11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.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> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24d6606683SLinus Torvalds 2563551ae0SDavid Gibson #include <asm/page.h> 2663551ae0SDavid Gibson #include <asm/pgtable.h> 2724669e58SAneesh Kumar K.V #include <asm/tlb.h> 2863551ae0SDavid Gibson 2924669e58SAneesh Kumar K.V #include <linux/io.h> 3063551ae0SDavid Gibson #include <linux/hugetlb.h> 319dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 329a305230SLee Schermerhorn #include <linux/node.h> 337835e98bSNick Piggin #include "internal.h" 341da177e4SLinus Torvalds 351da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 36396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 37396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 38a5516438SAndi Kleen 39c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 40e5ff2159SAndi Kleen unsigned int default_hstate_idx; 41e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 42e5ff2159SAndi Kleen 4353ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4453ba51d2SJon Tollefson 45e5ff2159SAndi Kleen /* for command line parsing */ 46e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 47e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 48e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 49e5ff2159SAndi Kleen 503935baa9SDavid Gibson /* 5131caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5231caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 533935baa9SDavid Gibson */ 54c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 550bd0f9fbSEric Paris 5690481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 5790481622SDavid Gibson { 5890481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 5990481622SDavid Gibson 6090481622SDavid Gibson spin_unlock(&spool->lock); 6190481622SDavid Gibson 6290481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 6390481622SDavid Gibson * remain, free the subpool the subpool remain */ 6490481622SDavid Gibson if (free) 6590481622SDavid Gibson kfree(spool); 6690481622SDavid Gibson } 6790481622SDavid Gibson 6890481622SDavid Gibson struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 6990481622SDavid Gibson { 7090481622SDavid Gibson struct hugepage_subpool *spool; 7190481622SDavid Gibson 7290481622SDavid Gibson spool = kmalloc(sizeof(*spool), GFP_KERNEL); 7390481622SDavid Gibson if (!spool) 7490481622SDavid Gibson return NULL; 7590481622SDavid Gibson 7690481622SDavid Gibson spin_lock_init(&spool->lock); 7790481622SDavid Gibson spool->count = 1; 7890481622SDavid Gibson spool->max_hpages = nr_blocks; 7990481622SDavid Gibson spool->used_hpages = 0; 8090481622SDavid Gibson 8190481622SDavid Gibson return spool; 8290481622SDavid Gibson } 8390481622SDavid Gibson 8490481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 8590481622SDavid Gibson { 8690481622SDavid Gibson spin_lock(&spool->lock); 8790481622SDavid Gibson BUG_ON(!spool->count); 8890481622SDavid Gibson spool->count--; 8990481622SDavid Gibson unlock_or_release_subpool(spool); 9090481622SDavid Gibson } 9190481622SDavid Gibson 9290481622SDavid Gibson static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 9390481622SDavid Gibson long delta) 9490481622SDavid Gibson { 9590481622SDavid Gibson int ret = 0; 9690481622SDavid Gibson 9790481622SDavid Gibson if (!spool) 9890481622SDavid Gibson return 0; 9990481622SDavid Gibson 10090481622SDavid Gibson spin_lock(&spool->lock); 10190481622SDavid Gibson if ((spool->used_hpages + delta) <= spool->max_hpages) { 10290481622SDavid Gibson spool->used_hpages += delta; 10390481622SDavid Gibson } else { 10490481622SDavid Gibson ret = -ENOMEM; 10590481622SDavid Gibson } 10690481622SDavid Gibson spin_unlock(&spool->lock); 10790481622SDavid Gibson 10890481622SDavid Gibson return ret; 10990481622SDavid Gibson } 11090481622SDavid Gibson 11190481622SDavid Gibson static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 11290481622SDavid Gibson long delta) 11390481622SDavid Gibson { 11490481622SDavid Gibson if (!spool) 11590481622SDavid Gibson return; 11690481622SDavid Gibson 11790481622SDavid Gibson spin_lock(&spool->lock); 11890481622SDavid Gibson spool->used_hpages -= delta; 11990481622SDavid Gibson /* If hugetlbfs_put_super couldn't free spool due to 12090481622SDavid Gibson * an outstanding quota reference, free it now. */ 12190481622SDavid Gibson unlock_or_release_subpool(spool); 12290481622SDavid Gibson } 12390481622SDavid Gibson 12490481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 12590481622SDavid Gibson { 12690481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 12790481622SDavid Gibson } 12890481622SDavid Gibson 12990481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 13090481622SDavid Gibson { 131496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 13290481622SDavid Gibson } 13390481622SDavid Gibson 134e7c4b0bfSAndy Whitcroft /* 13596822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 13696822904SAndy Whitcroft * across the pages in a mapping. 13784afd99bSAndy Whitcroft * 13884afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 139c748c262SJoonsoo Kim * and the hugetlb_instantiation_mutex. To access or modify a region the caller 14084afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 141c748c262SJoonsoo Kim * the hugetlb_instantiation_mutex: 14284afd99bSAndy Whitcroft * 14384afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 14484afd99bSAndy Whitcroft * or 14584afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 14684afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 14796822904SAndy Whitcroft */ 14896822904SAndy Whitcroft struct file_region { 14996822904SAndy Whitcroft struct list_head link; 15096822904SAndy Whitcroft long from; 15196822904SAndy Whitcroft long to; 15296822904SAndy Whitcroft }; 15396822904SAndy Whitcroft 15496822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 15596822904SAndy Whitcroft { 15696822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 15796822904SAndy Whitcroft 15896822904SAndy Whitcroft /* Locate the region we are either in or before. */ 15996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16096822904SAndy Whitcroft if (f <= rg->to) 16196822904SAndy Whitcroft break; 16296822904SAndy Whitcroft 16396822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 16496822904SAndy Whitcroft if (f > rg->from) 16596822904SAndy Whitcroft f = rg->from; 16696822904SAndy Whitcroft 16796822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 16896822904SAndy Whitcroft nrg = rg; 16996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 17096822904SAndy Whitcroft if (&rg->link == head) 17196822904SAndy Whitcroft break; 17296822904SAndy Whitcroft if (rg->from > t) 17396822904SAndy Whitcroft break; 17496822904SAndy Whitcroft 17596822904SAndy Whitcroft /* If this area reaches higher then extend our area to 17696822904SAndy Whitcroft * include it completely. If this is not the first area 17796822904SAndy Whitcroft * which we intend to reuse, free it. */ 17896822904SAndy Whitcroft if (rg->to > t) 17996822904SAndy Whitcroft t = rg->to; 18096822904SAndy Whitcroft if (rg != nrg) { 18196822904SAndy Whitcroft list_del(&rg->link); 18296822904SAndy Whitcroft kfree(rg); 18396822904SAndy Whitcroft } 18496822904SAndy Whitcroft } 18596822904SAndy Whitcroft nrg->from = f; 18696822904SAndy Whitcroft nrg->to = t; 18796822904SAndy Whitcroft return 0; 18896822904SAndy Whitcroft } 18996822904SAndy Whitcroft 19096822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 19196822904SAndy Whitcroft { 19296822904SAndy Whitcroft struct file_region *rg, *nrg; 19396822904SAndy Whitcroft long chg = 0; 19496822904SAndy Whitcroft 19596822904SAndy Whitcroft /* Locate the region we are before or in. */ 19696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 19796822904SAndy Whitcroft if (f <= rg->to) 19896822904SAndy Whitcroft break; 19996822904SAndy Whitcroft 20096822904SAndy Whitcroft /* If we are below the current region then a new region is required. 20196822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 20296822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 20396822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 20496822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 20596822904SAndy Whitcroft if (!nrg) 20696822904SAndy Whitcroft return -ENOMEM; 20796822904SAndy Whitcroft nrg->from = f; 20896822904SAndy Whitcroft nrg->to = f; 20996822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 21096822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 21196822904SAndy Whitcroft 21296822904SAndy Whitcroft return t - f; 21396822904SAndy Whitcroft } 21496822904SAndy Whitcroft 21596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 21696822904SAndy Whitcroft if (f > rg->from) 21796822904SAndy Whitcroft f = rg->from; 21896822904SAndy Whitcroft chg = t - f; 21996822904SAndy Whitcroft 22096822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 22196822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 22296822904SAndy Whitcroft if (&rg->link == head) 22396822904SAndy Whitcroft break; 22496822904SAndy Whitcroft if (rg->from > t) 22596822904SAndy Whitcroft return chg; 22696822904SAndy Whitcroft 22725985edcSLucas De Marchi /* We overlap with this area, if it extends further than 22896822904SAndy Whitcroft * us then we must extend ourselves. Account for its 22996822904SAndy Whitcroft * existing reservation. */ 23096822904SAndy Whitcroft if (rg->to > t) { 23196822904SAndy Whitcroft chg += rg->to - t; 23296822904SAndy Whitcroft t = rg->to; 23396822904SAndy Whitcroft } 23496822904SAndy Whitcroft chg -= rg->to - rg->from; 23596822904SAndy Whitcroft } 23696822904SAndy Whitcroft return chg; 23796822904SAndy Whitcroft } 23896822904SAndy Whitcroft 23996822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 24096822904SAndy Whitcroft { 24196822904SAndy Whitcroft struct file_region *rg, *trg; 24296822904SAndy Whitcroft long chg = 0; 24396822904SAndy Whitcroft 24496822904SAndy Whitcroft /* Locate the region we are either in or before. */ 24596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 24696822904SAndy Whitcroft if (end <= rg->to) 24796822904SAndy Whitcroft break; 24896822904SAndy Whitcroft if (&rg->link == head) 24996822904SAndy Whitcroft return 0; 25096822904SAndy Whitcroft 25196822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 25296822904SAndy Whitcroft if (end > rg->from) { 25396822904SAndy Whitcroft chg = rg->to - end; 25496822904SAndy Whitcroft rg->to = end; 25596822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 25696822904SAndy Whitcroft } 25796822904SAndy Whitcroft 25896822904SAndy Whitcroft /* Drop any remaining regions. */ 25996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 26096822904SAndy Whitcroft if (&rg->link == head) 26196822904SAndy Whitcroft break; 26296822904SAndy Whitcroft chg += rg->to - rg->from; 26396822904SAndy Whitcroft list_del(&rg->link); 26496822904SAndy Whitcroft kfree(rg); 26596822904SAndy Whitcroft } 26696822904SAndy Whitcroft return chg; 26796822904SAndy Whitcroft } 26896822904SAndy Whitcroft 26984afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 27084afd99bSAndy Whitcroft { 27184afd99bSAndy Whitcroft struct file_region *rg; 27284afd99bSAndy Whitcroft long chg = 0; 27384afd99bSAndy Whitcroft 27484afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 27584afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 276f2135a4aSWang Sheng-Hui long seg_from; 277f2135a4aSWang Sheng-Hui long seg_to; 27884afd99bSAndy Whitcroft 27984afd99bSAndy Whitcroft if (rg->to <= f) 28084afd99bSAndy Whitcroft continue; 28184afd99bSAndy Whitcroft if (rg->from >= t) 28284afd99bSAndy Whitcroft break; 28384afd99bSAndy Whitcroft 28484afd99bSAndy Whitcroft seg_from = max(rg->from, f); 28584afd99bSAndy Whitcroft seg_to = min(rg->to, t); 28684afd99bSAndy Whitcroft 28784afd99bSAndy Whitcroft chg += seg_to - seg_from; 28884afd99bSAndy Whitcroft } 28984afd99bSAndy Whitcroft 29084afd99bSAndy Whitcroft return chg; 29184afd99bSAndy Whitcroft } 29284afd99bSAndy Whitcroft 29396822904SAndy Whitcroft /* 294e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 295e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 296e7c4b0bfSAndy Whitcroft */ 297a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 298a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 299e7c4b0bfSAndy Whitcroft { 300a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 301a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 302e7c4b0bfSAndy Whitcroft } 303e7c4b0bfSAndy Whitcroft 3040fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 3050fe6e20bSNaoya Horiguchi unsigned long address) 3060fe6e20bSNaoya Horiguchi { 3070fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 3080fe6e20bSNaoya Horiguchi } 3090fe6e20bSNaoya Horiguchi 31084afd99bSAndy Whitcroft /* 31108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 31208fba699SMel Gorman * cases this will be same size as used by the page table entries. 31308fba699SMel Gorman */ 31408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 31508fba699SMel Gorman { 31608fba699SMel Gorman struct hstate *hstate; 31708fba699SMel Gorman 31808fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 31908fba699SMel Gorman return PAGE_SIZE; 32008fba699SMel Gorman 32108fba699SMel Gorman hstate = hstate_vma(vma); 32208fba699SMel Gorman 3232415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 32408fba699SMel Gorman } 325f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 32608fba699SMel Gorman 32708fba699SMel Gorman /* 3283340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 3293340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 3303340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 3313340289dSMel Gorman * function is required. 3323340289dSMel Gorman */ 3333340289dSMel Gorman #ifndef vma_mmu_pagesize 3343340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 3353340289dSMel Gorman { 3363340289dSMel Gorman return vma_kernel_pagesize(vma); 3373340289dSMel Gorman } 3383340289dSMel Gorman #endif 3393340289dSMel Gorman 3403340289dSMel Gorman /* 34184afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 34284afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 34384afd99bSAndy Whitcroft * alignment. 34484afd99bSAndy Whitcroft */ 34584afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 34684afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 34704f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 34884afd99bSAndy Whitcroft 349a1e78772SMel Gorman /* 350a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 351a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 352a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 353a1e78772SMel Gorman * 354a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 355a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 356a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 357a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 35884afd99bSAndy Whitcroft * 35984afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 36084afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 36184afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 36284afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 36384afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 36484afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 36584afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 36684afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 367a1e78772SMel Gorman */ 368e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 369e7c4b0bfSAndy Whitcroft { 370e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 371e7c4b0bfSAndy Whitcroft } 372e7c4b0bfSAndy Whitcroft 373e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 374e7c4b0bfSAndy Whitcroft unsigned long value) 375e7c4b0bfSAndy Whitcroft { 376e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 377e7c4b0bfSAndy Whitcroft } 378e7c4b0bfSAndy Whitcroft 37984afd99bSAndy Whitcroft struct resv_map { 38084afd99bSAndy Whitcroft struct kref refs; 38184afd99bSAndy Whitcroft struct list_head regions; 38284afd99bSAndy Whitcroft }; 38384afd99bSAndy Whitcroft 3842a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 38584afd99bSAndy Whitcroft { 38684afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 38784afd99bSAndy Whitcroft if (!resv_map) 38884afd99bSAndy Whitcroft return NULL; 38984afd99bSAndy Whitcroft 39084afd99bSAndy Whitcroft kref_init(&resv_map->refs); 39184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 39284afd99bSAndy Whitcroft 39384afd99bSAndy Whitcroft return resv_map; 39484afd99bSAndy Whitcroft } 39584afd99bSAndy Whitcroft 3962a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 39784afd99bSAndy Whitcroft { 39884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 39984afd99bSAndy Whitcroft 40084afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 40184afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 40284afd99bSAndy Whitcroft kfree(resv_map); 40384afd99bSAndy Whitcroft } 40484afd99bSAndy Whitcroft 40584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 406a1e78772SMel Gorman { 407a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 408f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 40984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 41084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 4112a4b3dedSHarvey Harrison return NULL; 412a1e78772SMel Gorman } 413a1e78772SMel Gorman 41484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 415a1e78772SMel Gorman { 416a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 417f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 418a1e78772SMel Gorman 41984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 42084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 42104f2cbe3SMel Gorman } 42204f2cbe3SMel Gorman 42304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 42404f2cbe3SMel Gorman { 42504f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 426f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 427e7c4b0bfSAndy Whitcroft 428e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 42904f2cbe3SMel Gorman } 43004f2cbe3SMel Gorman 43104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 43204f2cbe3SMel Gorman { 43304f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 434e7c4b0bfSAndy Whitcroft 435e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 436a1e78772SMel Gorman } 437a1e78772SMel Gorman 43804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 439a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 440a1e78772SMel Gorman { 441a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 442f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 443a1e78772SMel Gorman vma->vm_private_data = (void *)0; 444a1e78772SMel Gorman } 445a1e78772SMel Gorman 446a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 447af0ed73eSJoonsoo Kim static int vma_has_reserves(struct vm_area_struct *vma, long chg) 448a1e78772SMel Gorman { 449af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 450af0ed73eSJoonsoo Kim /* 451af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 452af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 453af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 454af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 455af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 456af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 457af0ed73eSJoonsoo Kim * properly, so add work-around here. 458af0ed73eSJoonsoo Kim */ 459af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 460af0ed73eSJoonsoo Kim return 1; 461af0ed73eSJoonsoo Kim else 46272231b03SJoonsoo Kim return 0; 463af0ed73eSJoonsoo Kim } 464a63884e9SJoonsoo Kim 465a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 466f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 467a1e78772SMel Gorman return 1; 468a63884e9SJoonsoo Kim 469a63884e9SJoonsoo Kim /* 470a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 471a63884e9SJoonsoo Kim * private mappings. 472a63884e9SJoonsoo Kim */ 4737f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4747f09ca51SMel Gorman return 1; 475a63884e9SJoonsoo Kim 4767f09ca51SMel Gorman return 0; 477a1e78772SMel Gorman } 478a1e78772SMel Gorman 4790ebabb41SNaoya Horiguchi static void copy_gigantic_page(struct page *dst, struct page *src) 4800ebabb41SNaoya Horiguchi { 4810ebabb41SNaoya Horiguchi int i; 4820ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4830ebabb41SNaoya Horiguchi struct page *dst_base = dst; 4840ebabb41SNaoya Horiguchi struct page *src_base = src; 4850ebabb41SNaoya Horiguchi 4860ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); ) { 4870ebabb41SNaoya Horiguchi cond_resched(); 4880ebabb41SNaoya Horiguchi copy_highpage(dst, src); 4890ebabb41SNaoya Horiguchi 4900ebabb41SNaoya Horiguchi i++; 4910ebabb41SNaoya Horiguchi dst = mem_map_next(dst, dst_base, i); 4920ebabb41SNaoya Horiguchi src = mem_map_next(src, src_base, i); 4930ebabb41SNaoya Horiguchi } 4940ebabb41SNaoya Horiguchi } 4950ebabb41SNaoya Horiguchi 4960ebabb41SNaoya Horiguchi void copy_huge_page(struct page *dst, struct page *src) 4970ebabb41SNaoya Horiguchi { 4980ebabb41SNaoya Horiguchi int i; 4990ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 5000ebabb41SNaoya Horiguchi 5010ebabb41SNaoya Horiguchi if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 5020ebabb41SNaoya Horiguchi copy_gigantic_page(dst, src); 5030ebabb41SNaoya Horiguchi return; 5040ebabb41SNaoya Horiguchi } 5050ebabb41SNaoya Horiguchi 5060ebabb41SNaoya Horiguchi might_sleep(); 5070ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); i++) { 5080ebabb41SNaoya Horiguchi cond_resched(); 5090ebabb41SNaoya Horiguchi copy_highpage(dst + i, src + i); 5100ebabb41SNaoya Horiguchi } 5110ebabb41SNaoya Horiguchi } 5120ebabb41SNaoya Horiguchi 513a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 5141da177e4SLinus Torvalds { 5151da177e4SLinus Torvalds int nid = page_to_nid(page); 5160edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 517a5516438SAndi Kleen h->free_huge_pages++; 518a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 5191da177e4SLinus Torvalds } 5201da177e4SLinus Torvalds 521bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 522bf50bab2SNaoya Horiguchi { 523bf50bab2SNaoya Horiguchi struct page *page; 524bf50bab2SNaoya Horiguchi 525bf50bab2SNaoya Horiguchi if (list_empty(&h->hugepage_freelists[nid])) 526bf50bab2SNaoya Horiguchi return NULL; 527bf50bab2SNaoya Horiguchi page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); 5280edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 529a9869b83SNaoya Horiguchi set_page_refcounted(page); 530bf50bab2SNaoya Horiguchi h->free_huge_pages--; 531bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 532bf50bab2SNaoya Horiguchi return page; 533bf50bab2SNaoya Horiguchi } 534bf50bab2SNaoya Horiguchi 535a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 536a5516438SAndi Kleen struct vm_area_struct *vma, 537af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 538af0ed73eSJoonsoo Kim long chg) 5391da177e4SLinus Torvalds { 540b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 541480eccf9SLee Schermerhorn struct mempolicy *mpol; 54219770b32SMel Gorman nodemask_t *nodemask; 543c0ff7453SMiao Xie struct zonelist *zonelist; 544dd1a239fSMel Gorman struct zone *zone; 545dd1a239fSMel Gorman struct zoneref *z; 546cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 5471da177e4SLinus Torvalds 548a1e78772SMel Gorman /* 549a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 550a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 551a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 552a1e78772SMel Gorman */ 553af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 554a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 555c0ff7453SMiao Xie goto err; 556a1e78772SMel Gorman 55704f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 558a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 5596eab04a8SJustin P. Mattock goto err; 56004f2cbe3SMel Gorman 5619966c4bbSJoonsoo Kim retry_cpuset: 5629966c4bbSJoonsoo Kim cpuset_mems_cookie = get_mems_allowed(); 5639966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 5649966c4bbSJoonsoo Kim htlb_alloc_mask, &mpol, &nodemask); 5659966c4bbSJoonsoo Kim 56619770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 56719770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 568bf50bab2SNaoya Horiguchi if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { 569bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 570bf50bab2SNaoya Horiguchi if (page) { 571af0ed73eSJoonsoo Kim if (avoid_reserve) 572af0ed73eSJoonsoo Kim break; 573af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 574af0ed73eSJoonsoo Kim break; 575af0ed73eSJoonsoo Kim 57607443a85SJoonsoo Kim SetPagePrivate(page); 577a63884e9SJoonsoo Kim h->resv_huge_pages--; 5785ab3ee7bSKen Chen break; 5791da177e4SLinus Torvalds } 5803abf7afdSAndrew Morton } 581bf50bab2SNaoya Horiguchi } 582cc9a6c87SMel Gorman 583cc9a6c87SMel Gorman mpol_cond_put(mpol); 584cc9a6c87SMel Gorman if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) 585cc9a6c87SMel Gorman goto retry_cpuset; 586cc9a6c87SMel Gorman return page; 587cc9a6c87SMel Gorman 588c0ff7453SMiao Xie err: 589cc9a6c87SMel Gorman return NULL; 5901da177e4SLinus Torvalds } 5911da177e4SLinus Torvalds 592a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5936af2acb6SAdam Litke { 5946af2acb6SAdam Litke int i; 595a5516438SAndi Kleen 59618229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 59718229df5SAndy Whitcroft 598a5516438SAndi Kleen h->nr_huge_pages--; 599a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 600a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 60132f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 60232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 60332f84528SChris Forbes 1 << PG_active | 1 << PG_reserved | 6046af2acb6SAdam Litke 1 << PG_private | 1 << PG_writeback); 6056af2acb6SAdam Litke } 6069dd540e2SAneesh Kumar K.V VM_BUG_ON(hugetlb_cgroup_from_page(page)); 6076af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 6086af2acb6SAdam Litke set_page_refcounted(page); 6097f2e9525SGerald Schaefer arch_release_hugepage(page); 610a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 6116af2acb6SAdam Litke } 6126af2acb6SAdam Litke 613e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 614e5ff2159SAndi Kleen { 615e5ff2159SAndi Kleen struct hstate *h; 616e5ff2159SAndi Kleen 617e5ff2159SAndi Kleen for_each_hstate(h) { 618e5ff2159SAndi Kleen if (huge_page_size(h) == size) 619e5ff2159SAndi Kleen return h; 620e5ff2159SAndi Kleen } 621e5ff2159SAndi Kleen return NULL; 622e5ff2159SAndi Kleen } 623e5ff2159SAndi Kleen 62427a85ef1SDavid Gibson static void free_huge_page(struct page *page) 62527a85ef1SDavid Gibson { 626a5516438SAndi Kleen /* 627a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 628a5516438SAndi Kleen * compound page destructor. 629a5516438SAndi Kleen */ 630e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 6317893d1d5SAdam Litke int nid = page_to_nid(page); 63290481622SDavid Gibson struct hugepage_subpool *spool = 63390481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 63407443a85SJoonsoo Kim bool restore_reserve; 63527a85ef1SDavid Gibson 636e5df70abSAndy Whitcroft set_page_private(page, 0); 63723be7468SMel Gorman page->mapping = NULL; 6387893d1d5SAdam Litke BUG_ON(page_count(page)); 6390fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 64007443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 64127a85ef1SDavid Gibson 64227a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 6436d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 6446d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 64507443a85SJoonsoo Kim if (restore_reserve) 64607443a85SJoonsoo Kim h->resv_huge_pages++; 64707443a85SJoonsoo Kim 648aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 6490edaecfaSAneesh Kumar K.V /* remove the page from active list */ 6500edaecfaSAneesh Kumar K.V list_del(&page->lru); 651a5516438SAndi Kleen update_and_free_page(h, page); 652a5516438SAndi Kleen h->surplus_huge_pages--; 653a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 6547893d1d5SAdam Litke } else { 6555d3a551cSWill Deacon arch_clear_hugepage_flags(page); 656a5516438SAndi Kleen enqueue_huge_page(h, page); 6577893d1d5SAdam Litke } 65827a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 65990481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 66027a85ef1SDavid Gibson } 66127a85ef1SDavid Gibson 662a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 663b7ba30c6SAndi Kleen { 6640edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 665b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 666b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 6679dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 668a5516438SAndi Kleen h->nr_huge_pages++; 669a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 670b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 671b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 672b7ba30c6SAndi Kleen } 673b7ba30c6SAndi Kleen 67420a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 67520a0307cSWu Fengguang { 67620a0307cSWu Fengguang int i; 67720a0307cSWu Fengguang int nr_pages = 1 << order; 67820a0307cSWu Fengguang struct page *p = page + 1; 67920a0307cSWu Fengguang 68020a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 68120a0307cSWu Fengguang set_compound_order(page, order); 68220a0307cSWu Fengguang __SetPageHead(page); 68320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 68420a0307cSWu Fengguang __SetPageTail(p); 68558a84aa9SYouquan Song set_page_count(p, 0); 68620a0307cSWu Fengguang p->first_page = page; 68720a0307cSWu Fengguang } 68820a0307cSWu Fengguang } 68920a0307cSWu Fengguang 6907795912cSAndrew Morton /* 6917795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 6927795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 6937795912cSAndrew Morton * details. 6947795912cSAndrew Morton */ 69520a0307cSWu Fengguang int PageHuge(struct page *page) 69620a0307cSWu Fengguang { 69720a0307cSWu Fengguang compound_page_dtor *dtor; 69820a0307cSWu Fengguang 69920a0307cSWu Fengguang if (!PageCompound(page)) 70020a0307cSWu Fengguang return 0; 70120a0307cSWu Fengguang 70220a0307cSWu Fengguang page = compound_head(page); 70320a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 70420a0307cSWu Fengguang 70520a0307cSWu Fengguang return dtor == free_huge_page; 70620a0307cSWu Fengguang } 70743131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 70843131e14SNaoya Horiguchi 70913d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 71013d60f4bSZhang Yi { 71113d60f4bSZhang Yi struct page *page_head = compound_head(page); 71213d60f4bSZhang Yi pgoff_t index = page_index(page_head); 71313d60f4bSZhang Yi unsigned long compound_idx; 71413d60f4bSZhang Yi 71513d60f4bSZhang Yi if (!PageHuge(page_head)) 71613d60f4bSZhang Yi return page_index(page); 71713d60f4bSZhang Yi 71813d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 71913d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 72013d60f4bSZhang Yi else 72113d60f4bSZhang Yi compound_idx = page - page_head; 72213d60f4bSZhang Yi 72313d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 72413d60f4bSZhang Yi } 72513d60f4bSZhang Yi 726a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 7271da177e4SLinus Torvalds { 7281da177e4SLinus Torvalds struct page *page; 729f96efd58SJoe Jin 730aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 731aa888a74SAndi Kleen return NULL; 732aa888a74SAndi Kleen 7336484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 734551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 735551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 736a5516438SAndi Kleen huge_page_order(h)); 7371da177e4SLinus Torvalds if (page) { 7387f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 739caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 7407b8ee84dSHarvey Harrison return NULL; 7417f2e9525SGerald Schaefer } 742a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 7431da177e4SLinus Torvalds } 74463b4613cSNishanth Aravamudan 74563b4613cSNishanth Aravamudan return page; 74663b4613cSNishanth Aravamudan } 74763b4613cSNishanth Aravamudan 7485ced66c9SAndi Kleen /* 7496ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 7506ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 7516ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7526ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 7536ae11b27SLee Schermerhorn * node for alloc or free. 7549a76db09SLee Schermerhorn */ 7556ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7569a76db09SLee Schermerhorn { 7576ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 7589a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 7596ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 7609a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 7619a76db09SLee Schermerhorn 7629a76db09SLee Schermerhorn return nid; 7639a76db09SLee Schermerhorn } 7649a76db09SLee Schermerhorn 7656ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 7665ced66c9SAndi Kleen { 7676ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 7686ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 7699a76db09SLee Schermerhorn return nid; 7705ced66c9SAndi Kleen } 7715ced66c9SAndi Kleen 7726ae11b27SLee Schermerhorn /* 7736ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 7746ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 7756ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 7766ae11b27SLee Schermerhorn * mask. 7776ae11b27SLee Schermerhorn */ 7786ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 7796ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 7806ae11b27SLee Schermerhorn { 7816ae11b27SLee Schermerhorn int nid; 7826ae11b27SLee Schermerhorn 7836ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7846ae11b27SLee Schermerhorn 7856ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 7866ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 7876ae11b27SLee Schermerhorn 7886ae11b27SLee Schermerhorn return nid; 7896ae11b27SLee Schermerhorn } 7906ae11b27SLee Schermerhorn 791e8c5c824SLee Schermerhorn /* 7926ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7936ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7946ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7956ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 796e8c5c824SLee Schermerhorn */ 7976ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 798e8c5c824SLee Schermerhorn { 7996ae11b27SLee Schermerhorn int nid; 8009a76db09SLee Schermerhorn 8016ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 8026ae11b27SLee Schermerhorn 8036ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 8046ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 8056ae11b27SLee Schermerhorn 8069a76db09SLee Schermerhorn return nid; 807e8c5c824SLee Schermerhorn } 808e8c5c824SLee Schermerhorn 809b2261026SJoonsoo Kim #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 810b2261026SJoonsoo Kim for (nr_nodes = nodes_weight(*mask); \ 811b2261026SJoonsoo Kim nr_nodes > 0 && \ 812b2261026SJoonsoo Kim ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 813b2261026SJoonsoo Kim nr_nodes--) 814b2261026SJoonsoo Kim 815b2261026SJoonsoo Kim #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 816b2261026SJoonsoo Kim for (nr_nodes = nodes_weight(*mask); \ 817b2261026SJoonsoo Kim nr_nodes > 0 && \ 818b2261026SJoonsoo Kim ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 819b2261026SJoonsoo Kim nr_nodes--) 820b2261026SJoonsoo Kim 821b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 822b2261026SJoonsoo Kim { 823b2261026SJoonsoo Kim struct page *page; 824b2261026SJoonsoo Kim int nr_nodes, node; 825b2261026SJoonsoo Kim int ret = 0; 826b2261026SJoonsoo Kim 827b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 828b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 829b2261026SJoonsoo Kim if (page) { 830b2261026SJoonsoo Kim ret = 1; 831b2261026SJoonsoo Kim break; 832b2261026SJoonsoo Kim } 833b2261026SJoonsoo Kim } 834b2261026SJoonsoo Kim 835b2261026SJoonsoo Kim if (ret) 836b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 837b2261026SJoonsoo Kim else 838b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 839b2261026SJoonsoo Kim 840b2261026SJoonsoo Kim return ret; 841b2261026SJoonsoo Kim } 842b2261026SJoonsoo Kim 843e8c5c824SLee Schermerhorn /* 844e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 845e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 846e8c5c824SLee Schermerhorn * balanced over allowed nodes. 847e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 848e8c5c824SLee Schermerhorn */ 8496ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 8506ae11b27SLee Schermerhorn bool acct_surplus) 851e8c5c824SLee Schermerhorn { 852b2261026SJoonsoo Kim int nr_nodes, node; 853e8c5c824SLee Schermerhorn int ret = 0; 854e8c5c824SLee Schermerhorn 855b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 856685f3457SLee Schermerhorn /* 857685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 858685f3457SLee Schermerhorn * nodes with surplus pages. 859685f3457SLee Schermerhorn */ 860b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 861b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 862e8c5c824SLee Schermerhorn struct page *page = 863b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 864e8c5c824SLee Schermerhorn struct page, lru); 865e8c5c824SLee Schermerhorn list_del(&page->lru); 866e8c5c824SLee Schermerhorn h->free_huge_pages--; 867b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 868685f3457SLee Schermerhorn if (acct_surplus) { 869685f3457SLee Schermerhorn h->surplus_huge_pages--; 870b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 871685f3457SLee Schermerhorn } 872e8c5c824SLee Schermerhorn update_and_free_page(h, page); 873e8c5c824SLee Schermerhorn ret = 1; 8749a76db09SLee Schermerhorn break; 875e8c5c824SLee Schermerhorn } 876b2261026SJoonsoo Kim } 877e8c5c824SLee Schermerhorn 878e8c5c824SLee Schermerhorn return ret; 879e8c5c824SLee Schermerhorn } 880e8c5c824SLee Schermerhorn 881bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 8827893d1d5SAdam Litke { 8837893d1d5SAdam Litke struct page *page; 884bf50bab2SNaoya Horiguchi unsigned int r_nid; 8857893d1d5SAdam Litke 886aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 887aa888a74SAndi Kleen return NULL; 888aa888a74SAndi Kleen 889d1c3fb1fSNishanth Aravamudan /* 890d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 891d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 892d1c3fb1fSNishanth Aravamudan * overcommit 893d1c3fb1fSNishanth Aravamudan * 894d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 895d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 896d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 897d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 898d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 899d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 900d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 901d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 902d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 903d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 904d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 905d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 906d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 907d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 908d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 909d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 910d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 911d1c3fb1fSNishanth Aravamudan */ 912d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 913a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 914d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 915d1c3fb1fSNishanth Aravamudan return NULL; 916d1c3fb1fSNishanth Aravamudan } else { 917a5516438SAndi Kleen h->nr_huge_pages++; 918a5516438SAndi Kleen h->surplus_huge_pages++; 919d1c3fb1fSNishanth Aravamudan } 920d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 921d1c3fb1fSNishanth Aravamudan 922bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 923551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 924551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 925a5516438SAndi Kleen huge_page_order(h)); 926bf50bab2SNaoya Horiguchi else 927bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 928bf50bab2SNaoya Horiguchi htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 929bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 930d1c3fb1fSNishanth Aravamudan 931caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 932caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 933ea5768c7SHillf Danton page = NULL; 934caff3a2cSGerald Schaefer } 935caff3a2cSGerald Schaefer 9367893d1d5SAdam Litke spin_lock(&hugetlb_lock); 937d1c3fb1fSNishanth Aravamudan if (page) { 9380edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 939bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 940d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 9419dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 942d1c3fb1fSNishanth Aravamudan /* 943d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 944d1c3fb1fSNishanth Aravamudan */ 945bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 946bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 9473b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 948d1c3fb1fSNishanth Aravamudan } else { 949a5516438SAndi Kleen h->nr_huge_pages--; 950a5516438SAndi Kleen h->surplus_huge_pages--; 9513b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 9527893d1d5SAdam Litke } 953d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 9547893d1d5SAdam Litke 9557893d1d5SAdam Litke return page; 9567893d1d5SAdam Litke } 9577893d1d5SAdam Litke 958e4e574b7SAdam Litke /* 959bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 960bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 961bf50bab2SNaoya Horiguchi * address of error page. 962bf50bab2SNaoya Horiguchi */ 963bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 964bf50bab2SNaoya Horiguchi { 9654ef91848SJoonsoo Kim struct page *page = NULL; 966bf50bab2SNaoya Horiguchi 967bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 9684ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 969bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 970bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 971bf50bab2SNaoya Horiguchi 97294ae8ba7SAneesh Kumar K.V if (!page) 973bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 974bf50bab2SNaoya Horiguchi 975bf50bab2SNaoya Horiguchi return page; 976bf50bab2SNaoya Horiguchi } 977bf50bab2SNaoya Horiguchi 978bf50bab2SNaoya Horiguchi /* 97925985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 980e4e574b7SAdam Litke * of size 'delta'. 981e4e574b7SAdam Litke */ 982a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 983e4e574b7SAdam Litke { 984e4e574b7SAdam Litke struct list_head surplus_list; 985e4e574b7SAdam Litke struct page *page, *tmp; 986e4e574b7SAdam Litke int ret, i; 987e4e574b7SAdam Litke int needed, allocated; 98828073b02SHillf Danton bool alloc_ok = true; 989e4e574b7SAdam Litke 990a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 991ac09b3a1SAdam Litke if (needed <= 0) { 992a5516438SAndi Kleen h->resv_huge_pages += delta; 993e4e574b7SAdam Litke return 0; 994ac09b3a1SAdam Litke } 995e4e574b7SAdam Litke 996e4e574b7SAdam Litke allocated = 0; 997e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 998e4e574b7SAdam Litke 999e4e574b7SAdam Litke ret = -ENOMEM; 1000e4e574b7SAdam Litke retry: 1001e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1002e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1003bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 100428073b02SHillf Danton if (!page) { 100528073b02SHillf Danton alloc_ok = false; 100628073b02SHillf Danton break; 100728073b02SHillf Danton } 1008e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1009e4e574b7SAdam Litke } 101028073b02SHillf Danton allocated += i; 1011e4e574b7SAdam Litke 1012e4e574b7SAdam Litke /* 1013e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1014e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1015e4e574b7SAdam Litke */ 1016e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1017a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1018a5516438SAndi Kleen (h->free_huge_pages + allocated); 101928073b02SHillf Danton if (needed > 0) { 102028073b02SHillf Danton if (alloc_ok) 1021e4e574b7SAdam Litke goto retry; 102228073b02SHillf Danton /* 102328073b02SHillf Danton * We were not able to allocate enough pages to 102428073b02SHillf Danton * satisfy the entire reservation so we free what 102528073b02SHillf Danton * we've allocated so far. 102628073b02SHillf Danton */ 102728073b02SHillf Danton goto free; 102828073b02SHillf Danton } 1029e4e574b7SAdam Litke /* 1030e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 103125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1032e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1033ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1034ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1035ac09b3a1SAdam Litke * before they are reserved. 1036e4e574b7SAdam Litke */ 1037e4e574b7SAdam Litke needed += allocated; 1038a5516438SAndi Kleen h->resv_huge_pages += delta; 1039e4e574b7SAdam Litke ret = 0; 1040a9869b83SNaoya Horiguchi 104119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 104219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 104319fc3f0aSAdam Litke if ((--needed) < 0) 104419fc3f0aSAdam Litke break; 1045a9869b83SNaoya Horiguchi /* 1046a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1047a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1048a9869b83SNaoya Horiguchi */ 1049a9869b83SNaoya Horiguchi put_page_testzero(page); 1050a9869b83SNaoya Horiguchi VM_BUG_ON(page_count(page)); 1051a5516438SAndi Kleen enqueue_huge_page(h, page); 105219fc3f0aSAdam Litke } 105328073b02SHillf Danton free: 1054b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 105519fc3f0aSAdam Litke 105619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1057c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1058a9869b83SNaoya Horiguchi put_page(page); 105919fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1060e4e574b7SAdam Litke 1061e4e574b7SAdam Litke return ret; 1062e4e574b7SAdam Litke } 1063e4e574b7SAdam Litke 1064e4e574b7SAdam Litke /* 1065e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1066e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1067e4e574b7SAdam Litke * never used. 1068685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1069e4e574b7SAdam Litke */ 1070a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1071a5516438SAndi Kleen unsigned long unused_resv_pages) 1072e4e574b7SAdam Litke { 1073e4e574b7SAdam Litke unsigned long nr_pages; 1074e4e574b7SAdam Litke 1075ac09b3a1SAdam Litke /* Uncommit the reservation */ 1076a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1077ac09b3a1SAdam Litke 1078aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1079aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1080aa888a74SAndi Kleen return; 1081aa888a74SAndi Kleen 1082a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1083e4e574b7SAdam Litke 1084685f3457SLee Schermerhorn /* 1085685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 10869b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 10879b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 10889b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 10899b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 10909b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1091685f3457SLee Schermerhorn */ 1092685f3457SLee Schermerhorn while (nr_pages--) { 10938cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1094685f3457SLee Schermerhorn break; 1095e4e574b7SAdam Litke } 1096e4e574b7SAdam Litke } 1097e4e574b7SAdam Litke 1098c37f9fb1SAndy Whitcroft /* 1099c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1100c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 110190481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 110290481622SDavid Gibson * can occur. Where any new reservation would be required the 110390481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 110490481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 110590481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 110690481622SDavid Gibson * failure. 1107c37f9fb1SAndy Whitcroft */ 1108e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1109a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1110c37f9fb1SAndy Whitcroft { 1111c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1112c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1113c37f9fb1SAndy Whitcroft 1114f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1115a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1116c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 1117c37f9fb1SAndy Whitcroft idx, idx + 1); 1118c37f9fb1SAndy Whitcroft 111984afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1120c37f9fb1SAndy Whitcroft return 1; 1121c37f9fb1SAndy Whitcroft 112284afd99bSAndy Whitcroft } else { 1123e2f17d94SRoel Kluin long err; 1124a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1125f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 112684afd99bSAndy Whitcroft 1127f522c3acSJoonsoo Kim err = region_chg(&resv->regions, idx, idx + 1); 112884afd99bSAndy Whitcroft if (err < 0) 112984afd99bSAndy Whitcroft return err; 1130c37f9fb1SAndy Whitcroft return 0; 1131c37f9fb1SAndy Whitcroft } 113284afd99bSAndy Whitcroft } 1133a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1134a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1135c37f9fb1SAndy Whitcroft { 1136c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1137c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1138c37f9fb1SAndy Whitcroft 1139f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1140a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1141c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 114284afd99bSAndy Whitcroft 114384afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1144a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1145f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 114684afd99bSAndy Whitcroft 114784afd99bSAndy Whitcroft /* Mark this page used in the map. */ 1148f522c3acSJoonsoo Kim region_add(&resv->regions, idx, idx + 1); 1149c37f9fb1SAndy Whitcroft } 1150c37f9fb1SAndy Whitcroft } 1151c37f9fb1SAndy Whitcroft 1152348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 115304f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1154348ea204SAdam Litke { 115590481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1156a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1157348ea204SAdam Litke struct page *page; 1158e2f17d94SRoel Kluin long chg; 11596d76dcf4SAneesh Kumar K.V int ret, idx; 11606d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 11612fc39cecSAdam Litke 11626d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1163a1e78772SMel Gorman /* 116490481622SDavid Gibson * Processes that did not create the mapping will have no 116590481622SDavid Gibson * reserves and will not have accounted against subpool 116690481622SDavid Gibson * limit. Check that the subpool limit can be made before 116790481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 116890481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 116990481622SDavid Gibson * mapping overlaps. 1170a1e78772SMel Gorman */ 1171a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1172c37f9fb1SAndy Whitcroft if (chg < 0) 117376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 11748bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 11758bb3f12eSJoonsoo Kim if (hugepage_subpool_get_pages(spool, 1)) 117676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 117790d8b7e6SAdam Litke 11786d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 11796d76dcf4SAneesh Kumar K.V if (ret) { 11808bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 11818bb3f12eSJoonsoo Kim hugepage_subpool_put_pages(spool, 1); 11826d76dcf4SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 11836d76dcf4SAneesh Kumar K.V } 1184a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1185af0ed73eSJoonsoo Kim page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 118681a6fcaeSJoonsoo Kim if (!page) { 118794ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1188bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1189a1e78772SMel Gorman if (!page) { 11906d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_cgroup(idx, 11916d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), 11926d76dcf4SAneesh Kumar K.V h_cg); 11938bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 11948bb3f12eSJoonsoo Kim hugepage_subpool_put_pages(spool, 1); 119576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 1196a1e78772SMel Gorman } 119779dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 119879dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 119981a6fcaeSJoonsoo Kim /* Fall through */ 1200a1e78772SMel Gorman } 120181a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 120281a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1203a1e78772SMel Gorman 120490481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1205a1e78772SMel Gorman 1206a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 12077893d1d5SAdam Litke return page; 1208b45b5bd6SDavid Gibson } 1209b45b5bd6SDavid Gibson 121091f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1211aa888a74SAndi Kleen { 1212aa888a74SAndi Kleen struct huge_bootmem_page *m; 1213b2261026SJoonsoo Kim int nr_nodes, node; 1214aa888a74SAndi Kleen 1215b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1216aa888a74SAndi Kleen void *addr; 1217aa888a74SAndi Kleen 1218b2261026SJoonsoo Kim addr = __alloc_bootmem_node_nopanic(NODE_DATA(node), 1219aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1220aa888a74SAndi Kleen 1221aa888a74SAndi Kleen if (addr) { 1222aa888a74SAndi Kleen /* 1223aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1224aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1225aa888a74SAndi Kleen * puts them into the mem_map). 1226aa888a74SAndi Kleen */ 1227aa888a74SAndi Kleen m = addr; 1228aa888a74SAndi Kleen goto found; 1229aa888a74SAndi Kleen } 1230aa888a74SAndi Kleen } 1231aa888a74SAndi Kleen return 0; 1232aa888a74SAndi Kleen 1233aa888a74SAndi Kleen found: 1234aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1235aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1236aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1237aa888a74SAndi Kleen m->hstate = h; 1238aa888a74SAndi Kleen return 1; 1239aa888a74SAndi Kleen } 1240aa888a74SAndi Kleen 124118229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 124218229df5SAndy Whitcroft { 124318229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 124418229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 124518229df5SAndy Whitcroft else 124618229df5SAndy Whitcroft prep_compound_page(page, order); 124718229df5SAndy Whitcroft } 124818229df5SAndy Whitcroft 1249aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1250aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1251aa888a74SAndi Kleen { 1252aa888a74SAndi Kleen struct huge_bootmem_page *m; 1253aa888a74SAndi Kleen 1254aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1255aa888a74SAndi Kleen struct hstate *h = m->hstate; 1256ee8f248dSBecky Bruce struct page *page; 1257ee8f248dSBecky Bruce 1258ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1259ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 1260ee8f248dSBecky Bruce free_bootmem_late((unsigned long)m, 1261ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1262ee8f248dSBecky Bruce #else 1263ee8f248dSBecky Bruce page = virt_to_page(m); 1264ee8f248dSBecky Bruce #endif 1265aa888a74SAndi Kleen __ClearPageReserved(page); 1266aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 126718229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1268aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1269b0320c7bSRafael Aquini /* 1270b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1271b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1272b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1273b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1274b0320c7bSRafael Aquini */ 1275b0320c7bSRafael Aquini if (h->order > (MAX_ORDER - 1)) 12763dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 1277aa888a74SAndi Kleen } 1278aa888a74SAndi Kleen } 1279aa888a74SAndi Kleen 12808faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 12811da177e4SLinus Torvalds { 12821da177e4SLinus Torvalds unsigned long i; 12831da177e4SLinus Torvalds 1284e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1285aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1286aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1287aa888a74SAndi Kleen break; 12889b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 12898cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 12901da177e4SLinus Torvalds break; 12911da177e4SLinus Torvalds } 12928faa8b07SAndi Kleen h->max_huge_pages = i; 1293e5ff2159SAndi Kleen } 1294e5ff2159SAndi Kleen 1295e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1296e5ff2159SAndi Kleen { 1297e5ff2159SAndi Kleen struct hstate *h; 1298e5ff2159SAndi Kleen 1299e5ff2159SAndi Kleen for_each_hstate(h) { 13008faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 13018faa8b07SAndi Kleen if (h->order < MAX_ORDER) 13028faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1303e5ff2159SAndi Kleen } 1304e5ff2159SAndi Kleen } 1305e5ff2159SAndi Kleen 13064abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 13074abd32dbSAndi Kleen { 13084abd32dbSAndi Kleen if (n >= (1UL << 30)) 13094abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 13104abd32dbSAndi Kleen else if (n >= (1UL << 20)) 13114abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 13124abd32dbSAndi Kleen else 13134abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 13144abd32dbSAndi Kleen return buf; 13154abd32dbSAndi Kleen } 13164abd32dbSAndi Kleen 1317e5ff2159SAndi Kleen static void __init report_hugepages(void) 1318e5ff2159SAndi Kleen { 1319e5ff2159SAndi Kleen struct hstate *h; 1320e5ff2159SAndi Kleen 1321e5ff2159SAndi Kleen for_each_hstate(h) { 13224abd32dbSAndi Kleen char buf[32]; 1323ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 13244abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 13254abd32dbSAndi Kleen h->free_huge_pages); 1326e5ff2159SAndi Kleen } 1327e5ff2159SAndi Kleen } 1328e5ff2159SAndi Kleen 13291da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 13306ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 13316ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13321da177e4SLinus Torvalds { 13334415cc8dSChristoph Lameter int i; 13344415cc8dSChristoph Lameter 1335aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1336aa888a74SAndi Kleen return; 1337aa888a74SAndi Kleen 13386ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 13391da177e4SLinus Torvalds struct page *page, *next; 1340a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1341a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1342a5516438SAndi Kleen if (count >= h->nr_huge_pages) 13436b0c880dSAdam Litke return; 13441da177e4SLinus Torvalds if (PageHighMem(page)) 13451da177e4SLinus Torvalds continue; 13461da177e4SLinus Torvalds list_del(&page->lru); 1347e5ff2159SAndi Kleen update_and_free_page(h, page); 1348a5516438SAndi Kleen h->free_huge_pages--; 1349a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 13501da177e4SLinus Torvalds } 13511da177e4SLinus Torvalds } 13521da177e4SLinus Torvalds } 13531da177e4SLinus Torvalds #else 13546ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 13556ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13561da177e4SLinus Torvalds { 13571da177e4SLinus Torvalds } 13581da177e4SLinus Torvalds #endif 13591da177e4SLinus Torvalds 136020a0307cSWu Fengguang /* 136120a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 136220a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 136320a0307cSWu Fengguang * Returns 1 if an adjustment was made. 136420a0307cSWu Fengguang */ 13656ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 13666ae11b27SLee Schermerhorn int delta) 136720a0307cSWu Fengguang { 1368b2261026SJoonsoo Kim int nr_nodes, node; 136920a0307cSWu Fengguang 137020a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 137120a0307cSWu Fengguang 1372e8c5c824SLee Schermerhorn if (delta < 0) { 1373b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1374b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 1375b2261026SJoonsoo Kim goto found; 1376b2261026SJoonsoo Kim } 1377b2261026SJoonsoo Kim } else { 1378b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1379b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 1380b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 1381b2261026SJoonsoo Kim goto found; 1382e8c5c824SLee Schermerhorn } 13839a76db09SLee Schermerhorn } 1384b2261026SJoonsoo Kim return 0; 138520a0307cSWu Fengguang 1386b2261026SJoonsoo Kim found: 138720a0307cSWu Fengguang h->surplus_huge_pages += delta; 1388b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 1389b2261026SJoonsoo Kim return 1; 139020a0307cSWu Fengguang } 139120a0307cSWu Fengguang 1392a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 13936ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 13946ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13951da177e4SLinus Torvalds { 13967893d1d5SAdam Litke unsigned long min_count, ret; 13971da177e4SLinus Torvalds 1398aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1399aa888a74SAndi Kleen return h->max_huge_pages; 1400aa888a74SAndi Kleen 14017893d1d5SAdam Litke /* 14027893d1d5SAdam Litke * Increase the pool size 14037893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 14047893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1405d1c3fb1fSNishanth Aravamudan * 1406d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1407d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1408d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1409d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1410d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 14117893d1d5SAdam Litke */ 14121da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1413a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 14146ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 14157893d1d5SAdam Litke break; 14167893d1d5SAdam Litke } 14177893d1d5SAdam Litke 1418a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 14197893d1d5SAdam Litke /* 14207893d1d5SAdam Litke * If this allocation races such that we no longer need the 14217893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 14227893d1d5SAdam Litke * and reducing the surplus. 14237893d1d5SAdam Litke */ 14247893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 14256ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 14267893d1d5SAdam Litke spin_lock(&hugetlb_lock); 14277893d1d5SAdam Litke if (!ret) 14287893d1d5SAdam Litke goto out; 14297893d1d5SAdam Litke 1430536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1431536240f2SMel Gorman if (signal_pending(current)) 1432536240f2SMel Gorman goto out; 14337893d1d5SAdam Litke } 14347893d1d5SAdam Litke 14357893d1d5SAdam Litke /* 14367893d1d5SAdam Litke * Decrease the pool size 14377893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 14387893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 14397893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 14407893d1d5SAdam Litke * to the desired size as pages become free. 1441d1c3fb1fSNishanth Aravamudan * 1442d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1443d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1444d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1445d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1446d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1447d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1448d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 14497893d1d5SAdam Litke */ 1450a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 14516b0c880dSAdam Litke min_count = max(count, min_count); 14526ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1453a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 14546ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 14551da177e4SLinus Torvalds break; 14561da177e4SLinus Torvalds } 1457a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 14586ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 14597893d1d5SAdam Litke break; 14607893d1d5SAdam Litke } 14617893d1d5SAdam Litke out: 1462a5516438SAndi Kleen ret = persistent_huge_pages(h); 14631da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 14647893d1d5SAdam Litke return ret; 14651da177e4SLinus Torvalds } 14661da177e4SLinus Torvalds 1467a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1468a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1469a3437870SNishanth Aravamudan 1470a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1471a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1472a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1473a3437870SNishanth Aravamudan 1474a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1475a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1476a3437870SNishanth Aravamudan 14779a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 14789a305230SLee Schermerhorn 14799a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1480a3437870SNishanth Aravamudan { 1481a3437870SNishanth Aravamudan int i; 14829a305230SLee Schermerhorn 1483a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 14849a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 14859a305230SLee Schermerhorn if (nidp) 14869a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1487a3437870SNishanth Aravamudan return &hstates[i]; 14889a305230SLee Schermerhorn } 14899a305230SLee Schermerhorn 14909a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1491a3437870SNishanth Aravamudan } 1492a3437870SNishanth Aravamudan 149306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1494a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1495a3437870SNishanth Aravamudan { 14969a305230SLee Schermerhorn struct hstate *h; 14979a305230SLee Schermerhorn unsigned long nr_huge_pages; 14989a305230SLee Schermerhorn int nid; 14999a305230SLee Schermerhorn 15009a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15019a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15029a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 15039a305230SLee Schermerhorn else 15049a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 15059a305230SLee Schermerhorn 15069a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1507a3437870SNishanth Aravamudan } 1508adbe8726SEric B Munson 150906808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 151006808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 151106808b08SLee Schermerhorn const char *buf, size_t len) 1512a3437870SNishanth Aravamudan { 1513a3437870SNishanth Aravamudan int err; 15149a305230SLee Schermerhorn int nid; 151506808b08SLee Schermerhorn unsigned long count; 15169a305230SLee Schermerhorn struct hstate *h; 1517bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1518a3437870SNishanth Aravamudan 15193dbb95f7SJingoo Han err = kstrtoul(buf, 10, &count); 152073ae31e5SEric B Munson if (err) 1521adbe8726SEric B Munson goto out; 1522a3437870SNishanth Aravamudan 15239a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1524adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1525adbe8726SEric B Munson err = -EINVAL; 1526adbe8726SEric B Munson goto out; 1527adbe8726SEric B Munson } 1528adbe8726SEric B Munson 15299a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 15309a305230SLee Schermerhorn /* 15319a305230SLee Schermerhorn * global hstate attribute 15329a305230SLee Schermerhorn */ 15339a305230SLee Schermerhorn if (!(obey_mempolicy && 15349a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 153506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 15368cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 153706808b08SLee Schermerhorn } 15389a305230SLee Schermerhorn } else if (nodes_allowed) { 15399a305230SLee Schermerhorn /* 15409a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 15419a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 15429a305230SLee Schermerhorn */ 15439a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 15449a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 15459a305230SLee Schermerhorn } else 15468cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 15479a305230SLee Schermerhorn 154806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1549a3437870SNishanth Aravamudan 15508cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 155106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 155206808b08SLee Schermerhorn 155306808b08SLee Schermerhorn return len; 1554adbe8726SEric B Munson out: 1555adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1556adbe8726SEric B Munson return err; 155706808b08SLee Schermerhorn } 155806808b08SLee Schermerhorn 155906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 156006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 156106808b08SLee Schermerhorn { 156206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 156306808b08SLee Schermerhorn } 156406808b08SLee Schermerhorn 156506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 156606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 156706808b08SLee Schermerhorn { 156806808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1569a3437870SNishanth Aravamudan } 1570a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1571a3437870SNishanth Aravamudan 157206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 157306808b08SLee Schermerhorn 157406808b08SLee Schermerhorn /* 157506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 157606808b08SLee Schermerhorn * huge page alloc/free. 157706808b08SLee Schermerhorn */ 157806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 157906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 158006808b08SLee Schermerhorn { 158106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 158206808b08SLee Schermerhorn } 158306808b08SLee Schermerhorn 158406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 158506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 158606808b08SLee Schermerhorn { 158706808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 158806808b08SLee Schermerhorn } 158906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 159006808b08SLee Schermerhorn #endif 159106808b08SLee Schermerhorn 159206808b08SLee Schermerhorn 1593a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1594a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1595a3437870SNishanth Aravamudan { 15969a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1597a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1598a3437870SNishanth Aravamudan } 1599adbe8726SEric B Munson 1600a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1601a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1602a3437870SNishanth Aravamudan { 1603a3437870SNishanth Aravamudan int err; 1604a3437870SNishanth Aravamudan unsigned long input; 16059a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1606a3437870SNishanth Aravamudan 1607adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1608adbe8726SEric B Munson return -EINVAL; 1609adbe8726SEric B Munson 16103dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 1611a3437870SNishanth Aravamudan if (err) 161273ae31e5SEric B Munson return err; 1613a3437870SNishanth Aravamudan 1614a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1615a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1616a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1617a3437870SNishanth Aravamudan 1618a3437870SNishanth Aravamudan return count; 1619a3437870SNishanth Aravamudan } 1620a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1621a3437870SNishanth Aravamudan 1622a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1623a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1624a3437870SNishanth Aravamudan { 16259a305230SLee Schermerhorn struct hstate *h; 16269a305230SLee Schermerhorn unsigned long free_huge_pages; 16279a305230SLee Schermerhorn int nid; 16289a305230SLee Schermerhorn 16299a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16309a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16319a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 16329a305230SLee Schermerhorn else 16339a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 16349a305230SLee Schermerhorn 16359a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1636a3437870SNishanth Aravamudan } 1637a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1638a3437870SNishanth Aravamudan 1639a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1640a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1641a3437870SNishanth Aravamudan { 16429a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1643a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1644a3437870SNishanth Aravamudan } 1645a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1646a3437870SNishanth Aravamudan 1647a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1648a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1649a3437870SNishanth Aravamudan { 16509a305230SLee Schermerhorn struct hstate *h; 16519a305230SLee Schermerhorn unsigned long surplus_huge_pages; 16529a305230SLee Schermerhorn int nid; 16539a305230SLee Schermerhorn 16549a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16559a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16569a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 16579a305230SLee Schermerhorn else 16589a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 16599a305230SLee Schermerhorn 16609a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1661a3437870SNishanth Aravamudan } 1662a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1663a3437870SNishanth Aravamudan 1664a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1665a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1666a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1667a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1668a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1669a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 167006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 167106808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 167206808b08SLee Schermerhorn #endif 1673a3437870SNishanth Aravamudan NULL, 1674a3437870SNishanth Aravamudan }; 1675a3437870SNishanth Aravamudan 1676a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1677a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1678a3437870SNishanth Aravamudan }; 1679a3437870SNishanth Aravamudan 1680094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 16819a305230SLee Schermerhorn struct kobject **hstate_kobjs, 16829a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1683a3437870SNishanth Aravamudan { 1684a3437870SNishanth Aravamudan int retval; 1685972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1686a3437870SNishanth Aravamudan 16879a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 16889a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1689a3437870SNishanth Aravamudan return -ENOMEM; 1690a3437870SNishanth Aravamudan 16919a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1692a3437870SNishanth Aravamudan if (retval) 16939a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1694a3437870SNishanth Aravamudan 1695a3437870SNishanth Aravamudan return retval; 1696a3437870SNishanth Aravamudan } 1697a3437870SNishanth Aravamudan 1698a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1699a3437870SNishanth Aravamudan { 1700a3437870SNishanth Aravamudan struct hstate *h; 1701a3437870SNishanth Aravamudan int err; 1702a3437870SNishanth Aravamudan 1703a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1704a3437870SNishanth Aravamudan if (!hugepages_kobj) 1705a3437870SNishanth Aravamudan return; 1706a3437870SNishanth Aravamudan 1707a3437870SNishanth Aravamudan for_each_hstate(h) { 17089a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 17099a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1710a3437870SNishanth Aravamudan if (err) 1711ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 1712a3437870SNishanth Aravamudan } 1713a3437870SNishanth Aravamudan } 1714a3437870SNishanth Aravamudan 17159a305230SLee Schermerhorn #ifdef CONFIG_NUMA 17169a305230SLee Schermerhorn 17179a305230SLee Schermerhorn /* 17189a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 171910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 172010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 172110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 17229a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 17239a305230SLee Schermerhorn */ 17249a305230SLee Schermerhorn struct node_hstate { 17259a305230SLee Schermerhorn struct kobject *hugepages_kobj; 17269a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 17279a305230SLee Schermerhorn }; 17289a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 17299a305230SLee Schermerhorn 17309a305230SLee Schermerhorn /* 173110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 17329a305230SLee Schermerhorn */ 17339a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 17349a305230SLee Schermerhorn &nr_hugepages_attr.attr, 17359a305230SLee Schermerhorn &free_hugepages_attr.attr, 17369a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 17379a305230SLee Schermerhorn NULL, 17389a305230SLee Schermerhorn }; 17399a305230SLee Schermerhorn 17409a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 17419a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 17429a305230SLee Schermerhorn }; 17439a305230SLee Schermerhorn 17449a305230SLee Schermerhorn /* 174510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 17469a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 17479a305230SLee Schermerhorn */ 17489a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17499a305230SLee Schermerhorn { 17509a305230SLee Schermerhorn int nid; 17519a305230SLee Schermerhorn 17529a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 17539a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 17549a305230SLee Schermerhorn int i; 17559a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 17569a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 17579a305230SLee Schermerhorn if (nidp) 17589a305230SLee Schermerhorn *nidp = nid; 17599a305230SLee Schermerhorn return &hstates[i]; 17609a305230SLee Schermerhorn } 17619a305230SLee Schermerhorn } 17629a305230SLee Schermerhorn 17639a305230SLee Schermerhorn BUG(); 17649a305230SLee Schermerhorn return NULL; 17659a305230SLee Schermerhorn } 17669a305230SLee Schermerhorn 17679a305230SLee Schermerhorn /* 176810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 17699a305230SLee Schermerhorn * No-op if no hstate attributes attached. 17709a305230SLee Schermerhorn */ 17713cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 17729a305230SLee Schermerhorn { 17739a305230SLee Schermerhorn struct hstate *h; 177410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 17759a305230SLee Schermerhorn 17769a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 17779b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 17789a305230SLee Schermerhorn 1779972dc4deSAneesh Kumar K.V for_each_hstate(h) { 1780972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 1781972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 1782972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 1783972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 1784972dc4deSAneesh Kumar K.V } 17859a305230SLee Schermerhorn } 17869a305230SLee Schermerhorn 17879a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 17889a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 17899a305230SLee Schermerhorn } 17909a305230SLee Schermerhorn 17919a305230SLee Schermerhorn /* 179210fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 17939a305230SLee Schermerhorn * that have them. 17949a305230SLee Schermerhorn */ 17959a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 17969a305230SLee Schermerhorn { 17979a305230SLee Schermerhorn int nid; 17989a305230SLee Schermerhorn 17999a305230SLee Schermerhorn /* 180010fbcf4cSKay Sievers * disable node device registrations. 18019a305230SLee Schermerhorn */ 18029a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 18039a305230SLee Schermerhorn 18049a305230SLee Schermerhorn /* 18059a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 18069a305230SLee Schermerhorn */ 18079a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 18088732794bSWen Congyang hugetlb_unregister_node(node_devices[nid]); 18099a305230SLee Schermerhorn } 18109a305230SLee Schermerhorn 18119a305230SLee Schermerhorn /* 181210fbcf4cSKay Sievers * Register hstate attributes for a single node device. 18139a305230SLee Schermerhorn * No-op if attributes already registered. 18149a305230SLee Schermerhorn */ 18153cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 18169a305230SLee Schermerhorn { 18179a305230SLee Schermerhorn struct hstate *h; 181810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 18199a305230SLee Schermerhorn int err; 18209a305230SLee Schermerhorn 18219a305230SLee Schermerhorn if (nhs->hugepages_kobj) 18229a305230SLee Schermerhorn return; /* already allocated */ 18239a305230SLee Schermerhorn 18249a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 182510fbcf4cSKay Sievers &node->dev.kobj); 18269a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 18279a305230SLee Schermerhorn return; 18289a305230SLee Schermerhorn 18299a305230SLee Schermerhorn for_each_hstate(h) { 18309a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 18319a305230SLee Schermerhorn nhs->hstate_kobjs, 18329a305230SLee Schermerhorn &per_node_hstate_attr_group); 18339a305230SLee Schermerhorn if (err) { 1834ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 183510fbcf4cSKay Sievers h->name, node->dev.id); 18369a305230SLee Schermerhorn hugetlb_unregister_node(node); 18379a305230SLee Schermerhorn break; 18389a305230SLee Schermerhorn } 18399a305230SLee Schermerhorn } 18409a305230SLee Schermerhorn } 18419a305230SLee Schermerhorn 18429a305230SLee Schermerhorn /* 18439b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 184410fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 184510fbcf4cSKay Sievers * registered their associated device by this time. 18469a305230SLee Schermerhorn */ 18479a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 18489a305230SLee Schermerhorn { 18499a305230SLee Schermerhorn int nid; 18509a305230SLee Schermerhorn 18518cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 18528732794bSWen Congyang struct node *node = node_devices[nid]; 185310fbcf4cSKay Sievers if (node->dev.id == nid) 18549a305230SLee Schermerhorn hugetlb_register_node(node); 18559a305230SLee Schermerhorn } 18569a305230SLee Schermerhorn 18579a305230SLee Schermerhorn /* 185810fbcf4cSKay Sievers * Let the node device driver know we're here so it can 18599a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 18609a305230SLee Schermerhorn */ 18619a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 18629a305230SLee Schermerhorn hugetlb_unregister_node); 18639a305230SLee Schermerhorn } 18649a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 18659a305230SLee Schermerhorn 18669a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 18679a305230SLee Schermerhorn { 18689a305230SLee Schermerhorn BUG(); 18699a305230SLee Schermerhorn if (nidp) 18709a305230SLee Schermerhorn *nidp = -1; 18719a305230SLee Schermerhorn return NULL; 18729a305230SLee Schermerhorn } 18739a305230SLee Schermerhorn 18749a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 18759a305230SLee Schermerhorn 18769a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 18779a305230SLee Schermerhorn 18789a305230SLee Schermerhorn #endif 18799a305230SLee Schermerhorn 1880a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1881a3437870SNishanth Aravamudan { 1882a3437870SNishanth Aravamudan struct hstate *h; 1883a3437870SNishanth Aravamudan 18849a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 18859a305230SLee Schermerhorn 1886a3437870SNishanth Aravamudan for_each_hstate(h) { 1887972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 1888a3437870SNishanth Aravamudan } 1889a3437870SNishanth Aravamudan 1890a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1891a3437870SNishanth Aravamudan } 1892a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1893a3437870SNishanth Aravamudan 1894a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1895a3437870SNishanth Aravamudan { 18960ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 18970ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 18980ef89d25SBenjamin Herrenschmidt * there is no such support 18990ef89d25SBenjamin Herrenschmidt */ 19000ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 19010ef89d25SBenjamin Herrenschmidt return 0; 1902a3437870SNishanth Aravamudan 1903e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1904e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1905e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1906a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1907a3437870SNishanth Aravamudan } 1908972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1909e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1910e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1911a3437870SNishanth Aravamudan 1912a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1913aa888a74SAndi Kleen gather_bootmem_prealloc(); 1914a3437870SNishanth Aravamudan report_hugepages(); 1915a3437870SNishanth Aravamudan 1916a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 19179a305230SLee Schermerhorn hugetlb_register_all_nodes(); 19187179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 19199a305230SLee Schermerhorn 1920a3437870SNishanth Aravamudan return 0; 1921a3437870SNishanth Aravamudan } 1922a3437870SNishanth Aravamudan module_init(hugetlb_init); 1923a3437870SNishanth Aravamudan 1924a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1925a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1926a3437870SNishanth Aravamudan { 1927a3437870SNishanth Aravamudan struct hstate *h; 19288faa8b07SAndi Kleen unsigned long i; 19298faa8b07SAndi Kleen 1930a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1931ffb22af5SAndrew Morton pr_warning("hugepagesz= specified twice, ignoring\n"); 1932a3437870SNishanth Aravamudan return; 1933a3437870SNishanth Aravamudan } 193447d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 1935a3437870SNishanth Aravamudan BUG_ON(order == 0); 193647d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 1937a3437870SNishanth Aravamudan h->order = order; 1938a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 19398faa8b07SAndi Kleen h->nr_huge_pages = 0; 19408faa8b07SAndi Kleen h->free_huge_pages = 0; 19418faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 19428faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 19430edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 19448cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 19458cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 1946a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1947a3437870SNishanth Aravamudan huge_page_size(h)/1024); 19488faa8b07SAndi Kleen 1949a3437870SNishanth Aravamudan parsed_hstate = h; 1950a3437870SNishanth Aravamudan } 1951a3437870SNishanth Aravamudan 1952e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1953a3437870SNishanth Aravamudan { 1954a3437870SNishanth Aravamudan unsigned long *mhp; 19558faa8b07SAndi Kleen static unsigned long *last_mhp; 1956a3437870SNishanth Aravamudan 1957a3437870SNishanth Aravamudan /* 195847d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 1959a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1960a3437870SNishanth Aravamudan */ 196147d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 1962a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1963a3437870SNishanth Aravamudan else 1964a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1965a3437870SNishanth Aravamudan 19668faa8b07SAndi Kleen if (mhp == last_mhp) { 1967ffb22af5SAndrew Morton pr_warning("hugepages= specified twice without " 19688faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 19698faa8b07SAndi Kleen return 1; 19708faa8b07SAndi Kleen } 19718faa8b07SAndi Kleen 1972a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1973a3437870SNishanth Aravamudan *mhp = 0; 1974a3437870SNishanth Aravamudan 19758faa8b07SAndi Kleen /* 19768faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 19778faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 19788faa8b07SAndi Kleen * use the bootmem allocator. 19798faa8b07SAndi Kleen */ 198047d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 19818faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 19828faa8b07SAndi Kleen 19838faa8b07SAndi Kleen last_mhp = mhp; 19848faa8b07SAndi Kleen 1985a3437870SNishanth Aravamudan return 1; 1986a3437870SNishanth Aravamudan } 1987e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1988e11bfbfcSNick Piggin 1989e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1990e11bfbfcSNick Piggin { 1991e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1992e11bfbfcSNick Piggin return 1; 1993e11bfbfcSNick Piggin } 1994e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1995a3437870SNishanth Aravamudan 19968a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 19978a213460SNishanth Aravamudan { 19988a213460SNishanth Aravamudan int node; 19998a213460SNishanth Aravamudan unsigned int nr = 0; 20008a213460SNishanth Aravamudan 20018a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 20028a213460SNishanth Aravamudan nr += array[node]; 20038a213460SNishanth Aravamudan 20048a213460SNishanth Aravamudan return nr; 20058a213460SNishanth Aravamudan } 20068a213460SNishanth Aravamudan 20078a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 200806808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 200906808b08SLee Schermerhorn struct ctl_table *table, int write, 201006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 20111da177e4SLinus Torvalds { 2012e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2013e5ff2159SAndi Kleen unsigned long tmp; 201408d4a246SMichal Hocko int ret; 2015e5ff2159SAndi Kleen 2016e5ff2159SAndi Kleen tmp = h->max_huge_pages; 2017e5ff2159SAndi Kleen 2018adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2019adbe8726SEric B Munson return -EINVAL; 2020adbe8726SEric B Munson 2021e5ff2159SAndi Kleen table->data = &tmp; 2022e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 202308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 202408d4a246SMichal Hocko if (ret) 202508d4a246SMichal Hocko goto out; 2026e5ff2159SAndi Kleen 202706808b08SLee Schermerhorn if (write) { 2028bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2029bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 203006808b08SLee Schermerhorn if (!(obey_mempolicy && 203106808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 203206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 20338cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 203406808b08SLee Schermerhorn } 203506808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 203606808b08SLee Schermerhorn 20378cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 203806808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 203906808b08SLee Schermerhorn } 204008d4a246SMichal Hocko out: 204108d4a246SMichal Hocko return ret; 20421da177e4SLinus Torvalds } 2043396faf03SMel Gorman 204406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 204506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 204606808b08SLee Schermerhorn { 204706808b08SLee Schermerhorn 204806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 204906808b08SLee Schermerhorn buffer, length, ppos); 205006808b08SLee Schermerhorn } 205106808b08SLee Schermerhorn 205206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 205306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 205406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 205506808b08SLee Schermerhorn { 205606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 205706808b08SLee Schermerhorn buffer, length, ppos); 205806808b08SLee Schermerhorn } 205906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 206006808b08SLee Schermerhorn 2061396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 20628d65af78SAlexey Dobriyan void __user *buffer, 2063396faf03SMel Gorman size_t *length, loff_t *ppos) 2064396faf03SMel Gorman { 20658d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 2066396faf03SMel Gorman if (hugepages_treat_as_movable) 2067396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 2068396faf03SMel Gorman else 2069396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 2070396faf03SMel Gorman return 0; 2071396faf03SMel Gorman } 2072396faf03SMel Gorman 2073a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 20748d65af78SAlexey Dobriyan void __user *buffer, 2075a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2076a3d0c6aaSNishanth Aravamudan { 2077a5516438SAndi Kleen struct hstate *h = &default_hstate; 2078e5ff2159SAndi Kleen unsigned long tmp; 207908d4a246SMichal Hocko int ret; 2080e5ff2159SAndi Kleen 2081e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2082e5ff2159SAndi Kleen 2083adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2084adbe8726SEric B Munson return -EINVAL; 2085adbe8726SEric B Munson 2086e5ff2159SAndi Kleen table->data = &tmp; 2087e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 208808d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 208908d4a246SMichal Hocko if (ret) 209008d4a246SMichal Hocko goto out; 2091e5ff2159SAndi Kleen 2092e5ff2159SAndi Kleen if (write) { 2093064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2094e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2095a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2096e5ff2159SAndi Kleen } 209708d4a246SMichal Hocko out: 209808d4a246SMichal Hocko return ret; 2099a3d0c6aaSNishanth Aravamudan } 2100a3d0c6aaSNishanth Aravamudan 21011da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 21021da177e4SLinus Torvalds 2103e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 21041da177e4SLinus Torvalds { 2105a5516438SAndi Kleen struct hstate *h = &default_hstate; 2106e1759c21SAlexey Dobriyan seq_printf(m, 21071da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 21081da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2109b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 21107893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 21114f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2112a5516438SAndi Kleen h->nr_huge_pages, 2113a5516438SAndi Kleen h->free_huge_pages, 2114a5516438SAndi Kleen h->resv_huge_pages, 2115a5516438SAndi Kleen h->surplus_huge_pages, 2116a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 21171da177e4SLinus Torvalds } 21181da177e4SLinus Torvalds 21191da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 21201da177e4SLinus Torvalds { 2121a5516438SAndi Kleen struct hstate *h = &default_hstate; 21221da177e4SLinus Torvalds return sprintf(buf, 21231da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2124a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2125a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2126a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2127a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2128a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 21291da177e4SLinus Torvalds } 21301da177e4SLinus Torvalds 2131949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2132949f7ec5SDavid Rientjes { 2133949f7ec5SDavid Rientjes struct hstate *h; 2134949f7ec5SDavid Rientjes int nid; 2135949f7ec5SDavid Rientjes 2136949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2137949f7ec5SDavid Rientjes for_each_hstate(h) 2138949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2139949f7ec5SDavid Rientjes nid, 2140949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2141949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2142949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2143949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2144949f7ec5SDavid Rientjes } 2145949f7ec5SDavid Rientjes 21461da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 21471da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 21481da177e4SLinus Torvalds { 2149d0028588SWanpeng Li struct hstate *h; 2150d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2151d0028588SWanpeng Li 2152d0028588SWanpeng Li for_each_hstate(h) 2153d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2154d0028588SWanpeng Li return nr_total_pages; 21551da177e4SLinus Torvalds } 21561da177e4SLinus Torvalds 2157a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2158fc1b8a73SMel Gorman { 2159fc1b8a73SMel Gorman int ret = -ENOMEM; 2160fc1b8a73SMel Gorman 2161fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2162fc1b8a73SMel Gorman /* 2163fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2164fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2165fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2166fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2167fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2168fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2169fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2170fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2171fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2172fc1b8a73SMel Gorman * 2173fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2174fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2175fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2176fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2177fc1b8a73SMel Gorman * semantics that cpuset has. 2178fc1b8a73SMel Gorman */ 2179fc1b8a73SMel Gorman if (delta > 0) { 2180a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2181fc1b8a73SMel Gorman goto out; 2182fc1b8a73SMel Gorman 2183a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2184a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2185fc1b8a73SMel Gorman goto out; 2186fc1b8a73SMel Gorman } 2187fc1b8a73SMel Gorman } 2188fc1b8a73SMel Gorman 2189fc1b8a73SMel Gorman ret = 0; 2190fc1b8a73SMel Gorman if (delta < 0) 2191a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2192fc1b8a73SMel Gorman 2193fc1b8a73SMel Gorman out: 2194fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2195fc1b8a73SMel Gorman return ret; 2196fc1b8a73SMel Gorman } 2197fc1b8a73SMel Gorman 219884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 219984afd99bSAndy Whitcroft { 2200f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 220184afd99bSAndy Whitcroft 220284afd99bSAndy Whitcroft /* 220384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 220484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 220584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 220625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 220784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 220884afd99bSAndy Whitcroft * new reference here without additional locking. 220984afd99bSAndy Whitcroft */ 2210f522c3acSJoonsoo Kim if (resv) 2211f522c3acSJoonsoo Kim kref_get(&resv->refs); 221284afd99bSAndy Whitcroft } 221384afd99bSAndy Whitcroft 2214c50ac050SDave Hansen static void resv_map_put(struct vm_area_struct *vma) 2215c50ac050SDave Hansen { 2216f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 2217c50ac050SDave Hansen 2218f522c3acSJoonsoo Kim if (!resv) 2219c50ac050SDave Hansen return; 2220f522c3acSJoonsoo Kim kref_put(&resv->refs, resv_map_release); 2221c50ac050SDave Hansen } 2222c50ac050SDave Hansen 2223a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2224a1e78772SMel Gorman { 2225a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2226f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 222790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 222884afd99bSAndy Whitcroft unsigned long reserve; 222984afd99bSAndy Whitcroft unsigned long start; 223084afd99bSAndy Whitcroft unsigned long end; 223184afd99bSAndy Whitcroft 2232f522c3acSJoonsoo Kim if (resv) { 2233a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2234a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 223584afd99bSAndy Whitcroft 223684afd99bSAndy Whitcroft reserve = (end - start) - 2237f522c3acSJoonsoo Kim region_count(&resv->regions, start, end); 223884afd99bSAndy Whitcroft 2239c50ac050SDave Hansen resv_map_put(vma); 224084afd99bSAndy Whitcroft 22417251ff78SAdam Litke if (reserve) { 2242a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 224390481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 22447251ff78SAdam Litke } 2245a1e78772SMel Gorman } 224684afd99bSAndy Whitcroft } 2247a1e78772SMel Gorman 22481da177e4SLinus Torvalds /* 22491da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 22501da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 22511da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 22521da177e4SLinus Torvalds * this far. 22531da177e4SLinus Torvalds */ 2254d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 22551da177e4SLinus Torvalds { 22561da177e4SLinus Torvalds BUG(); 2257d0217ac0SNick Piggin return 0; 22581da177e4SLinus Torvalds } 22591da177e4SLinus Torvalds 2260f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2261d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 226284afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2263a1e78772SMel Gorman .close = hugetlb_vm_op_close, 22641da177e4SLinus Torvalds }; 22651da177e4SLinus Torvalds 22661e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 22671e8f889bSDavid Gibson int writable) 226863551ae0SDavid Gibson { 226963551ae0SDavid Gibson pte_t entry; 227063551ae0SDavid Gibson 22711e8f889bSDavid Gibson if (writable) { 2272106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2273106c992aSGerald Schaefer vma->vm_page_prot))); 227463551ae0SDavid Gibson } else { 2275106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 2276106c992aSGerald Schaefer vma->vm_page_prot)); 227763551ae0SDavid Gibson } 227863551ae0SDavid Gibson entry = pte_mkyoung(entry); 227963551ae0SDavid Gibson entry = pte_mkhuge(entry); 2280d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 228163551ae0SDavid Gibson 228263551ae0SDavid Gibson return entry; 228363551ae0SDavid Gibson } 228463551ae0SDavid Gibson 22851e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 22861e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 22871e8f889bSDavid Gibson { 22881e8f889bSDavid Gibson pte_t entry; 22891e8f889bSDavid Gibson 2290106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 229132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 22924b3073e1SRussell King update_mmu_cache(vma, address, ptep); 22931e8f889bSDavid Gibson } 22941e8f889bSDavid Gibson 22951e8f889bSDavid Gibson 229663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 229763551ae0SDavid Gibson struct vm_area_struct *vma) 229863551ae0SDavid Gibson { 229963551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 230063551ae0SDavid Gibson struct page *ptepage; 23011c59827dSHugh Dickins unsigned long addr; 23021e8f889bSDavid Gibson int cow; 2303a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2304a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 23051e8f889bSDavid Gibson 23061e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 230763551ae0SDavid Gibson 2308a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2309c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2310c74df32cSHugh Dickins if (!src_pte) 2311c74df32cSHugh Dickins continue; 2312a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 231363551ae0SDavid Gibson if (!dst_pte) 231463551ae0SDavid Gibson goto nomem; 2315c5c99429SLarry Woodman 2316c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2317c5c99429SLarry Woodman if (dst_pte == src_pte) 2318c5c99429SLarry Woodman continue; 2319c5c99429SLarry Woodman 2320c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 232146478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 23227f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 23231e8f889bSDavid Gibson if (cow) 23247f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 23257f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 232663551ae0SDavid Gibson ptepage = pte_page(entry); 232763551ae0SDavid Gibson get_page(ptepage); 23280fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 232963551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 23301c59827dSHugh Dickins } 23311c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2332c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 233363551ae0SDavid Gibson } 233463551ae0SDavid Gibson return 0; 233563551ae0SDavid Gibson 233663551ae0SDavid Gibson nomem: 233763551ae0SDavid Gibson return -ENOMEM; 233863551ae0SDavid Gibson } 233963551ae0SDavid Gibson 2340290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2341290408d4SNaoya Horiguchi { 2342290408d4SNaoya Horiguchi swp_entry_t swp; 2343290408d4SNaoya Horiguchi 2344290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2345290408d4SNaoya Horiguchi return 0; 2346290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 234732f84528SChris Forbes if (non_swap_entry(swp) && is_migration_entry(swp)) 2348290408d4SNaoya Horiguchi return 1; 234932f84528SChris Forbes else 2350290408d4SNaoya Horiguchi return 0; 2351290408d4SNaoya Horiguchi } 2352290408d4SNaoya Horiguchi 2353fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2354fd6a03edSNaoya Horiguchi { 2355fd6a03edSNaoya Horiguchi swp_entry_t swp; 2356fd6a03edSNaoya Horiguchi 2357fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2358fd6a03edSNaoya Horiguchi return 0; 2359fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 236032f84528SChris Forbes if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2361fd6a03edSNaoya Horiguchi return 1; 236232f84528SChris Forbes else 2363fd6a03edSNaoya Horiguchi return 0; 2364fd6a03edSNaoya Horiguchi } 2365fd6a03edSNaoya Horiguchi 236624669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 236724669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 236824669e58SAneesh Kumar K.V struct page *ref_page) 236963551ae0SDavid Gibson { 237024669e58SAneesh Kumar K.V int force_flush = 0; 237163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 237263551ae0SDavid Gibson unsigned long address; 2373c7546f8fSDavid Gibson pte_t *ptep; 237463551ae0SDavid Gibson pte_t pte; 237563551ae0SDavid Gibson struct page *page; 2376a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2377a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 23782ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 23792ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 2380a5516438SAndi Kleen 238163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2382a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2383a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 238463551ae0SDavid Gibson 238524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 23862ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 238724669e58SAneesh Kumar K.V again: 2388508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2389a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2390c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2391c7546f8fSDavid Gibson if (!ptep) 2392c7546f8fSDavid Gibson continue; 2393c7546f8fSDavid Gibson 239439dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 239539dde65cSChen, Kenneth W continue; 239639dde65cSChen, Kenneth W 23976629326bSHillf Danton pte = huge_ptep_get(ptep); 23986629326bSHillf Danton if (huge_pte_none(pte)) 23996629326bSHillf Danton continue; 24006629326bSHillf Danton 24016629326bSHillf Danton /* 24026629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 24036629326bSHillf Danton */ 24048c4894c6SNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 2405106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 24066629326bSHillf Danton continue; 24078c4894c6SNaoya Horiguchi } 24086629326bSHillf Danton 24096629326bSHillf Danton page = pte_page(pte); 241004f2cbe3SMel Gorman /* 241104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 241204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 241304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 241404f2cbe3SMel Gorman */ 241504f2cbe3SMel Gorman if (ref_page) { 241604f2cbe3SMel Gorman if (page != ref_page) 241704f2cbe3SMel Gorman continue; 241804f2cbe3SMel Gorman 241904f2cbe3SMel Gorman /* 242004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 242104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 242204f2cbe3SMel Gorman * looking like data was lost 242304f2cbe3SMel Gorman */ 242404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 242504f2cbe3SMel Gorman } 242604f2cbe3SMel Gorman 2427c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 242824669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 2429106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 24306649a386SKen Chen set_page_dirty(page); 24319e81130bSHillf Danton 243224669e58SAneesh Kumar K.V page_remove_rmap(page); 243324669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 243424669e58SAneesh Kumar K.V if (force_flush) 243524669e58SAneesh Kumar K.V break; 24369e81130bSHillf Danton /* Bail out after unmapping reference page if supplied */ 24379e81130bSHillf Danton if (ref_page) 24389e81130bSHillf Danton break; 243963551ae0SDavid Gibson } 2440cd2934a3SAl Viro spin_unlock(&mm->page_table_lock); 244124669e58SAneesh Kumar K.V /* 244224669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 244324669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 244424669e58SAneesh Kumar K.V * and page-free while holding it. 244524669e58SAneesh Kumar K.V */ 244624669e58SAneesh Kumar K.V if (force_flush) { 244724669e58SAneesh Kumar K.V force_flush = 0; 244824669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 244924669e58SAneesh Kumar K.V if (address < end && !ref_page) 245024669e58SAneesh Kumar K.V goto again; 2451fe1668aeSChen, Kenneth W } 24522ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 245324669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 24541da177e4SLinus Torvalds } 245563551ae0SDavid Gibson 2456d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2457d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2458d833352aSMel Gorman unsigned long end, struct page *ref_page) 2459d833352aSMel Gorman { 2460d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2461d833352aSMel Gorman 2462d833352aSMel Gorman /* 2463d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2464d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2465d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2466d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2467d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2468d833352aSMel Gorman * is to clear it before releasing the i_mmap_mutex. This works 2469d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2470d833352aSMel Gorman * destroyed and the i_mmap_mutex is held. 2471d833352aSMel Gorman */ 2472d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2473d833352aSMel Gorman } 2474d833352aSMel Gorman 2475502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 247604f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2477502717f4SChen, Kenneth W { 247824669e58SAneesh Kumar K.V struct mm_struct *mm; 247924669e58SAneesh Kumar K.V struct mmu_gather tlb; 248024669e58SAneesh Kumar K.V 248124669e58SAneesh Kumar K.V mm = vma->vm_mm; 248224669e58SAneesh Kumar K.V 24832b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 248424669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 248524669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2486502717f4SChen, Kenneth W } 2487502717f4SChen, Kenneth W 248804f2cbe3SMel Gorman /* 248904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 249004f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 249104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 249204f2cbe3SMel Gorman * same region. 249304f2cbe3SMel Gorman */ 24942a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 24952a4b3dedSHarvey Harrison struct page *page, unsigned long address) 249604f2cbe3SMel Gorman { 24977526674dSAdam Litke struct hstate *h = hstate_vma(vma); 249804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 249904f2cbe3SMel Gorman struct address_space *mapping; 250004f2cbe3SMel Gorman pgoff_t pgoff; 250104f2cbe3SMel Gorman 250204f2cbe3SMel Gorman /* 250304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 250404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 250504f2cbe3SMel Gorman */ 25067526674dSAdam Litke address = address & huge_page_mask(h); 250736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 250836e4f20aSMichal Hocko vma->vm_pgoff; 2509496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 251004f2cbe3SMel Gorman 25114eb2b1dcSMel Gorman /* 25124eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 25134eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 25144eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 25154eb2b1dcSMel Gorman */ 25163d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 25176b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 251804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 251904f2cbe3SMel Gorman if (iter_vma == vma) 252004f2cbe3SMel Gorman continue; 252104f2cbe3SMel Gorman 252204f2cbe3SMel Gorman /* 252304f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 252404f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 252504f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 252604f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 252704f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 252804f2cbe3SMel Gorman */ 252904f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 253024669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 253124669e58SAneesh Kumar K.V address + huge_page_size(h), page); 253204f2cbe3SMel Gorman } 25333d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 253404f2cbe3SMel Gorman 253504f2cbe3SMel Gorman return 1; 253604f2cbe3SMel Gorman } 253704f2cbe3SMel Gorman 25380fe6e20bSNaoya Horiguchi /* 25390fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2540ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2541ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2542ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 25430fe6e20bSNaoya Horiguchi */ 25441e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 254504f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 254604f2cbe3SMel Gorman struct page *pagecache_page) 25471e8f889bSDavid Gibson { 2548a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 25491e8f889bSDavid Gibson struct page *old_page, *new_page; 255004f2cbe3SMel Gorman int outside_reserve = 0; 25512ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 25522ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 25531e8f889bSDavid Gibson 25541e8f889bSDavid Gibson old_page = pte_page(pte); 25551e8f889bSDavid Gibson 255604f2cbe3SMel Gorman retry_avoidcopy: 25571e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 25581e8f889bSDavid Gibson * and just make the page writable */ 255937a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 25600fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 25611e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 256283c54070SNick Piggin return 0; 25631e8f889bSDavid Gibson } 25641e8f889bSDavid Gibson 256504f2cbe3SMel Gorman /* 256604f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 256704f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 256804f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 256904f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 257004f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 257104f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 257204f2cbe3SMel Gorman * of the full address range. 257304f2cbe3SMel Gorman */ 25745944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 257504f2cbe3SMel Gorman old_page != pagecache_page) 257604f2cbe3SMel Gorman outside_reserve = 1; 257704f2cbe3SMel Gorman 25781e8f889bSDavid Gibson page_cache_get(old_page); 2579b76c8cfbSLarry Woodman 2580b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2581b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 258204f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 25831e8f889bSDavid Gibson 25842fc39cecSAdam Litke if (IS_ERR(new_page)) { 258576dcee75SAneesh Kumar K.V long err = PTR_ERR(new_page); 25861e8f889bSDavid Gibson page_cache_release(old_page); 258704f2cbe3SMel Gorman 258804f2cbe3SMel Gorman /* 258904f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 259004f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 259104f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 259204f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 259304f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 259404f2cbe3SMel Gorman */ 259504f2cbe3SMel Gorman if (outside_reserve) { 259604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 259704f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 259804f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2599b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2600a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2601a734bcc8SHillf Danton if (likely(pte_same(huge_ptep_get(ptep), pte))) 260204f2cbe3SMel Gorman goto retry_avoidcopy; 2603a734bcc8SHillf Danton /* 2604a734bcc8SHillf Danton * race occurs while re-acquiring page_table_lock, and 2605a734bcc8SHillf Danton * our job is done. 2606a734bcc8SHillf Danton */ 2607a734bcc8SHillf Danton return 0; 260804f2cbe3SMel Gorman } 260904f2cbe3SMel Gorman WARN_ON_ONCE(1); 261004f2cbe3SMel Gorman } 261104f2cbe3SMel Gorman 2612b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2613b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 261476dcee75SAneesh Kumar K.V if (err == -ENOMEM) 261576dcee75SAneesh Kumar K.V return VM_FAULT_OOM; 261676dcee75SAneesh Kumar K.V else 261776dcee75SAneesh Kumar K.V return VM_FAULT_SIGBUS; 26181e8f889bSDavid Gibson } 26191e8f889bSDavid Gibson 26200fe6e20bSNaoya Horiguchi /* 26210fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 26220fe6e20bSNaoya Horiguchi * anon_vma prepared. 26230fe6e20bSNaoya Horiguchi */ 262444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2625ea4039a3SHillf Danton page_cache_release(new_page); 2626ea4039a3SHillf Danton page_cache_release(old_page); 262744e2aa93SDean Nelson /* Caller expects lock to be held */ 262844e2aa93SDean Nelson spin_lock(&mm->page_table_lock); 26290fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 263044e2aa93SDean Nelson } 26310fe6e20bSNaoya Horiguchi 263247ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 263347ad8475SAndrea Arcangeli pages_per_huge_page(h)); 26340ed361deSNick Piggin __SetPageUptodate(new_page); 26351e8f889bSDavid Gibson 26362ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 26372ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 26382ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2639b76c8cfbSLarry Woodman /* 2640b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2641b76c8cfbSLarry Woodman * before the page tables are altered 2642b76c8cfbSLarry Woodman */ 2643b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2644a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 26457f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 264607443a85SJoonsoo Kim ClearPagePrivate(new_page); 264707443a85SJoonsoo Kim 26481e8f889bSDavid Gibson /* Break COW */ 26498fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 26501e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 26511e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 26520fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2653cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 26541e8f889bSDavid Gibson /* Make the old page be freed below */ 26551e8f889bSDavid Gibson new_page = old_page; 26561e8f889bSDavid Gibson } 26572ec74c3eSSagi Grimberg spin_unlock(&mm->page_table_lock); 26582ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 26591e8f889bSDavid Gibson page_cache_release(new_page); 26601e8f889bSDavid Gibson page_cache_release(old_page); 26618312034fSJoonsoo Kim 26628312034fSJoonsoo Kim /* Caller expects lock to be held */ 26638312034fSJoonsoo Kim spin_lock(&mm->page_table_lock); 266483c54070SNick Piggin return 0; 26651e8f889bSDavid Gibson } 26661e8f889bSDavid Gibson 266704f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2668a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2669a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 267004f2cbe3SMel Gorman { 267104f2cbe3SMel Gorman struct address_space *mapping; 2672e7c4b0bfSAndy Whitcroft pgoff_t idx; 267304f2cbe3SMel Gorman 267404f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2675a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 267604f2cbe3SMel Gorman 267704f2cbe3SMel Gorman return find_lock_page(mapping, idx); 267804f2cbe3SMel Gorman } 267904f2cbe3SMel Gorman 26803ae77f43SHugh Dickins /* 26813ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 26823ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 26833ae77f43SHugh Dickins */ 26843ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 26852a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 26862a15efc9SHugh Dickins { 26872a15efc9SHugh Dickins struct address_space *mapping; 26882a15efc9SHugh Dickins pgoff_t idx; 26892a15efc9SHugh Dickins struct page *page; 26902a15efc9SHugh Dickins 26912a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 26922a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 26932a15efc9SHugh Dickins 26942a15efc9SHugh Dickins page = find_get_page(mapping, idx); 26952a15efc9SHugh Dickins if (page) 26962a15efc9SHugh Dickins put_page(page); 26972a15efc9SHugh Dickins return page != NULL; 26982a15efc9SHugh Dickins } 26992a15efc9SHugh Dickins 2700a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2701788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2702ac9b9c66SHugh Dickins { 2703a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2704ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2705409eb8c2SHillf Danton int anon_rmap = 0; 2706e7c4b0bfSAndy Whitcroft pgoff_t idx; 27074c887265SAdam Litke unsigned long size; 27084c887265SAdam Litke struct page *page; 27094c887265SAdam Litke struct address_space *mapping; 27101e8f889bSDavid Gibson pte_t new_pte; 27114c887265SAdam Litke 271204f2cbe3SMel Gorman /* 271304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 271404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 271525985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 271604f2cbe3SMel Gorman */ 271704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2718ffb22af5SAndrew Morton pr_warning("PID %d killed due to inadequate hugepage pool\n", 271904f2cbe3SMel Gorman current->pid); 272004f2cbe3SMel Gorman return ret; 272104f2cbe3SMel Gorman } 272204f2cbe3SMel Gorman 27234c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2724a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 27254c887265SAdam Litke 27264c887265SAdam Litke /* 27274c887265SAdam Litke * Use page lock to guard against racing truncation 27284c887265SAdam Litke * before we get page_table_lock. 27294c887265SAdam Litke */ 27306bda666aSChristoph Lameter retry: 27316bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 27326bda666aSChristoph Lameter if (!page) { 2733a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2734ebed4bfcSHugh Dickins if (idx >= size) 2735ebed4bfcSHugh Dickins goto out; 273604f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 27372fc39cecSAdam Litke if (IS_ERR(page)) { 273876dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 273976dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 274076dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 274176dcee75SAneesh Kumar K.V else 274276dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 27436bda666aSChristoph Lameter goto out; 27446bda666aSChristoph Lameter } 274547ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 27460ed361deSNick Piggin __SetPageUptodate(page); 2747ac9b9c66SHugh Dickins 2748f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 27496bda666aSChristoph Lameter int err; 275045c682a6SKen Chen struct inode *inode = mapping->host; 27516bda666aSChristoph Lameter 27526bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 27536bda666aSChristoph Lameter if (err) { 27546bda666aSChristoph Lameter put_page(page); 27556bda666aSChristoph Lameter if (err == -EEXIST) 27566bda666aSChristoph Lameter goto retry; 27576bda666aSChristoph Lameter goto out; 27586bda666aSChristoph Lameter } 275907443a85SJoonsoo Kim ClearPagePrivate(page); 276045c682a6SKen Chen 276145c682a6SKen Chen spin_lock(&inode->i_lock); 2762a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 276345c682a6SKen Chen spin_unlock(&inode->i_lock); 276423be7468SMel Gorman } else { 27656bda666aSChristoph Lameter lock_page(page); 27660fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 27670fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 27680fe6e20bSNaoya Horiguchi goto backout_unlocked; 276923be7468SMel Gorman } 2770409eb8c2SHillf Danton anon_rmap = 1; 27710fe6e20bSNaoya Horiguchi } 27720fe6e20bSNaoya Horiguchi } else { 277357303d80SAndy Whitcroft /* 2774998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2775998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2776998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2777fd6a03edSNaoya Horiguchi */ 2778fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2779aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2780972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2781fd6a03edSNaoya Horiguchi goto backout_unlocked; 27826bda666aSChristoph Lameter } 2783998b4382SNaoya Horiguchi } 27841e8f889bSDavid Gibson 278557303d80SAndy Whitcroft /* 278657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 278757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 278857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 278957303d80SAndy Whitcroft * the spinlock. 279057303d80SAndy Whitcroft */ 2791788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 27922b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 27932b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 27942b26736cSAndy Whitcroft goto backout_unlocked; 27952b26736cSAndy Whitcroft } 279657303d80SAndy Whitcroft 2797ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2798a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 27994c887265SAdam Litke if (idx >= size) 28004c887265SAdam Litke goto backout; 28014c887265SAdam Litke 280283c54070SNick Piggin ret = 0; 28037f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 28044c887265SAdam Litke goto backout; 28054c887265SAdam Litke 280607443a85SJoonsoo Kim if (anon_rmap) { 280707443a85SJoonsoo Kim ClearPagePrivate(page); 2808409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 280907443a85SJoonsoo Kim } 2810409eb8c2SHillf Danton else 2811409eb8c2SHillf Danton page_dup_rmap(page); 28121e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 28131e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 28141e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 28151e8f889bSDavid Gibson 2816788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 28171e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 281804f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 28191e8f889bSDavid Gibson } 28201e8f889bSDavid Gibson 2821ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 28224c887265SAdam Litke unlock_page(page); 28234c887265SAdam Litke out: 2824ac9b9c66SHugh Dickins return ret; 28254c887265SAdam Litke 28264c887265SAdam Litke backout: 28274c887265SAdam Litke spin_unlock(&mm->page_table_lock); 28282b26736cSAndy Whitcroft backout_unlocked: 28294c887265SAdam Litke unlock_page(page); 28304c887265SAdam Litke put_page(page); 28314c887265SAdam Litke goto out; 2832ac9b9c66SHugh Dickins } 2833ac9b9c66SHugh Dickins 283486e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2835788c7df4SHugh Dickins unsigned long address, unsigned int flags) 283686e5216fSAdam Litke { 283786e5216fSAdam Litke pte_t *ptep; 283886e5216fSAdam Litke pte_t entry; 28391e8f889bSDavid Gibson int ret; 28400fe6e20bSNaoya Horiguchi struct page *page = NULL; 284157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 28423935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2843a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 284486e5216fSAdam Litke 28451e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 28461e16a539SKAMEZAWA Hiroyuki 2847fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2848fd6a03edSNaoya Horiguchi if (ptep) { 2849fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2850290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 285130dad309SNaoya Horiguchi migration_entry_wait_huge(mm, ptep); 2852290408d4SNaoya Horiguchi return 0; 2853290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2854aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2855972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2856fd6a03edSNaoya Horiguchi } 2857fd6a03edSNaoya Horiguchi 2858a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 285986e5216fSAdam Litke if (!ptep) 286086e5216fSAdam Litke return VM_FAULT_OOM; 286186e5216fSAdam Litke 28623935baa9SDavid Gibson /* 28633935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 28643935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 28653935baa9SDavid Gibson * the same page in the page cache. 28663935baa9SDavid Gibson */ 28673935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 28687f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 28697f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2870788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2871b4d1d99fSDavid Gibson goto out_mutex; 28723935baa9SDavid Gibson } 287386e5216fSAdam Litke 287483c54070SNick Piggin ret = 0; 28751e8f889bSDavid Gibson 287657303d80SAndy Whitcroft /* 287757303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 287857303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 287957303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 288057303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 288157303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 288257303d80SAndy Whitcroft * consumed. 288357303d80SAndy Whitcroft */ 2884106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 28852b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28862b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2887b4d1d99fSDavid Gibson goto out_mutex; 28882b26736cSAndy Whitcroft } 288957303d80SAndy Whitcroft 2890f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 289157303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 289257303d80SAndy Whitcroft vma, address); 289357303d80SAndy Whitcroft } 289457303d80SAndy Whitcroft 289556c9cfb1SNaoya Horiguchi /* 289656c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 289756c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 289856c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 289956c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 290056c9cfb1SNaoya Horiguchi * so no worry about deadlock. 290156c9cfb1SNaoya Horiguchi */ 29020fe6e20bSNaoya Horiguchi page = pte_page(entry); 290366aebce7SChris Metcalf get_page(page); 290456c9cfb1SNaoya Horiguchi if (page != pagecache_page) 29050fe6e20bSNaoya Horiguchi lock_page(page); 29060fe6e20bSNaoya Horiguchi 29071e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 29081e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2909b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2910b4d1d99fSDavid Gibson goto out_page_table_lock; 2911b4d1d99fSDavid Gibson 2912b4d1d99fSDavid Gibson 2913788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2914106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 291557303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 291657303d80SAndy Whitcroft pagecache_page); 2917b4d1d99fSDavid Gibson goto out_page_table_lock; 2918b4d1d99fSDavid Gibson } 2919106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 2920b4d1d99fSDavid Gibson } 2921b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2922788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2923788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 29244b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2925b4d1d99fSDavid Gibson 2926b4d1d99fSDavid Gibson out_page_table_lock: 29271e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 292857303d80SAndy Whitcroft 292957303d80SAndy Whitcroft if (pagecache_page) { 293057303d80SAndy Whitcroft unlock_page(pagecache_page); 293157303d80SAndy Whitcroft put_page(pagecache_page); 293257303d80SAndy Whitcroft } 29331f64d69cSDean Nelson if (page != pagecache_page) 293456c9cfb1SNaoya Horiguchi unlock_page(page); 293566aebce7SChris Metcalf put_page(page); 293657303d80SAndy Whitcroft 2937b4d1d99fSDavid Gibson out_mutex: 29383935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 29391e8f889bSDavid Gibson 29401e8f889bSDavid Gibson return ret; 294186e5216fSAdam Litke } 294286e5216fSAdam Litke 294328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 294463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 294528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 294628a35716SMichel Lespinasse long i, unsigned int flags) 294763551ae0SDavid Gibson { 2948d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2949d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 295028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 2951a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 295263551ae0SDavid Gibson 29531c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 295463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 295563551ae0SDavid Gibson pte_t *pte; 29562a15efc9SHugh Dickins int absent; 295763551ae0SDavid Gibson struct page *page; 295863551ae0SDavid Gibson 29594c887265SAdam Litke /* 29604c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 29612a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 29624c887265SAdam Litke * first, for the page indexing below to work. 29634c887265SAdam Litke */ 2964a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 29652a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 296663551ae0SDavid Gibson 29672a15efc9SHugh Dickins /* 29682a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 29693ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 29703ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 29713ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 29723ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 29732a15efc9SHugh Dickins */ 29743ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 29753ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 29762a15efc9SHugh Dickins remainder = 0; 29772a15efc9SHugh Dickins break; 29782a15efc9SHugh Dickins } 29792a15efc9SHugh Dickins 29809cc3a5bdSNaoya Horiguchi /* 29819cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 29829cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 29839cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 29849cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 29859cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 29869cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 29879cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 29889cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 29899cc3a5bdSNaoya Horiguchi */ 29909cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 2991106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 2992106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 29934c887265SAdam Litke int ret; 29944c887265SAdam Litke 29954c887265SAdam Litke spin_unlock(&mm->page_table_lock); 29962a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 29972a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 29984c887265SAdam Litke spin_lock(&mm->page_table_lock); 2999a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 30004c887265SAdam Litke continue; 30014c887265SAdam Litke 30021c59827dSHugh Dickins remainder = 0; 30031c59827dSHugh Dickins break; 30041c59827dSHugh Dickins } 300563551ae0SDavid Gibson 3006a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 30077f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3008d5d4b0aaSChen, Kenneth W same_page: 3009d6692183SChen, Kenneth W if (pages) { 301069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 30114b2e38adSKOSAKI Motohiro get_page(pages[i]); 3012d6692183SChen, Kenneth W } 301363551ae0SDavid Gibson 301463551ae0SDavid Gibson if (vmas) 301563551ae0SDavid Gibson vmas[i] = vma; 301663551ae0SDavid Gibson 301763551ae0SDavid Gibson vaddr += PAGE_SIZE; 3018d5d4b0aaSChen, Kenneth W ++pfn_offset; 301963551ae0SDavid Gibson --remainder; 302063551ae0SDavid Gibson ++i; 3021d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3022a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3023d5d4b0aaSChen, Kenneth W /* 3024d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3025d5d4b0aaSChen, Kenneth W * of this compound page. 3026d5d4b0aaSChen, Kenneth W */ 3027d5d4b0aaSChen, Kenneth W goto same_page; 3028d5d4b0aaSChen, Kenneth W } 302963551ae0SDavid Gibson } 30301c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 303128a35716SMichel Lespinasse *nr_pages = remainder; 303263551ae0SDavid Gibson *position = vaddr; 303363551ae0SDavid Gibson 30342a15efc9SHugh Dickins return i ? i : -EFAULT; 303563551ae0SDavid Gibson } 30368f860591SZhang, Yanmin 30377da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 30388f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 30398f860591SZhang, Yanmin { 30408f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 30418f860591SZhang, Yanmin unsigned long start = address; 30428f860591SZhang, Yanmin pte_t *ptep; 30438f860591SZhang, Yanmin pte_t pte; 3044a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 30457da4d641SPeter Zijlstra unsigned long pages = 0; 30468f860591SZhang, Yanmin 30478f860591SZhang, Yanmin BUG_ON(address >= end); 30488f860591SZhang, Yanmin flush_cache_range(vma, address, end); 30498f860591SZhang, Yanmin 30503d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 30518f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 3052a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 30538f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 30548f860591SZhang, Yanmin if (!ptep) 30558f860591SZhang, Yanmin continue; 30567da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 30577da4d641SPeter Zijlstra pages++; 305839dde65cSChen, Kenneth W continue; 30597da4d641SPeter Zijlstra } 30607f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 30618f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3062106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3063be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 30648f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 30657da4d641SPeter Zijlstra pages++; 30668f860591SZhang, Yanmin } 30678f860591SZhang, Yanmin } 30688f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 3069d833352aSMel Gorman /* 3070d833352aSMel Gorman * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3071d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3072d833352aSMel Gorman * once we release i_mmap_mutex, another task can do the final put_page 3073d833352aSMel Gorman * and that page table be reused and filled with junk. 3074d833352aSMel Gorman */ 30758f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 3076d833352aSMel Gorman mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 30777da4d641SPeter Zijlstra 30787da4d641SPeter Zijlstra return pages << h->order; 30798f860591SZhang, Yanmin } 30808f860591SZhang, Yanmin 3081a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3082a1e78772SMel Gorman long from, long to, 30835a6fe125SMel Gorman struct vm_area_struct *vma, 3084ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3085e4e574b7SAdam Litke { 308617c9d12eSMel Gorman long ret, chg; 3087a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 308890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 3089e4e574b7SAdam Litke 3090a1e78772SMel Gorman /* 309117c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 309217c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 309390481622SDavid Gibson * without using reserves 309417c9d12eSMel Gorman */ 3095ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 309617c9d12eSMel Gorman return 0; 309717c9d12eSMel Gorman 309817c9d12eSMel Gorman /* 3099a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3100a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3101a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3102a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3103a1e78772SMel Gorman */ 3104f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 3105e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 31065a6fe125SMel Gorman else { 31075a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 31085a6fe125SMel Gorman if (!resv_map) 31095a6fe125SMel Gorman return -ENOMEM; 31105a6fe125SMel Gorman 311117c9d12eSMel Gorman chg = to - from; 311217c9d12eSMel Gorman 31135a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 31145a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 31155a6fe125SMel Gorman } 31165a6fe125SMel Gorman 3117c50ac050SDave Hansen if (chg < 0) { 3118c50ac050SDave Hansen ret = chg; 3119c50ac050SDave Hansen goto out_err; 3120c50ac050SDave Hansen } 312117c9d12eSMel Gorman 312290481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3123c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3124c50ac050SDave Hansen ret = -ENOSPC; 3125c50ac050SDave Hansen goto out_err; 3126c50ac050SDave Hansen } 312717c9d12eSMel Gorman 312817c9d12eSMel Gorman /* 312917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 313090481622SDavid Gibson * Hand the pages back to the subpool if there are not 313117c9d12eSMel Gorman */ 313217c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 313317c9d12eSMel Gorman if (ret < 0) { 313490481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3135c50ac050SDave Hansen goto out_err; 313617c9d12eSMel Gorman } 313717c9d12eSMel Gorman 313817c9d12eSMel Gorman /* 313917c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 314017c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 314117c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 314217c9d12eSMel Gorman * the reservation was and the page cache tells how much of 314317c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 314417c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 314517c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 314617c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 314717c9d12eSMel Gorman * else has to be done for private mappings here 314817c9d12eSMel Gorman */ 3149f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 315017c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 3151a43a8c39SChen, Kenneth W return 0; 3152c50ac050SDave Hansen out_err: 31534523e145SDave Hansen if (vma) 3154c50ac050SDave Hansen resv_map_put(vma); 3155c50ac050SDave Hansen return ret; 3156a43a8c39SChen, Kenneth W } 3157a43a8c39SChen, Kenneth W 3158a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3159a43a8c39SChen, Kenneth W { 3160a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 3161a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 316290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 316345c682a6SKen Chen 316445c682a6SKen Chen spin_lock(&inode->i_lock); 3165e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 316645c682a6SKen Chen spin_unlock(&inode->i_lock); 316745c682a6SKen Chen 316890481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3169a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3170a43a8c39SChen, Kenneth W } 317193f70f90SNaoya Horiguchi 31723212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 31733212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 31743212b535SSteve Capper struct vm_area_struct *vma, 31753212b535SSteve Capper unsigned long addr, pgoff_t idx) 31763212b535SSteve Capper { 31773212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 31783212b535SSteve Capper svma->vm_start; 31793212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 31803212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 31813212b535SSteve Capper 31823212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 31833212b535SSteve Capper unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 31843212b535SSteve Capper unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 31853212b535SSteve Capper 31863212b535SSteve Capper /* 31873212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 31883212b535SSteve Capper * page table page. 31893212b535SSteve Capper */ 31903212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 31913212b535SSteve Capper vm_flags != svm_flags || 31923212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 31933212b535SSteve Capper return 0; 31943212b535SSteve Capper 31953212b535SSteve Capper return saddr; 31963212b535SSteve Capper } 31973212b535SSteve Capper 31983212b535SSteve Capper static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 31993212b535SSteve Capper { 32003212b535SSteve Capper unsigned long base = addr & PUD_MASK; 32013212b535SSteve Capper unsigned long end = base + PUD_SIZE; 32023212b535SSteve Capper 32033212b535SSteve Capper /* 32043212b535SSteve Capper * check on proper vm_flags and page table alignment 32053212b535SSteve Capper */ 32063212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 32073212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 32083212b535SSteve Capper return 1; 32093212b535SSteve Capper return 0; 32103212b535SSteve Capper } 32113212b535SSteve Capper 32123212b535SSteve Capper /* 32133212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 32143212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 32153212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 32163212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 32173212b535SSteve Capper * pud has to be populated inside the same i_mmap_mutex section - otherwise 32183212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 32193212b535SSteve Capper * bad pmd for sharing. 32203212b535SSteve Capper */ 32213212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 32223212b535SSteve Capper { 32233212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 32243212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 32253212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 32263212b535SSteve Capper vma->vm_pgoff; 32273212b535SSteve Capper struct vm_area_struct *svma; 32283212b535SSteve Capper unsigned long saddr; 32293212b535SSteve Capper pte_t *spte = NULL; 32303212b535SSteve Capper pte_t *pte; 32313212b535SSteve Capper 32323212b535SSteve Capper if (!vma_shareable(vma, addr)) 32333212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 32343212b535SSteve Capper 32353212b535SSteve Capper mutex_lock(&mapping->i_mmap_mutex); 32363212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 32373212b535SSteve Capper if (svma == vma) 32383212b535SSteve Capper continue; 32393212b535SSteve Capper 32403212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 32413212b535SSteve Capper if (saddr) { 32423212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 32433212b535SSteve Capper if (spte) { 32443212b535SSteve Capper get_page(virt_to_page(spte)); 32453212b535SSteve Capper break; 32463212b535SSteve Capper } 32473212b535SSteve Capper } 32483212b535SSteve Capper } 32493212b535SSteve Capper 32503212b535SSteve Capper if (!spte) 32513212b535SSteve Capper goto out; 32523212b535SSteve Capper 32533212b535SSteve Capper spin_lock(&mm->page_table_lock); 32543212b535SSteve Capper if (pud_none(*pud)) 32553212b535SSteve Capper pud_populate(mm, pud, 32563212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 32573212b535SSteve Capper else 32583212b535SSteve Capper put_page(virt_to_page(spte)); 32593212b535SSteve Capper spin_unlock(&mm->page_table_lock); 32603212b535SSteve Capper out: 32613212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 32623212b535SSteve Capper mutex_unlock(&mapping->i_mmap_mutex); 32633212b535SSteve Capper return pte; 32643212b535SSteve Capper } 32653212b535SSteve Capper 32663212b535SSteve Capper /* 32673212b535SSteve Capper * unmap huge page backed by shared pte. 32683212b535SSteve Capper * 32693212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 32703212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 32713212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 32723212b535SSteve Capper * 32733212b535SSteve Capper * called with vma->vm_mm->page_table_lock held. 32743212b535SSteve Capper * 32753212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 32763212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 32773212b535SSteve Capper */ 32783212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 32793212b535SSteve Capper { 32803212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 32813212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 32823212b535SSteve Capper 32833212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 32843212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 32853212b535SSteve Capper return 0; 32863212b535SSteve Capper 32873212b535SSteve Capper pud_clear(pud); 32883212b535SSteve Capper put_page(virt_to_page(ptep)); 32893212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 32903212b535SSteve Capper return 1; 32913212b535SSteve Capper } 32929e5fc74cSSteve Capper #define want_pmd_share() (1) 32939e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 32949e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 32959e5fc74cSSteve Capper { 32969e5fc74cSSteve Capper return NULL; 32979e5fc74cSSteve Capper } 32989e5fc74cSSteve Capper #define want_pmd_share() (0) 32993212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 33003212b535SSteve Capper 33019e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 33029e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 33039e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 33049e5fc74cSSteve Capper { 33059e5fc74cSSteve Capper pgd_t *pgd; 33069e5fc74cSSteve Capper pud_t *pud; 33079e5fc74cSSteve Capper pte_t *pte = NULL; 33089e5fc74cSSteve Capper 33099e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 33109e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 33119e5fc74cSSteve Capper if (pud) { 33129e5fc74cSSteve Capper if (sz == PUD_SIZE) { 33139e5fc74cSSteve Capper pte = (pte_t *)pud; 33149e5fc74cSSteve Capper } else { 33159e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 33169e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 33179e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 33189e5fc74cSSteve Capper else 33199e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 33209e5fc74cSSteve Capper } 33219e5fc74cSSteve Capper } 33229e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 33239e5fc74cSSteve Capper 33249e5fc74cSSteve Capper return pte; 33259e5fc74cSSteve Capper } 33269e5fc74cSSteve Capper 33279e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 33289e5fc74cSSteve Capper { 33299e5fc74cSSteve Capper pgd_t *pgd; 33309e5fc74cSSteve Capper pud_t *pud; 33319e5fc74cSSteve Capper pmd_t *pmd = NULL; 33329e5fc74cSSteve Capper 33339e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 33349e5fc74cSSteve Capper if (pgd_present(*pgd)) { 33359e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 33369e5fc74cSSteve Capper if (pud_present(*pud)) { 33379e5fc74cSSteve Capper if (pud_huge(*pud)) 33389e5fc74cSSteve Capper return (pte_t *)pud; 33399e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 33409e5fc74cSSteve Capper } 33419e5fc74cSSteve Capper } 33429e5fc74cSSteve Capper return (pte_t *) pmd; 33439e5fc74cSSteve Capper } 33449e5fc74cSSteve Capper 33459e5fc74cSSteve Capper struct page * 33469e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 33479e5fc74cSSteve Capper pmd_t *pmd, int write) 33489e5fc74cSSteve Capper { 33499e5fc74cSSteve Capper struct page *page; 33509e5fc74cSSteve Capper 33519e5fc74cSSteve Capper page = pte_page(*(pte_t *)pmd); 33529e5fc74cSSteve Capper if (page) 33539e5fc74cSSteve Capper page += ((address & ~PMD_MASK) >> PAGE_SHIFT); 33549e5fc74cSSteve Capper return page; 33559e5fc74cSSteve Capper } 33569e5fc74cSSteve Capper 33579e5fc74cSSteve Capper struct page * 33589e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 33599e5fc74cSSteve Capper pud_t *pud, int write) 33609e5fc74cSSteve Capper { 33619e5fc74cSSteve Capper struct page *page; 33629e5fc74cSSteve Capper 33639e5fc74cSSteve Capper page = pte_page(*(pte_t *)pud); 33649e5fc74cSSteve Capper if (page) 33659e5fc74cSSteve Capper page += ((address & ~PUD_MASK) >> PAGE_SHIFT); 33669e5fc74cSSteve Capper return page; 33679e5fc74cSSteve Capper } 33689e5fc74cSSteve Capper 33699e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 33709e5fc74cSSteve Capper 33719e5fc74cSSteve Capper /* Can be overriden by architectures */ 33729e5fc74cSSteve Capper __attribute__((weak)) struct page * 33739e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 33749e5fc74cSSteve Capper pud_t *pud, int write) 33759e5fc74cSSteve Capper { 33769e5fc74cSSteve Capper BUG(); 33779e5fc74cSSteve Capper return NULL; 33789e5fc74cSSteve Capper } 33799e5fc74cSSteve Capper 33809e5fc74cSSteve Capper #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 33819e5fc74cSSteve Capper 3382d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3383d5bd9106SAndi Kleen 33846de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 33856de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 33866de2b1aaSNaoya Horiguchi { 33876de2b1aaSNaoya Horiguchi struct page *page; 33886de2b1aaSNaoya Horiguchi struct page *tmp; 33896de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 33906de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 33916de2b1aaSNaoya Horiguchi 33926de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 33936de2b1aaSNaoya Horiguchi if (page == hpage) 33946de2b1aaSNaoya Horiguchi return 1; 33956de2b1aaSNaoya Horiguchi return 0; 33966de2b1aaSNaoya Horiguchi } 33976de2b1aaSNaoya Horiguchi 339893f70f90SNaoya Horiguchi /* 339993f70f90SNaoya Horiguchi * This function is called from memory failure code. 340093f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 340193f70f90SNaoya Horiguchi */ 34026de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 340393f70f90SNaoya Horiguchi { 340493f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 340593f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 34066de2b1aaSNaoya Horiguchi int ret = -EBUSY; 340793f70f90SNaoya Horiguchi 340893f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 34096de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 341056f2fb14SNaoya Horiguchi /* 341156f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 341256f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 341356f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 341456f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 341556f2fb14SNaoya Horiguchi */ 341656f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 34178c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 341893f70f90SNaoya Horiguchi h->free_huge_pages--; 341993f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 34206de2b1aaSNaoya Horiguchi ret = 0; 342193f70f90SNaoya Horiguchi } 34226de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 34236de2b1aaSNaoya Horiguchi return ret; 34246de2b1aaSNaoya Horiguchi } 34256de2b1aaSNaoya Horiguchi #endif 342631caf665SNaoya Horiguchi 342731caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 342831caf665SNaoya Horiguchi { 342931caf665SNaoya Horiguchi VM_BUG_ON(!PageHead(page)); 343031caf665SNaoya Horiguchi if (!get_page_unless_zero(page)) 343131caf665SNaoya Horiguchi return false; 343231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 343331caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 343431caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 343531caf665SNaoya Horiguchi return true; 343631caf665SNaoya Horiguchi } 343731caf665SNaoya Horiguchi 343831caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 343931caf665SNaoya Horiguchi { 344031caf665SNaoya Horiguchi VM_BUG_ON(!PageHead(page)); 344131caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 344231caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 344331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 344431caf665SNaoya Horiguchi put_page(page); 344531caf665SNaoya Horiguchi } 3446