocko retry_cpuset: 10723e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 10733e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 107494310cbcSAnshuman Khandual struct page *page; 107594310cbcSAnshuman Khandual 10763e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 10773e59fcb0SMichal Hocko continue; 10783e59fcb0SMichal Hocko /* 10793e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 10803e59fcb0SMichal Hocko * zone aware 10813e59fcb0SMichal Hocko */ 10823e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 10833e59fcb0SMichal Hocko continue; 10843e59fcb0SMichal Hocko node = zone_to_nid(zone); 108594310cbcSAnshuman Khandual 108694310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 108794310cbcSAnshuman Khandual if (page) 108894310cbcSAnshuman Khandual return page; 108994310cbcSAnshuman Khandual } 10903e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 10913e59fcb0SMichal Hocko goto retry_cpuset; 10923e59fcb0SMichal Hocko 109394310cbcSAnshuman Khandual return NULL; 109494310cbcSAnshuman Khandual } 109594310cbcSAnshuman Khandual 109686cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 109786cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 109886cdb465SNaoya Horiguchi { 10997ed2c31dSAnshuman Khandual if (hugepage_movable_supported(h)) 110086cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 110186cdb465SNaoya Horiguchi else 110286cdb465SNaoya Horiguchi return GFP_HIGHUSER; 110386cdb465SNaoya Horiguchi } 110486cdb465SNaoya Horiguchi 1105a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1106a5516438SAndi Kleen struct vm_area_struct *vma, 1107af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1108af0ed73eSJoonsoo Kim long chg) 11091da177e4SLinus Torvalds { 11103e59fcb0SMichal Hocko struct page *page; 1111480eccf9SLee Schermerhorn struct mempolicy *mpol; 111204ec6264SVlastimil Babka gfp_t gfp_mask; 11133e59fcb0SMichal Hocko nodemask_t *nodemask; 111404ec6264SVlastimil Babka int nid; 11151da177e4SLinus Torvalds 1116a1e78772SMel Gorman /* 1117a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1118a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1119a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1120a1e78772SMel Gorman */ 1121af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1122a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1123c0ff7453SMiao Xie goto err; 1124a1e78772SMel Gorman 112504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1126a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11276eab04a8SJustin P. Mattock goto err; 112804f2cbe3SMel Gorman 112904ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 113004ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11313e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11323e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 113307443a85SJoonsoo Kim SetPagePrivate(page); 1134a63884e9SJoonsoo Kim h->resv_huge_pages--; 1135bf50bab2SNaoya Horiguchi } 1136cc9a6c87SMel Gorman 1137cc9a6c87SMel Gorman mpol_cond_put(mpol); 1138cc9a6c87SMel Gorman return page; 1139cc9a6c87SMel Gorman 1140c0ff7453SMiao Xie err: 1141cc9a6c87SMel Gorman return NULL; 11421da177e4SLinus Torvalds } 11431da177e4SLinus Torvalds 11441cac6f2cSLuiz Capitulino /* 11451cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11461cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11471cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11481cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11491cac6f2cSLuiz Capitulino * node for alloc or free. 11501cac6f2cSLuiz Capitulino */ 11511cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11521cac6f2cSLuiz Capitulino { 11530edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11541cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11551cac6f2cSLuiz Capitulino 11561cac6f2cSLuiz Capitulino return nid; 11571cac6f2cSLuiz Capitulino } 11581cac6f2cSLuiz Capitulino 11591cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11601cac6f2cSLuiz Capitulino { 11611cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11621cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11631cac6f2cSLuiz Capitulino return nid; 11641cac6f2cSLuiz Capitulino } 11651cac6f2cSLuiz Capitulino 11661cac6f2cSLuiz Capitulino /* 11671cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11681cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11691cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11701cac6f2cSLuiz Capitulino * mask. 11711cac6f2cSLuiz Capitulino */ 11721cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 11731cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 11741cac6f2cSLuiz Capitulino { 11751cac6f2cSLuiz Capitulino int nid; 11761cac6f2cSLuiz Capitulino 11771cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11781cac6f2cSLuiz Capitulino 11791cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 11801cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 11811cac6f2cSLuiz Capitulino 11821cac6f2cSLuiz Capitulino return nid; 11831cac6f2cSLuiz Capitulino } 11841cac6f2cSLuiz Capitulino 11851cac6f2cSLuiz Capitulino /* 11861cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 11871cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 11881cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 11891cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 11901cac6f2cSLuiz Capitulino */ 11911cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 11921cac6f2cSLuiz Capitulino { 11931cac6f2cSLuiz Capitulino int nid; 11941cac6f2cSLuiz Capitulino 11951cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11961cac6f2cSLuiz Capitulino 11971cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 11981cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 11991cac6f2cSLuiz Capitulino 12001cac6f2cSLuiz Capitulino return nid; 12011cac6f2cSLuiz Capitulino } 12021cac6f2cSLuiz Capitulino 12031cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12041cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12051cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12061cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12071cac6f2cSLuiz Capitulino nr_nodes--) 12081cac6f2cSLuiz Capitulino 12091cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12101cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12111cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12121cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12131cac6f2cSLuiz Capitulino nr_nodes--) 12141cac6f2cSLuiz Capitulino 1215e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1216944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1217d00181b9SKirill A. Shutemov unsigned int order) 1218944d9fecSLuiz Capitulino { 1219944d9fecSLuiz Capitulino int i; 1220944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1221944d9fecSLuiz Capitulino struct page *p = page + 1; 1222944d9fecSLuiz Capitulino 1223c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 122447e29d32SJohn Hubbard if (hpage_pincount_available(page)) 122547e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 122647e29d32SJohn Hubbard 1227944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12281d798ca3SKirill A. Shutemov clear_compound_head(p); 1229944d9fecSLuiz Capitulino set_page_refcounted(p); 1230944d9fecSLuiz Capitulino } 1231944d9fecSLuiz Capitulino 1232944d9fecSLuiz Capitulino set_compound_order(page, 0); 1233944d9fecSLuiz Capitulino __ClearPageHead(page); 1234944d9fecSLuiz Capitulino } 1235944d9fecSLuiz Capitulino 1236d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1237944d9fecSLuiz Capitulino { 1238cf11e85fSRoman Gushchin /* 1239cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1240cf11e85fSRoman Gushchin * cma_release() returns false. 1241cf11e85fSRoman Gushchin */ 1242dbda8feaSBarry Song #ifdef CONFIG_CMA 1243dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1244cf11e85fSRoman Gushchin return; 1245dbda8feaSBarry Song #endif 1246cf11e85fSRoman Gushchin 1247944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1248944d9fecSLuiz Capitulino } 1249944d9fecSLuiz Capitulino 12504eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1251d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1252d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1253944d9fecSLuiz Capitulino { 12545e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1255944d9fecSLuiz Capitulino 1256dbda8feaSBarry Song #ifdef CONFIG_CMA 1257dbda8feaSBarry Song { 1258cf11e85fSRoman Gushchin struct page *page; 1259cf11e85fSRoman Gushchin int node; 1260cf11e85fSRoman Gushchin 1261cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1262cf11e85fSRoman Gushchin if (!hugetlb_cma[node]) 1263cf11e85fSRoman Gushchin continue; 1264cf11e85fSRoman Gushchin 1265cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1266cf11e85fSRoman Gushchin huge_page_order(h), true); 1267cf11e85fSRoman Gushchin if (page) 1268cf11e85fSRoman Gushchin return page; 1269cf11e85fSRoman Gushchin } 1270cf11e85fSRoman Gushchin } 1271dbda8feaSBarry Song #endif 1272cf11e85fSRoman Gushchin 12735e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1274944d9fecSLuiz Capitulino } 1275944d9fecSLuiz Capitulino 1276944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1277d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 12784eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 12794eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 12804eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 12814eb0716eSAlexandre Ghiti { 12824eb0716eSAlexandre Ghiti return NULL; 12834eb0716eSAlexandre Ghiti } 12844eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1285944d9fecSLuiz Capitulino 1286e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1287d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 12884eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 12894eb0716eSAlexandre Ghiti { 12904eb0716eSAlexandre Ghiti return NULL; 12914eb0716eSAlexandre Ghiti } 1292d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1293944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1294d00181b9SKirill A. Shutemov unsigned int order) { } 1295944d9fecSLuiz Capitulino #endif 1296944d9fecSLuiz Capitulino 1297a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 12986af2acb6SAdam Litke { 12996af2acb6SAdam Litke int i; 1300a5516438SAndi Kleen 13014eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1302944d9fecSLuiz Capitulino return; 130318229df5SAndy Whitcroft 1304a5516438SAndi Kleen h->nr_huge_pages--; 1305a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1306a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 130732f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 130832f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1309a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1310a7407a27SLuiz Capitulino 1 << PG_writeback); 13116af2acb6SAdam Litke } 1312309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13131adc4d41SMina Almasry VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1314f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13156af2acb6SAdam Litke set_page_refcounted(page); 1316944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1317cf11e85fSRoman Gushchin /* 1318cf11e85fSRoman Gushchin * Temporarily drop the hugetlb_lock, because 1319cf11e85fSRoman Gushchin * we might block in free_gigantic_page(). 1320cf11e85fSRoman Gushchin */ 1321cf11e85fSRoman Gushchin spin_unlock(&hugetlb_lock); 1322944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1323944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1324cf11e85fSRoman Gushchin spin_lock(&hugetlb_lock); 1325944d9fecSLuiz Capitulino } else { 1326a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13276af2acb6SAdam Litke } 1328944d9fecSLuiz Capitulino } 13296af2acb6SAdam Litke 1330e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1331e5ff2159SAndi Kleen { 1332e5ff2159SAndi Kleen struct hstate *h; 1333e5ff2159SAndi Kleen 1334e5ff2159SAndi Kleen for_each_hstate(h) { 1335e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1336e5ff2159SAndi Kleen return h; 1337e5ff2159SAndi Kleen } 1338e5ff2159SAndi Kleen return NULL; 1339e5ff2159SAndi Kleen } 1340e5ff2159SAndi Kleen 1341bcc54222SNaoya Horiguchi /* 1342bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1343bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1344bcc54222SNaoya Horiguchi * 1345bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1346bcc54222SNaoya Horiguchi */ 1347bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1348bcc54222SNaoya Horiguchi { 1349bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1350bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1351bcc54222SNaoya Horiguchi } 1352bcc54222SNaoya Horiguchi 1353bcc54222SNaoya Horiguchi /* never called for tail page */ 1354bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1355bcc54222SNaoya Horiguchi { 1356bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1357bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1358bcc54222SNaoya Horiguchi } 1359bcc54222SNaoya Horiguchi 1360bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1361bcc54222SNaoya Horiguchi { 1362bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1363bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1364bcc54222SNaoya Horiguchi } 1365bcc54222SNaoya Horiguchi 1366ab5ac90aSMichal Hocko /* 1367ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1368ab5ac90aSMichal Hocko * code 1369ab5ac90aSMichal Hocko */ 1370ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1371ab5ac90aSMichal Hocko { 1372ab5ac90aSMichal Hocko if (!PageHuge(page)) 1373ab5ac90aSMichal Hocko return false; 1374ab5ac90aSMichal Hocko 1375ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1376ab5ac90aSMichal Hocko } 1377ab5ac90aSMichal Hocko 1378ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1379ab5ac90aSMichal Hocko { 1380ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1381ab5ac90aSMichal Hocko } 1382ab5ac90aSMichal Hocko 1383ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1384ab5ac90aSMichal Hocko { 1385ab5ac90aSMichal Hocko page[2].mapping = NULL; 1386ab5ac90aSMichal Hocko } 1387ab5ac90aSMichal Hocko 1388c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 138927a85ef1SDavid Gibson { 1390a5516438SAndi Kleen /* 1391a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1392a5516438SAndi Kleen * compound page destructor. 1393a5516438SAndi Kleen */ 1394e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 13957893d1d5SAdam Litke int nid = page_to_nid(page); 139690481622SDavid Gibson struct hugepage_subpool *spool = 139790481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 139807443a85SJoonsoo Kim bool restore_reserve; 139927a85ef1SDavid Gibson 1400b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1401b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14028ace22bcSYongkai Wu 14038ace22bcSYongkai Wu set_page_private(page, 0); 14048ace22bcSYongkai Wu page->mapping = NULL; 140507443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 140616c794b4SJoonsoo Kim ClearPagePrivate(page); 140727a85ef1SDavid Gibson 14081c5ecae3SMike Kravetz /* 14090919e1b6SMike Kravetz * If PagePrivate() was set on page, page allocation consumed a 14100919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14110919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14120919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14130919e1b6SMike Kravetz * reservtion, do not call hugepage_subpool_put_pages() as this will 14140919e1b6SMike Kravetz * remove the reserved page from the subpool. 14150919e1b6SMike Kravetz */ 14160919e1b6SMike Kravetz if (!restore_reserve) { 14170919e1b6SMike Kravetz /* 14180919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14190919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14200919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14210919e1b6SMike Kravetz * operation. 14221c5ecae3SMike Kravetz */ 14231c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14241c5ecae3SMike Kravetz restore_reserve = true; 14250919e1b6SMike Kravetz } 14261c5ecae3SMike Kravetz 142727a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1428bcc54222SNaoya Horiguchi clear_page_huge_active(page); 14296d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14306d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 143108cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 143208cf9fafSMina Almasry pages_per_huge_page(h), page); 143307443a85SJoonsoo Kim if (restore_reserve) 143407443a85SJoonsoo Kim h->resv_huge_pages++; 143507443a85SJoonsoo Kim 1436ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1437ab5ac90aSMichal Hocko list_del(&page->lru); 1438ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1439ab5ac90aSMichal Hocko update_and_free_page(h, page); 1440ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14410edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14420edaecfaSAneesh Kumar K.V list_del(&page->lru); 1443a5516438SAndi Kleen update_and_free_page(h, page); 1444a5516438SAndi Kleen h->surplus_huge_pages--; 1445a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 14467893d1d5SAdam Litke } else { 14475d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1448a5516438SAndi Kleen enqueue_huge_page(h, page); 14497893d1d5SAdam Litke } 145027a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 145127a85ef1SDavid Gibson } 145227a85ef1SDavid Gibson 1453c77c0a8aSWaiman Long /* 1454c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1455c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1456c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1457c77c0a8aSWaiman Long * 1458c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1459c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1460c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1461c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1462c77c0a8aSWaiman Long */ 1463c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1464c77c0a8aSWaiman Long 1465c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1466c77c0a8aSWaiman Long { 1467c77c0a8aSWaiman Long struct llist_node *node; 1468c77c0a8aSWaiman Long struct page *page; 1469c77c0a8aSWaiman Long 1470c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1471c77c0a8aSWaiman Long 1472c77c0a8aSWaiman Long while (node) { 1473c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1474c77c0a8aSWaiman Long struct page, mapping); 1475c77c0a8aSWaiman Long node = node->next; 1476c77c0a8aSWaiman Long __free_huge_page(page); 1477c77c0a8aSWaiman Long } 1478c77c0a8aSWaiman Long } 1479c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1480c77c0a8aSWaiman Long 1481c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1482c77c0a8aSWaiman Long { 1483c77c0a8aSWaiman Long /* 1484c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1485c77c0a8aSWaiman Long */ 1486c77c0a8aSWaiman Long if (!in_task()) { 1487c77c0a8aSWaiman Long /* 1488c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1489c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1490c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1491c77c0a8aSWaiman Long */ 1492c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1493c77c0a8aSWaiman Long &hpage_freelist)) 1494c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1495c77c0a8aSWaiman Long return; 1496c77c0a8aSWaiman Long } 1497c77c0a8aSWaiman Long 1498c77c0a8aSWaiman Long __free_huge_page(page); 1499c77c0a8aSWaiman Long } 1500c77c0a8aSWaiman Long 1501a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1502b7ba30c6SAndi Kleen { 15030edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1504f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1505b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 15069dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15071adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1508a5516438SAndi Kleen h->nr_huge_pages++; 1509a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1510b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1511b7ba30c6SAndi Kleen } 1512b7ba30c6SAndi Kleen 1513d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 151420a0307cSWu Fengguang { 151520a0307cSWu Fengguang int i; 151620a0307cSWu Fengguang int nr_pages = 1 << order; 151720a0307cSWu Fengguang struct page *p = page + 1; 151820a0307cSWu Fengguang 151920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 152020a0307cSWu Fengguang set_compound_order(page, order); 1521ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1522de09d31dSKirill A. Shutemov __SetPageHead(page); 152320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1524ef5a22beSAndrea Arcangeli /* 1525ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1526ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1527ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15287c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1529ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1530ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1531ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1532ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1533ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1534ef5a22beSAndrea Arcangeli * after get_user_pages(). 1535ef5a22beSAndrea Arcangeli */ 1536ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 153758a84aa9SYouquan Song set_page_count(p, 0); 15381d798ca3SKirill A. Shutemov set_compound_head(p, page); 153920a0307cSWu Fengguang } 1540b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 154147e29d32SJohn Hubbard 154247e29d32SJohn Hubbard if (hpage_pincount_available(page)) 154347e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 154420a0307cSWu Fengguang } 154520a0307cSWu Fengguang 15467795912cSAndrew Morton /* 15477795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15487795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15497795912cSAndrew Morton * details. 15507795912cSAndrew Morton */ 155120a0307cSWu Fengguang int PageHuge(struct page *page) 155220a0307cSWu Fengguang { 155320a0307cSWu Fengguang if (!PageCompound(page)) 155420a0307cSWu Fengguang return 0; 155520a0307cSWu Fengguang 155620a0307cSWu Fengguang page = compound_head(page); 1557f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 155820a0307cSWu Fengguang } 155943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 156043131e14SNaoya Horiguchi 156127c73ae7SAndrea Arcangeli /* 156227c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 156327c73ae7SAndrea Arcangeli * normal or transparent huge pages. 156427c73ae7SAndrea Arcangeli */ 156527c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 156627c73ae7SAndrea Arcangeli { 156727c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 156827c73ae7SAndrea Arcangeli return 0; 156927c73ae7SAndrea Arcangeli 1570d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 157127c73ae7SAndrea Arcangeli } 157227c73ae7SAndrea Arcangeli 1573c0d0381aSMike Kravetz /* 1574c0d0381aSMike Kravetz * Find address_space associated with hugetlbfs page. 1575c0d0381aSMike Kravetz * Upon entry page is locked and page 'was' mapped although mapped state 1576c0d0381aSMike Kravetz * could change. If necessary, use anon_vma to find vma and associated 1577c0d0381aSMike Kravetz * address space. The returned mapping may be stale, but it can not be 1578c0d0381aSMike Kravetz * invalid as page lock (which is held) is required to destroy mapping. 1579c0d0381aSMike Kravetz */ 1580c0d0381aSMike Kravetz static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) 1581c0d0381aSMike Kravetz { 1582c0d0381aSMike Kravetz struct anon_vma *anon_vma; 1583c0d0381aSMike Kravetz pgoff_t pgoff_start, pgoff_end; 1584c0d0381aSMike Kravetz struct anon_vma_chain *avc; 1585c0d0381aSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1586c0d0381aSMike Kravetz 1587c0d0381aSMike Kravetz /* Simple file based mapping */ 1588c0d0381aSMike Kravetz if (mapping) 1589c0d0381aSMike Kravetz return mapping; 1590c0d0381aSMike Kravetz 1591c0d0381aSMike Kravetz /* 1592c0d0381aSMike Kravetz * Even anonymous hugetlbfs mappings are associated with an 1593c0d0381aSMike Kravetz * underlying hugetlbfs file (see hugetlb_file_setup in mmap 1594c0d0381aSMike Kravetz * code). Find a vma associated with the anonymous vma, and 1595c0d0381aSMike Kravetz * use the file pointer to get address_space. 1596c0d0381aSMike Kravetz */ 1597c0d0381aSMike Kravetz anon_vma = page_lock_anon_vma_read(hpage); 1598c0d0381aSMike Kravetz if (!anon_vma) 1599c0d0381aSMike Kravetz return mapping; /* NULL */ 1600c0d0381aSMike Kravetz 1601c0d0381aSMike Kravetz /* Use first found vma */ 1602c0d0381aSMike Kravetz pgoff_start = page_to_pgoff(hpage); 16031139d336SMike Kravetz pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1; 1604c0d0381aSMike Kravetz anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 1605c0d0381aSMike Kravetz pgoff_start, pgoff_end) { 1606c0d0381aSMike Kravetz struct vm_area_struct *vma = avc->vma; 1607c0d0381aSMike Kravetz 1608c0d0381aSMike Kravetz mapping = vma->vm_file->f_mapping; 1609c0d0381aSMike Kravetz break; 1610c0d0381aSMike Kravetz } 1611c0d0381aSMike Kravetz 1612c0d0381aSMike Kravetz anon_vma_unlock_read(anon_vma); 1613c0d0381aSMike Kravetz return mapping; 1614c0d0381aSMike Kravetz } 1615c0d0381aSMike Kravetz 1616c0d0381aSMike Kravetz /* 1617c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1618c0d0381aSMike Kravetz * 1619c0d0381aSMike Kravetz * Upon entry, the page is locked which allows us to find the mapping 1620c0d0381aSMike Kravetz * even in the case of an anon page. However, locking order dictates 1621c0d0381aSMike Kravetz * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs 1622c0d0381aSMike Kravetz * specific. So, we first try to lock the sema while still holding the 1623c0d0381aSMike Kravetz * page lock. If this works, great! If not, then we need to drop the 1624c0d0381aSMike Kravetz * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of 1625c0d0381aSMike Kravetz * course, need to revalidate state along the way. 1626c0d0381aSMike Kravetz */ 1627c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1628c0d0381aSMike Kravetz { 1629c0d0381aSMike Kravetz struct address_space *mapping, *mapping2; 1630c0d0381aSMike Kravetz 1631c0d0381aSMike Kravetz mapping = _get_hugetlb_page_mapping(hpage); 1632c0d0381aSMike Kravetz retry: 1633c0d0381aSMike Kravetz if (!mapping) 1634c0d0381aSMike Kravetz return mapping; 1635c0d0381aSMike Kravetz 1636c0d0381aSMike Kravetz /* 1637c0d0381aSMike Kravetz * If no contention, take lock and return 1638c0d0381aSMike Kravetz */ 1639c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1640c0d0381aSMike Kravetz return mapping; 1641c0d0381aSMike Kravetz 1642c0d0381aSMike Kravetz /* 1643c0d0381aSMike Kravetz * Must drop page lock and wait on mapping sema. 1644c0d0381aSMike Kravetz * Note: Once page lock is dropped, mapping could become invalid. 1645c0d0381aSMike Kravetz * As a hack, increase map count until we lock page again. 1646c0d0381aSMike Kravetz */ 1647c0d0381aSMike Kravetz atomic_inc(&hpage->_mapcount); 1648c0d0381aSMike Kravetz unlock_page(hpage); 1649c0d0381aSMike Kravetz i_mmap_lock_write(mapping); 1650c0d0381aSMike Kravetz lock_page(hpage); 1651c0d0381aSMike Kravetz atomic_add_negative(-1, &hpage->_mapcount); 1652c0d0381aSMike Kravetz 1653c0d0381aSMike Kravetz /* verify page is still mapped */ 1654c0d0381aSMike Kravetz if (!page_mapped(hpage)) { 1655c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1656c0d0381aSMike Kravetz return NULL; 1657c0d0381aSMike Kravetz } 1658c0d0381aSMike Kravetz 1659c0d0381aSMike Kravetz /* 1660c0d0381aSMike Kravetz * Get address space again and verify it is the same one 1661c0d0381aSMike Kravetz * we locked. If not, drop lock and retry. 1662c0d0381aSMike Kravetz */ 1663c0d0381aSMike Kravetz mapping2 = _get_hugetlb_page_mapping(hpage); 1664c0d0381aSMike Kravetz if (mapping2 != mapping) { 1665c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1666c0d0381aSMike Kravetz mapping = mapping2; 1667c0d0381aSMike Kravetz goto retry; 1668c0d0381aSMike Kravetz } 1669c0d0381aSMike Kravetz 1670c0d0381aSMike Kravetz return mapping; 1671c0d0381aSMike Kravetz } 1672c0d0381aSMike Kravetz 167313d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 167413d60f4bSZhang Yi { 167513d60f4bSZhang Yi struct page *page_head = compound_head(page); 167613d60f4bSZhang Yi pgoff_t index = page_index(page_head); 167713d60f4bSZhang Yi unsigned long compound_idx; 167813d60f4bSZhang Yi 167913d60f4bSZhang Yi if (!PageHuge(page_head)) 168013d60f4bSZhang Yi return page_index(page); 168113d60f4bSZhang Yi 168213d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 168313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 168413d60f4bSZhang Yi else 168513d60f4bSZhang Yi compound_idx = page - page_head; 168613d60f4bSZhang Yi 168713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 168813d60f4bSZhang Yi } 168913d60f4bSZhang Yi 16900c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1691f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1692f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16931da177e4SLinus Torvalds { 1694af0fb9dfSMichal Hocko int order = huge_page_order(h); 16951da177e4SLinus Torvalds struct page *page; 1696f60858f9SMike Kravetz bool alloc_try_hard = true; 1697f96efd58SJoe Jin 1698f60858f9SMike Kravetz /* 1699f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1700f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1701f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1702f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1703f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1704f60858f9SMike Kravetz */ 1705f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1706f60858f9SMike Kravetz alloc_try_hard = false; 1707f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1708f60858f9SMike Kravetz if (alloc_try_hard) 1709f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1710af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1711af0fb9dfSMichal Hocko nid = numa_mem_id(); 1712af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1713af0fb9dfSMichal Hocko if (page) 1714af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1715af0fb9dfSMichal Hocko else 1716af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 171763b4613cSNishanth Aravamudan 1718f60858f9SMike Kravetz /* 1719f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1720f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1721f60858f9SMike Kravetz * normal 'try hard' allocations. 1722f60858f9SMike Kravetz */ 1723f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1724f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1725f60858f9SMike Kravetz 1726f60858f9SMike Kravetz /* 1727f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1728f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1729f60858f9SMike Kravetz * overall state change. 1730f60858f9SMike Kravetz */ 1731f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1732f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1733f60858f9SMike Kravetz 173463b4613cSNishanth Aravamudan return page; 173563b4613cSNishanth Aravamudan } 173663b4613cSNishanth Aravamudan 1737af0fb9dfSMichal Hocko /* 17380c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 17390c397daeSMichal Hocko * should use this function to get new hugetlb pages 17400c397daeSMichal Hocko */ 17410c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1742f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1743f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17440c397daeSMichal Hocko { 17450c397daeSMichal Hocko struct page *page; 17460c397daeSMichal Hocko 17470c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17480c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 17490c397daeSMichal Hocko else 17500c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1751f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 17520c397daeSMichal Hocko if (!page) 17530c397daeSMichal Hocko return NULL; 17540c397daeSMichal Hocko 17550c397daeSMichal Hocko if (hstate_is_gigantic(h)) 17560c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 17570c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 17580c397daeSMichal Hocko 17590c397daeSMichal Hocko return page; 17600c397daeSMichal Hocko } 17610c397daeSMichal Hocko 17620c397daeSMichal Hocko /* 1763af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1764af0fb9dfSMichal Hocko * manner. 1765af0fb9dfSMichal Hocko */ 1766f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1767f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1768b2261026SJoonsoo Kim { 1769b2261026SJoonsoo Kim struct page *page; 1770b2261026SJoonsoo Kim int nr_nodes, node; 1771af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1772b2261026SJoonsoo Kim 1773b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1774f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1775f60858f9SMike Kravetz node_alloc_noretry); 1776af0fb9dfSMichal Hocko if (page) 1777b2261026SJoonsoo Kim break; 1778b2261026SJoonsoo Kim } 1779b2261026SJoonsoo Kim 1780af0fb9dfSMichal Hocko if (!page) 1781af0fb9dfSMichal Hocko return 0; 1782b2261026SJoonsoo Kim 1783af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1784af0fb9dfSMichal Hocko 1785af0fb9dfSMichal Hocko return 1; 1786b2261026SJoonsoo Kim } 1787b2261026SJoonsoo Kim 1788e8c5c824SLee Schermerhorn /* 1789e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1790e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1791e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1792e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1793e8c5c824SLee Schermerhorn */ 17946ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17956ae11b27SLee Schermerhorn bool acct_surplus) 1796e8c5c824SLee Schermerhorn { 1797b2261026SJoonsoo Kim int nr_nodes, node; 1798e8c5c824SLee Schermerhorn int ret = 0; 1799e8c5c824SLee Schermerhorn 1800b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1801685f3457SLee Schermerhorn /* 1802685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1803685f3457SLee Schermerhorn * nodes with surplus pages. 1804685f3457SLee Schermerhorn */ 1805b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1806b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1807e8c5c824SLee Schermerhorn struct page *page = 1808b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1809e8c5c824SLee Schermerhorn struct page, lru); 1810e8c5c824SLee Schermerhorn list_del(&page->lru); 1811e8c5c824SLee Schermerhorn h->free_huge_pages--; 1812b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1813685f3457SLee Schermerhorn if (acct_surplus) { 1814685f3457SLee Schermerhorn h->surplus_huge_pages--; 1815b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1816685f3457SLee Schermerhorn } 1817e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1818e8c5c824SLee Schermerhorn ret = 1; 18199a76db09SLee Schermerhorn break; 1820e8c5c824SLee Schermerhorn } 1821b2261026SJoonsoo Kim } 1822e8c5c824SLee Schermerhorn 1823e8c5c824SLee Schermerhorn return ret; 1824e8c5c824SLee Schermerhorn } 1825e8c5c824SLee Schermerhorn 1826c8721bbbSNaoya Horiguchi /* 1827c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1828faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1829faf53defSNaoya Horiguchi * This function returns values like below: 1830faf53defSNaoya Horiguchi * 1831faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1832faf53defSNaoya Horiguchi * (allocated or reserved.) 1833faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1834faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1835c8721bbbSNaoya Horiguchi */ 1836c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1837c8721bbbSNaoya Horiguchi { 18386bc9b564SNaoya Horiguchi int rc = -EBUSY; 1839082d5b6bSGerald Schaefer 1840faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1841faf53defSNaoya Horiguchi if (!PageHuge(page)) 1842faf53defSNaoya Horiguchi return 0; 1843faf53defSNaoya Horiguchi 1844c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1845faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1846faf53defSNaoya Horiguchi rc = 0; 1847faf53defSNaoya Horiguchi goto out; 1848faf53defSNaoya Horiguchi } 1849faf53defSNaoya Horiguchi 1850faf53defSNaoya Horiguchi if (!page_count(page)) { 18512247bb33SGerald Schaefer struct page *head = compound_head(page); 18522247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 18532247bb33SGerald Schaefer int nid = page_to_nid(head); 18546bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1855082d5b6bSGerald Schaefer goto out; 1856c3114a84SAnshuman Khandual /* 1857c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1858c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1859c3114a84SAnshuman Khandual */ 1860c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1861c3114a84SAnshuman Khandual SetPageHWPoison(page); 1862c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1863c3114a84SAnshuman Khandual } 18642247bb33SGerald Schaefer list_del(&head->lru); 1865c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1866c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1867c1470b33Szhong jiang h->max_huge_pages--; 18682247bb33SGerald Schaefer update_and_free_page(h, head); 18696bc9b564SNaoya Horiguchi rc = 0; 1870c8721bbbSNaoya Horiguchi } 1871082d5b6bSGerald Schaefer out: 1872c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1873082d5b6bSGerald Schaefer return rc; 1874c8721bbbSNaoya Horiguchi } 1875c8721bbbSNaoya Horiguchi 1876c8721bbbSNaoya Horiguchi /* 1877c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1878c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18792247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18802247bb33SGerald Schaefer * part of it lies within the given range. 1881082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1882082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1883c8721bbbSNaoya Horiguchi */ 1884082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1885c8721bbbSNaoya Horiguchi { 1886c8721bbbSNaoya Horiguchi unsigned long pfn; 1887eb03aa00SGerald Schaefer struct page *page; 1888082d5b6bSGerald Schaefer int rc = 0; 1889c8721bbbSNaoya Horiguchi 1890d0177639SLi Zhong if (!hugepages_supported()) 1891082d5b6bSGerald Schaefer return rc; 1892d0177639SLi Zhong 1893eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1894eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1895eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1896eb03aa00SGerald Schaefer if (rc) 1897082d5b6bSGerald Schaefer break; 1898eb03aa00SGerald Schaefer } 1899082d5b6bSGerald Schaefer 1900082d5b6bSGerald Schaefer return rc; 1901c8721bbbSNaoya Horiguchi } 1902c8721bbbSNaoya Horiguchi 1903ab5ac90aSMichal Hocko /* 1904ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1905ab5ac90aSMichal Hocko */ 19060c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1907aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 19087893d1d5SAdam Litke { 19099980d744SMichal Hocko struct page *page = NULL; 19107893d1d5SAdam Litke 1911bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1912aa888a74SAndi Kleen return NULL; 1913aa888a74SAndi Kleen 1914d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 19159980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 19169980d744SMichal Hocko goto out_unlock; 1917d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1918d1c3fb1fSNishanth Aravamudan 1919f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 19209980d744SMichal Hocko if (!page) 19210c397daeSMichal Hocko return NULL; 1922d1c3fb1fSNishanth Aravamudan 19237893d1d5SAdam Litke spin_lock(&hugetlb_lock); 19249980d744SMichal Hocko /* 19259980d744SMichal Hocko * We could have raced with the pool size change. 19269980d744SMichal Hocko * Double check that and simply deallocate the new page 19279980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 19289980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 19299980d744SMichal Hocko * codeflow 19309980d744SMichal Hocko */ 19319980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 19329980d744SMichal Hocko SetPageHugeTemporary(page); 19332bf753e6SKai Shen spin_unlock(&hugetlb_lock); 19349980d744SMichal Hocko put_page(page); 19352bf753e6SKai Shen return NULL; 19369980d744SMichal Hocko } else { 19379980d744SMichal Hocko h->surplus_huge_pages++; 19384704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 19397893d1d5SAdam Litke } 19409980d744SMichal Hocko 19419980d744SMichal Hocko out_unlock: 1942d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 19437893d1d5SAdam Litke 19447893d1d5SAdam Litke return page; 19457893d1d5SAdam Litke } 19467893d1d5SAdam Litke 19479a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1948ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1949ab5ac90aSMichal Hocko { 1950ab5ac90aSMichal Hocko struct page *page; 1951ab5ac90aSMichal Hocko 1952ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1953ab5ac90aSMichal Hocko return NULL; 1954ab5ac90aSMichal Hocko 1955f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1956ab5ac90aSMichal Hocko if (!page) 1957ab5ac90aSMichal Hocko return NULL; 1958ab5ac90aSMichal Hocko 1959ab5ac90aSMichal Hocko /* 1960ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1961ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1962ab5ac90aSMichal Hocko */ 1963ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1964ab5ac90aSMichal Hocko 1965ab5ac90aSMichal Hocko return page; 1966ab5ac90aSMichal Hocko } 1967ab5ac90aSMichal Hocko 1968e4e574b7SAdam Litke /* 1969099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1970099730d6SDave Hansen */ 1971e0ec90eeSDave Hansen static 19720c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1973099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1974099730d6SDave Hansen { 1975aaf14e40SMichal Hocko struct page *page; 1976aaf14e40SMichal Hocko struct mempolicy *mpol; 1977aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1978aaf14e40SMichal Hocko int nid; 1979aaf14e40SMichal Hocko nodemask_t *nodemask; 1980aaf14e40SMichal Hocko 1981aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19820c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1983aaf14e40SMichal Hocko mpol_cond_put(mpol); 1984aaf14e40SMichal Hocko 1985aaf14e40SMichal Hocko return page; 1986099730d6SDave Hansen } 1987099730d6SDave Hansen 1988ab5ac90aSMichal Hocko /* page migration callback function */ 1989bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1990bf50bab2SNaoya Horiguchi { 1991aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 19924ef91848SJoonsoo Kim struct page *page = NULL; 1993bf50bab2SNaoya Horiguchi 1994aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1995aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1996aaf14e40SMichal Hocko 1997bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 19984ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 19993e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 2000bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 2001bf50bab2SNaoya Horiguchi 200294ae8ba7SAneesh Kumar K.V if (!page) 20030c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 2004bf50bab2SNaoya Horiguchi 2005bf50bab2SNaoya Horiguchi return page; 2006bf50bab2SNaoya Horiguchi } 2007bf50bab2SNaoya Horiguchi 2008ab5ac90aSMichal Hocko /* page migration callback function */ 20093e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 20103e59fcb0SMichal Hocko nodemask_t *nmask) 20114db9b2efSMichal Hocko { 2012aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 20134db9b2efSMichal Hocko 20144db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 20154db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 20163e59fcb0SMichal Hocko struct page *page; 20173e59fcb0SMichal Hocko 20183e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 20193e59fcb0SMichal Hocko if (page) { 20203e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 20213e59fcb0SMichal Hocko return page; 20224db9b2efSMichal Hocko } 20234db9b2efSMichal Hocko } 20244db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 20254db9b2efSMichal Hocko 20260c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 20274db9b2efSMichal Hocko } 20284db9b2efSMichal Hocko 2029ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2030389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2031389c8178SMichal Hocko unsigned long address) 2032ebd63723SMichal Hocko { 2033ebd63723SMichal Hocko struct mempolicy *mpol; 2034ebd63723SMichal Hocko nodemask_t *nodemask; 2035ebd63723SMichal Hocko struct page *page; 2036ebd63723SMichal Hocko gfp_t gfp_mask; 2037ebd63723SMichal Hocko int node; 2038ebd63723SMichal Hocko 2039ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2040ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2041ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 2042ebd63723SMichal Hocko mpol_cond_put(mpol); 2043ebd63723SMichal Hocko 2044ebd63723SMichal Hocko return page; 2045ebd63723SMichal Hocko } 2046ebd63723SMichal Hocko 2047bf50bab2SNaoya Horiguchi /* 204825985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2049e4e574b7SAdam Litke * of size 'delta'. 2050e4e574b7SAdam Litke */ 2051a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 20521b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2053e4e574b7SAdam Litke { 2054e4e574b7SAdam Litke struct list_head surplus_list; 2055e4e574b7SAdam Litke struct page *page, *tmp; 2056e4e574b7SAdam Litke int ret, i; 2057e4e574b7SAdam Litke int needed, allocated; 205828073b02SHillf Danton bool alloc_ok = true; 2059e4e574b7SAdam Litke 2060a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2061ac09b3a1SAdam Litke if (needed <= 0) { 2062a5516438SAndi Kleen h->resv_huge_pages += delta; 2063e4e574b7SAdam Litke return 0; 2064ac09b3a1SAdam Litke } 2065e4e574b7SAdam Litke 2066e4e574b7SAdam Litke allocated = 0; 2067e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2068e4e574b7SAdam Litke 2069e4e574b7SAdam Litke ret = -ENOMEM; 2070e4e574b7SAdam Litke retry: 2071e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 2072e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 20730c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2074aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 207528073b02SHillf Danton if (!page) { 207628073b02SHillf Danton alloc_ok = false; 207728073b02SHillf Danton break; 207828073b02SHillf Danton } 2079e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 208069ed779aSDavid Rientjes cond_resched(); 2081e4e574b7SAdam Litke } 208228073b02SHillf Danton allocated += i; 2083e4e574b7SAdam Litke 2084e4e574b7SAdam Litke /* 2085e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2086e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2087e4e574b7SAdam Litke */ 2088e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 2089a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2090a5516438SAndi Kleen (h->free_huge_pages + allocated); 209128073b02SHillf Danton if (needed > 0) { 209228073b02SHillf Danton if (alloc_ok) 2093e4e574b7SAdam Litke goto retry; 209428073b02SHillf Danton /* 209528073b02SHillf Danton * We were not able to allocate enough pages to 209628073b02SHillf Danton * satisfy the entire reservation so we free what 209728073b02SHillf Danton * we've allocated so far. 209828073b02SHillf Danton */ 209928073b02SHillf Danton goto free; 210028073b02SHillf Danton } 2101e4e574b7SAdam Litke /* 2102e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 210325985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2104e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2105ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2106ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2107ac09b3a1SAdam Litke * before they are reserved. 2108e4e574b7SAdam Litke */ 2109e4e574b7SAdam Litke needed += allocated; 2110a5516438SAndi Kleen h->resv_huge_pages += delta; 2111e4e574b7SAdam Litke ret = 0; 2112a9869b83SNaoya Horiguchi 211319fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 211419fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 211519fc3f0aSAdam Litke if ((--needed) < 0) 211619fc3f0aSAdam Litke break; 2117a9869b83SNaoya Horiguchi /* 2118a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2119a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2120a9869b83SNaoya Horiguchi */ 2121a9869b83SNaoya Horiguchi put_page_testzero(page); 2122309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 2123a5516438SAndi Kleen enqueue_huge_page(h, page); 212419fc3f0aSAdam Litke } 212528073b02SHillf Danton free: 2126b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 212719fc3f0aSAdam Litke 212819fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2129c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2130a9869b83SNaoya Horiguchi put_page(page); 213119fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2132e4e574b7SAdam Litke 2133e4e574b7SAdam Litke return ret; 2134e4e574b7SAdam Litke } 2135e4e574b7SAdam Litke 2136e4e574b7SAdam Litke /* 2137e5bbc8a6SMike Kravetz * This routine has two main purposes: 2138e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2139e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2140e5bbc8a6SMike Kravetz * to the associated reservation map. 2141e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2142e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2143e5bbc8a6SMike Kravetz * 2144e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2145e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2146e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2147e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2148e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2149e4e574b7SAdam Litke */ 2150a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2151a5516438SAndi Kleen unsigned long unused_resv_pages) 2152e4e574b7SAdam Litke { 2153e4e574b7SAdam Litke unsigned long nr_pages; 2154e4e574b7SAdam Litke 2155aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2156bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2157e5bbc8a6SMike Kravetz goto out; 2158aa888a74SAndi Kleen 2159e5bbc8a6SMike Kravetz /* 2160e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2161e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2162e5bbc8a6SMike Kravetz */ 2163a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2164e4e574b7SAdam Litke 2165685f3457SLee Schermerhorn /* 2166685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 21679b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 21689b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 21699b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 21709b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 21719b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2172e5bbc8a6SMike Kravetz * 2173e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2174e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2175e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2176685f3457SLee Schermerhorn */ 2177685f3457SLee Schermerhorn while (nr_pages--) { 2178e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2179e5bbc8a6SMike Kravetz unused_resv_pages--; 21808cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2181e5bbc8a6SMike Kravetz goto out; 21827848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2183e4e574b7SAdam Litke } 2184e5bbc8a6SMike Kravetz 2185e5bbc8a6SMike Kravetz out: 2186e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2187e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2188e4e574b7SAdam Litke } 2189e4e574b7SAdam Litke 21905e911373SMike Kravetz 2191c37f9fb1SAndy Whitcroft /* 2192feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21935e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2194cf3ad20bSMike Kravetz * 2195cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2196cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2197cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2198cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2199cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2200feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2201feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2202feba16e2SMike Kravetz * is called in such cases. 2203cf3ad20bSMike Kravetz * 2204cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2205cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2206cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2207cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2208cf3ad20bSMike Kravetz * take appropriate action. 220996b96a96SMike Kravetz * 221096b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 221196b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 221296b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 221396b96a96SMike Kravetz * reservation exists. 2214c37f9fb1SAndy Whitcroft */ 22155e911373SMike Kravetz enum vma_resv_mode { 22165e911373SMike Kravetz VMA_NEEDS_RESV, 22175e911373SMike Kravetz VMA_COMMIT_RESV, 2218feba16e2SMike Kravetz VMA_END_RESV, 221996b96a96SMike Kravetz VMA_ADD_RESV, 22205e911373SMike Kravetz }; 2221cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2222cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 22235e911373SMike Kravetz enum vma_resv_mode mode) 2224c37f9fb1SAndy Whitcroft { 22254e35f483SJoonsoo Kim struct resv_map *resv; 22264e35f483SJoonsoo Kim pgoff_t idx; 2227cf3ad20bSMike Kravetz long ret; 22280db9d74eSMina Almasry long dummy_out_regions_needed; 2229c37f9fb1SAndy Whitcroft 22304e35f483SJoonsoo Kim resv = vma_resv_map(vma); 22314e35f483SJoonsoo Kim if (!resv) 2232c37f9fb1SAndy Whitcroft return 1; 2233c37f9fb1SAndy Whitcroft 22344e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 22355e911373SMike Kravetz switch (mode) { 22365e911373SMike Kravetz case VMA_NEEDS_RESV: 22370db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 22380db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 22390db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 22400db9d74eSMina Almasry * ever require 1 region. 22410db9d74eSMina Almasry */ 22420db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 22435e911373SMike Kravetz break; 22445e911373SMike Kravetz case VMA_COMMIT_RESV: 2245075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22460db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22470db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22485e911373SMike Kravetz break; 2249feba16e2SMike Kravetz case VMA_END_RESV: 22500db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 22515e911373SMike Kravetz ret = 0; 22525e911373SMike Kravetz break; 225396b96a96SMike Kravetz case VMA_ADD_RESV: 22540db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2255075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 22560db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 22570db9d74eSMina Almasry VM_BUG_ON(ret < 0); 22580db9d74eSMina Almasry } else { 22590db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 226096b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 226196b96a96SMike Kravetz } 226296b96a96SMike Kravetz break; 22635e911373SMike Kravetz default: 22645e911373SMike Kravetz BUG(); 22655e911373SMike Kravetz } 226684afd99bSAndy Whitcroft 22674e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2268cf3ad20bSMike Kravetz return ret; 226967961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 227067961f9dSMike Kravetz /* 227167961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 227267961f9dSMike Kravetz * However, a file associated with mapping could have been 227367961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 227467961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 227567961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 227667961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 227767961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 227867961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 227967961f9dSMike Kravetz * reservation has already been consumed. As a result, the 228067961f9dSMike Kravetz * return value of this routine is the opposite of the 228167961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 228267961f9dSMike Kravetz */ 228367961f9dSMike Kravetz if (ret) 228467961f9dSMike Kravetz return 0; 228567961f9dSMike Kravetz else 228667961f9dSMike Kravetz return 1; 228767961f9dSMike Kravetz } 22884e35f483SJoonsoo Kim else 2289cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 229084afd99bSAndy Whitcroft } 2291cf3ad20bSMike Kravetz 2292cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2293a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2294c37f9fb1SAndy Whitcroft { 22955e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2296cf3ad20bSMike Kravetz } 2297c37f9fb1SAndy Whitcroft 2298cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2299cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2300cf3ad20bSMike Kravetz { 23015e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 23025e911373SMike Kravetz } 23035e911373SMike Kravetz 2304feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 23055e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 23065e911373SMike Kravetz { 2307feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2308c37f9fb1SAndy Whitcroft } 2309c37f9fb1SAndy Whitcroft 231096b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 231196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 231296b96a96SMike Kravetz { 231396b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 231496b96a96SMike Kravetz } 231596b96a96SMike Kravetz 231696b96a96SMike Kravetz /* 231796b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 231896b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 231996b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 232096b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 232196b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 232296b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 232396b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 232496b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 232596b96a96SMike Kravetz * to be made by free_huge_page. 232696b96a96SMike Kravetz */ 232796b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 232896b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 232996b96a96SMike Kravetz struct page *page) 233096b96a96SMike Kravetz { 233196b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 233296b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 233396b96a96SMike Kravetz 233496b96a96SMike Kravetz if (unlikely(rc < 0)) { 233596b96a96SMike Kravetz /* 233696b96a96SMike Kravetz * Rare out of memory condition in reserve map 233796b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 233896b96a96SMike Kravetz * global reserve count will not be incremented 233996b96a96SMike Kravetz * by free_huge_page. This will make it appear 234096b96a96SMike Kravetz * as though the reservation for this page was 234196b96a96SMike Kravetz * consumed. This may prevent the task from 234296b96a96SMike Kravetz * faulting in the page at a later time. This 234396b96a96SMike Kravetz * is better than inconsistent global huge page 234496b96a96SMike Kravetz * accounting of reserve counts. 234596b96a96SMike Kravetz */ 234696b96a96SMike Kravetz ClearPagePrivate(page); 234796b96a96SMike Kravetz } else if (rc) { 234896b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 234996b96a96SMike Kravetz if (unlikely(rc < 0)) 235096b96a96SMike Kravetz /* 235196b96a96SMike Kravetz * See above comment about rare out of 235296b96a96SMike Kravetz * memory condition. 235396b96a96SMike Kravetz */ 235496b96a96SMike Kravetz ClearPagePrivate(page); 235596b96a96SMike Kravetz } else 235696b96a96SMike Kravetz vma_end_reservation(h, vma, address); 235796b96a96SMike Kravetz } 235896b96a96SMike Kravetz } 235996b96a96SMike Kravetz 236070c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 236104f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2362348ea204SAdam Litke { 236390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2364a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2365348ea204SAdam Litke struct page *page; 2366d85f69b0SMike Kravetz long map_chg, map_commit; 2367d85f69b0SMike Kravetz long gbl_chg; 23686d76dcf4SAneesh Kumar K.V int ret, idx; 23696d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 237008cf9fafSMina Almasry bool deferred_reserve; 23712fc39cecSAdam Litke 23726d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2373a1e78772SMel Gorman /* 2374d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2375d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2376d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2377a1e78772SMel Gorman */ 2378d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2379d85f69b0SMike Kravetz if (map_chg < 0) 238076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2381d85f69b0SMike Kravetz 2382d85f69b0SMike Kravetz /* 2383d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2384d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2385d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2386d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2387d85f69b0SMike Kravetz * checked against any subpool limit. 2388d85f69b0SMike Kravetz */ 2389d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2390d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2391d85f69b0SMike Kravetz if (gbl_chg < 0) { 2392feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 239376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 23945e911373SMike Kravetz } 239590d8b7e6SAdam Litke 2396d85f69b0SMike Kravetz /* 2397d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2398d85f69b0SMike Kravetz * map, there could be reservations associated with the 2399d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2400d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2401d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2402d85f69b0SMike Kravetz * the subpool reservations. 2403d85f69b0SMike Kravetz */ 2404d85f69b0SMike Kravetz if (avoid_reserve) 2405d85f69b0SMike Kravetz gbl_chg = 1; 2406d85f69b0SMike Kravetz } 2407d85f69b0SMike Kravetz 240808cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 240908cf9fafSMina Almasry */ 241008cf9fafSMina Almasry deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); 241108cf9fafSMina Almasry if (deferred_reserve) { 241208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 241308cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 24148f34af6fSJianyu Zhan if (ret) 24158f34af6fSJianyu Zhan goto out_subpool_put; 241608cf9fafSMina Almasry } 241708cf9fafSMina Almasry 241808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 241908cf9fafSMina Almasry if (ret) 242008cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 24218f34af6fSJianyu Zhan 2422a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2423d85f69b0SMike Kravetz /* 2424d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2425d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2426d85f69b0SMike Kravetz * a reservation exists for the allocation. 2427d85f69b0SMike Kravetz */ 2428d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 242981a6fcaeSJoonsoo Kim if (!page) { 243094ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 24310c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 24328f34af6fSJianyu Zhan if (!page) 24338f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2434a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2435a88c7695SNaoya Horiguchi SetPagePrivate(page); 2436a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2437a88c7695SNaoya Horiguchi } 243879dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 243979dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 244081a6fcaeSJoonsoo Kim /* Fall through */ 2441a1e78772SMel Gorman } 244281a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 244308cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 244408cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 244508cf9fafSMina Almasry */ 244608cf9fafSMina Almasry if (deferred_reserve) { 244708cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 244808cf9fafSMina Almasry h_cg, page); 244908cf9fafSMina Almasry } 245008cf9fafSMina Almasry 245181a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2452a1e78772SMel Gorman 245390481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2454a1e78772SMel Gorman 2455d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2456d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 245733039678SMike Kravetz /* 245833039678SMike Kravetz * The page was added to the reservation map between 245933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 246033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 246133039678SMike Kravetz * Adjust for the subpool count incremented above AND 246233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 246333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 246433039678SMike Kravetz * no longer applies. 246533039678SMike Kravetz */ 246633039678SMike Kravetz long rsv_adjust; 246733039678SMike Kravetz 246833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 246933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 247033039678SMike Kravetz } 24717893d1d5SAdam Litke return page; 24728f34af6fSJianyu Zhan 24738f34af6fSJianyu Zhan out_uncharge_cgroup: 24748f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 247508cf9fafSMina Almasry out_uncharge_cgroup_reservation: 247608cf9fafSMina Almasry if (deferred_reserve) 247708cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 247808cf9fafSMina Almasry h_cg); 24798f34af6fSJianyu Zhan out_subpool_put: 2480d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 24818f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2482feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 24838f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2484b45b5bd6SDavid Gibson } 2485b45b5bd6SDavid Gibson 2486e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2487e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2488e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2489aa888a74SAndi Kleen { 2490aa888a74SAndi Kleen struct huge_bootmem_page *m; 2491b2261026SJoonsoo Kim int nr_nodes, node; 2492aa888a74SAndi Kleen 2493b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2494aa888a74SAndi Kleen void *addr; 2495aa888a74SAndi Kleen 2496eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 24978b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 249897ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2499aa888a74SAndi Kleen if (addr) { 2500aa888a74SAndi Kleen /* 2501aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2502aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2503aa888a74SAndi Kleen * puts them into the mem_map). 2504aa888a74SAndi Kleen */ 2505aa888a74SAndi Kleen m = addr; 2506aa888a74SAndi Kleen goto found; 2507aa888a74SAndi Kleen } 2508aa888a74SAndi Kleen } 2509aa888a74SAndi Kleen return 0; 2510aa888a74SAndi Kleen 2511aa888a74SAndi Kleen found: 2512df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2513aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2514330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2515aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2516aa888a74SAndi Kleen m->hstate = h; 2517aa888a74SAndi Kleen return 1; 2518aa888a74SAndi Kleen } 2519aa888a74SAndi Kleen 2520d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2521d00181b9SKirill A. Shutemov unsigned int order) 252218229df5SAndy Whitcroft { 252318229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 252418229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 252518229df5SAndy Whitcroft else 252618229df5SAndy Whitcroft prep_compound_page(page, order); 252718229df5SAndy Whitcroft } 252818229df5SAndy Whitcroft 2529aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2530aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2531aa888a74SAndi Kleen { 2532aa888a74SAndi Kleen struct huge_bootmem_page *m; 2533aa888a74SAndi Kleen 2534aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 253540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2536aa888a74SAndi Kleen struct hstate *h = m->hstate; 2537ee8f248dSBecky Bruce 2538aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 253918229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2540ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2541aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2542af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2543af0fb9dfSMichal Hocko 2544b0320c7bSRafael Aquini /* 2545b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2546b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2547b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2548b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2549b0320c7bSRafael Aquini */ 2550bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 25513dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2552520495feSCannon Matthews cond_resched(); 2553aa888a74SAndi Kleen } 2554aa888a74SAndi Kleen } 2555aa888a74SAndi Kleen 25568faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 25571da177e4SLinus Torvalds { 25581da177e4SLinus Torvalds unsigned long i; 2559f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2560f60858f9SMike Kravetz 2561f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2562f60858f9SMike Kravetz /* 2563f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2564f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2565f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2566f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2567f60858f9SMike Kravetz */ 2568f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2569f60858f9SMike Kravetz GFP_KERNEL); 2570f60858f9SMike Kravetz } else { 2571f60858f9SMike Kravetz /* allocations done at boot time */ 2572f60858f9SMike Kravetz node_alloc_noretry = NULL; 2573f60858f9SMike Kravetz } 2574f60858f9SMike Kravetz 2575f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2576f60858f9SMike Kravetz if (node_alloc_noretry) 2577f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 25781da177e4SLinus Torvalds 2579e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2580bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2581dbda8feaSBarry Song if (hugetlb_cma_size) { 2582cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 2583cf11e85fSRoman Gushchin break; 2584cf11e85fSRoman Gushchin } 2585aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2586aa888a74SAndi Kleen break; 25870c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2588f60858f9SMike Kravetz &node_states[N_MEMORY], 2589f60858f9SMike Kravetz node_alloc_noretry)) 25901da177e4SLinus Torvalds break; 259169ed779aSDavid Rientjes cond_resched(); 25921da177e4SLinus Torvalds } 2593d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2594d715cf80SLiam R. Howlett char buf[32]; 2595d715cf80SLiam R. Howlett 2596c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2597d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2598d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 25998faa8b07SAndi Kleen h->max_huge_pages = i; 2600e5ff2159SAndi Kleen } 2601f60858f9SMike Kravetz 2602f60858f9SMike Kravetz kfree(node_alloc_noretry); 2603d715cf80SLiam R. Howlett } 2604e5ff2159SAndi Kleen 2605e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2606e5ff2159SAndi Kleen { 2607e5ff2159SAndi Kleen struct hstate *h; 2608e5ff2159SAndi Kleen 2609e5ff2159SAndi Kleen for_each_hstate(h) { 2610641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2611641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2612641844f5SNaoya Horiguchi 26138faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2614bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 26158faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2616e5ff2159SAndi Kleen } 2617641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2618e5ff2159SAndi Kleen } 2619e5ff2159SAndi Kleen 2620e5ff2159SAndi Kleen static void __init report_hugepages(void) 2621e5ff2159SAndi Kleen { 2622e5ff2159SAndi Kleen struct hstate *h; 2623e5ff2159SAndi Kleen 2624e5ff2159SAndi Kleen for_each_hstate(h) { 26254abd32dbSAndi Kleen char buf[32]; 2626c6247f72SMatthew Wilcox 2627c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2628ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2629c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2630e5ff2159SAndi Kleen } 2631e5ff2159SAndi Kleen } 2632e5ff2159SAndi Kleen 26331da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 26346ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 26356ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26361da177e4SLinus Torvalds { 26374415cc8dSChristoph Lameter int i; 26384415cc8dSChristoph Lameter 2639bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2640aa888a74SAndi Kleen return; 2641aa888a74SAndi Kleen 26426ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 26431da177e4SLinus Torvalds struct page *page, *next; 2644a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2645a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2646a5516438SAndi Kleen if (count >= h->nr_huge_pages) 26476b0c880dSAdam Litke return; 26481da177e4SLinus Torvalds if (PageHighMem(page)) 26491da177e4SLinus Torvalds continue; 26501da177e4SLinus Torvalds list_del(&page->lru); 2651e5ff2159SAndi Kleen update_and_free_page(h, page); 2652a5516438SAndi Kleen h->free_huge_pages--; 2653a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 26541da177e4SLinus Torvalds } 26551da177e4SLinus Torvalds } 26561da177e4SLinus Torvalds } 26571da177e4SLinus Torvalds #else 26586ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 26596ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26601da177e4SLinus Torvalds { 26611da177e4SLinus Torvalds } 26621da177e4SLinus Torvalds #endif 26631da177e4SLinus Torvalds 266420a0307cSWu Fengguang /* 266520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 266620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 266720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 266820a0307cSWu Fengguang */ 26696ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 26706ae11b27SLee Schermerhorn int delta) 267120a0307cSWu Fengguang { 2672b2261026SJoonsoo Kim int nr_nodes, node; 267320a0307cSWu Fengguang 267420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 267520a0307cSWu Fengguang 2676e8c5c824SLee Schermerhorn if (delta < 0) { 2677b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2678b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2679b2261026SJoonsoo Kim goto found; 2680b2261026SJoonsoo Kim } 2681b2261026SJoonsoo Kim } else { 2682b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2683b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2684b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2685b2261026SJoonsoo Kim goto found; 2686e8c5c824SLee Schermerhorn } 26879a76db09SLee Schermerhorn } 2688b2261026SJoonsoo Kim return 0; 268920a0307cSWu Fengguang 2690b2261026SJoonsoo Kim found: 269120a0307cSWu Fengguang h->surplus_huge_pages += delta; 2692b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2693b2261026SJoonsoo Kim return 1; 269420a0307cSWu Fengguang } 269520a0307cSWu Fengguang 2696a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2697fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 26986ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26991da177e4SLinus Torvalds { 27007893d1d5SAdam Litke unsigned long min_count, ret; 2701f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2702f60858f9SMike Kravetz 2703f60858f9SMike Kravetz /* 2704f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2705f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2706f60858f9SMike Kravetz * the requested huge pages. 2707f60858f9SMike Kravetz */ 2708f60858f9SMike Kravetz if (node_alloc_noretry) 2709f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2710f60858f9SMike Kravetz else 2711f60858f9SMike Kravetz return -ENOMEM; 27121da177e4SLinus Torvalds 27134eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 27144eb0716eSAlexandre Ghiti 27154eb0716eSAlexandre Ghiti /* 2716fd875dcaSMike Kravetz * Check for a node specific request. 2717fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2718fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2719fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2720fd875dcaSMike Kravetz */ 2721fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2722fd875dcaSMike Kravetz unsigned long old_count = count; 2723fd875dcaSMike Kravetz 2724fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2725fd875dcaSMike Kravetz /* 2726fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2727fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2728fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2729fd875dcaSMike Kravetz * largest possible value to align with their intention. 2730fd875dcaSMike Kravetz */ 2731fd875dcaSMike Kravetz if (count < old_count) 2732fd875dcaSMike Kravetz count = ULONG_MAX; 2733fd875dcaSMike Kravetz } 2734fd875dcaSMike Kravetz 2735fd875dcaSMike Kravetz /* 27364eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 27374eb0716eSAlexandre Ghiti * page range allocation. 27384eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 27394eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 27404eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 27414eb0716eSAlexandre Ghiti */ 27424eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 27434eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 27444eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2745f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 27464eb0716eSAlexandre Ghiti return -EINVAL; 27474eb0716eSAlexandre Ghiti } 27484eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 27494eb0716eSAlexandre Ghiti } 2750aa888a74SAndi Kleen 27517893d1d5SAdam Litke /* 27527893d1d5SAdam Litke * Increase the pool size 27537893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 27547893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2755d1c3fb1fSNishanth Aravamudan * 27560c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2757d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2758d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2759d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2760d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 27617893d1d5SAdam Litke */ 2762a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 27636ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 27647893d1d5SAdam Litke break; 27657893d1d5SAdam Litke } 27667893d1d5SAdam Litke 2767a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 27687893d1d5SAdam Litke /* 27697893d1d5SAdam Litke * If this allocation races such that we no longer need the 27707893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 27717893d1d5SAdam Litke * and reducing the surplus. 27727893d1d5SAdam Litke */ 27737893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2774649920c6SJia He 2775649920c6SJia He /* yield cpu to avoid soft lockup */ 2776649920c6SJia He cond_resched(); 2777649920c6SJia He 2778f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2779f60858f9SMike Kravetz node_alloc_noretry); 27807893d1d5SAdam Litke spin_lock(&hugetlb_lock); 27817893d1d5SAdam Litke if (!ret) 27827893d1d5SAdam Litke goto out; 27837893d1d5SAdam Litke 2784536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2785536240f2SMel Gorman if (signal_pending(current)) 2786536240f2SMel Gorman goto out; 27877893d1d5SAdam Litke } 27887893d1d5SAdam Litke 27897893d1d5SAdam Litke /* 27907893d1d5SAdam Litke * Decrease the pool size 27917893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27927893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27937893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27947893d1d5SAdam Litke * to the desired size as pages become free. 2795d1c3fb1fSNishanth Aravamudan * 2796d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2797d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2798d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27990c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2800d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2801d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2802d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 28037893d1d5SAdam Litke */ 2804a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 28056b0c880dSAdam Litke min_count = max(count, min_count); 28066ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2807a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 28086ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 28091da177e4SLinus Torvalds break; 281055f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 28111da177e4SLinus Torvalds } 2812a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 28136ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 28147893d1d5SAdam Litke break; 28157893d1d5SAdam Litke } 28167893d1d5SAdam Litke out: 28174eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 28181da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 28194eb0716eSAlexandre Ghiti 2820f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2821f60858f9SMike Kravetz 28224eb0716eSAlexandre Ghiti return 0; 28231da177e4SLinus Torvalds } 28241da177e4SLinus Torvalds 2825a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2826a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2827a3437870SNishanth Aravamudan 2828a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2829a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2830a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2831a3437870SNishanth Aravamudan 2832a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2833a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2834a3437870SNishanth Aravamudan 28359a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 28369a305230SLee Schermerhorn 28379a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2838a3437870SNishanth Aravamudan { 2839a3437870SNishanth Aravamudan int i; 28409a305230SLee Schermerhorn 2841a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 28429a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 28439a305230SLee Schermerhorn if (nidp) 28449a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2845a3437870SNishanth Aravamudan return &hstates[i]; 28469a305230SLee Schermerhorn } 28479a305230SLee Schermerhorn 28489a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2849a3437870SNishanth Aravamudan } 2850a3437870SNishanth Aravamudan 285106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2852a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2853a3437870SNishanth Aravamudan { 28549a305230SLee Schermerhorn struct hstate *h; 28559a305230SLee Schermerhorn unsigned long nr_huge_pages; 28569a305230SLee Schermerhorn int nid; 28579a305230SLee Schermerhorn 28589a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 28599a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 28609a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 28619a305230SLee Schermerhorn else 28629a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 28639a305230SLee Schermerhorn 28649a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2865a3437870SNishanth Aravamudan } 2866adbe8726SEric B Munson 2867238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2868238d3c13SDavid Rientjes struct hstate *h, int nid, 2869238d3c13SDavid Rientjes unsigned long count, size_t len) 2870a3437870SNishanth Aravamudan { 2871a3437870SNishanth Aravamudan int err; 28722d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2873a3437870SNishanth Aravamudan 28742d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 28752d0adf7eSOscar Salvador return -EINVAL; 2876adbe8726SEric B Munson 28779a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 28789a305230SLee Schermerhorn /* 28799a305230SLee Schermerhorn * global hstate attribute 28809a305230SLee Schermerhorn */ 28819a305230SLee Schermerhorn if (!(obey_mempolicy && 28822d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 28832d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 28842d0adf7eSOscar Salvador else 28852d0adf7eSOscar Salvador n_mask = &nodes_allowed; 28862d0adf7eSOscar Salvador } else { 28879a305230SLee Schermerhorn /* 2888fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2889fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28909a305230SLee Schermerhorn */ 28912d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28922d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2893fd875dcaSMike Kravetz } 28949a305230SLee Schermerhorn 28952d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 289606808b08SLee Schermerhorn 28974eb0716eSAlexandre Ghiti return err ? err : len; 289806808b08SLee Schermerhorn } 289906808b08SLee Schermerhorn 2900238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2901238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2902238d3c13SDavid Rientjes size_t len) 2903238d3c13SDavid Rientjes { 2904238d3c13SDavid Rientjes struct hstate *h; 2905238d3c13SDavid Rientjes unsigned long count; 2906238d3c13SDavid Rientjes int nid; 2907238d3c13SDavid Rientjes int err; 2908238d3c13SDavid Rientjes 2909238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2910238d3c13SDavid Rientjes if (err) 2911238d3c13SDavid Rientjes return err; 2912238d3c13SDavid Rientjes 2913238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2914238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2915238d3c13SDavid Rientjes } 2916238d3c13SDavid Rientjes 291706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 291806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 291906808b08SLee Schermerhorn { 292006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 292106808b08SLee Schermerhorn } 292206808b08SLee Schermerhorn 292306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 292406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 292506808b08SLee Schermerhorn { 2926238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2927a3437870SNishanth Aravamudan } 2928a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2929a3437870SNishanth Aravamudan 293006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 293106808b08SLee Schermerhorn 293206808b08SLee Schermerhorn /* 293306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 293406808b08SLee Schermerhorn * huge page alloc/free. 293506808b08SLee Schermerhorn */ 293606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 293706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 293806808b08SLee Schermerhorn { 293906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 294006808b08SLee Schermerhorn } 294106808b08SLee Schermerhorn 294206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 294306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 294406808b08SLee Schermerhorn { 2945238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 294606808b08SLee Schermerhorn } 294706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 294806808b08SLee Schermerhorn #endif 294906808b08SLee Schermerhorn 295006808b08SLee Schermerhorn 2951a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2952a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2953a3437870SNishanth Aravamudan { 29549a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2955a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2956a3437870SNishanth Aravamudan } 2957adbe8726SEric B Munson 2958a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2959a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2960a3437870SNishanth Aravamudan { 2961a3437870SNishanth Aravamudan int err; 2962a3437870SNishanth Aravamudan unsigned long input; 29639a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2964a3437870SNishanth Aravamudan 2965bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2966adbe8726SEric B Munson return -EINVAL; 2967adbe8726SEric B Munson 29683dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2969a3437870SNishanth Aravamudan if (err) 297073ae31e5SEric B Munson return err; 2971a3437870SNishanth Aravamudan 2972a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2973a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2974a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2975a3437870SNishanth Aravamudan 2976a3437870SNishanth Aravamudan return count; 2977a3437870SNishanth Aravamudan } 2978a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2979a3437870SNishanth Aravamudan 2980a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2981a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2982a3437870SNishanth Aravamudan { 29839a305230SLee Schermerhorn struct hstate *h; 29849a305230SLee Schermerhorn unsigned long free_huge_pages; 29859a305230SLee Schermerhorn int nid; 29869a305230SLee Schermerhorn 29879a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29889a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29899a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29909a305230SLee Schermerhorn else 29919a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29929a305230SLee Schermerhorn 29939a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2994a3437870SNishanth Aravamudan } 2995a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2996a3437870SNishanth Aravamudan 2997a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2998a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2999a3437870SNishanth Aravamudan { 30009a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3001a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 3002a3437870SNishanth Aravamudan } 3003a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3004a3437870SNishanth Aravamudan 3005a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3006a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3007a3437870SNishanth Aravamudan { 30089a305230SLee Schermerhorn struct hstate *h; 30099a305230SLee Schermerhorn unsigned long surplus_huge_pages; 30109a305230SLee Schermerhorn int nid; 30119a305230SLee Schermerhorn 30129a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 30139a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 30149a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 30159a305230SLee Schermerhorn else 30169a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 30179a305230SLee Schermerhorn 30189a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 3019a3437870SNishanth Aravamudan } 3020a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3021a3437870SNishanth Aravamudan 3022a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3023a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3024a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3025a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3026a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3027a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 302806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 302906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 303006808b08SLee Schermerhorn #endif 3031a3437870SNishanth Aravamudan NULL, 3032a3437870SNishanth Aravamudan }; 3033a3437870SNishanth Aravamudan 303467e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3035a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3036a3437870SNishanth Aravamudan }; 3037a3437870SNishanth Aravamudan 3038094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 30399a305230SLee Schermerhorn struct kobject **hstate_kobjs, 304067e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3041a3437870SNishanth Aravamudan { 3042a3437870SNishanth Aravamudan int retval; 3043972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3044a3437870SNishanth Aravamudan 30459a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 30469a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3047a3437870SNishanth Aravamudan return -ENOMEM; 3048a3437870SNishanth Aravamudan 30499a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3050a3437870SNishanth Aravamudan if (retval) 30519a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3052a3437870SNishanth Aravamudan 3053a3437870SNishanth Aravamudan return retval; 3054a3437870SNishanth Aravamudan } 3055a3437870SNishanth Aravamudan 3056a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3057a3437870SNishanth Aravamudan { 3058a3437870SNishanth Aravamudan struct hstate *h; 3059a3437870SNishanth Aravamudan int err; 3060a3437870SNishanth Aravamudan 3061a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3062a3437870SNishanth Aravamudan if (!hugepages_kobj) 3063a3437870SNishanth Aravamudan return; 3064a3437870SNishanth Aravamudan 3065a3437870SNishanth Aravamudan for_each_hstate(h) { 30669a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 30679a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3068a3437870SNishanth Aravamudan if (err) 3069282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3070a3437870SNishanth Aravamudan } 3071a3437870SNishanth Aravamudan } 3072a3437870SNishanth Aravamudan 30739a305230SLee Schermerhorn #ifdef CONFIG_NUMA 30749a305230SLee Schermerhorn 30759a305230SLee Schermerhorn /* 30769a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 307710fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 307810fbcf4cSKay Sievers * index of a node device or _hstate == node id. 307910fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 30809a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 30819a305230SLee Schermerhorn */ 30829a305230SLee Schermerhorn struct node_hstate { 30839a305230SLee Schermerhorn struct kobject *hugepages_kobj; 30849a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30859a305230SLee Schermerhorn }; 3086b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30879a305230SLee Schermerhorn 30889a305230SLee Schermerhorn /* 308910fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30909a305230SLee Schermerhorn */ 30919a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30929a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30939a305230SLee Schermerhorn &free_hugepages_attr.attr, 30949a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30959a305230SLee Schermerhorn NULL, 30969a305230SLee Schermerhorn }; 30979a305230SLee Schermerhorn 309867e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30999a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 31009a305230SLee Schermerhorn }; 31019a305230SLee Schermerhorn 31029a305230SLee Schermerhorn /* 310310fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 31049a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 31059a305230SLee Schermerhorn */ 31069a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31079a305230SLee Schermerhorn { 31089a305230SLee Schermerhorn int nid; 31099a305230SLee Schermerhorn 31109a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 31119a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 31129a305230SLee Schermerhorn int i; 31139a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 31149a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 31159a305230SLee Schermerhorn if (nidp) 31169a305230SLee Schermerhorn *nidp = nid; 31179a305230SLee Schermerhorn return &hstates[i]; 31189a305230SLee Schermerhorn } 31199a305230SLee Schermerhorn } 31209a305230SLee Schermerhorn 31219a305230SLee Schermerhorn BUG(); 31229a305230SLee Schermerhorn return NULL; 31239a305230SLee Schermerhorn } 31249a305230SLee Schermerhorn 31259a305230SLee Schermerhorn /* 312610fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 31279a305230SLee Schermerhorn * No-op if no hstate attributes attached. 31289a305230SLee Schermerhorn */ 31293cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 31309a305230SLee Schermerhorn { 31319a305230SLee Schermerhorn struct hstate *h; 313210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 31339a305230SLee Schermerhorn 31349a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31359b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 31369a305230SLee Schermerhorn 3137972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3138972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3139972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3140972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3141972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3142972dc4deSAneesh Kumar K.V } 31439a305230SLee Schermerhorn } 31449a305230SLee Schermerhorn 31459a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 31469a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 31479a305230SLee Schermerhorn } 31489a305230SLee Schermerhorn 31499a305230SLee Schermerhorn 31509a305230SLee Schermerhorn /* 315110fbcf4cSKay Sievers * Register hstate attributes for a single node device. 31529a305230SLee Schermerhorn * No-op if attributes already registered. 31539a305230SLee Schermerhorn */ 31543cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 31559a305230SLee Schermerhorn { 31569a305230SLee Schermerhorn struct hstate *h; 315710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 31589a305230SLee Schermerhorn int err; 31599a305230SLee Schermerhorn 31609a305230SLee Schermerhorn if (nhs->hugepages_kobj) 31619a305230SLee Schermerhorn return; /* already allocated */ 31629a305230SLee Schermerhorn 31639a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 316410fbcf4cSKay Sievers &node->dev.kobj); 31659a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 31669a305230SLee Schermerhorn return; 31679a305230SLee Schermerhorn 31689a305230SLee Schermerhorn for_each_hstate(h) { 31699a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 31709a305230SLee Schermerhorn nhs->hstate_kobjs, 31719a305230SLee Schermerhorn &per_node_hstate_attr_group); 31729a305230SLee Schermerhorn if (err) { 3173282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 317410fbcf4cSKay Sievers h->name, node->dev.id); 31759a305230SLee Schermerhorn hugetlb_unregister_node(node); 31769a305230SLee Schermerhorn break; 31779a305230SLee Schermerhorn } 31789a305230SLee Schermerhorn } 31799a305230SLee Schermerhorn } 31809a305230SLee Schermerhorn 31819a305230SLee Schermerhorn /* 31829b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 318310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 318410fbcf4cSKay Sievers * registered their associated device by this time. 31859a305230SLee Schermerhorn */ 31867d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31879a305230SLee Schermerhorn { 31889a305230SLee Schermerhorn int nid; 31899a305230SLee Schermerhorn 31908cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31918732794bSWen Congyang struct node *node = node_devices[nid]; 319210fbcf4cSKay Sievers if (node->dev.id == nid) 31939a305230SLee Schermerhorn hugetlb_register_node(node); 31949a305230SLee Schermerhorn } 31959a305230SLee Schermerhorn 31969a305230SLee Schermerhorn /* 319710fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31989a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31999a305230SLee Schermerhorn */ 32009a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 32019a305230SLee Schermerhorn hugetlb_unregister_node); 32029a305230SLee Schermerhorn } 32039a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 32049a305230SLee Schermerhorn 32059a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 32069a305230SLee Schermerhorn { 32079a305230SLee Schermerhorn BUG(); 32089a305230SLee Schermerhorn if (nidp) 32099a305230SLee Schermerhorn *nidp = -1; 32109a305230SLee Schermerhorn return NULL; 32119a305230SLee Schermerhorn } 32129a305230SLee Schermerhorn 32139a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 32149a305230SLee Schermerhorn 32159a305230SLee Schermerhorn #endif 32169a305230SLee Schermerhorn 3217a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3218a3437870SNishanth Aravamudan { 32198382d914SDavidlohr Bueso int i; 32208382d914SDavidlohr Bueso 3221c2833a5bSMike Kravetz if (!hugepages_supported()) { 3222c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3223c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 32240ef89d25SBenjamin Herrenschmidt return 0; 3225d715cf80SLiam R. Howlett } 3226d715cf80SLiam R. Howlett 3227282f4214SMike Kravetz /* 3228282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3229282f4214SMike Kravetz * architectures depend on setup being done here. 3230282f4214SMike Kravetz */ 3231a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3232282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3233282f4214SMike Kravetz /* 3234282f4214SMike Kravetz * If we did not parse a default huge page size, set 3235282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3236282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3237282f4214SMike Kravetz * specified, set that here as well. 3238282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3239282f4214SMike Kravetz * setting. A warning will be printed in this case. 3240282f4214SMike Kravetz */ 3241282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3242f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3243282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3244282f4214SMike Kravetz char buf[32]; 3245282f4214SMike Kravetz 3246282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3247282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3248282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3249282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3250282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3251282f4214SMike Kravetz default_hstate_max_huge_pages); 3252282f4214SMike Kravetz } 3253282f4214SMike Kravetz default_hstate.max_huge_pages = 3254282f4214SMike Kravetz default_hstate_max_huge_pages; 3255282f4214SMike Kravetz } 3256f8b74815SVaishali Thakkar } 3257a3437870SNishanth Aravamudan 3258cf11e85fSRoman Gushchin hugetlb_cma_check(); 3259a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3260aa888a74SAndi Kleen gather_bootmem_prealloc(); 3261a3437870SNishanth Aravamudan report_hugepages(); 3262a3437870SNishanth Aravamudan 3263a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 32649a305230SLee Schermerhorn hugetlb_register_all_nodes(); 32657179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 32669a305230SLee Schermerhorn 32678382d914SDavidlohr Bueso #ifdef CONFIG_SMP 32688382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 32698382d914SDavidlohr Bueso #else 32708382d914SDavidlohr Bueso num_fault_mutexes = 1; 32718382d914SDavidlohr Bueso #endif 3272c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 32736da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 32746da2ec56SKees Cook GFP_KERNEL); 3275c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 32768382d914SDavidlohr Bueso 32778382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3278c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3279a3437870SNishanth Aravamudan return 0; 3280a3437870SNishanth Aravamudan } 32813e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3282a3437870SNishanth Aravamudan 3283ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3284ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32859fee021dSVaishali Thakkar { 3286ae94da89SMike Kravetz return size == HPAGE_SIZE; 32879fee021dSVaishali Thakkar } 32889fee021dSVaishali Thakkar 3289d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3290a3437870SNishanth Aravamudan { 3291a3437870SNishanth Aravamudan struct hstate *h; 32928faa8b07SAndi Kleen unsigned long i; 32938faa8b07SAndi Kleen 3294a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3295a3437870SNishanth Aravamudan return; 3296a3437870SNishanth Aravamudan } 329747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3298a3437870SNishanth Aravamudan BUG_ON(order == 0); 329947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3300a3437870SNishanth Aravamudan h->order = order; 3301a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 33028faa8b07SAndi Kleen h->nr_huge_pages = 0; 33038faa8b07SAndi Kleen h->free_huge_pages = 0; 33048faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 33058faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 33060edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 330754f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 330854f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3309a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3310a3437870SNishanth Aravamudan huge_page_size(h)/1024); 33118faa8b07SAndi Kleen 3312a3437870SNishanth Aravamudan parsed_hstate = h; 3313a3437870SNishanth Aravamudan } 3314a3437870SNishanth Aravamudan 3315282f4214SMike Kravetz /* 3316282f4214SMike Kravetz * hugepages command line processing 3317282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3318282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3319282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3320282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3321282f4214SMike Kravetz */ 3322282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3323a3437870SNishanth Aravamudan { 3324a3437870SNishanth Aravamudan unsigned long *mhp; 33258faa8b07SAndi Kleen static unsigned long *last_mhp; 3326a3437870SNishanth Aravamudan 33279fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3328282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 33299fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3330282f4214SMike Kravetz return 0; 33319fee021dSVaishali Thakkar } 3332282f4214SMike Kravetz 3333a3437870SNishanth Aravamudan /* 3334282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3335282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3336282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3337282f4214SMike Kravetz * default_hugepagesz. 3338a3437870SNishanth Aravamudan */ 33399fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3340a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3341a3437870SNishanth Aravamudan else 3342a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3343a3437870SNishanth Aravamudan 33448faa8b07SAndi Kleen if (mhp == last_mhp) { 3345282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3346282f4214SMike Kravetz return 0; 33478faa8b07SAndi Kleen } 33488faa8b07SAndi Kleen 3349a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3350a3437870SNishanth Aravamudan *mhp = 0; 3351a3437870SNishanth Aravamudan 33528faa8b07SAndi Kleen /* 33538faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 33548faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 33558faa8b07SAndi Kleen * use the bootmem allocator. 33568faa8b07SAndi Kleen */ 335747d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 33588faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 33598faa8b07SAndi Kleen 33608faa8b07SAndi Kleen last_mhp = mhp; 33618faa8b07SAndi Kleen 3362a3437870SNishanth Aravamudan return 1; 3363a3437870SNishanth Aravamudan } 3364282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3365e11bfbfcSNick Piggin 3366282f4214SMike Kravetz /* 3367282f4214SMike Kravetz * hugepagesz command line processing 3368282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3369282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3370282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3371282f4214SMike Kravetz * hugepagesz argument was valid. 3372282f4214SMike Kravetz */ 3373359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3374e11bfbfcSNick Piggin { 3375359f2544SMike Kravetz unsigned long size; 3376282f4214SMike Kravetz struct hstate *h; 3377282f4214SMike Kravetz 3378282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3379359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3380359f2544SMike Kravetz 3381359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3382282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3383359f2544SMike Kravetz return 0; 3384359f2544SMike Kravetz } 3385359f2544SMike Kravetz 3386282f4214SMike Kravetz h = size_to_hstate(size); 3387282f4214SMike Kravetz if (h) { 3388282f4214SMike Kravetz /* 3389282f4214SMike Kravetz * hstate for this size already exists. This is normally 3390282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3391282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3392282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3393282f4214SMike Kravetz * previously specified. 3394282f4214SMike Kravetz */ 3395282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3396282f4214SMike Kravetz default_hstate.max_huge_pages) { 3397282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 339838237830SMike Kravetz return 0; 339938237830SMike Kravetz } 340038237830SMike Kravetz 3401282f4214SMike Kravetz /* 3402282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3403282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3404282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3405282f4214SMike Kravetz */ 3406282f4214SMike Kravetz parsed_hstate = h; 3407282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3408e11bfbfcSNick Piggin return 1; 3409e11bfbfcSNick Piggin } 3410282f4214SMike Kravetz 3411359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3412282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3413359f2544SMike Kravetz return 1; 3414359f2544SMike Kravetz } 3415359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3416359f2544SMike Kravetz 3417282f4214SMike Kravetz /* 3418282f4214SMike Kravetz * default_hugepagesz command line input 3419282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3420282f4214SMike Kravetz */ 3421ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3422e11bfbfcSNick Piggin { 3423ae94da89SMike Kravetz unsigned long size; 3424ae94da89SMike Kravetz 3425282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3426282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3427282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3428282f4214SMike Kravetz return 0; 3429282f4214SMike Kravetz } 3430282f4214SMike Kravetz 3431282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3432282f4214SMike Kravetz 3433282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3434282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3435282f4214SMike Kravetz return 0; 3436282f4214SMike Kravetz } 3437282f4214SMike Kravetz 3438282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3439282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3440282f4214SMike Kravetz parsed_default_hugepagesz = true; 3441282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3442282f4214SMike Kravetz 3443282f4214SMike Kravetz /* 3444282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3445282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3446282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3447282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3448282f4214SMike Kravetz * allocated here from bootmem allocator. 3449282f4214SMike Kravetz */ 3450282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3451282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3452282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3453282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3454282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3455282f4214SMike Kravetz } 3456282f4214SMike Kravetz 3457e11bfbfcSNick Piggin return 1; 3458e11bfbfcSNick Piggin } 3459ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3460a3437870SNishanth Aravamudan 34618a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 34628a213460SNishanth Aravamudan { 34638a213460SNishanth Aravamudan int node; 34648a213460SNishanth Aravamudan unsigned int nr = 0; 34658a213460SNishanth Aravamudan 34668a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 34678a213460SNishanth Aravamudan nr += array[node]; 34688a213460SNishanth Aravamudan 34698a213460SNishanth Aravamudan return nr; 34708a213460SNishanth Aravamudan } 34718a213460SNishanth Aravamudan 34728a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 347306808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 347406808b08SLee Schermerhorn struct ctl_table *table, int write, 347532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34761da177e4SLinus Torvalds { 3477e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3478238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 347908d4a246SMichal Hocko int ret; 3480e5ff2159SAndi Kleen 3481457c1b27SNishanth Aravamudan if (!hugepages_supported()) 348286613628SJan Stancek return -EOPNOTSUPP; 3483457c1b27SNishanth Aravamudan 3484e5ff2159SAndi Kleen table->data = &tmp; 3485e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 348608d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 348708d4a246SMichal Hocko if (ret) 348808d4a246SMichal Hocko goto out; 3489e5ff2159SAndi Kleen 3490238d3c13SDavid Rientjes if (write) 3491238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3492238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 349308d4a246SMichal Hocko out: 349408d4a246SMichal Hocko return ret; 34951da177e4SLinus Torvalds } 3496396faf03SMel Gorman 349706808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 349832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 349906808b08SLee Schermerhorn { 350006808b08SLee Schermerhorn 350106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 350206808b08SLee Schermerhorn buffer, length, ppos); 350306808b08SLee Schermerhorn } 350406808b08SLee Schermerhorn 350506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 350606808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 350732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 350806808b08SLee Schermerhorn { 350906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 351006808b08SLee Schermerhorn buffer, length, ppos); 351106808b08SLee Schermerhorn } 351206808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 351306808b08SLee Schermerhorn 3514a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 351532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3516a3d0c6aaSNishanth Aravamudan { 3517a5516438SAndi Kleen struct hstate *h = &default_hstate; 3518e5ff2159SAndi Kleen unsigned long tmp; 351908d4a246SMichal Hocko int ret; 3520e5ff2159SAndi Kleen 3521457c1b27SNishanth Aravamudan if (!hugepages_supported()) 352286613628SJan Stancek return -EOPNOTSUPP; 3523457c1b27SNishanth Aravamudan 3524e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3525e5ff2159SAndi Kleen 3526bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3527adbe8726SEric B Munson return -EINVAL; 3528adbe8726SEric B Munson 3529e5ff2159SAndi Kleen table->data = &tmp; 3530e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 353108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 353208d4a246SMichal Hocko if (ret) 353308d4a246SMichal Hocko goto out; 3534e5ff2159SAndi Kleen 3535e5ff2159SAndi Kleen if (write) { 3536064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3537e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3538a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3539e5ff2159SAndi Kleen } 354008d4a246SMichal Hocko out: 354108d4a246SMichal Hocko return ret; 3542a3d0c6aaSNishanth Aravamudan } 3543a3d0c6aaSNishanth Aravamudan 35441da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 35451da177e4SLinus Torvalds 3546e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 35471da177e4SLinus Torvalds { 3548fcb2b0c5SRoman Gushchin struct hstate *h; 3549fcb2b0c5SRoman Gushchin unsigned long total = 0; 3550fcb2b0c5SRoman Gushchin 3551457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3552457c1b27SNishanth Aravamudan return; 3553fcb2b0c5SRoman Gushchin 3554fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3555fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3556fcb2b0c5SRoman Gushchin 3557fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3558fcb2b0c5SRoman Gushchin 3559fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3560e1759c21SAlexey Dobriyan seq_printf(m, 35611da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 35621da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3563b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35647893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35654f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3566fcb2b0c5SRoman Gushchin count, 3567a5516438SAndi Kleen h->free_huge_pages, 3568a5516438SAndi Kleen h->resv_huge_pages, 3569a5516438SAndi Kleen h->surplus_huge_pages, 3570fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3571fcb2b0c5SRoman Gushchin } 3572fcb2b0c5SRoman Gushchin 3573fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 35741da177e4SLinus Torvalds } 35751da177e4SLinus Torvalds 35761da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 35771da177e4SLinus Torvalds { 3578a5516438SAndi Kleen struct hstate *h = &default_hstate; 3579457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3580457c1b27SNishanth Aravamudan return 0; 35811da177e4SLinus Torvalds return sprintf(buf, 35821da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3583a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3584a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3585a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3586a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3587a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35881da177e4SLinus Torvalds } 35891da177e4SLinus Torvalds 3590949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3591949f7ec5SDavid Rientjes { 3592949f7ec5SDavid Rientjes struct hstate *h; 3593949f7ec5SDavid Rientjes int nid; 3594949f7ec5SDavid Rientjes 3595457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3596457c1b27SNishanth Aravamudan return; 3597457c1b27SNishanth Aravamudan 3598949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3599949f7ec5SDavid Rientjes for_each_hstate(h) 3600949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3601949f7ec5SDavid Rientjes nid, 3602949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3603949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3604949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3605949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3606949f7ec5SDavid Rientjes } 3607949f7ec5SDavid Rientjes 36085d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 36095d317b2bSNaoya Horiguchi { 36105d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 36115d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 36125d317b2bSNaoya Horiguchi } 36135d317b2bSNaoya Horiguchi 36141da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 36151da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 36161da177e4SLinus Torvalds { 3617d0028588SWanpeng Li struct hstate *h; 3618d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3619d0028588SWanpeng Li 3620d0028588SWanpeng Li for_each_hstate(h) 3621d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3622d0028588SWanpeng Li return nr_total_pages; 36231da177e4SLinus Torvalds } 36241da177e4SLinus Torvalds 3625a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3626fc1b8a73SMel Gorman { 3627fc1b8a73SMel Gorman int ret = -ENOMEM; 3628fc1b8a73SMel Gorman 3629fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3630fc1b8a73SMel Gorman /* 3631fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3632fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3633fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3634fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3635fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3636fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3637fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3638fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3639fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3640fc1b8a73SMel Gorman * 3641fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3642fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3643fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3644fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3645fc1b8a73SMel Gorman * semantics that cpuset has. 3646fc1b8a73SMel Gorman */ 3647fc1b8a73SMel Gorman if (delta > 0) { 3648a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3649fc1b8a73SMel Gorman goto out; 3650fc1b8a73SMel Gorman 3651a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3652a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3653fc1b8a73SMel Gorman goto out; 3654fc1b8a73SMel Gorman } 3655fc1b8a73SMel Gorman } 3656fc1b8a73SMel Gorman 3657fc1b8a73SMel Gorman ret = 0; 3658fc1b8a73SMel Gorman if (delta < 0) 3659a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3660fc1b8a73SMel Gorman 3661fc1b8a73SMel Gorman out: 3662fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3663fc1b8a73SMel Gorman return ret; 3664fc1b8a73SMel Gorman } 3665fc1b8a73SMel Gorman 366684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 366784afd99bSAndy Whitcroft { 3668f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 366984afd99bSAndy Whitcroft 367084afd99bSAndy Whitcroft /* 367184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 367284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 367384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 367425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 367584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 367684afd99bSAndy Whitcroft * new reference here without additional locking. 367784afd99bSAndy Whitcroft */ 36784e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3679f522c3acSJoonsoo Kim kref_get(&resv->refs); 368084afd99bSAndy Whitcroft } 368184afd99bSAndy Whitcroft 3682a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3683a1e78772SMel Gorman { 3684a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3685f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 368690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36874e35f483SJoonsoo Kim unsigned long reserve, start, end; 36881c5ecae3SMike Kravetz long gbl_reserve; 368984afd99bSAndy Whitcroft 36904e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36914e35f483SJoonsoo Kim return; 36924e35f483SJoonsoo Kim 3693a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3694a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 369584afd99bSAndy Whitcroft 36964e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3697e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36987251ff78SAdam Litke if (reserve) { 36991c5ecae3SMike Kravetz /* 37001c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 37011c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 37021c5ecae3SMike Kravetz */ 37031c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 37041c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 37057251ff78SAdam Litke } 3706e9fe92aeSMina Almasry 3707e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3708a1e78772SMel Gorman } 3709a1e78772SMel Gorman 371031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 371131383c68SDan Williams { 371231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 371331383c68SDan Williams return -EINVAL; 371431383c68SDan Williams return 0; 371531383c68SDan Williams } 371631383c68SDan Williams 371705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 371805ea8860SDan Williams { 371905ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 372005ea8860SDan Williams 372105ea8860SDan Williams return 1UL << huge_page_shift(hstate); 372205ea8860SDan Williams } 372305ea8860SDan Williams 37241da177e4SLinus Torvalds /* 37251da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 37261da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 37271da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 37281da177e4SLinus Torvalds * this far. 37291da177e4SLinus Torvalds */ 3730b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 37311da177e4SLinus Torvalds { 37321da177e4SLinus Torvalds BUG(); 3733d0217ac0SNick Piggin return 0; 37341da177e4SLinus Torvalds } 37351da177e4SLinus Torvalds 3736eec3636aSJane Chu /* 3737eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3738eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3739eec3636aSJane Chu * This is because under System V memory model, mappings created via 3740eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3741eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3742eec3636aSJane Chu */ 3743f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3744d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 374584afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3746a1e78772SMel Gorman .close = hugetlb_vm_op_close, 374731383c68SDan Williams .split = hugetlb_vm_op_split, 374805ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 37491da177e4SLinus Torvalds }; 37501da177e4SLinus Torvalds 37511e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 37521e8f889bSDavid Gibson int writable) 375363551ae0SDavid Gibson { 375463551ae0SDavid Gibson pte_t entry; 375563551ae0SDavid Gibson 37561e8f889bSDavid Gibson if (writable) { 3757106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3758106c992aSGerald Schaefer vma->vm_page_prot))); 375963551ae0SDavid Gibson } else { 3760106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3761106c992aSGerald Schaefer vma->vm_page_prot)); 376263551ae0SDavid Gibson } 376363551ae0SDavid Gibson entry = pte_mkyoung(entry); 376463551ae0SDavid Gibson entry = pte_mkhuge(entry); 3765d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 376663551ae0SDavid Gibson 376763551ae0SDavid Gibson return entry; 376863551ae0SDavid Gibson } 376963551ae0SDavid Gibson 37701e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37711e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37721e8f889bSDavid Gibson { 37731e8f889bSDavid Gibson pte_t entry; 37741e8f889bSDavid Gibson 3775106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 377632f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37774b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37781e8f889bSDavid Gibson } 37791e8f889bSDavid Gibson 3780d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37814a705fefSNaoya Horiguchi { 37824a705fefSNaoya Horiguchi swp_entry_t swp; 37834a705fefSNaoya Horiguchi 37844a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3785d5ed7444SAneesh Kumar K.V return false; 37864a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 37874a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3788d5ed7444SAneesh Kumar K.V return true; 37894a705fefSNaoya Horiguchi else 3790d5ed7444SAneesh Kumar K.V return false; 37914a705fefSNaoya Horiguchi } 37924a705fefSNaoya Horiguchi 37934a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 37944a705fefSNaoya Horiguchi { 37954a705fefSNaoya Horiguchi swp_entry_t swp; 37964a705fefSNaoya Horiguchi 37974a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37984a705fefSNaoya Horiguchi return 0; 37994a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 38004a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 38014a705fefSNaoya Horiguchi return 1; 38024a705fefSNaoya Horiguchi else 38034a705fefSNaoya Horiguchi return 0; 38044a705fefSNaoya Horiguchi } 38051e8f889bSDavid Gibson 380663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 380763551ae0SDavid Gibson struct vm_area_struct *vma) 380863551ae0SDavid Gibson { 38095e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 381063551ae0SDavid Gibson struct page *ptepage; 38111c59827dSHugh Dickins unsigned long addr; 38121e8f889bSDavid Gibson int cow; 3813a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3814a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3815c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3816ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3817e8569dd2SAndreas Sandberg int ret = 0; 38181e8f889bSDavid Gibson 38191e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 382063551ae0SDavid Gibson 3821ac46d4f3SJérôme Glisse if (cow) { 38227269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 38236f4f13e8SJérôme Glisse vma->vm_start, 3824ac46d4f3SJérôme Glisse vma->vm_end); 3825ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3826c0d0381aSMike Kravetz } else { 3827c0d0381aSMike Kravetz /* 3828c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3829c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3830c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3831c0d0381aSMike Kravetz * huge_pmd_unshare. 3832c0d0381aSMike Kravetz */ 3833c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3834ac46d4f3SJérôme Glisse } 3835e8569dd2SAndreas Sandberg 3836a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3837cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 38387868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3839c74df32cSHugh Dickins if (!src_pte) 3840c74df32cSHugh Dickins continue; 3841a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3842e8569dd2SAndreas Sandberg if (!dst_pte) { 3843e8569dd2SAndreas Sandberg ret = -ENOMEM; 3844e8569dd2SAndreas Sandberg break; 3845e8569dd2SAndreas Sandberg } 3846c5c99429SLarry Woodman 38475e41540cSMike Kravetz /* 38485e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38495e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38505e41540cSMike Kravetz * 38515e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38525e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38535e41540cSMike Kravetz * Check here before taking page table lock, and once again 38545e41540cSMike Kravetz * after taking the lock below. 38555e41540cSMike Kravetz */ 38565e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38575e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3858c5c99429SLarry Woodman continue; 3859c5c99429SLarry Woodman 3860cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3861cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3862cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38634a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38645e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38655e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38665e41540cSMike Kravetz /* 38675e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38685e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38695e41540cSMike Kravetz * sharing with another vma. 38705e41540cSMike Kravetz */ 38714a705fefSNaoya Horiguchi ; 38724a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38734a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38744a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38754a705fefSNaoya Horiguchi 38764a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38774a705fefSNaoya Horiguchi /* 38784a705fefSNaoya Horiguchi * COW mappings require pages in both 38794a705fefSNaoya Horiguchi * parent and child to be set to read. 38804a705fefSNaoya Horiguchi */ 38814a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38824a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3883e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3884e5251fd4SPunit Agrawal entry, sz); 38854a705fefSNaoya Horiguchi } 3886e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38874a705fefSNaoya Horiguchi } else { 388834ee645eSJoerg Roedel if (cow) { 38890f10851eSJérôme Glisse /* 38900f10851eSJérôme Glisse * No need to notify as we are downgrading page 38910f10851eSJérôme Glisse * table protection not changing it to point 38920f10851eSJérôme Glisse * to a new page. 38930f10851eSJérôme Glisse * 3894ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 38950f10851eSJérôme Glisse */ 38967f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 389734ee645eSJoerg Roedel } 38980253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 389963551ae0SDavid Gibson ptepage = pte_page(entry); 390063551ae0SDavid Gibson get_page(ptepage); 390153f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 390263551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 39035d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 39041c59827dSHugh Dickins } 3905cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3906cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 390763551ae0SDavid Gibson } 390863551ae0SDavid Gibson 3909e8569dd2SAndreas Sandberg if (cow) 3910ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3911c0d0381aSMike Kravetz else 3912c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3913e8569dd2SAndreas Sandberg 3914e8569dd2SAndreas Sandberg return ret; 391563551ae0SDavid Gibson } 391663551ae0SDavid Gibson 391724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 391824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 391924669e58SAneesh Kumar K.V struct page *ref_page) 392063551ae0SDavid Gibson { 392163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 392263551ae0SDavid Gibson unsigned long address; 3923c7546f8fSDavid Gibson pte_t *ptep; 392463551ae0SDavid Gibson pte_t pte; 3925cb900f41SKirill A. Shutemov spinlock_t *ptl; 392663551ae0SDavid Gibson struct page *page; 3927a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3928a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3929ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3930a5516438SAndi Kleen 393163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3932a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3933a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 393463551ae0SDavid Gibson 393507e32661SAneesh Kumar K.V /* 393607e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 393707e32661SAneesh Kumar K.V * to huge page. 393807e32661SAneesh Kumar K.V */ 3939ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 394024669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3941dff11abeSMike Kravetz 3942dff11abeSMike Kravetz /* 3943dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3944dff11abeSMike Kravetz */ 39456f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39466f4f13e8SJérôme Glisse end); 3947ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3948ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3949569f48b8SHillf Danton address = start; 3950569f48b8SHillf Danton for (; address < end; address += sz) { 39517868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3952c7546f8fSDavid Gibson if (!ptep) 3953c7546f8fSDavid Gibson continue; 3954c7546f8fSDavid Gibson 3955cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 395631d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 395731d49da5SAneesh Kumar K.V spin_unlock(ptl); 3958dff11abeSMike Kravetz /* 3959dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3960dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3961dff11abeSMike Kravetz */ 396231d49da5SAneesh Kumar K.V continue; 396331d49da5SAneesh Kumar K.V } 396439dde65cSChen, Kenneth W 39656629326bSHillf Danton pte = huge_ptep_get(ptep); 396631d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 396731d49da5SAneesh Kumar K.V spin_unlock(ptl); 396831d49da5SAneesh Kumar K.V continue; 396931d49da5SAneesh Kumar K.V } 39706629326bSHillf Danton 39716629326bSHillf Danton /* 39729fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39739fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39746629326bSHillf Danton */ 39759fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39769386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 397731d49da5SAneesh Kumar K.V spin_unlock(ptl); 397831d49da5SAneesh Kumar K.V continue; 39798c4894c6SNaoya Horiguchi } 39806629326bSHillf Danton 39816629326bSHillf Danton page = pte_page(pte); 398204f2cbe3SMel Gorman /* 398304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 398404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 398504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 398604f2cbe3SMel Gorman */ 398704f2cbe3SMel Gorman if (ref_page) { 398831d49da5SAneesh Kumar K.V if (page != ref_page) { 398931d49da5SAneesh Kumar K.V spin_unlock(ptl); 399031d49da5SAneesh Kumar K.V continue; 399131d49da5SAneesh Kumar K.V } 399204f2cbe3SMel Gorman /* 399304f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 399404f2cbe3SMel Gorman * future faults in this VMA will fail rather than 399504f2cbe3SMel Gorman * looking like data was lost 399604f2cbe3SMel Gorman */ 399704f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 399804f2cbe3SMel Gorman } 399904f2cbe3SMel Gorman 4000c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4001b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4002106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 40036649a386SKen Chen set_page_dirty(page); 40049e81130bSHillf Danton 40055d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4006d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 400731d49da5SAneesh Kumar K.V 4008cb900f41SKirill A. Shutemov spin_unlock(ptl); 4009e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 401024669e58SAneesh Kumar K.V /* 401131d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 401224669e58SAneesh Kumar K.V */ 401331d49da5SAneesh Kumar K.V if (ref_page) 401431d49da5SAneesh Kumar K.V break; 4015fe1668aeSChen, Kenneth W } 4016ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 401724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 40181da177e4SLinus Torvalds } 401963551ae0SDavid Gibson 4020d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4021d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4022d833352aSMel Gorman unsigned long end, struct page *ref_page) 4023d833352aSMel Gorman { 4024d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4025d833352aSMel Gorman 4026d833352aSMel Gorman /* 4027d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4028d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4029d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4030d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4031d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4032c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4033d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4034c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4035d833352aSMel Gorman */ 4036d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4037d833352aSMel Gorman } 4038d833352aSMel Gorman 4039502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 404004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4041502717f4SChen, Kenneth W { 404224669e58SAneesh Kumar K.V struct mm_struct *mm; 404324669e58SAneesh Kumar K.V struct mmu_gather tlb; 4044dff11abeSMike Kravetz unsigned long tlb_start = start; 4045dff11abeSMike Kravetz unsigned long tlb_end = end; 4046dff11abeSMike Kravetz 4047dff11abeSMike Kravetz /* 4048dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 4049dff11abeSMike Kravetz * start/end for worst case tlb flushing. 4050dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 4051dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 4052dff11abeSMike Kravetz * the largest possible range. 4053dff11abeSMike Kravetz */ 4054dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 405524669e58SAneesh Kumar K.V 405624669e58SAneesh Kumar K.V mm = vma->vm_mm; 405724669e58SAneesh Kumar K.V 4058dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 405924669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4060dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 4061502717f4SChen, Kenneth W } 4062502717f4SChen, Kenneth W 406304f2cbe3SMel Gorman /* 406404f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 406504f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 406604f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 406704f2cbe3SMel Gorman * same region. 406804f2cbe3SMel Gorman */ 40692f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40702a4b3dedSHarvey Harrison struct page *page, unsigned long address) 407104f2cbe3SMel Gorman { 40727526674dSAdam Litke struct hstate *h = hstate_vma(vma); 407304f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 407404f2cbe3SMel Gorman struct address_space *mapping; 407504f2cbe3SMel Gorman pgoff_t pgoff; 407604f2cbe3SMel Gorman 407704f2cbe3SMel Gorman /* 407804f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 407904f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 408004f2cbe3SMel Gorman */ 40817526674dSAdam Litke address = address & huge_page_mask(h); 408236e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 408336e4f20aSMichal Hocko vma->vm_pgoff; 408493c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 408504f2cbe3SMel Gorman 40864eb2b1dcSMel Gorman /* 40874eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40884eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40894eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40904eb2b1dcSMel Gorman */ 409183cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40926b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 409304f2cbe3SMel Gorman /* Do not unmap the current VMA */ 409404f2cbe3SMel Gorman if (iter_vma == vma) 409504f2cbe3SMel Gorman continue; 409604f2cbe3SMel Gorman 409704f2cbe3SMel Gorman /* 40982f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40992f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 41002f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 41012f84a899SMel Gorman */ 41022f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 41032f84a899SMel Gorman continue; 41042f84a899SMel Gorman 41052f84a899SMel Gorman /* 410604f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 410704f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 410804f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 410904f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 411004f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 411104f2cbe3SMel Gorman */ 411204f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 411324669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 411424669e58SAneesh Kumar K.V address + huge_page_size(h), page); 411504f2cbe3SMel Gorman } 411683cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 411704f2cbe3SMel Gorman } 411804f2cbe3SMel Gorman 41190fe6e20bSNaoya Horiguchi /* 41200fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4121ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4122ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4123ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 41240fe6e20bSNaoya Horiguchi */ 41252b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4126974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4127cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 41281e8f889bSDavid Gibson { 41293999f52eSAneesh Kumar K.V pte_t pte; 4130a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41311e8f889bSDavid Gibson struct page *old_page, *new_page; 41322b740303SSouptick Joarder int outside_reserve = 0; 41332b740303SSouptick Joarder vm_fault_t ret = 0; 4134974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4135ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 41361e8f889bSDavid Gibson 41373999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 41381e8f889bSDavid Gibson old_page = pte_page(pte); 41391e8f889bSDavid Gibson 414004f2cbe3SMel Gorman retry_avoidcopy: 41411e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 41421e8f889bSDavid Gibson * and just make the page writable */ 414337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 41445a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 41455b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 414683c54070SNick Piggin return 0; 41471e8f889bSDavid Gibson } 41481e8f889bSDavid Gibson 414904f2cbe3SMel Gorman /* 415004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 415104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 415204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 415304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 415404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 415504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 415604f2cbe3SMel Gorman * of the full address range. 415704f2cbe3SMel Gorman */ 41585944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 415904f2cbe3SMel Gorman old_page != pagecache_page) 416004f2cbe3SMel Gorman outside_reserve = 1; 416104f2cbe3SMel Gorman 416209cbfeafSKirill A. Shutemov get_page(old_page); 4163b76c8cfbSLarry Woodman 4164ad4404a2SDavidlohr Bueso /* 4165ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4166ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4167ad4404a2SDavidlohr Bueso */ 4168cb900f41SKirill A. Shutemov spin_unlock(ptl); 41695b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41701e8f889bSDavid Gibson 41712fc39cecSAdam Litke if (IS_ERR(new_page)) { 417204f2cbe3SMel Gorman /* 417304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 417404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 417504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 417604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 417704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 417804f2cbe3SMel Gorman */ 417904f2cbe3SMel Gorman if (outside_reserve) { 418009cbfeafSKirill A. Shutemov put_page(old_page); 418104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 41825b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 418304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4184cb900f41SKirill A. Shutemov spin_lock(ptl); 41855b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4186a9af0c5dSNaoya Horiguchi if (likely(ptep && 4187a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 418804f2cbe3SMel Gorman goto retry_avoidcopy; 4189a734bcc8SHillf Danton /* 4190cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4191cb900f41SKirill A. Shutemov * lock, and our job is done. 4192a734bcc8SHillf Danton */ 4193a734bcc8SHillf Danton return 0; 419404f2cbe3SMel Gorman } 419504f2cbe3SMel Gorman 41962b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4197ad4404a2SDavidlohr Bueso goto out_release_old; 41981e8f889bSDavid Gibson } 41991e8f889bSDavid Gibson 42000fe6e20bSNaoya Horiguchi /* 42010fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 42020fe6e20bSNaoya Horiguchi * anon_vma prepared. 42030fe6e20bSNaoya Horiguchi */ 420444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4205ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4206ad4404a2SDavidlohr Bueso goto out_release_all; 420744e2aa93SDean Nelson } 42080fe6e20bSNaoya Horiguchi 4209974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 421047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 42110ed361deSNick Piggin __SetPageUptodate(new_page); 42121e8f889bSDavid Gibson 42137269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 42146f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4215ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4216ad4404a2SDavidlohr Bueso 4217b76c8cfbSLarry Woodman /* 4218cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4219b76c8cfbSLarry Woodman * before the page tables are altered 4220b76c8cfbSLarry Woodman */ 4221cb900f41SKirill A. Shutemov spin_lock(ptl); 42225b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4223a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 422407443a85SJoonsoo Kim ClearPagePrivate(new_page); 422507443a85SJoonsoo Kim 42261e8f889bSDavid Gibson /* Break COW */ 42275b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4228ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 42295b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 42301e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4231d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 42325b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 4233cb6acd01SMike Kravetz set_page_huge_active(new_page); 42341e8f889bSDavid Gibson /* Make the old page be freed below */ 42351e8f889bSDavid Gibson new_page = old_page; 42361e8f889bSDavid Gibson } 4237cb900f41SKirill A. Shutemov spin_unlock(ptl); 4238ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4239ad4404a2SDavidlohr Bueso out_release_all: 42405b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 424109cbfeafSKirill A. Shutemov put_page(new_page); 4242ad4404a2SDavidlohr Bueso out_release_old: 424309cbfeafSKirill A. Shutemov put_page(old_page); 42448312034fSJoonsoo Kim 4245ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4246ad4404a2SDavidlohr Bueso return ret; 42471e8f889bSDavid Gibson } 42481e8f889bSDavid Gibson 424904f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4250a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4251a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 425204f2cbe3SMel Gorman { 425304f2cbe3SMel Gorman struct address_space *mapping; 4254e7c4b0bfSAndy Whitcroft pgoff_t idx; 425504f2cbe3SMel Gorman 425604f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4257a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 425804f2cbe3SMel Gorman 425904f2cbe3SMel Gorman return find_lock_page(mapping, idx); 426004f2cbe3SMel Gorman } 426104f2cbe3SMel Gorman 42623ae77f43SHugh Dickins /* 42633ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42643ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42653ae77f43SHugh Dickins */ 42663ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42672a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42682a15efc9SHugh Dickins { 42692a15efc9SHugh Dickins struct address_space *mapping; 42702a15efc9SHugh Dickins pgoff_t idx; 42712a15efc9SHugh Dickins struct page *page; 42722a15efc9SHugh Dickins 42732a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42742a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42752a15efc9SHugh Dickins 42762a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42772a15efc9SHugh Dickins if (page) 42782a15efc9SHugh Dickins put_page(page); 42792a15efc9SHugh Dickins return page != NULL; 42802a15efc9SHugh Dickins } 42812a15efc9SHugh Dickins 4282ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4283ab76ad54SMike Kravetz pgoff_t idx) 4284ab76ad54SMike Kravetz { 4285ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4286ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4287ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4288ab76ad54SMike Kravetz 4289ab76ad54SMike Kravetz if (err) 4290ab76ad54SMike Kravetz return err; 4291ab76ad54SMike Kravetz ClearPagePrivate(page); 4292ab76ad54SMike Kravetz 429322146c3cSMike Kravetz /* 429422146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 429522146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 429622146c3cSMike Kravetz */ 429722146c3cSMike Kravetz set_page_dirty(page); 429822146c3cSMike Kravetz 4299ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4300ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4301ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4302ab76ad54SMike Kravetz return 0; 4303ab76ad54SMike Kravetz } 4304ab76ad54SMike Kravetz 43052b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 43062b740303SSouptick Joarder struct vm_area_struct *vma, 43078382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4308788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4309ac9b9c66SHugh Dickins { 4310a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43112b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4312409eb8c2SHillf Danton int anon_rmap = 0; 43134c887265SAdam Litke unsigned long size; 43144c887265SAdam Litke struct page *page; 43151e8f889bSDavid Gibson pte_t new_pte; 4316cb900f41SKirill A. Shutemov spinlock_t *ptl; 4317285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4318cb6acd01SMike Kravetz bool new_page = false; 43194c887265SAdam Litke 432004f2cbe3SMel Gorman /* 432104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 432204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 432325985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 432404f2cbe3SMel Gorman */ 432504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4326910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 432704f2cbe3SMel Gorman current->pid); 432804f2cbe3SMel Gorman return ret; 432904f2cbe3SMel Gorman } 433004f2cbe3SMel Gorman 43314c887265SAdam Litke /* 433287bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 433387bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 433487bf91d3SMike Kravetz * once for faults beyond end of file. 43354c887265SAdam Litke */ 4336a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4337ebed4bfcSHugh Dickins if (idx >= size) 4338ebed4bfcSHugh Dickins goto out; 43391a1aad8aSMike Kravetz 434087bf91d3SMike Kravetz retry: 434187bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 434287bf91d3SMike Kravetz if (!page) { 43431a1aad8aSMike Kravetz /* 43441a1aad8aSMike Kravetz * Check for page in userfault range 43451a1aad8aSMike Kravetz */ 43461a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43471a1aad8aSMike Kravetz u32 hash; 43481a1aad8aSMike Kravetz struct vm_fault vmf = { 43491a1aad8aSMike Kravetz .vma = vma, 4350285b8dcaSHuang Ying .address = haddr, 43511a1aad8aSMike Kravetz .flags = flags, 43521a1aad8aSMike Kravetz /* 43531a1aad8aSMike Kravetz * Hard to debug if it ends up being 43541a1aad8aSMike Kravetz * used by a callee that assumes 43551a1aad8aSMike Kravetz * something about the other 43561a1aad8aSMike Kravetz * uninitialized fields... same as in 43571a1aad8aSMike Kravetz * memory.c 43581a1aad8aSMike Kravetz */ 43591a1aad8aSMike Kravetz }; 43601a1aad8aSMike Kravetz 43611a1aad8aSMike Kravetz /* 4362c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4363c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4364c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43651a1aad8aSMike Kravetz */ 4366188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43671a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4368c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43691a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4370c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 43711a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43721a1aad8aSMike Kravetz goto out; 43731a1aad8aSMike Kravetz } 43741a1aad8aSMike Kravetz 4375285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43762fc39cecSAdam Litke if (IS_ERR(page)) { 43774643d67eSMike Kravetz /* 43784643d67eSMike Kravetz * Returning error will result in faulting task being 43794643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43804643d67eSMike Kravetz * tasks from racing to fault in the same page which 43814643d67eSMike Kravetz * could result in false unable to allocate errors. 43824643d67eSMike Kravetz * Page migration does not take the fault mutex, but 43834643d67eSMike Kravetz * does a clear then write of pte's under page table 43844643d67eSMike Kravetz * lock. Page fault code could race with migration, 43854643d67eSMike Kravetz * notice the clear pte and try to allocate a page 43864643d67eSMike Kravetz * here. Before returning error, get ptl and make 43874643d67eSMike Kravetz * sure there really is no pte entry. 43884643d67eSMike Kravetz */ 43894643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 43904643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 43914643d67eSMike Kravetz ret = 0; 43924643d67eSMike Kravetz spin_unlock(ptl); 43934643d67eSMike Kravetz goto out; 43944643d67eSMike Kravetz } 43954643d67eSMike Kravetz spin_unlock(ptl); 43962b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 43976bda666aSChristoph Lameter goto out; 43986bda666aSChristoph Lameter } 439947ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 44000ed361deSNick Piggin __SetPageUptodate(page); 4401cb6acd01SMike Kravetz new_page = true; 4402ac9b9c66SHugh Dickins 4403f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4404ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 44056bda666aSChristoph Lameter if (err) { 44066bda666aSChristoph Lameter put_page(page); 44076bda666aSChristoph Lameter if (err == -EEXIST) 44086bda666aSChristoph Lameter goto retry; 44096bda666aSChristoph Lameter goto out; 44106bda666aSChristoph Lameter } 441123be7468SMel Gorman } else { 44126bda666aSChristoph Lameter lock_page(page); 44130fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 44140fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 44150fe6e20bSNaoya Horiguchi goto backout_unlocked; 441623be7468SMel Gorman } 4417409eb8c2SHillf Danton anon_rmap = 1; 44180fe6e20bSNaoya Horiguchi } 44190fe6e20bSNaoya Horiguchi } else { 442057303d80SAndy Whitcroft /* 4421998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4422998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4423998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4424fd6a03edSNaoya Horiguchi */ 4425fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 4426aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 4427972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4428fd6a03edSNaoya Horiguchi goto backout_unlocked; 44296bda666aSChristoph Lameter } 4430998b4382SNaoya Horiguchi } 44311e8f889bSDavid Gibson 443257303d80SAndy Whitcroft /* 443357303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 443457303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 443557303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 443657303d80SAndy Whitcroft * the spinlock. 443757303d80SAndy Whitcroft */ 44385e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4439285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44402b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44412b26736cSAndy Whitcroft goto backout_unlocked; 44422b26736cSAndy Whitcroft } 44435e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4444285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44455e911373SMike Kravetz } 444657303d80SAndy Whitcroft 44478bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 444883c54070SNick Piggin ret = 0; 44497f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44504c887265SAdam Litke goto backout; 44514c887265SAdam Litke 445207443a85SJoonsoo Kim if (anon_rmap) { 445307443a85SJoonsoo Kim ClearPagePrivate(page); 4454285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4455ac714904SChoi Gi-yong } else 445653f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44571e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44581e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4459285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44601e8f889bSDavid Gibson 44615d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4462788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44631e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4464974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44651e8f889bSDavid Gibson } 44661e8f889bSDavid Gibson 4467cb900f41SKirill A. Shutemov spin_unlock(ptl); 4468cb6acd01SMike Kravetz 4469cb6acd01SMike Kravetz /* 4470cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 4471cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 4472cb6acd01SMike Kravetz * isolated for migration. 4473cb6acd01SMike Kravetz */ 4474cb6acd01SMike Kravetz if (new_page) 4475cb6acd01SMike Kravetz set_page_huge_active(page); 4476cb6acd01SMike Kravetz 44774c887265SAdam Litke unlock_page(page); 44784c887265SAdam Litke out: 4479ac9b9c66SHugh Dickins return ret; 44804c887265SAdam Litke 44814c887265SAdam Litke backout: 4482cb900f41SKirill A. Shutemov spin_unlock(ptl); 44832b26736cSAndy Whitcroft backout_unlocked: 44844c887265SAdam Litke unlock_page(page); 4485285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 44864c887265SAdam Litke put_page(page); 44874c887265SAdam Litke goto out; 4488ac9b9c66SHugh Dickins } 4489ac9b9c66SHugh Dickins 44908382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4491188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44928382d914SDavidlohr Bueso { 44938382d914SDavidlohr Bueso unsigned long key[2]; 44948382d914SDavidlohr Bueso u32 hash; 44958382d914SDavidlohr Bueso 44968382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 44978382d914SDavidlohr Bueso key[1] = idx; 44988382d914SDavidlohr Bueso 449955254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 45008382d914SDavidlohr Bueso 45018382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 45028382d914SDavidlohr Bueso } 45038382d914SDavidlohr Bueso #else 45048382d914SDavidlohr Bueso /* 45058382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 45068382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 45078382d914SDavidlohr Bueso */ 4508188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45098382d914SDavidlohr Bueso { 45108382d914SDavidlohr Bueso return 0; 45118382d914SDavidlohr Bueso } 45128382d914SDavidlohr Bueso #endif 45138382d914SDavidlohr Bueso 45142b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4515788c7df4SHugh Dickins unsigned long address, unsigned int flags) 451686e5216fSAdam Litke { 45178382d914SDavidlohr Bueso pte_t *ptep, entry; 4518cb900f41SKirill A. Shutemov spinlock_t *ptl; 45192b740303SSouptick Joarder vm_fault_t ret; 45208382d914SDavidlohr Bueso u32 hash; 45218382d914SDavidlohr Bueso pgoff_t idx; 45220fe6e20bSNaoya Horiguchi struct page *page = NULL; 452357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4524a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45258382d914SDavidlohr Bueso struct address_space *mapping; 45260f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4527285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 452886e5216fSAdam Litke 4529285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4530fd6a03edSNaoya Horiguchi if (ptep) { 4531c0d0381aSMike Kravetz /* 4532c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4533c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4534c0d0381aSMike Kravetz * not actually modifying content here. 4535c0d0381aSMike Kravetz */ 4536fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4537290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4538cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4539290408d4SNaoya Horiguchi return 0; 4540290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4541aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4542972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4543ddeaab32SMike Kravetz } else { 4544b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4545ddeaab32SMike Kravetz if (!ptep) 4546b43a9990SMike Kravetz return VM_FAULT_OOM; 4547b43a9990SMike Kravetz } 45488382d914SDavidlohr Bueso 4549c0d0381aSMike Kravetz /* 4550c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 455187bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 455287bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 455387bf91d3SMike Kravetz * and making the ptep no longer valid. 455487bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4555c0d0381aSMike Kravetz * 4556c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4557c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4558c0d0381aSMike Kravetz * something has changed. 4559c0d0381aSMike Kravetz */ 4560ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4561c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4562c0d0381aSMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4563c0d0381aSMike Kravetz if (!ptep) { 4564c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4565c0d0381aSMike Kravetz return VM_FAULT_OOM; 4566c0d0381aSMike Kravetz } 4567ddeaab32SMike Kravetz 45683935baa9SDavid Gibson /* 45693935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45703935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45713935baa9SDavid Gibson * the same page in the page cache. 45723935baa9SDavid Gibson */ 4573c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4574188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4575c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45768382d914SDavidlohr Bueso 45777f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45787f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45798382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4580b4d1d99fSDavid Gibson goto out_mutex; 45813935baa9SDavid Gibson } 458286e5216fSAdam Litke 458383c54070SNick Piggin ret = 0; 45841e8f889bSDavid Gibson 458557303d80SAndy Whitcroft /* 45860f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45870f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45887c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45897c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 45907c8de358SEthon Paul * properly handle it. 45910f792cf9SNaoya Horiguchi */ 45920f792cf9SNaoya Horiguchi if (!pte_present(entry)) 45930f792cf9SNaoya Horiguchi goto out_mutex; 45940f792cf9SNaoya Horiguchi 45950f792cf9SNaoya Horiguchi /* 459657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 459757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 459857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 459957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 460057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 460157303d80SAndy Whitcroft * consumed. 460257303d80SAndy Whitcroft */ 4603106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4604285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 46052b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4606b4d1d99fSDavid Gibson goto out_mutex; 46072b26736cSAndy Whitcroft } 46085e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4609285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 461057303d80SAndy Whitcroft 4611f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 461257303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4613285b8dcaSHuang Ying vma, haddr); 461457303d80SAndy Whitcroft } 461557303d80SAndy Whitcroft 46160f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 46170fe6e20bSNaoya Horiguchi 46181e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4619b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4620cb900f41SKirill A. Shutemov goto out_ptl; 4621b4d1d99fSDavid Gibson 46220f792cf9SNaoya Horiguchi /* 46230f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 46240f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 46250f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 46260f792cf9SNaoya Horiguchi */ 46270f792cf9SNaoya Horiguchi page = pte_page(entry); 46280f792cf9SNaoya Horiguchi if (page != pagecache_page) 46290f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 46300f792cf9SNaoya Horiguchi need_wait_lock = 1; 46310f792cf9SNaoya Horiguchi goto out_ptl; 46320f792cf9SNaoya Horiguchi } 46330f792cf9SNaoya Horiguchi 46340f792cf9SNaoya Horiguchi get_page(page); 4635b4d1d99fSDavid Gibson 4636788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4637106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4638974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4639cb900f41SKirill A. Shutemov pagecache_page, ptl); 46400f792cf9SNaoya Horiguchi goto out_put_page; 4641b4d1d99fSDavid Gibson } 4642106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4643b4d1d99fSDavid Gibson } 4644b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4645285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4646788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4647285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46480f792cf9SNaoya Horiguchi out_put_page: 46490f792cf9SNaoya Horiguchi if (page != pagecache_page) 46500f792cf9SNaoya Horiguchi unlock_page(page); 46510f792cf9SNaoya Horiguchi put_page(page); 4652cb900f41SKirill A. Shutemov out_ptl: 4653cb900f41SKirill A. Shutemov spin_unlock(ptl); 465457303d80SAndy Whitcroft 465557303d80SAndy Whitcroft if (pagecache_page) { 465657303d80SAndy Whitcroft unlock_page(pagecache_page); 465757303d80SAndy Whitcroft put_page(pagecache_page); 465857303d80SAndy Whitcroft } 4659b4d1d99fSDavid Gibson out_mutex: 4660c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4661c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46620f792cf9SNaoya Horiguchi /* 46630f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46640f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46650f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46660f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46670f792cf9SNaoya Horiguchi * here without taking refcount. 46680f792cf9SNaoya Horiguchi */ 46690f792cf9SNaoya Horiguchi if (need_wait_lock) 46700f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46711e8f889bSDavid Gibson return ret; 467286e5216fSAdam Litke } 467386e5216fSAdam Litke 46748fb5debcSMike Kravetz /* 46758fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46768fb5debcSMike Kravetz * modifications for huge pages. 46778fb5debcSMike Kravetz */ 46788fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46798fb5debcSMike Kravetz pte_t *dst_pte, 46808fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46818fb5debcSMike Kravetz unsigned long dst_addr, 46828fb5debcSMike Kravetz unsigned long src_addr, 46838fb5debcSMike Kravetz struct page **pagep) 46848fb5debcSMike Kravetz { 46851e392147SAndrea Arcangeli struct address_space *mapping; 46861e392147SAndrea Arcangeli pgoff_t idx; 46871e392147SAndrea Arcangeli unsigned long size; 46881c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46898fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 46908fb5debcSMike Kravetz pte_t _dst_pte; 46918fb5debcSMike Kravetz spinlock_t *ptl; 46928fb5debcSMike Kravetz int ret; 46938fb5debcSMike Kravetz struct page *page; 46948fb5debcSMike Kravetz 46958fb5debcSMike Kravetz if (!*pagep) { 46968fb5debcSMike Kravetz ret = -ENOMEM; 46978fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 46988fb5debcSMike Kravetz if (IS_ERR(page)) 46998fb5debcSMike Kravetz goto out; 47008fb5debcSMike Kravetz 47018fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 47028fb5debcSMike Kravetz (const void __user *) src_addr, 4703810a56b9SMike Kravetz pages_per_huge_page(h), false); 47048fb5debcSMike Kravetz 4705c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 47068fb5debcSMike Kravetz if (unlikely(ret)) { 47079e368259SAndrea Arcangeli ret = -ENOENT; 47088fb5debcSMike Kravetz *pagep = page; 47098fb5debcSMike Kravetz /* don't free the page */ 47108fb5debcSMike Kravetz goto out; 47118fb5debcSMike Kravetz } 47128fb5debcSMike Kravetz } else { 47138fb5debcSMike Kravetz page = *pagep; 47148fb5debcSMike Kravetz *pagep = NULL; 47158fb5debcSMike Kravetz } 47168fb5debcSMike Kravetz 47178fb5debcSMike Kravetz /* 47188fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 47198fb5debcSMike Kravetz * preceding stores to the page contents become visible before 47208fb5debcSMike Kravetz * the set_pte_at() write. 47218fb5debcSMike Kravetz */ 47228fb5debcSMike Kravetz __SetPageUptodate(page); 47238fb5debcSMike Kravetz 47241e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 47251e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 47261e392147SAndrea Arcangeli 47271c9e8defSMike Kravetz /* 47281c9e8defSMike Kravetz * If shared, add to page cache 47291c9e8defSMike Kravetz */ 47301c9e8defSMike Kravetz if (vm_shared) { 47311e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47321e392147SAndrea Arcangeli ret = -EFAULT; 47331e392147SAndrea Arcangeli if (idx >= size) 47341e392147SAndrea Arcangeli goto out_release_nounlock; 47351c9e8defSMike Kravetz 47361e392147SAndrea Arcangeli /* 47371e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 47381e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 47391e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47401e392147SAndrea Arcangeli * the caller. 47411e392147SAndrea Arcangeli */ 47421c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47431c9e8defSMike Kravetz if (ret) 47441c9e8defSMike Kravetz goto out_release_nounlock; 47451c9e8defSMike Kravetz } 47461c9e8defSMike Kravetz 47478fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47488fb5debcSMike Kravetz spin_lock(ptl); 47498fb5debcSMike Kravetz 47501e392147SAndrea Arcangeli /* 47511e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47521e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47531e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47541e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47551e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47561e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47571e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47581e392147SAndrea Arcangeli */ 47591e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47601e392147SAndrea Arcangeli ret = -EFAULT; 47611e392147SAndrea Arcangeli if (idx >= size) 47621e392147SAndrea Arcangeli goto out_release_unlock; 47631e392147SAndrea Arcangeli 47648fb5debcSMike Kravetz ret = -EEXIST; 47658fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47668fb5debcSMike Kravetz goto out_release_unlock; 47678fb5debcSMike Kravetz 47681c9e8defSMike Kravetz if (vm_shared) { 47691c9e8defSMike Kravetz page_dup_rmap(page, true); 47701c9e8defSMike Kravetz } else { 47718fb5debcSMike Kravetz ClearPagePrivate(page); 47728fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47731c9e8defSMike Kravetz } 47748fb5debcSMike Kravetz 47758fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47768fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47778fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47788fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47798fb5debcSMike Kravetz 47808fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47818fb5debcSMike Kravetz 47828fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47838fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47848fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47858fb5debcSMike Kravetz 47868fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47878fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47888fb5debcSMike Kravetz 47898fb5debcSMike Kravetz spin_unlock(ptl); 4790cb6acd01SMike Kravetz set_page_huge_active(page); 47911c9e8defSMike Kravetz if (vm_shared) 47921c9e8defSMike Kravetz unlock_page(page); 47938fb5debcSMike Kravetz ret = 0; 47948fb5debcSMike Kravetz out: 47958fb5debcSMike Kravetz return ret; 47968fb5debcSMike Kravetz out_release_unlock: 47978fb5debcSMike Kravetz spin_unlock(ptl); 47981c9e8defSMike Kravetz if (vm_shared) 47991c9e8defSMike Kravetz unlock_page(page); 48005af10dfdSAndrea Arcangeli out_release_nounlock: 48018fb5debcSMike Kravetz put_page(page); 48028fb5debcSMike Kravetz goto out; 48038fb5debcSMike Kravetz } 48048fb5debcSMike Kravetz 480528a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 480663551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 480728a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 48084f6da934SPeter Xu long i, unsigned int flags, int *locked) 480963551ae0SDavid Gibson { 4810d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4811d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 481228a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4813a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48142be7cfedSDaniel Jordan int err = -EFAULT; 481563551ae0SDavid Gibson 481663551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 481763551ae0SDavid Gibson pte_t *pte; 4818cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 48192a15efc9SHugh Dickins int absent; 482063551ae0SDavid Gibson struct page *page; 482163551ae0SDavid Gibson 48224c887265SAdam Litke /* 482302057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 482402057967SDavid Rientjes * potentially allocating memory. 482502057967SDavid Rientjes */ 4826fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 482702057967SDavid Rientjes remainder = 0; 482802057967SDavid Rientjes break; 482902057967SDavid Rientjes } 483002057967SDavid Rientjes 483102057967SDavid Rientjes /* 48324c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48332a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48344c887265SAdam Litke * first, for the page indexing below to work. 4835cb900f41SKirill A. Shutemov * 4836cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48374c887265SAdam Litke */ 48387868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48397868a208SPunit Agrawal huge_page_size(h)); 4840cb900f41SKirill A. Shutemov if (pte) 4841cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48422a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 484363551ae0SDavid Gibson 48442a15efc9SHugh Dickins /* 48452a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48463ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48473ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48483ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48493ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48502a15efc9SHugh Dickins */ 48513ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48523ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4853cb900f41SKirill A. Shutemov if (pte) 4854cb900f41SKirill A. Shutemov spin_unlock(ptl); 48552a15efc9SHugh Dickins remainder = 0; 48562a15efc9SHugh Dickins break; 48572a15efc9SHugh Dickins } 48582a15efc9SHugh Dickins 48599cc3a5bdSNaoya Horiguchi /* 48609cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48619cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48629cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 48639cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 48649cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 48659cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 48669cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 48679cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 48689cc3a5bdSNaoya Horiguchi */ 48699cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4870106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4871106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 48722b740303SSouptick Joarder vm_fault_t ret; 487387ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 48744c887265SAdam Litke 4875cb900f41SKirill A. Shutemov if (pte) 4876cb900f41SKirill A. Shutemov spin_unlock(ptl); 487787ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 487887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 48794f6da934SPeter Xu if (locked) 488071335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 488171335f37SPeter Xu FAULT_FLAG_KILLABLE; 488287ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 488387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 488487ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 488587ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 48864426e945SPeter Xu /* 48874426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 48884426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 48894426e945SPeter Xu */ 489087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 489187ffc118SAndrea Arcangeli } 489287ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 489387ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 48942be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 48951c59827dSHugh Dickins remainder = 0; 48961c59827dSHugh Dickins break; 48971c59827dSHugh Dickins } 489887ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 48994f6da934SPeter Xu if (locked && 49001ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 49014f6da934SPeter Xu *locked = 0; 490287ffc118SAndrea Arcangeli *nr_pages = 0; 490387ffc118SAndrea Arcangeli /* 490487ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 490587ffc118SAndrea Arcangeli * error, it will return zero 490687ffc118SAndrea Arcangeli * instead. 490787ffc118SAndrea Arcangeli * 490887ffc118SAndrea Arcangeli * No need to update "position" as the 490987ffc118SAndrea Arcangeli * caller will not check it after 491087ffc118SAndrea Arcangeli * *nr_pages is set to 0. 491187ffc118SAndrea Arcangeli */ 491287ffc118SAndrea Arcangeli return i; 491387ffc118SAndrea Arcangeli } 491487ffc118SAndrea Arcangeli continue; 491587ffc118SAndrea Arcangeli } 491663551ae0SDavid Gibson 4917a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 49187f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 49198fde12caSLinus Torvalds 49208fde12caSLinus Torvalds /* 4921acbfb087SZhigang Lu * If subpage information not requested, update counters 4922acbfb087SZhigang Lu * and skip the same_page loop below. 4923acbfb087SZhigang Lu */ 4924acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4925acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4926acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4927acbfb087SZhigang Lu vaddr += huge_page_size(h); 4928acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4929acbfb087SZhigang Lu i += pages_per_huge_page(h); 4930acbfb087SZhigang Lu spin_unlock(ptl); 4931acbfb087SZhigang Lu continue; 4932acbfb087SZhigang Lu } 4933acbfb087SZhigang Lu 4934d5d4b0aaSChen, Kenneth W same_page: 4935d6692183SChen, Kenneth W if (pages) { 493669d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 49373faa52c0SJohn Hubbard /* 49383faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: 49393faa52c0SJohn Hubbard * a) we hold the ptl lock, and b) we've just checked 49403faa52c0SJohn Hubbard * that the huge page is present in the page tables. If 49413faa52c0SJohn Hubbard * the huge page is present, then the tail pages must 49423faa52c0SJohn Hubbard * also be present. The ptl prevents the head page and 49433faa52c0SJohn Hubbard * tail pages from being rearranged in any way. So this 49443faa52c0SJohn Hubbard * page must be available at this point, unless the page 49453faa52c0SJohn Hubbard * refcount overflowed: 49463faa52c0SJohn Hubbard */ 49473faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { 49483faa52c0SJohn Hubbard spin_unlock(ptl); 49493faa52c0SJohn Hubbard remainder = 0; 49503faa52c0SJohn Hubbard err = -ENOMEM; 49513faa52c0SJohn Hubbard break; 49523faa52c0SJohn Hubbard } 4953d6692183SChen, Kenneth W } 495463551ae0SDavid Gibson 495563551ae0SDavid Gibson if (vmas) 495663551ae0SDavid Gibson vmas[i] = vma; 495763551ae0SDavid Gibson 495863551ae0SDavid Gibson vaddr += PAGE_SIZE; 4959d5d4b0aaSChen, Kenneth W ++pfn_offset; 496063551ae0SDavid Gibson --remainder; 496163551ae0SDavid Gibson ++i; 4962d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4963a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4964d5d4b0aaSChen, Kenneth W /* 4965d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4966d5d4b0aaSChen, Kenneth W * of this compound page. 4967d5d4b0aaSChen, Kenneth W */ 4968d5d4b0aaSChen, Kenneth W goto same_page; 4969d5d4b0aaSChen, Kenneth W } 4970cb900f41SKirill A. Shutemov spin_unlock(ptl); 497163551ae0SDavid Gibson } 497228a35716SMichel Lespinasse *nr_pages = remainder; 497387ffc118SAndrea Arcangeli /* 497487ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 497587ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 497687ffc118SAndrea Arcangeli * branch. 497787ffc118SAndrea Arcangeli */ 497863551ae0SDavid Gibson *position = vaddr; 497963551ae0SDavid Gibson 49802be7cfedSDaniel Jordan return i ? i : err; 498163551ae0SDavid Gibson } 49828f860591SZhang, Yanmin 49835491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 49845491ae7bSAneesh Kumar K.V /* 49855491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 49865491ae7bSAneesh Kumar K.V * implement this. 49875491ae7bSAneesh Kumar K.V */ 49885491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 49895491ae7bSAneesh Kumar K.V #endif 49905491ae7bSAneesh Kumar K.V 49917da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 49928f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 49938f860591SZhang, Yanmin { 49948f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 49958f860591SZhang, Yanmin unsigned long start = address; 49968f860591SZhang, Yanmin pte_t *ptep; 49978f860591SZhang, Yanmin pte_t pte; 4998a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49997da4d641SPeter Zijlstra unsigned long pages = 0; 5000dff11abeSMike Kravetz bool shared_pmd = false; 5001ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5002dff11abeSMike Kravetz 5003dff11abeSMike Kravetz /* 5004dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5005ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5006dff11abeSMike Kravetz * range if PMD sharing is possible. 5007dff11abeSMike Kravetz */ 50087269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 50097269f999SJérôme Glisse 0, vma, mm, start, end); 5010ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50118f860591SZhang, Yanmin 50128f860591SZhang, Yanmin BUG_ON(address >= end); 5013ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 50148f860591SZhang, Yanmin 5015ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 501683cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5017a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5018cb900f41SKirill A. Shutemov spinlock_t *ptl; 50197868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 50208f860591SZhang, Yanmin if (!ptep) 50218f860591SZhang, Yanmin continue; 5022cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 50237da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 50247da4d641SPeter Zijlstra pages++; 5025cb900f41SKirill A. Shutemov spin_unlock(ptl); 5026dff11abeSMike Kravetz shared_pmd = true; 502739dde65cSChen, Kenneth W continue; 50287da4d641SPeter Zijlstra } 5029a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5030a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5031a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5032a8bda28dSNaoya Horiguchi continue; 5033a8bda28dSNaoya Horiguchi } 5034a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5035a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5036a8bda28dSNaoya Horiguchi 5037a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5038a8bda28dSNaoya Horiguchi pte_t newpte; 5039a8bda28dSNaoya Horiguchi 5040a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5041a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5042e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5043e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5044a8bda28dSNaoya Horiguchi pages++; 5045a8bda28dSNaoya Horiguchi } 5046a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5047a8bda28dSNaoya Horiguchi continue; 5048a8bda28dSNaoya Horiguchi } 5049a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5050023bdd00SAneesh Kumar K.V pte_t old_pte; 5051023bdd00SAneesh Kumar K.V 5052023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5053023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5054be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5055023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50567da4d641SPeter Zijlstra pages++; 50578f860591SZhang, Yanmin } 5058cb900f41SKirill A. Shutemov spin_unlock(ptl); 50598f860591SZhang, Yanmin } 5060d833352aSMel Gorman /* 5061c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5062d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5063c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5064dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5065dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5066d833352aSMel Gorman */ 5067dff11abeSMike Kravetz if (shared_pmd) 5068ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5069dff11abeSMike Kravetz else 50705491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50710f10851eSJérôme Glisse /* 50720f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 50730f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 50740f10851eSJérôme Glisse * 5075ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50760f10851eSJérôme Glisse */ 507783cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5078ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50797da4d641SPeter Zijlstra 50807da4d641SPeter Zijlstra return pages << h->order; 50818f860591SZhang, Yanmin } 50828f860591SZhang, Yanmin 5083a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 5084a1e78772SMel Gorman long from, long to, 50855a6fe125SMel Gorman struct vm_area_struct *vma, 5086ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5087e4e574b7SAdam Litke { 50880db9d74eSMina Almasry long ret, chg, add = -1; 5089a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 509090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 50919119a41eSJoonsoo Kim struct resv_map *resv_map; 5092075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 50930db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5094e4e574b7SAdam Litke 509563489f8eSMike Kravetz /* This should never happen */ 509663489f8eSMike Kravetz if (from > to) { 509763489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 509863489f8eSMike Kravetz return -EINVAL; 509963489f8eSMike Kravetz } 510063489f8eSMike Kravetz 5101a1e78772SMel Gorman /* 510217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 510317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 510490481622SDavid Gibson * without using reserves 510517c9d12eSMel Gorman */ 5106ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 510717c9d12eSMel Gorman return 0; 510817c9d12eSMel Gorman 510917c9d12eSMel Gorman /* 5110a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5111a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5112a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5113a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5114a1e78772SMel Gorman */ 51159119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5116f27a5136SMike Kravetz /* 5117f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5118f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5119f27a5136SMike Kravetz * hugetlbfs_get_inode). 5120f27a5136SMike Kravetz */ 51214e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 51229119a41eSJoonsoo Kim 51230db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 51249119a41eSJoonsoo Kim 51259119a41eSJoonsoo Kim } else { 5126e9fe92aeSMina Almasry /* Private mapping. */ 51279119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51285a6fe125SMel Gorman if (!resv_map) 51295a6fe125SMel Gorman return -ENOMEM; 51305a6fe125SMel Gorman 513117c9d12eSMel Gorman chg = to - from; 513217c9d12eSMel Gorman 51335a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51345a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51355a6fe125SMel Gorman } 51365a6fe125SMel Gorman 5137c50ac050SDave Hansen if (chg < 0) { 5138c50ac050SDave Hansen ret = chg; 5139c50ac050SDave Hansen goto out_err; 5140c50ac050SDave Hansen } 514117c9d12eSMel Gorman 5142075a61d0SMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 5143075a61d0SMina Almasry hstate_index(h), chg * pages_per_huge_page(h), &h_cg); 5144075a61d0SMina Almasry 5145075a61d0SMina Almasry if (ret < 0) { 5146075a61d0SMina Almasry ret = -ENOMEM; 5147075a61d0SMina Almasry goto out_err; 5148075a61d0SMina Almasry } 5149075a61d0SMina Almasry 5150075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5151075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5152075a61d0SMina Almasry * of the resv_map. 5153075a61d0SMina Almasry */ 5154075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5155075a61d0SMina Almasry } 5156075a61d0SMina Almasry 51571c5ecae3SMike Kravetz /* 51581c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51591c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51601c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51611c5ecae3SMike Kravetz */ 51621c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 51631c5ecae3SMike Kravetz if (gbl_reserve < 0) { 5164c50ac050SDave Hansen ret = -ENOSPC; 5165075a61d0SMina Almasry goto out_uncharge_cgroup; 5166c50ac050SDave Hansen } 516717c9d12eSMel Gorman 516817c9d12eSMel Gorman /* 516917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 517090481622SDavid Gibson * Hand the pages back to the subpool if there are not 517117c9d12eSMel Gorman */ 51721c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 517317c9d12eSMel Gorman if (ret < 0) { 5174075a61d0SMina Almasry goto out_put_pages; 517517c9d12eSMel Gorman } 517617c9d12eSMel Gorman 517717c9d12eSMel Gorman /* 517817c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 517917c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 518017c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 518117c9d12eSMel Gorman * the reservation was and the page cache tells how much of 518217c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 518317c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 518417c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 518517c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 518617c9d12eSMel Gorman * else has to be done for private mappings here 518717c9d12eSMel Gorman */ 518833039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5189075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 519033039678SMike Kravetz 51910db9d74eSMina Almasry if (unlikely(add < 0)) { 51920db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5193075a61d0SMina Almasry goto out_put_pages; 51940db9d74eSMina Almasry } else if (unlikely(chg > add)) { 519533039678SMike Kravetz /* 519633039678SMike Kravetz * pages in this range were added to the reserve 519733039678SMike Kravetz * map between region_chg and region_add. This 519833039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 519933039678SMike Kravetz * the subpool and reserve counts modified above 520033039678SMike Kravetz * based on the difference. 520133039678SMike Kravetz */ 520233039678SMike Kravetz long rsv_adjust; 520333039678SMike Kravetz 5204075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5205075a61d0SMina Almasry hstate_index(h), 5206075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5207075a61d0SMina Almasry 520833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 520933039678SMike Kravetz chg - add); 521033039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 521133039678SMike Kravetz } 521233039678SMike Kravetz } 5213a43a8c39SChen, Kenneth W return 0; 5214075a61d0SMina Almasry out_put_pages: 5215075a61d0SMina Almasry /* put back original number of pages, chg */ 5216075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5217075a61d0SMina Almasry out_uncharge_cgroup: 5218075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5219075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5220c50ac050SDave Hansen out_err: 52215e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 52220db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 52230db9d74eSMina Almasry * region_add failed or didn't run. 52240db9d74eSMina Almasry */ 52250db9d74eSMina Almasry if (chg >= 0 && add < 0) 52260db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5227f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5228f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 5229c50ac050SDave Hansen return ret; 5230a43a8c39SChen, Kenneth W } 5231a43a8c39SChen, Kenneth W 5232b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5233b5cec28dSMike Kravetz long freed) 5234a43a8c39SChen, Kenneth W { 5235a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52364e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52379119a41eSJoonsoo Kim long chg = 0; 523890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52391c5ecae3SMike Kravetz long gbl_reserve; 524045c682a6SKen Chen 5241f27a5136SMike Kravetz /* 5242f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5243f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5244f27a5136SMike Kravetz */ 5245b5cec28dSMike Kravetz if (resv_map) { 5246b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5247b5cec28dSMike Kravetz /* 5248b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5249b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5250b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5251b5cec28dSMike Kravetz */ 5252b5cec28dSMike Kravetz if (chg < 0) 5253b5cec28dSMike Kravetz return chg; 5254b5cec28dSMike Kravetz } 5255b5cec28dSMike Kravetz 525645c682a6SKen Chen spin_lock(&inode->i_lock); 5257e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 525845c682a6SKen Chen spin_unlock(&inode->i_lock); 525945c682a6SKen Chen 52601c5ecae3SMike Kravetz /* 52611c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52621c5ecae3SMike Kravetz * reservations to be released may be adjusted. 52631c5ecae3SMike Kravetz */ 52641c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52651c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5266b5cec28dSMike Kravetz 5267b5cec28dSMike Kravetz return 0; 5268a43a8c39SChen, Kenneth W } 526993f70f90SNaoya Horiguchi 52703212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 52713212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 52723212b535SSteve Capper struct vm_area_struct *vma, 52733212b535SSteve Capper unsigned long addr, pgoff_t idx) 52743212b535SSteve Capper { 52753212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52763212b535SSteve Capper svma->vm_start; 52773212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52783212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 52793212b535SSteve Capper 52803212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5281de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5282de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 52833212b535SSteve Capper 52843212b535SSteve Capper /* 52853212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 52863212b535SSteve Capper * page table page. 52873212b535SSteve Capper */ 52883212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 52893212b535SSteve Capper vm_flags != svm_flags || 52903212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 52913212b535SSteve Capper return 0; 52923212b535SSteve Capper 52933212b535SSteve Capper return saddr; 52943212b535SSteve Capper } 52953212b535SSteve Capper 529631aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 52973212b535SSteve Capper { 52983212b535SSteve Capper unsigned long base = addr & PUD_MASK; 52993212b535SSteve Capper unsigned long end = base + PUD_SIZE; 53003212b535SSteve Capper 53013212b535SSteve Capper /* 53023212b535SSteve Capper * check on proper vm_flags and page table alignment 53033212b535SSteve Capper */ 5304017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 530531aafb45SNicholas Krause return true; 530631aafb45SNicholas Krause return false; 53073212b535SSteve Capper } 53083212b535SSteve Capper 53093212b535SSteve Capper /* 5310017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5311017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5312017b1660SMike Kravetz * shared pmd mappings. 5313017b1660SMike Kravetz */ 5314017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5315017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5316017b1660SMike Kravetz { 5317353b2de4SMateusz Nosek unsigned long check_addr; 5318017b1660SMike Kravetz 5319017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 5320017b1660SMike Kravetz return; 5321017b1660SMike Kravetz 5322017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 5323017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 5324017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 5325017b1660SMike Kravetz 5326017b1660SMike Kravetz /* 5327017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 5328017b1660SMike Kravetz */ 5329017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 5330017b1660SMike Kravetz if (a_start < *start) 5331017b1660SMike Kravetz *start = a_start; 5332017b1660SMike Kravetz if (a_end > *end) 5333017b1660SMike Kravetz *end = a_end; 5334017b1660SMike Kravetz } 5335017b1660SMike Kravetz } 5336017b1660SMike Kravetz } 5337017b1660SMike Kravetz 5338017b1660SMike Kravetz /* 53393212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53403212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53413212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5342c0d0381aSMike Kravetz * code much cleaner. 5343c0d0381aSMike Kravetz * 5344c0d0381aSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode. 5345c0d0381aSMike Kravetz * For hugetlbfs, this prevents removal of any page table entries associated 5346c0d0381aSMike Kravetz * with the address space. This is important as we are setting up sharing 5347c0d0381aSMike Kravetz * based on existing page table entries (mappings). 53483212b535SSteve Capper */ 53493212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 53503212b535SSteve Capper { 53513212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 53523212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 53533212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 53543212b535SSteve Capper vma->vm_pgoff; 53553212b535SSteve Capper struct vm_area_struct *svma; 53563212b535SSteve Capper unsigned long saddr; 53573212b535SSteve Capper pte_t *spte = NULL; 53583212b535SSteve Capper pte_t *pte; 5359cb900f41SKirill A. Shutemov spinlock_t *ptl; 53603212b535SSteve Capper 53613212b535SSteve Capper if (!vma_shareable(vma, addr)) 53623212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 53633212b535SSteve Capper 53643212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 53653212b535SSteve Capper if (svma == vma) 53663212b535SSteve Capper continue; 53673212b535SSteve Capper 53683212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 53693212b535SSteve Capper if (saddr) { 53707868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 53717868a208SPunit Agrawal vma_mmu_pagesize(svma)); 53723212b535SSteve Capper if (spte) { 53733212b535SSteve Capper get_page(virt_to_page(spte)); 53743212b535SSteve Capper break; 53753212b535SSteve Capper } 53763212b535SSteve Capper } 53773212b535SSteve Capper } 53783212b535SSteve Capper 53793212b535SSteve Capper if (!spte) 53803212b535SSteve Capper goto out; 53813212b535SSteve Capper 53828bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5383dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 53843212b535SSteve Capper pud_populate(mm, pud, 53853212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5386c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5387dc6c9a35SKirill A. Shutemov } else { 53883212b535SSteve Capper put_page(virt_to_page(spte)); 5389dc6c9a35SKirill A. Shutemov } 5390cb900f41SKirill A. Shutemov spin_unlock(ptl); 53913212b535SSteve Capper out: 53923212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 53933212b535SSteve Capper return pte; 53943212b535SSteve Capper } 53953212b535SSteve Capper 53963212b535SSteve Capper /* 53973212b535SSteve Capper * unmap huge page backed by shared pte. 53983212b535SSteve Capper * 53993212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 54003212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 54013212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 54023212b535SSteve Capper * 5403c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 54043212b535SSteve Capper * 54053212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 54063212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 54073212b535SSteve Capper */ 54083212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 54093212b535SSteve Capper { 54103212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5411c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5412c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 54133212b535SSteve Capper 54143212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 54153212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 54163212b535SSteve Capper return 0; 54173212b535SSteve Capper 54183212b535SSteve Capper pud_clear(pud); 54193212b535SSteve Capper put_page(virt_to_page(ptep)); 5420dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54213212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54223212b535SSteve Capper return 1; 54233212b535SSteve Capper } 54249e5fc74cSSteve Capper #define want_pmd_share() (1) 54259e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54269e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 54279e5fc74cSSteve Capper { 54289e5fc74cSSteve Capper return NULL; 54299e5fc74cSSteve Capper } 5430e81f2d22SZhang Zhen 5431e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 5432e81f2d22SZhang Zhen { 5433e81f2d22SZhang Zhen return 0; 5434e81f2d22SZhang Zhen } 5435017b1660SMike Kravetz 5436017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5437017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5438017b1660SMike Kravetz { 5439017b1660SMike Kravetz } 54409e5fc74cSSteve Capper #define want_pmd_share() (0) 54413212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54423212b535SSteve Capper 54439e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 54449e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 54459e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54469e5fc74cSSteve Capper { 54479e5fc74cSSteve Capper pgd_t *pgd; 5448c2febafcSKirill A. Shutemov p4d_t *p4d; 54499e5fc74cSSteve Capper pud_t *pud; 54509e5fc74cSSteve Capper pte_t *pte = NULL; 54519e5fc74cSSteve Capper 54529e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5453f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5454f4f0a3d8SKirill A. Shutemov if (!p4d) 5455f4f0a3d8SKirill A. Shutemov return NULL; 5456c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 54579e5fc74cSSteve Capper if (pud) { 54589e5fc74cSSteve Capper if (sz == PUD_SIZE) { 54599e5fc74cSSteve Capper pte = (pte_t *)pud; 54609e5fc74cSSteve Capper } else { 54619e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 54629e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 54639e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 54649e5fc74cSSteve Capper else 54659e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54669e5fc74cSSteve Capper } 54679e5fc74cSSteve Capper } 54684e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 54699e5fc74cSSteve Capper 54709e5fc74cSSteve Capper return pte; 54719e5fc74cSSteve Capper } 54729e5fc74cSSteve Capper 54739b19df29SPunit Agrawal /* 54749b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 54759b19df29SPunit Agrawal * entry at address @addr 54769b19df29SPunit Agrawal * 54778ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 54788ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 54799b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 54809b19df29SPunit Agrawal * table. 54819b19df29SPunit Agrawal */ 54827868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 54837868a208SPunit Agrawal unsigned long addr, unsigned long sz) 54849e5fc74cSSteve Capper { 54859e5fc74cSSteve Capper pgd_t *pgd; 5486c2febafcSKirill A. Shutemov p4d_t *p4d; 54878ac0b81aSLi Xinhai pud_t *pud; 54888ac0b81aSLi Xinhai pmd_t *pmd; 54899e5fc74cSSteve Capper 54909e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5491c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5492c2febafcSKirill A. Shutemov return NULL; 5493c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5494c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5495c2febafcSKirill A. Shutemov return NULL; 54969b19df29SPunit Agrawal 5497c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 54988ac0b81aSLi Xinhai if (sz == PUD_SIZE) 54998ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 55009e5fc74cSSteve Capper return (pte_t *)pud; 55018ac0b81aSLi Xinhai if (!pud_present(*pud)) 55028ac0b81aSLi Xinhai return NULL; 55038ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 55049b19df29SPunit Agrawal 55059e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 55068ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 55079e5fc74cSSteve Capper return (pte_t *)pmd; 55089e5fc74cSSteve Capper } 55099e5fc74cSSteve Capper 551061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 551161f77edaSNaoya Horiguchi 551261f77edaSNaoya Horiguchi /* 551361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 551461f77edaSNaoya Horiguchi * behavior. 551561f77edaSNaoya Horiguchi */ 551661f77edaSNaoya Horiguchi struct page * __weak 551761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 551861f77edaSNaoya Horiguchi int write) 551961f77edaSNaoya Horiguchi { 552061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 552161f77edaSNaoya Horiguchi } 552261f77edaSNaoya Horiguchi 552361f77edaSNaoya Horiguchi struct page * __weak 55244dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55254dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55264dc71451SAneesh Kumar K.V { 55274dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55284dc71451SAneesh Kumar K.V return NULL; 55294dc71451SAneesh Kumar K.V } 55304dc71451SAneesh Kumar K.V 55314dc71451SAneesh Kumar K.V struct page * __weak 55329e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5533e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55349e5fc74cSSteve Capper { 5535e66f17ffSNaoya Horiguchi struct page *page = NULL; 5536e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5537c9d398faSNaoya Horiguchi pte_t pte; 55383faa52c0SJohn Hubbard 55393faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55403faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55413faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55423faa52c0SJohn Hubbard return NULL; 55433faa52c0SJohn Hubbard 5544e66f17ffSNaoya Horiguchi retry: 5545e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5546e66f17ffSNaoya Horiguchi spin_lock(ptl); 5547e66f17ffSNaoya Horiguchi /* 5548e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5549e66f17ffSNaoya Horiguchi * unmapped from other threads. 5550e66f17ffSNaoya Horiguchi */ 5551e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5552e66f17ffSNaoya Horiguchi goto out; 5553c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5554c9d398faSNaoya Horiguchi if (pte_present(pte)) { 555597534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 55563faa52c0SJohn Hubbard /* 55573faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 55583faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 55593faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 55603faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 55613faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 55623faa52c0SJohn Hubbard * unless the page refcount overflowed: 55633faa52c0SJohn Hubbard */ 55643faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 55653faa52c0SJohn Hubbard page = NULL; 55663faa52c0SJohn Hubbard goto out; 55673faa52c0SJohn Hubbard } 5568e66f17ffSNaoya Horiguchi } else { 5569c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5570e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5571e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5572e66f17ffSNaoya Horiguchi goto retry; 5573e66f17ffSNaoya Horiguchi } 5574e66f17ffSNaoya Horiguchi /* 5575e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5576e66f17ffSNaoya Horiguchi * follow_page_mask(). 5577e66f17ffSNaoya Horiguchi */ 5578e66f17ffSNaoya Horiguchi } 5579e66f17ffSNaoya Horiguchi out: 5580e66f17ffSNaoya Horiguchi spin_unlock(ptl); 55819e5fc74cSSteve Capper return page; 55829e5fc74cSSteve Capper } 55839e5fc74cSSteve Capper 558461f77edaSNaoya Horiguchi struct page * __weak 55859e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5586e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 55879e5fc74cSSteve Capper { 55883faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5589e66f17ffSNaoya Horiguchi return NULL; 55909e5fc74cSSteve Capper 5591e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 55929e5fc74cSSteve Capper } 55939e5fc74cSSteve Capper 5594faaa5b62SAnshuman Khandual struct page * __weak 5595faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5596faaa5b62SAnshuman Khandual { 55973faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5598faaa5b62SAnshuman Khandual return NULL; 5599faaa5b62SAnshuman Khandual 5600faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5601faaa5b62SAnshuman Khandual } 5602faaa5b62SAnshuman Khandual 560331caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 560431caf665SNaoya Horiguchi { 5605bcc54222SNaoya Horiguchi bool ret = true; 5606bcc54222SNaoya Horiguchi 5607309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 560831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5609bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 5610bcc54222SNaoya Horiguchi ret = false; 5611bcc54222SNaoya Horiguchi goto unlock; 5612bcc54222SNaoya Horiguchi } 5613bcc54222SNaoya Horiguchi clear_page_huge_active(page); 561431caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5615bcc54222SNaoya Horiguchi unlock: 561631caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5617bcc54222SNaoya Horiguchi return ret; 561831caf665SNaoya Horiguchi } 561931caf665SNaoya Horiguchi 562031caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 562131caf665SNaoya Horiguchi { 5622309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 562331caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5624bcc54222SNaoya Horiguchi set_page_huge_active(page); 562531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 562631caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 562731caf665SNaoya Horiguchi put_page(page); 562831caf665SNaoya Horiguchi } 5629ab5ac90aSMichal Hocko 5630ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5631ab5ac90aSMichal Hocko { 5632ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5633ab5ac90aSMichal Hocko 5634ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5635ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5636ab5ac90aSMichal Hocko 5637ab5ac90aSMichal Hocko /* 5638ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5639ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5640ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5641ab5ac90aSMichal Hocko * the temporary status. 5642ab5ac90aSMichal Hocko * 5643ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5644ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5645ab5ac90aSMichal Hocko * the per-node's. 5646ab5ac90aSMichal Hocko */ 5647ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5648ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5649ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5650ab5ac90aSMichal Hocko 5651ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5652ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5653ab5ac90aSMichal Hocko 5654ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5655ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5656ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5657ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5658ab5ac90aSMichal Hocko } 5659ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5660ab5ac90aSMichal Hocko } 5661ab5ac90aSMichal Hocko } 5662cf11e85fSRoman Gushchin 5663cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5664cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5665cf11e85fSRoman Gushchin 5666cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5667cf11e85fSRoman Gushchin { 5668cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5669cf11e85fSRoman Gushchin return 0; 5670cf11e85fSRoman Gushchin } 5671cf11e85fSRoman Gushchin 5672cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5673cf11e85fSRoman Gushchin 5674cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5675cf11e85fSRoman Gushchin { 5676cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5677cf11e85fSRoman Gushchin int nid; 5678cf11e85fSRoman Gushchin 5679cf11e85fSRoman Gushchin cma_reserve_called = true; 5680cf11e85fSRoman Gushchin 5681cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5682cf11e85fSRoman Gushchin return; 5683cf11e85fSRoman Gushchin 5684cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5685cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5686cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5687cf11e85fSRoman Gushchin return; 5688cf11e85fSRoman Gushchin } 5689cf11e85fSRoman Gushchin 5690cf11e85fSRoman Gushchin /* 5691cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5692cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5693cf11e85fSRoman Gushchin */ 5694cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5695cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5696cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5697cf11e85fSRoman Gushchin 5698cf11e85fSRoman Gushchin reserved = 0; 5699cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5700cf11e85fSRoman Gushchin int res; 5701cf11e85fSRoman Gushchin 5702cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5703cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5704cf11e85fSRoman Gushchin 5705cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 5706cf11e85fSRoman Gushchin 0, false, "hugetlb", 5707cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5708cf11e85fSRoman Gushchin if (res) { 5709cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5710cf11e85fSRoman Gushchin res, nid); 5711cf11e85fSRoman Gushchin continue; 5712cf11e85fSRoman Gushchin } 5713cf11e85fSRoman Gushchin 5714cf11e85fSRoman Gushchin reserved += size; 5715cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5716cf11e85fSRoman Gushchin size / SZ_1M, nid); 5717cf11e85fSRoman Gushchin 5718cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5719cf11e85fSRoman Gushchin break; 5720cf11e85fSRoman Gushchin } 5721cf11e85fSRoman Gushchin } 5722cf11e85fSRoman Gushchin 5723cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5724cf11e85fSRoman Gushchin { 5725cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5726cf11e85fSRoman Gushchin return; 5727cf11e85fSRoman Gushchin 5728cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5729cf11e85fSRoman Gushchin } 5730cf11e85fSRoman Gushchin 5731cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5732