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 /* 513935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 523935baa9SDavid Gibson */ 53c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 540bd0f9fbSEric Paris 5590481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 5690481622SDavid Gibson { 5790481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 5890481622SDavid Gibson 5990481622SDavid Gibson spin_unlock(&spool->lock); 6090481622SDavid Gibson 6190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 6290481622SDavid Gibson * remain, free the subpool the subpool remain */ 6390481622SDavid Gibson if (free) 6490481622SDavid Gibson kfree(spool); 6590481622SDavid Gibson } 6690481622SDavid Gibson 6790481622SDavid Gibson struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 6890481622SDavid Gibson { 6990481622SDavid Gibson struct hugepage_subpool *spool; 7090481622SDavid Gibson 7190481622SDavid Gibson spool = kmalloc(sizeof(*spool), GFP_KERNEL); 7290481622SDavid Gibson if (!spool) 7390481622SDavid Gibson return NULL; 7490481622SDavid Gibson 7590481622SDavid Gibson spin_lock_init(&spool->lock); 7690481622SDavid Gibson spool->count = 1; 7790481622SDavid Gibson spool->max_hpages = nr_blocks; 7890481622SDavid Gibson spool->used_hpages = 0; 7990481622SDavid Gibson 8090481622SDavid Gibson return spool; 8190481622SDavid Gibson } 8290481622SDavid Gibson 8390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 8490481622SDavid Gibson { 8590481622SDavid Gibson spin_lock(&spool->lock); 8690481622SDavid Gibson BUG_ON(!spool->count); 8790481622SDavid Gibson spool->count--; 8890481622SDavid Gibson unlock_or_release_subpool(spool); 8990481622SDavid Gibson } 9090481622SDavid Gibson 9190481622SDavid Gibson static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 9290481622SDavid Gibson long delta) 9390481622SDavid Gibson { 9490481622SDavid Gibson int ret = 0; 9590481622SDavid Gibson 9690481622SDavid Gibson if (!spool) 9790481622SDavid Gibson return 0; 9890481622SDavid Gibson 9990481622SDavid Gibson spin_lock(&spool->lock); 10090481622SDavid Gibson if ((spool->used_hpages + delta) <= spool->max_hpages) { 10190481622SDavid Gibson spool->used_hpages += delta; 10290481622SDavid Gibson } else { 10390481622SDavid Gibson ret = -ENOMEM; 10490481622SDavid Gibson } 10590481622SDavid Gibson spin_unlock(&spool->lock); 10690481622SDavid Gibson 10790481622SDavid Gibson return ret; 10890481622SDavid Gibson } 10990481622SDavid Gibson 11090481622SDavid Gibson static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 11190481622SDavid Gibson long delta) 11290481622SDavid Gibson { 11390481622SDavid Gibson if (!spool) 11490481622SDavid Gibson return; 11590481622SDavid Gibson 11690481622SDavid Gibson spin_lock(&spool->lock); 11790481622SDavid Gibson spool->used_hpages -= delta; 11890481622SDavid Gibson /* If hugetlbfs_put_super couldn't free spool due to 11990481622SDavid Gibson * an outstanding quota reference, free it now. */ 12090481622SDavid Gibson unlock_or_release_subpool(spool); 12190481622SDavid Gibson } 12290481622SDavid Gibson 12390481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 12490481622SDavid Gibson { 12590481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 12690481622SDavid Gibson } 12790481622SDavid Gibson 12890481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 12990481622SDavid Gibson { 130496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 13190481622SDavid Gibson } 13290481622SDavid Gibson 133e7c4b0bfSAndy Whitcroft /* 13496822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 13596822904SAndy Whitcroft * across the pages in a mapping. 13684afd99bSAndy Whitcroft * 13784afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 13884afd99bSAndy Whitcroft * and the hugetlb_instantion_mutex. To access or modify a region the caller 13984afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 14084afd99bSAndy Whitcroft * the hugetlb_instantiation mutex: 14184afd99bSAndy Whitcroft * 14284afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 14384afd99bSAndy Whitcroft * or 14484afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 14584afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 14696822904SAndy Whitcroft */ 14796822904SAndy Whitcroft struct file_region { 14896822904SAndy Whitcroft struct list_head link; 14996822904SAndy Whitcroft long from; 15096822904SAndy Whitcroft long to; 15196822904SAndy Whitcroft }; 15296822904SAndy Whitcroft 15396822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 15496822904SAndy Whitcroft { 15596822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 15696822904SAndy Whitcroft 15796822904SAndy Whitcroft /* Locate the region we are either in or before. */ 15896822904SAndy Whitcroft list_for_each_entry(rg, head, link) 15996822904SAndy Whitcroft if (f <= rg->to) 16096822904SAndy Whitcroft break; 16196822904SAndy Whitcroft 16296822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 16396822904SAndy Whitcroft if (f > rg->from) 16496822904SAndy Whitcroft f = rg->from; 16596822904SAndy Whitcroft 16696822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 16796822904SAndy Whitcroft nrg = rg; 16896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 16996822904SAndy Whitcroft if (&rg->link == head) 17096822904SAndy Whitcroft break; 17196822904SAndy Whitcroft if (rg->from > t) 17296822904SAndy Whitcroft break; 17396822904SAndy Whitcroft 17496822904SAndy Whitcroft /* If this area reaches higher then extend our area to 17596822904SAndy Whitcroft * include it completely. If this is not the first area 17696822904SAndy Whitcroft * which we intend to reuse, free it. */ 17796822904SAndy Whitcroft if (rg->to > t) 17896822904SAndy Whitcroft t = rg->to; 17996822904SAndy Whitcroft if (rg != nrg) { 18096822904SAndy Whitcroft list_del(&rg->link); 18196822904SAndy Whitcroft kfree(rg); 18296822904SAndy Whitcroft } 18396822904SAndy Whitcroft } 18496822904SAndy Whitcroft nrg->from = f; 18596822904SAndy Whitcroft nrg->to = t; 18696822904SAndy Whitcroft return 0; 18796822904SAndy Whitcroft } 18896822904SAndy Whitcroft 18996822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 19096822904SAndy Whitcroft { 19196822904SAndy Whitcroft struct file_region *rg, *nrg; 19296822904SAndy Whitcroft long chg = 0; 19396822904SAndy Whitcroft 19496822904SAndy Whitcroft /* Locate the region we are before or in. */ 19596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 19696822904SAndy Whitcroft if (f <= rg->to) 19796822904SAndy Whitcroft break; 19896822904SAndy Whitcroft 19996822904SAndy Whitcroft /* If we are below the current region then a new region is required. 20096822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 20196822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 20296822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 20396822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 20496822904SAndy Whitcroft if (!nrg) 20596822904SAndy Whitcroft return -ENOMEM; 20696822904SAndy Whitcroft nrg->from = f; 20796822904SAndy Whitcroft nrg->to = f; 20896822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 20996822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 21096822904SAndy Whitcroft 21196822904SAndy Whitcroft return t - f; 21296822904SAndy Whitcroft } 21396822904SAndy Whitcroft 21496822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 21596822904SAndy Whitcroft if (f > rg->from) 21696822904SAndy Whitcroft f = rg->from; 21796822904SAndy Whitcroft chg = t - f; 21896822904SAndy Whitcroft 21996822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 22096822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 22196822904SAndy Whitcroft if (&rg->link == head) 22296822904SAndy Whitcroft break; 22396822904SAndy Whitcroft if (rg->from > t) 22496822904SAndy Whitcroft return chg; 22596822904SAndy Whitcroft 22625985edcSLucas De Marchi /* We overlap with this area, if it extends further than 22796822904SAndy Whitcroft * us then we must extend ourselves. Account for its 22896822904SAndy Whitcroft * existing reservation. */ 22996822904SAndy Whitcroft if (rg->to > t) { 23096822904SAndy Whitcroft chg += rg->to - t; 23196822904SAndy Whitcroft t = rg->to; 23296822904SAndy Whitcroft } 23396822904SAndy Whitcroft chg -= rg->to - rg->from; 23496822904SAndy Whitcroft } 23596822904SAndy Whitcroft return chg; 23696822904SAndy Whitcroft } 23796822904SAndy Whitcroft 23896822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 23996822904SAndy Whitcroft { 24096822904SAndy Whitcroft struct file_region *rg, *trg; 24196822904SAndy Whitcroft long chg = 0; 24296822904SAndy Whitcroft 24396822904SAndy Whitcroft /* Locate the region we are either in or before. */ 24496822904SAndy Whitcroft list_for_each_entry(rg, head, link) 24596822904SAndy Whitcroft if (end <= rg->to) 24696822904SAndy Whitcroft break; 24796822904SAndy Whitcroft if (&rg->link == head) 24896822904SAndy Whitcroft return 0; 24996822904SAndy Whitcroft 25096822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 25196822904SAndy Whitcroft if (end > rg->from) { 25296822904SAndy Whitcroft chg = rg->to - end; 25396822904SAndy Whitcroft rg->to = end; 25496822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 25596822904SAndy Whitcroft } 25696822904SAndy Whitcroft 25796822904SAndy Whitcroft /* Drop any remaining regions. */ 25896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 25996822904SAndy Whitcroft if (&rg->link == head) 26096822904SAndy Whitcroft break; 26196822904SAndy Whitcroft chg += rg->to - rg->from; 26296822904SAndy Whitcroft list_del(&rg->link); 26396822904SAndy Whitcroft kfree(rg); 26496822904SAndy Whitcroft } 26596822904SAndy Whitcroft return chg; 26696822904SAndy Whitcroft } 26796822904SAndy Whitcroft 26884afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 26984afd99bSAndy Whitcroft { 27084afd99bSAndy Whitcroft struct file_region *rg; 27184afd99bSAndy Whitcroft long chg = 0; 27284afd99bSAndy Whitcroft 27384afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 27484afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 275f2135a4aSWang Sheng-Hui long seg_from; 276f2135a4aSWang Sheng-Hui long seg_to; 27784afd99bSAndy Whitcroft 27884afd99bSAndy Whitcroft if (rg->to <= f) 27984afd99bSAndy Whitcroft continue; 28084afd99bSAndy Whitcroft if (rg->from >= t) 28184afd99bSAndy Whitcroft break; 28284afd99bSAndy Whitcroft 28384afd99bSAndy Whitcroft seg_from = max(rg->from, f); 28484afd99bSAndy Whitcroft seg_to = min(rg->to, t); 28584afd99bSAndy Whitcroft 28684afd99bSAndy Whitcroft chg += seg_to - seg_from; 28784afd99bSAndy Whitcroft } 28884afd99bSAndy Whitcroft 28984afd99bSAndy Whitcroft return chg; 29084afd99bSAndy Whitcroft } 29184afd99bSAndy Whitcroft 29296822904SAndy Whitcroft /* 293e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 294e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 295e7c4b0bfSAndy Whitcroft */ 296a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 297a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 298e7c4b0bfSAndy Whitcroft { 299a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 300a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 301e7c4b0bfSAndy Whitcroft } 302e7c4b0bfSAndy Whitcroft 3030fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 3040fe6e20bSNaoya Horiguchi unsigned long address) 3050fe6e20bSNaoya Horiguchi { 3060fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 3070fe6e20bSNaoya Horiguchi } 3080fe6e20bSNaoya Horiguchi 30984afd99bSAndy Whitcroft /* 31008fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 31108fba699SMel Gorman * cases this will be same size as used by the page table entries. 31208fba699SMel Gorman */ 31308fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 31408fba699SMel Gorman { 31508fba699SMel Gorman struct hstate *hstate; 31608fba699SMel Gorman 31708fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 31808fba699SMel Gorman return PAGE_SIZE; 31908fba699SMel Gorman 32008fba699SMel Gorman hstate = hstate_vma(vma); 32108fba699SMel Gorman 3222415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 32308fba699SMel Gorman } 324f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 32508fba699SMel Gorman 32608fba699SMel Gorman /* 3273340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 3283340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 3293340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 3303340289dSMel Gorman * function is required. 3313340289dSMel Gorman */ 3323340289dSMel Gorman #ifndef vma_mmu_pagesize 3333340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 3343340289dSMel Gorman { 3353340289dSMel Gorman return vma_kernel_pagesize(vma); 3363340289dSMel Gorman } 3373340289dSMel Gorman #endif 3383340289dSMel Gorman 3393340289dSMel Gorman /* 34084afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 34184afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 34284afd99bSAndy Whitcroft * alignment. 34384afd99bSAndy Whitcroft */ 34484afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 34584afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 34604f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 34784afd99bSAndy Whitcroft 348a1e78772SMel Gorman /* 349a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 350a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 351a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 352a1e78772SMel Gorman * 353a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 354a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 355a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 356a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 35784afd99bSAndy Whitcroft * 35884afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 35984afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 36084afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 36184afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 36284afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 36384afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 36484afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 36584afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 366a1e78772SMel Gorman */ 367e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 368e7c4b0bfSAndy Whitcroft { 369e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 370e7c4b0bfSAndy Whitcroft } 371e7c4b0bfSAndy Whitcroft 372e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 373e7c4b0bfSAndy Whitcroft unsigned long value) 374e7c4b0bfSAndy Whitcroft { 375e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 376e7c4b0bfSAndy Whitcroft } 377e7c4b0bfSAndy Whitcroft 37884afd99bSAndy Whitcroft struct resv_map { 37984afd99bSAndy Whitcroft struct kref refs; 38084afd99bSAndy Whitcroft struct list_head regions; 38184afd99bSAndy Whitcroft }; 38284afd99bSAndy Whitcroft 3832a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 38484afd99bSAndy Whitcroft { 38584afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 38684afd99bSAndy Whitcroft if (!resv_map) 38784afd99bSAndy Whitcroft return NULL; 38884afd99bSAndy Whitcroft 38984afd99bSAndy Whitcroft kref_init(&resv_map->refs); 39084afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 39184afd99bSAndy Whitcroft 39284afd99bSAndy Whitcroft return resv_map; 39384afd99bSAndy Whitcroft } 39484afd99bSAndy Whitcroft 3952a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 39684afd99bSAndy Whitcroft { 39784afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 39884afd99bSAndy Whitcroft 39984afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 40084afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 40184afd99bSAndy Whitcroft kfree(resv_map); 40284afd99bSAndy Whitcroft } 40384afd99bSAndy Whitcroft 40484afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 405a1e78772SMel Gorman { 406a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 407f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 40884afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 40984afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 4102a4b3dedSHarvey Harrison return NULL; 411a1e78772SMel Gorman } 412a1e78772SMel Gorman 41384afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 414a1e78772SMel Gorman { 415a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 416f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 417a1e78772SMel Gorman 41884afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 41984afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 42004f2cbe3SMel Gorman } 42104f2cbe3SMel Gorman 42204f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 42304f2cbe3SMel Gorman { 42404f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 425f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 426e7c4b0bfSAndy Whitcroft 427e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 42804f2cbe3SMel Gorman } 42904f2cbe3SMel Gorman 43004f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 43104f2cbe3SMel Gorman { 43204f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 433e7c4b0bfSAndy Whitcroft 434e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 435a1e78772SMel Gorman } 436a1e78772SMel Gorman 437a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 438a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 439a5516438SAndi Kleen struct vm_area_struct *vma) 440a1e78772SMel Gorman { 441c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 442c37f9fb1SAndy Whitcroft return; 443c37f9fb1SAndy Whitcroft 444f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 445a1e78772SMel Gorman /* Shared mappings always use reserves */ 446a5516438SAndi Kleen h->resv_huge_pages--; 44784afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 448a1e78772SMel Gorman /* 449a1e78772SMel Gorman * Only the process that called mmap() has reserves for 450a1e78772SMel Gorman * private mappings. 451a1e78772SMel Gorman */ 452a5516438SAndi Kleen h->resv_huge_pages--; 453a1e78772SMel Gorman } 454a1e78772SMel Gorman } 455a1e78772SMel Gorman 45604f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 457a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 458a1e78772SMel Gorman { 459a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 460f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 461a1e78772SMel Gorman vma->vm_private_data = (void *)0; 462a1e78772SMel Gorman } 463a1e78772SMel Gorman 464a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 4657f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 466a1e78772SMel Gorman { 467f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 468a1e78772SMel Gorman return 1; 4697f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4707f09ca51SMel Gorman return 1; 4717f09ca51SMel Gorman return 0; 472a1e78772SMel Gorman } 473a1e78772SMel Gorman 4740ebabb41SNaoya Horiguchi static void copy_gigantic_page(struct page *dst, struct page *src) 4750ebabb41SNaoya Horiguchi { 4760ebabb41SNaoya Horiguchi int i; 4770ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4780ebabb41SNaoya Horiguchi struct page *dst_base = dst; 4790ebabb41SNaoya Horiguchi struct page *src_base = src; 4800ebabb41SNaoya Horiguchi 4810ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); ) { 4820ebabb41SNaoya Horiguchi cond_resched(); 4830ebabb41SNaoya Horiguchi copy_highpage(dst, src); 4840ebabb41SNaoya Horiguchi 4850ebabb41SNaoya Horiguchi i++; 4860ebabb41SNaoya Horiguchi dst = mem_map_next(dst, dst_base, i); 4870ebabb41SNaoya Horiguchi src = mem_map_next(src, src_base, i); 4880ebabb41SNaoya Horiguchi } 4890ebabb41SNaoya Horiguchi } 4900ebabb41SNaoya Horiguchi 4910ebabb41SNaoya Horiguchi void copy_huge_page(struct page *dst, struct page *src) 4920ebabb41SNaoya Horiguchi { 4930ebabb41SNaoya Horiguchi int i; 4940ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4950ebabb41SNaoya Horiguchi 4960ebabb41SNaoya Horiguchi if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 4970ebabb41SNaoya Horiguchi copy_gigantic_page(dst, src); 4980ebabb41SNaoya Horiguchi return; 4990ebabb41SNaoya Horiguchi } 5000ebabb41SNaoya Horiguchi 5010ebabb41SNaoya Horiguchi might_sleep(); 5020ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); i++) { 5030ebabb41SNaoya Horiguchi cond_resched(); 5040ebabb41SNaoya Horiguchi copy_highpage(dst + i, src + i); 5050ebabb41SNaoya Horiguchi } 5060ebabb41SNaoya Horiguchi } 5070ebabb41SNaoya Horiguchi 508a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 5091da177e4SLinus Torvalds { 5101da177e4SLinus Torvalds int nid = page_to_nid(page); 5110edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 512a5516438SAndi Kleen h->free_huge_pages++; 513a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 5141da177e4SLinus Torvalds } 5151da177e4SLinus Torvalds 516bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 517bf50bab2SNaoya Horiguchi { 518bf50bab2SNaoya Horiguchi struct page *page; 519bf50bab2SNaoya Horiguchi 520bf50bab2SNaoya Horiguchi if (list_empty(&h->hugepage_freelists[nid])) 521bf50bab2SNaoya Horiguchi return NULL; 522bf50bab2SNaoya Horiguchi page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); 5230edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 524a9869b83SNaoya Horiguchi set_page_refcounted(page); 525bf50bab2SNaoya Horiguchi h->free_huge_pages--; 526bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 527bf50bab2SNaoya Horiguchi return page; 528bf50bab2SNaoya Horiguchi } 529bf50bab2SNaoya Horiguchi 530a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 531a5516438SAndi Kleen struct vm_area_struct *vma, 53204f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 5331da177e4SLinus Torvalds { 534b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 535480eccf9SLee Schermerhorn struct mempolicy *mpol; 53619770b32SMel Gorman nodemask_t *nodemask; 537c0ff7453SMiao Xie struct zonelist *zonelist; 538dd1a239fSMel Gorman struct zone *zone; 539dd1a239fSMel Gorman struct zoneref *z; 540cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 5411da177e4SLinus Torvalds 542cc9a6c87SMel Gorman retry_cpuset: 543cc9a6c87SMel Gorman cpuset_mems_cookie = get_mems_allowed(); 544c0ff7453SMiao Xie zonelist = huge_zonelist(vma, address, 545c0ff7453SMiao Xie htlb_alloc_mask, &mpol, &nodemask); 546a1e78772SMel Gorman /* 547a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 548a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 549a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 550a1e78772SMel Gorman */ 5517f09ca51SMel Gorman if (!vma_has_reserves(vma) && 552a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 553c0ff7453SMiao Xie goto err; 554a1e78772SMel Gorman 55504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 556a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 5576eab04a8SJustin P. Mattock goto err; 55804f2cbe3SMel Gorman 55919770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 56019770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 561bf50bab2SNaoya Horiguchi if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { 562bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 563bf50bab2SNaoya Horiguchi if (page) { 56404f2cbe3SMel Gorman if (!avoid_reserve) 565a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 5665ab3ee7bSKen Chen break; 5671da177e4SLinus Torvalds } 5683abf7afdSAndrew Morton } 569bf50bab2SNaoya Horiguchi } 570cc9a6c87SMel Gorman 571cc9a6c87SMel Gorman mpol_cond_put(mpol); 572cc9a6c87SMel Gorman if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) 573cc9a6c87SMel Gorman goto retry_cpuset; 574cc9a6c87SMel Gorman return page; 575cc9a6c87SMel Gorman 576c0ff7453SMiao Xie err: 57752cd3b07SLee Schermerhorn mpol_cond_put(mpol); 578cc9a6c87SMel Gorman return NULL; 5791da177e4SLinus Torvalds } 5801da177e4SLinus Torvalds 581a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5826af2acb6SAdam Litke { 5836af2acb6SAdam Litke int i; 584a5516438SAndi Kleen 58518229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 58618229df5SAndy Whitcroft 587a5516438SAndi Kleen h->nr_huge_pages--; 588a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 589a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 59032f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 59132f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 59232f84528SChris Forbes 1 << PG_active | 1 << PG_reserved | 5936af2acb6SAdam Litke 1 << PG_private | 1 << PG_writeback); 5946af2acb6SAdam Litke } 5959dd540e2SAneesh Kumar K.V VM_BUG_ON(hugetlb_cgroup_from_page(page)); 5966af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5976af2acb6SAdam Litke set_page_refcounted(page); 5987f2e9525SGerald Schaefer arch_release_hugepage(page); 599a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 6006af2acb6SAdam Litke } 6016af2acb6SAdam Litke 602e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 603e5ff2159SAndi Kleen { 604e5ff2159SAndi Kleen struct hstate *h; 605e5ff2159SAndi Kleen 606e5ff2159SAndi Kleen for_each_hstate(h) { 607e5ff2159SAndi Kleen if (huge_page_size(h) == size) 608e5ff2159SAndi Kleen return h; 609e5ff2159SAndi Kleen } 610e5ff2159SAndi Kleen return NULL; 611e5ff2159SAndi Kleen } 612e5ff2159SAndi Kleen 61327a85ef1SDavid Gibson static void free_huge_page(struct page *page) 61427a85ef1SDavid Gibson { 615a5516438SAndi Kleen /* 616a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 617a5516438SAndi Kleen * compound page destructor. 618a5516438SAndi Kleen */ 619e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 6207893d1d5SAdam Litke int nid = page_to_nid(page); 62190481622SDavid Gibson struct hugepage_subpool *spool = 62290481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 62327a85ef1SDavid Gibson 624e5df70abSAndy Whitcroft set_page_private(page, 0); 62523be7468SMel Gorman page->mapping = NULL; 6267893d1d5SAdam Litke BUG_ON(page_count(page)); 6270fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 62827a85ef1SDavid Gibson 62927a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 6306d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 6316d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 632aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 6330edaecfaSAneesh Kumar K.V /* remove the page from active list */ 6340edaecfaSAneesh Kumar K.V list_del(&page->lru); 635a5516438SAndi Kleen update_and_free_page(h, page); 636a5516438SAndi Kleen h->surplus_huge_pages--; 637a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 6387893d1d5SAdam Litke } else { 6395d3a551cSWill Deacon arch_clear_hugepage_flags(page); 640a5516438SAndi Kleen enqueue_huge_page(h, page); 6417893d1d5SAdam Litke } 64227a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 64390481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 64427a85ef1SDavid Gibson } 64527a85ef1SDavid Gibson 646a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 647b7ba30c6SAndi Kleen { 6480edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 649b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 650b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 6519dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 652a5516438SAndi Kleen h->nr_huge_pages++; 653a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 654b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 655b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 656b7ba30c6SAndi Kleen } 657b7ba30c6SAndi Kleen 65820a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 65920a0307cSWu Fengguang { 66020a0307cSWu Fengguang int i; 66120a0307cSWu Fengguang int nr_pages = 1 << order; 66220a0307cSWu Fengguang struct page *p = page + 1; 66320a0307cSWu Fengguang 66420a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 66520a0307cSWu Fengguang set_compound_order(page, order); 66620a0307cSWu Fengguang __SetPageHead(page); 66720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 66820a0307cSWu Fengguang __SetPageTail(p); 66958a84aa9SYouquan Song set_page_count(p, 0); 67020a0307cSWu Fengguang p->first_page = page; 67120a0307cSWu Fengguang } 67220a0307cSWu Fengguang } 67320a0307cSWu Fengguang 6747795912cSAndrew Morton /* 6757795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 6767795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 6777795912cSAndrew Morton * details. 6787795912cSAndrew Morton */ 67920a0307cSWu Fengguang int PageHuge(struct page *page) 68020a0307cSWu Fengguang { 68120a0307cSWu Fengguang compound_page_dtor *dtor; 68220a0307cSWu Fengguang 68320a0307cSWu Fengguang if (!PageCompound(page)) 68420a0307cSWu Fengguang return 0; 68520a0307cSWu Fengguang 68620a0307cSWu Fengguang page = compound_head(page); 68720a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 68820a0307cSWu Fengguang 68920a0307cSWu Fengguang return dtor == free_huge_page; 69020a0307cSWu Fengguang } 69143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 69243131e14SNaoya Horiguchi 69313d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 69413d60f4bSZhang Yi { 69513d60f4bSZhang Yi struct page *page_head = compound_head(page); 69613d60f4bSZhang Yi pgoff_t index = page_index(page_head); 69713d60f4bSZhang Yi unsigned long compound_idx; 69813d60f4bSZhang Yi 69913d60f4bSZhang Yi if (!PageHuge(page_head)) 70013d60f4bSZhang Yi return page_index(page); 70113d60f4bSZhang Yi 70213d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 70313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 70413d60f4bSZhang Yi else 70513d60f4bSZhang Yi compound_idx = page - page_head; 70613d60f4bSZhang Yi 70713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 70813d60f4bSZhang Yi } 70913d60f4bSZhang Yi 710a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 7111da177e4SLinus Torvalds { 7121da177e4SLinus Torvalds struct page *page; 713f96efd58SJoe Jin 714aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 715aa888a74SAndi Kleen return NULL; 716aa888a74SAndi Kleen 7176484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 718551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 719551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 720a5516438SAndi Kleen huge_page_order(h)); 7211da177e4SLinus Torvalds if (page) { 7227f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 723caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 7247b8ee84dSHarvey Harrison return NULL; 7257f2e9525SGerald Schaefer } 726a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 7271da177e4SLinus Torvalds } 72863b4613cSNishanth Aravamudan 72963b4613cSNishanth Aravamudan return page; 73063b4613cSNishanth Aravamudan } 73163b4613cSNishanth Aravamudan 7325ced66c9SAndi Kleen /* 7336ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 7346ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 7356ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7366ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 7376ae11b27SLee Schermerhorn * node for alloc or free. 7389a76db09SLee Schermerhorn */ 7396ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7409a76db09SLee Schermerhorn { 7416ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 7429a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 7436ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 7449a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 7459a76db09SLee Schermerhorn 7469a76db09SLee Schermerhorn return nid; 7479a76db09SLee Schermerhorn } 7489a76db09SLee Schermerhorn 7496ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 7505ced66c9SAndi Kleen { 7516ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 7526ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 7539a76db09SLee Schermerhorn return nid; 7545ced66c9SAndi Kleen } 7555ced66c9SAndi Kleen 7566ae11b27SLee Schermerhorn /* 7576ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 7586ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 7596ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 7606ae11b27SLee Schermerhorn * mask. 7616ae11b27SLee Schermerhorn */ 7626ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 7636ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 7646ae11b27SLee Schermerhorn { 7656ae11b27SLee Schermerhorn int nid; 7666ae11b27SLee Schermerhorn 7676ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7686ae11b27SLee Schermerhorn 7696ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 7706ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 7716ae11b27SLee Schermerhorn 7726ae11b27SLee Schermerhorn return nid; 7736ae11b27SLee Schermerhorn } 7746ae11b27SLee Schermerhorn 7756ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 77663b4613cSNishanth Aravamudan { 77763b4613cSNishanth Aravamudan struct page *page; 77863b4613cSNishanth Aravamudan int start_nid; 77963b4613cSNishanth Aravamudan int next_nid; 78063b4613cSNishanth Aravamudan int ret = 0; 78163b4613cSNishanth Aravamudan 7826ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 783e8c5c824SLee Schermerhorn next_nid = start_nid; 78463b4613cSNishanth Aravamudan 78563b4613cSNishanth Aravamudan do { 786e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 7879a76db09SLee Schermerhorn if (page) { 78863b4613cSNishanth Aravamudan ret = 1; 7899a76db09SLee Schermerhorn break; 7909a76db09SLee Schermerhorn } 7916ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 7929a76db09SLee Schermerhorn } while (next_nid != start_nid); 79363b4613cSNishanth Aravamudan 7943b116300SAdam Litke if (ret) 7953b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7963b116300SAdam Litke else 7973b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7983b116300SAdam Litke 79963b4613cSNishanth Aravamudan return ret; 8001da177e4SLinus Torvalds } 8011da177e4SLinus Torvalds 802e8c5c824SLee Schermerhorn /* 8036ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 8046ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 8056ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 8066ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 807e8c5c824SLee Schermerhorn */ 8086ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 809e8c5c824SLee Schermerhorn { 8106ae11b27SLee Schermerhorn int nid; 8119a76db09SLee Schermerhorn 8126ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 8136ae11b27SLee Schermerhorn 8146ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 8156ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 8166ae11b27SLee Schermerhorn 8179a76db09SLee Schermerhorn return nid; 818e8c5c824SLee Schermerhorn } 819e8c5c824SLee Schermerhorn 820e8c5c824SLee Schermerhorn /* 821e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 822e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 823e8c5c824SLee Schermerhorn * balanced over allowed nodes. 824e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 825e8c5c824SLee Schermerhorn */ 8266ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 8276ae11b27SLee Schermerhorn bool acct_surplus) 828e8c5c824SLee Schermerhorn { 829e8c5c824SLee Schermerhorn int start_nid; 830e8c5c824SLee Schermerhorn int next_nid; 831e8c5c824SLee Schermerhorn int ret = 0; 832e8c5c824SLee Schermerhorn 8336ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 834e8c5c824SLee Schermerhorn next_nid = start_nid; 835e8c5c824SLee Schermerhorn 836e8c5c824SLee Schermerhorn do { 837685f3457SLee Schermerhorn /* 838685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 839685f3457SLee Schermerhorn * nodes with surplus pages. 840685f3457SLee Schermerhorn */ 841685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 842685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 843e8c5c824SLee Schermerhorn struct page *page = 844e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 845e8c5c824SLee Schermerhorn struct page, lru); 846e8c5c824SLee Schermerhorn list_del(&page->lru); 847e8c5c824SLee Schermerhorn h->free_huge_pages--; 848e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 849685f3457SLee Schermerhorn if (acct_surplus) { 850685f3457SLee Schermerhorn h->surplus_huge_pages--; 851685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 852685f3457SLee Schermerhorn } 853e8c5c824SLee Schermerhorn update_and_free_page(h, page); 854e8c5c824SLee Schermerhorn ret = 1; 8559a76db09SLee Schermerhorn break; 856e8c5c824SLee Schermerhorn } 8576ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 8589a76db09SLee Schermerhorn } while (next_nid != start_nid); 859e8c5c824SLee Schermerhorn 860e8c5c824SLee Schermerhorn return ret; 861e8c5c824SLee Schermerhorn } 862e8c5c824SLee Schermerhorn 863bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 8647893d1d5SAdam Litke { 8657893d1d5SAdam Litke struct page *page; 866bf50bab2SNaoya Horiguchi unsigned int r_nid; 8677893d1d5SAdam Litke 868aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 869aa888a74SAndi Kleen return NULL; 870aa888a74SAndi Kleen 871d1c3fb1fSNishanth Aravamudan /* 872d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 873d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 874d1c3fb1fSNishanth Aravamudan * overcommit 875d1c3fb1fSNishanth Aravamudan * 876d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 877d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 878d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 879d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 880d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 881d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 882d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 883d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 884d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 885d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 886d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 887d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 888d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 889d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 890d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 891d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 892d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 893d1c3fb1fSNishanth Aravamudan */ 894d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 895a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 896d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 897d1c3fb1fSNishanth Aravamudan return NULL; 898d1c3fb1fSNishanth Aravamudan } else { 899a5516438SAndi Kleen h->nr_huge_pages++; 900a5516438SAndi Kleen h->surplus_huge_pages++; 901d1c3fb1fSNishanth Aravamudan } 902d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 903d1c3fb1fSNishanth Aravamudan 904bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 905551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 906551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 907a5516438SAndi Kleen huge_page_order(h)); 908bf50bab2SNaoya Horiguchi else 909bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 910bf50bab2SNaoya Horiguchi htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 911bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 912d1c3fb1fSNishanth Aravamudan 913caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 914caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 915ea5768c7SHillf Danton page = NULL; 916caff3a2cSGerald Schaefer } 917caff3a2cSGerald Schaefer 9187893d1d5SAdam Litke spin_lock(&hugetlb_lock); 919d1c3fb1fSNishanth Aravamudan if (page) { 9200edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 921bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 922d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 9239dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 924d1c3fb1fSNishanth Aravamudan /* 925d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 926d1c3fb1fSNishanth Aravamudan */ 927bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 928bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 9293b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 930d1c3fb1fSNishanth Aravamudan } else { 931a5516438SAndi Kleen h->nr_huge_pages--; 932a5516438SAndi Kleen h->surplus_huge_pages--; 9333b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 9347893d1d5SAdam Litke } 935d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 9367893d1d5SAdam Litke 9377893d1d5SAdam Litke return page; 9387893d1d5SAdam Litke } 9397893d1d5SAdam Litke 940e4e574b7SAdam Litke /* 941bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 942bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 943bf50bab2SNaoya Horiguchi * address of error page. 944bf50bab2SNaoya Horiguchi */ 945bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 946bf50bab2SNaoya Horiguchi { 947bf50bab2SNaoya Horiguchi struct page *page; 948bf50bab2SNaoya Horiguchi 949bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 950bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 951bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 952bf50bab2SNaoya Horiguchi 95394ae8ba7SAneesh Kumar K.V if (!page) 954bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 955bf50bab2SNaoya Horiguchi 956bf50bab2SNaoya Horiguchi return page; 957bf50bab2SNaoya Horiguchi } 958bf50bab2SNaoya Horiguchi 959bf50bab2SNaoya Horiguchi /* 96025985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 961e4e574b7SAdam Litke * of size 'delta'. 962e4e574b7SAdam Litke */ 963a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 964e4e574b7SAdam Litke { 965e4e574b7SAdam Litke struct list_head surplus_list; 966e4e574b7SAdam Litke struct page *page, *tmp; 967e4e574b7SAdam Litke int ret, i; 968e4e574b7SAdam Litke int needed, allocated; 96928073b02SHillf Danton bool alloc_ok = true; 970e4e574b7SAdam Litke 971a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 972ac09b3a1SAdam Litke if (needed <= 0) { 973a5516438SAndi Kleen h->resv_huge_pages += delta; 974e4e574b7SAdam Litke return 0; 975ac09b3a1SAdam Litke } 976e4e574b7SAdam Litke 977e4e574b7SAdam Litke allocated = 0; 978e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 979e4e574b7SAdam Litke 980e4e574b7SAdam Litke ret = -ENOMEM; 981e4e574b7SAdam Litke retry: 982e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 983e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 984bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 98528073b02SHillf Danton if (!page) { 98628073b02SHillf Danton alloc_ok = false; 98728073b02SHillf Danton break; 98828073b02SHillf Danton } 989e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 990e4e574b7SAdam Litke } 99128073b02SHillf Danton allocated += i; 992e4e574b7SAdam Litke 993e4e574b7SAdam Litke /* 994e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 995e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 996e4e574b7SAdam Litke */ 997e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 998a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 999a5516438SAndi Kleen (h->free_huge_pages + allocated); 100028073b02SHillf Danton if (needed > 0) { 100128073b02SHillf Danton if (alloc_ok) 1002e4e574b7SAdam Litke goto retry; 100328073b02SHillf Danton /* 100428073b02SHillf Danton * We were not able to allocate enough pages to 100528073b02SHillf Danton * satisfy the entire reservation so we free what 100628073b02SHillf Danton * we've allocated so far. 100728073b02SHillf Danton */ 100828073b02SHillf Danton goto free; 100928073b02SHillf Danton } 1010e4e574b7SAdam Litke /* 1011e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 101225985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1013e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1014ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1015ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1016ac09b3a1SAdam Litke * before they are reserved. 1017e4e574b7SAdam Litke */ 1018e4e574b7SAdam Litke needed += allocated; 1019a5516438SAndi Kleen h->resv_huge_pages += delta; 1020e4e574b7SAdam Litke ret = 0; 1021a9869b83SNaoya Horiguchi 102219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 102319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 102419fc3f0aSAdam Litke if ((--needed) < 0) 102519fc3f0aSAdam Litke break; 1026a9869b83SNaoya Horiguchi /* 1027a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1028a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1029a9869b83SNaoya Horiguchi */ 1030a9869b83SNaoya Horiguchi put_page_testzero(page); 1031a9869b83SNaoya Horiguchi VM_BUG_ON(page_count(page)); 1032a5516438SAndi Kleen enqueue_huge_page(h, page); 103319fc3f0aSAdam Litke } 103428073b02SHillf Danton free: 1035b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 103619fc3f0aSAdam Litke 103719fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 103819fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 1039e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1040a9869b83SNaoya Horiguchi put_page(page); 1041a9869b83SNaoya Horiguchi } 1042af767cbdSAdam Litke } 104319fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1044e4e574b7SAdam Litke 1045e4e574b7SAdam Litke return ret; 1046e4e574b7SAdam Litke } 1047e4e574b7SAdam Litke 1048e4e574b7SAdam Litke /* 1049e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1050e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1051e4e574b7SAdam Litke * never used. 1052685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1053e4e574b7SAdam Litke */ 1054a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1055a5516438SAndi Kleen unsigned long unused_resv_pages) 1056e4e574b7SAdam Litke { 1057e4e574b7SAdam Litke unsigned long nr_pages; 1058e4e574b7SAdam Litke 1059ac09b3a1SAdam Litke /* Uncommit the reservation */ 1060a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1061ac09b3a1SAdam Litke 1062aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1063aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1064aa888a74SAndi Kleen return; 1065aa888a74SAndi Kleen 1066a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1067e4e574b7SAdam Litke 1068685f3457SLee Schermerhorn /* 1069685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 10709b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 10719b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 10729b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 10739b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 10749b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1075685f3457SLee Schermerhorn */ 1076685f3457SLee Schermerhorn while (nr_pages--) { 10778cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1078685f3457SLee Schermerhorn break; 1079e4e574b7SAdam Litke } 1080e4e574b7SAdam Litke } 1081e4e574b7SAdam Litke 1082c37f9fb1SAndy Whitcroft /* 1083c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1084c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 108590481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 108690481622SDavid Gibson * can occur. Where any new reservation would be required the 108790481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 108890481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 108990481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 109090481622SDavid Gibson * failure. 1091c37f9fb1SAndy Whitcroft */ 1092e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1093a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1094c37f9fb1SAndy Whitcroft { 1095c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1096c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1097c37f9fb1SAndy Whitcroft 1098f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1099a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1100c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 1101c37f9fb1SAndy Whitcroft idx, idx + 1); 1102c37f9fb1SAndy Whitcroft 110384afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1104c37f9fb1SAndy Whitcroft return 1; 1105c37f9fb1SAndy Whitcroft 110684afd99bSAndy Whitcroft } else { 1107e2f17d94SRoel Kluin long err; 1108a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 110984afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 111084afd99bSAndy Whitcroft 111184afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 111284afd99bSAndy Whitcroft if (err < 0) 111384afd99bSAndy Whitcroft return err; 1114c37f9fb1SAndy Whitcroft return 0; 1115c37f9fb1SAndy Whitcroft } 111684afd99bSAndy Whitcroft } 1117a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1118a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1119c37f9fb1SAndy Whitcroft { 1120c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1121c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1122c37f9fb1SAndy Whitcroft 1123f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1124a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1125c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 112684afd99bSAndy Whitcroft 112784afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1128a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 112984afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 113084afd99bSAndy Whitcroft 113184afd99bSAndy Whitcroft /* Mark this page used in the map. */ 113284afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1133c37f9fb1SAndy Whitcroft } 1134c37f9fb1SAndy Whitcroft } 1135c37f9fb1SAndy Whitcroft 1136348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 113704f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1138348ea204SAdam Litke { 113990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1140a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1141348ea204SAdam Litke struct page *page; 1142e2f17d94SRoel Kluin long chg; 11436d76dcf4SAneesh Kumar K.V int ret, idx; 11446d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 11452fc39cecSAdam Litke 11466d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1147a1e78772SMel Gorman /* 114890481622SDavid Gibson * Processes that did not create the mapping will have no 114990481622SDavid Gibson * reserves and will not have accounted against subpool 115090481622SDavid Gibson * limit. Check that the subpool limit can be made before 115190481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 115290481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 115390481622SDavid Gibson * mapping overlaps. 1154a1e78772SMel Gorman */ 1155a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1156c37f9fb1SAndy Whitcroft if (chg < 0) 115776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1158c37f9fb1SAndy Whitcroft if (chg) 115990481622SDavid Gibson if (hugepage_subpool_get_pages(spool, chg)) 116076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 116190d8b7e6SAdam Litke 11626d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 11636d76dcf4SAneesh Kumar K.V if (ret) { 11646d76dcf4SAneesh Kumar K.V hugepage_subpool_put_pages(spool, chg); 11656d76dcf4SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 11666d76dcf4SAneesh Kumar K.V } 1167a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1168a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 116994ae8ba7SAneesh Kumar K.V if (page) { 117094ae8ba7SAneesh Kumar K.V /* update page cgroup details */ 117194ae8ba7SAneesh Kumar K.V hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), 117294ae8ba7SAneesh Kumar K.V h_cg, page); 1173a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 117494ae8ba7SAneesh Kumar K.V } else { 117594ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1176bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1177a1e78772SMel Gorman if (!page) { 11786d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_cgroup(idx, 11796d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), 11806d76dcf4SAneesh Kumar K.V h_cg); 118190481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 118276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 1183a1e78772SMel Gorman } 118479dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 118594ae8ba7SAneesh Kumar K.V hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), 118694ae8ba7SAneesh Kumar K.V h_cg, page); 118779dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 118879dbb236SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1189a1e78772SMel Gorman } 1190a1e78772SMel Gorman 119190481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1192a1e78772SMel Gorman 1193a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 11947893d1d5SAdam Litke return page; 1195b45b5bd6SDavid Gibson } 1196b45b5bd6SDavid Gibson 119791f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1198aa888a74SAndi Kleen { 1199aa888a74SAndi Kleen struct huge_bootmem_page *m; 12008cebfcd0SLai Jiangshan int nr_nodes = nodes_weight(node_states[N_MEMORY]); 1201aa888a74SAndi Kleen 1202aa888a74SAndi Kleen while (nr_nodes) { 1203aa888a74SAndi Kleen void *addr; 1204aa888a74SAndi Kleen 1205aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 12066ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 12078cebfcd0SLai Jiangshan &node_states[N_MEMORY])), 1208aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1209aa888a74SAndi Kleen 1210aa888a74SAndi Kleen if (addr) { 1211aa888a74SAndi Kleen /* 1212aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1213aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1214aa888a74SAndi Kleen * puts them into the mem_map). 1215aa888a74SAndi Kleen */ 1216aa888a74SAndi Kleen m = addr; 1217aa888a74SAndi Kleen goto found; 1218aa888a74SAndi Kleen } 1219aa888a74SAndi Kleen nr_nodes--; 1220aa888a74SAndi Kleen } 1221aa888a74SAndi Kleen return 0; 1222aa888a74SAndi Kleen 1223aa888a74SAndi Kleen found: 1224aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1225aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1226aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1227aa888a74SAndi Kleen m->hstate = h; 1228aa888a74SAndi Kleen return 1; 1229aa888a74SAndi Kleen } 1230aa888a74SAndi Kleen 123118229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 123218229df5SAndy Whitcroft { 123318229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 123418229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 123518229df5SAndy Whitcroft else 123618229df5SAndy Whitcroft prep_compound_page(page, order); 123718229df5SAndy Whitcroft } 123818229df5SAndy Whitcroft 1239aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1240aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1241aa888a74SAndi Kleen { 1242aa888a74SAndi Kleen struct huge_bootmem_page *m; 1243aa888a74SAndi Kleen 1244aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1245aa888a74SAndi Kleen struct hstate *h = m->hstate; 1246ee8f248dSBecky Bruce struct page *page; 1247ee8f248dSBecky Bruce 1248ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1249ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 1250ee8f248dSBecky Bruce free_bootmem_late((unsigned long)m, 1251ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1252ee8f248dSBecky Bruce #else 1253ee8f248dSBecky Bruce page = virt_to_page(m); 1254ee8f248dSBecky Bruce #endif 1255aa888a74SAndi Kleen __ClearPageReserved(page); 1256aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 125718229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1258aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1259b0320c7bSRafael Aquini /* 1260b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1261b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1262b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1263b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1264b0320c7bSRafael Aquini */ 1265b0320c7bSRafael Aquini if (h->order > (MAX_ORDER - 1)) 12663dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 1267aa888a74SAndi Kleen } 1268aa888a74SAndi Kleen } 1269aa888a74SAndi Kleen 12708faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 12711da177e4SLinus Torvalds { 12721da177e4SLinus Torvalds unsigned long i; 12731da177e4SLinus Torvalds 1274e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1275aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1276aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1277aa888a74SAndi Kleen break; 12789b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 12798cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 12801da177e4SLinus Torvalds break; 12811da177e4SLinus Torvalds } 12828faa8b07SAndi Kleen h->max_huge_pages = i; 1283e5ff2159SAndi Kleen } 1284e5ff2159SAndi Kleen 1285e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1286e5ff2159SAndi Kleen { 1287e5ff2159SAndi Kleen struct hstate *h; 1288e5ff2159SAndi Kleen 1289e5ff2159SAndi Kleen for_each_hstate(h) { 12908faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 12918faa8b07SAndi Kleen if (h->order < MAX_ORDER) 12928faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1293e5ff2159SAndi Kleen } 1294e5ff2159SAndi Kleen } 1295e5ff2159SAndi Kleen 12964abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 12974abd32dbSAndi Kleen { 12984abd32dbSAndi Kleen if (n >= (1UL << 30)) 12994abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 13004abd32dbSAndi Kleen else if (n >= (1UL << 20)) 13014abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 13024abd32dbSAndi Kleen else 13034abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 13044abd32dbSAndi Kleen return buf; 13054abd32dbSAndi Kleen } 13064abd32dbSAndi Kleen 1307e5ff2159SAndi Kleen static void __init report_hugepages(void) 1308e5ff2159SAndi Kleen { 1309e5ff2159SAndi Kleen struct hstate *h; 1310e5ff2159SAndi Kleen 1311e5ff2159SAndi Kleen for_each_hstate(h) { 13124abd32dbSAndi Kleen char buf[32]; 1313ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 13144abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 13154abd32dbSAndi Kleen h->free_huge_pages); 1316e5ff2159SAndi Kleen } 1317e5ff2159SAndi Kleen } 1318e5ff2159SAndi Kleen 13191da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 13206ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 13216ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13221da177e4SLinus Torvalds { 13234415cc8dSChristoph Lameter int i; 13244415cc8dSChristoph Lameter 1325aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1326aa888a74SAndi Kleen return; 1327aa888a74SAndi Kleen 13286ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 13291da177e4SLinus Torvalds struct page *page, *next; 1330a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1331a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1332a5516438SAndi Kleen if (count >= h->nr_huge_pages) 13336b0c880dSAdam Litke return; 13341da177e4SLinus Torvalds if (PageHighMem(page)) 13351da177e4SLinus Torvalds continue; 13361da177e4SLinus Torvalds list_del(&page->lru); 1337e5ff2159SAndi Kleen update_and_free_page(h, page); 1338a5516438SAndi Kleen h->free_huge_pages--; 1339a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 13401da177e4SLinus Torvalds } 13411da177e4SLinus Torvalds } 13421da177e4SLinus Torvalds } 13431da177e4SLinus Torvalds #else 13446ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 13456ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13461da177e4SLinus Torvalds { 13471da177e4SLinus Torvalds } 13481da177e4SLinus Torvalds #endif 13491da177e4SLinus Torvalds 135020a0307cSWu Fengguang /* 135120a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 135220a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 135320a0307cSWu Fengguang * Returns 1 if an adjustment was made. 135420a0307cSWu Fengguang */ 13556ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 13566ae11b27SLee Schermerhorn int delta) 135720a0307cSWu Fengguang { 1358e8c5c824SLee Schermerhorn int start_nid, next_nid; 135920a0307cSWu Fengguang int ret = 0; 136020a0307cSWu Fengguang 136120a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 136220a0307cSWu Fengguang 1363e8c5c824SLee Schermerhorn if (delta < 0) 13646ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1365e8c5c824SLee Schermerhorn else 13666ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1367e8c5c824SLee Schermerhorn next_nid = start_nid; 1368e8c5c824SLee Schermerhorn 1369e8c5c824SLee Schermerhorn do { 1370e8c5c824SLee Schermerhorn int nid = next_nid; 1371e8c5c824SLee Schermerhorn if (delta < 0) { 1372e8c5c824SLee Schermerhorn /* 1373e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1374e8c5c824SLee Schermerhorn */ 13759a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 13766ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 13776ae11b27SLee Schermerhorn nodes_allowed); 137820a0307cSWu Fengguang continue; 1379e8c5c824SLee Schermerhorn } 13809a76db09SLee Schermerhorn } 1381e8c5c824SLee Schermerhorn if (delta > 0) { 1382e8c5c824SLee Schermerhorn /* 1383e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1384e8c5c824SLee Schermerhorn */ 1385e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 13869a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 13876ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 13886ae11b27SLee Schermerhorn nodes_allowed); 138920a0307cSWu Fengguang continue; 1390e8c5c824SLee Schermerhorn } 13919a76db09SLee Schermerhorn } 139220a0307cSWu Fengguang 139320a0307cSWu Fengguang h->surplus_huge_pages += delta; 139420a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 139520a0307cSWu Fengguang ret = 1; 139620a0307cSWu Fengguang break; 1397e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 139820a0307cSWu Fengguang 139920a0307cSWu Fengguang return ret; 140020a0307cSWu Fengguang } 140120a0307cSWu Fengguang 1402a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 14036ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 14046ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 14051da177e4SLinus Torvalds { 14067893d1d5SAdam Litke unsigned long min_count, ret; 14071da177e4SLinus Torvalds 1408aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1409aa888a74SAndi Kleen return h->max_huge_pages; 1410aa888a74SAndi Kleen 14117893d1d5SAdam Litke /* 14127893d1d5SAdam Litke * Increase the pool size 14137893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 14147893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1415d1c3fb1fSNishanth Aravamudan * 1416d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1417d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1418d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1419d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1420d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 14217893d1d5SAdam Litke */ 14221da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1423a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 14246ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 14257893d1d5SAdam Litke break; 14267893d1d5SAdam Litke } 14277893d1d5SAdam Litke 1428a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 14297893d1d5SAdam Litke /* 14307893d1d5SAdam Litke * If this allocation races such that we no longer need the 14317893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 14327893d1d5SAdam Litke * and reducing the surplus. 14337893d1d5SAdam Litke */ 14347893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 14356ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 14367893d1d5SAdam Litke spin_lock(&hugetlb_lock); 14377893d1d5SAdam Litke if (!ret) 14387893d1d5SAdam Litke goto out; 14397893d1d5SAdam Litke 1440536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1441536240f2SMel Gorman if (signal_pending(current)) 1442536240f2SMel Gorman goto out; 14437893d1d5SAdam Litke } 14447893d1d5SAdam Litke 14457893d1d5SAdam Litke /* 14467893d1d5SAdam Litke * Decrease the pool size 14477893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 14487893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 14497893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 14507893d1d5SAdam Litke * to the desired size as pages become free. 1451d1c3fb1fSNishanth Aravamudan * 1452d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1453d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1454d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1455d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1456d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1457d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1458d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 14597893d1d5SAdam Litke */ 1460a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 14616b0c880dSAdam Litke min_count = max(count, min_count); 14626ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1463a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 14646ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 14651da177e4SLinus Torvalds break; 14661da177e4SLinus Torvalds } 1467a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 14686ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 14697893d1d5SAdam Litke break; 14707893d1d5SAdam Litke } 14717893d1d5SAdam Litke out: 1472a5516438SAndi Kleen ret = persistent_huge_pages(h); 14731da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 14747893d1d5SAdam Litke return ret; 14751da177e4SLinus Torvalds } 14761da177e4SLinus Torvalds 1477a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1478a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1479a3437870SNishanth Aravamudan 1480a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1481a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1482a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1483a3437870SNishanth Aravamudan 1484a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1485a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1486a3437870SNishanth Aravamudan 14879a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 14889a305230SLee Schermerhorn 14899a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1490a3437870SNishanth Aravamudan { 1491a3437870SNishanth Aravamudan int i; 14929a305230SLee Schermerhorn 1493a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 14949a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 14959a305230SLee Schermerhorn if (nidp) 14969a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1497a3437870SNishanth Aravamudan return &hstates[i]; 14989a305230SLee Schermerhorn } 14999a305230SLee Schermerhorn 15009a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1501a3437870SNishanth Aravamudan } 1502a3437870SNishanth Aravamudan 150306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1504a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1505a3437870SNishanth Aravamudan { 15069a305230SLee Schermerhorn struct hstate *h; 15079a305230SLee Schermerhorn unsigned long nr_huge_pages; 15089a305230SLee Schermerhorn int nid; 15099a305230SLee Schermerhorn 15109a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15119a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15129a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 15139a305230SLee Schermerhorn else 15149a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 15159a305230SLee Schermerhorn 15169a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1517a3437870SNishanth Aravamudan } 1518adbe8726SEric B Munson 151906808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 152006808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 152106808b08SLee Schermerhorn const char *buf, size_t len) 1522a3437870SNishanth Aravamudan { 1523a3437870SNishanth Aravamudan int err; 15249a305230SLee Schermerhorn int nid; 152506808b08SLee Schermerhorn unsigned long count; 15269a305230SLee Schermerhorn struct hstate *h; 1527bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1528a3437870SNishanth Aravamudan 15293dbb95f7SJingoo Han err = kstrtoul(buf, 10, &count); 153073ae31e5SEric B Munson if (err) 1531adbe8726SEric B Munson goto out; 1532a3437870SNishanth Aravamudan 15339a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1534adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1535adbe8726SEric B Munson err = -EINVAL; 1536adbe8726SEric B Munson goto out; 1537adbe8726SEric B Munson } 1538adbe8726SEric B Munson 15399a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 15409a305230SLee Schermerhorn /* 15419a305230SLee Schermerhorn * global hstate attribute 15429a305230SLee Schermerhorn */ 15439a305230SLee Schermerhorn if (!(obey_mempolicy && 15449a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 154506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 15468cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 154706808b08SLee Schermerhorn } 15489a305230SLee Schermerhorn } else if (nodes_allowed) { 15499a305230SLee Schermerhorn /* 15509a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 15519a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 15529a305230SLee Schermerhorn */ 15539a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 15549a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 15559a305230SLee Schermerhorn } else 15568cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 15579a305230SLee Schermerhorn 155806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1559a3437870SNishanth Aravamudan 15608cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 156106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 156206808b08SLee Schermerhorn 156306808b08SLee Schermerhorn return len; 1564adbe8726SEric B Munson out: 1565adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1566adbe8726SEric B Munson return err; 156706808b08SLee Schermerhorn } 156806808b08SLee Schermerhorn 156906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 157006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 157106808b08SLee Schermerhorn { 157206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 157306808b08SLee Schermerhorn } 157406808b08SLee Schermerhorn 157506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 157606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 157706808b08SLee Schermerhorn { 157806808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1579a3437870SNishanth Aravamudan } 1580a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1581a3437870SNishanth Aravamudan 158206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 158306808b08SLee Schermerhorn 158406808b08SLee Schermerhorn /* 158506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 158606808b08SLee Schermerhorn * huge page alloc/free. 158706808b08SLee Schermerhorn */ 158806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 158906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 159006808b08SLee Schermerhorn { 159106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 159206808b08SLee Schermerhorn } 159306808b08SLee Schermerhorn 159406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 159506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 159606808b08SLee Schermerhorn { 159706808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 159806808b08SLee Schermerhorn } 159906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 160006808b08SLee Schermerhorn #endif 160106808b08SLee Schermerhorn 160206808b08SLee Schermerhorn 1603a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1604a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1605a3437870SNishanth Aravamudan { 16069a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1607a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1608a3437870SNishanth Aravamudan } 1609adbe8726SEric B Munson 1610a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1611a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1612a3437870SNishanth Aravamudan { 1613a3437870SNishanth Aravamudan int err; 1614a3437870SNishanth Aravamudan unsigned long input; 16159a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1616a3437870SNishanth Aravamudan 1617adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1618adbe8726SEric B Munson return -EINVAL; 1619adbe8726SEric B Munson 16203dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 1621a3437870SNishanth Aravamudan if (err) 162273ae31e5SEric B Munson return err; 1623a3437870SNishanth Aravamudan 1624a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1625a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1626a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1627a3437870SNishanth Aravamudan 1628a3437870SNishanth Aravamudan return count; 1629a3437870SNishanth Aravamudan } 1630a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1631a3437870SNishanth Aravamudan 1632a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1633a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1634a3437870SNishanth Aravamudan { 16359a305230SLee Schermerhorn struct hstate *h; 16369a305230SLee Schermerhorn unsigned long free_huge_pages; 16379a305230SLee Schermerhorn int nid; 16389a305230SLee Schermerhorn 16399a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16409a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16419a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 16429a305230SLee Schermerhorn else 16439a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 16449a305230SLee Schermerhorn 16459a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1646a3437870SNishanth Aravamudan } 1647a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1648a3437870SNishanth Aravamudan 1649a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1650a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1651a3437870SNishanth Aravamudan { 16529a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1653a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1654a3437870SNishanth Aravamudan } 1655a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1656a3437870SNishanth Aravamudan 1657a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1658a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1659a3437870SNishanth Aravamudan { 16609a305230SLee Schermerhorn struct hstate *h; 16619a305230SLee Schermerhorn unsigned long surplus_huge_pages; 16629a305230SLee Schermerhorn int nid; 16639a305230SLee Schermerhorn 16649a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16659a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16669a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 16679a305230SLee Schermerhorn else 16689a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 16699a305230SLee Schermerhorn 16709a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1671a3437870SNishanth Aravamudan } 1672a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1673a3437870SNishanth Aravamudan 1674a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1675a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1676a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1677a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1678a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1679a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 168006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 168106808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 168206808b08SLee Schermerhorn #endif 1683a3437870SNishanth Aravamudan NULL, 1684a3437870SNishanth Aravamudan }; 1685a3437870SNishanth Aravamudan 1686a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1687a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1688a3437870SNishanth Aravamudan }; 1689a3437870SNishanth Aravamudan 1690094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 16919a305230SLee Schermerhorn struct kobject **hstate_kobjs, 16929a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1693a3437870SNishanth Aravamudan { 1694a3437870SNishanth Aravamudan int retval; 1695972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1696a3437870SNishanth Aravamudan 16979a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 16989a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1699a3437870SNishanth Aravamudan return -ENOMEM; 1700a3437870SNishanth Aravamudan 17019a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1702a3437870SNishanth Aravamudan if (retval) 17039a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1704a3437870SNishanth Aravamudan 1705a3437870SNishanth Aravamudan return retval; 1706a3437870SNishanth Aravamudan } 1707a3437870SNishanth Aravamudan 1708a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1709a3437870SNishanth Aravamudan { 1710a3437870SNishanth Aravamudan struct hstate *h; 1711a3437870SNishanth Aravamudan int err; 1712a3437870SNishanth Aravamudan 1713a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1714a3437870SNishanth Aravamudan if (!hugepages_kobj) 1715a3437870SNishanth Aravamudan return; 1716a3437870SNishanth Aravamudan 1717a3437870SNishanth Aravamudan for_each_hstate(h) { 17189a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 17199a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1720a3437870SNishanth Aravamudan if (err) 1721ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 1722a3437870SNishanth Aravamudan } 1723a3437870SNishanth Aravamudan } 1724a3437870SNishanth Aravamudan 17259a305230SLee Schermerhorn #ifdef CONFIG_NUMA 17269a305230SLee Schermerhorn 17279a305230SLee Schermerhorn /* 17289a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 172910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 173010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 173110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 17329a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 17339a305230SLee Schermerhorn */ 17349a305230SLee Schermerhorn struct node_hstate { 17359a305230SLee Schermerhorn struct kobject *hugepages_kobj; 17369a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 17379a305230SLee Schermerhorn }; 17389a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 17399a305230SLee Schermerhorn 17409a305230SLee Schermerhorn /* 174110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 17429a305230SLee Schermerhorn */ 17439a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 17449a305230SLee Schermerhorn &nr_hugepages_attr.attr, 17459a305230SLee Schermerhorn &free_hugepages_attr.attr, 17469a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 17479a305230SLee Schermerhorn NULL, 17489a305230SLee Schermerhorn }; 17499a305230SLee Schermerhorn 17509a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 17519a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 17529a305230SLee Schermerhorn }; 17539a305230SLee Schermerhorn 17549a305230SLee Schermerhorn /* 175510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 17569a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 17579a305230SLee Schermerhorn */ 17589a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17599a305230SLee Schermerhorn { 17609a305230SLee Schermerhorn int nid; 17619a305230SLee Schermerhorn 17629a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 17639a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 17649a305230SLee Schermerhorn int i; 17659a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 17669a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 17679a305230SLee Schermerhorn if (nidp) 17689a305230SLee Schermerhorn *nidp = nid; 17699a305230SLee Schermerhorn return &hstates[i]; 17709a305230SLee Schermerhorn } 17719a305230SLee Schermerhorn } 17729a305230SLee Schermerhorn 17739a305230SLee Schermerhorn BUG(); 17749a305230SLee Schermerhorn return NULL; 17759a305230SLee Schermerhorn } 17769a305230SLee Schermerhorn 17779a305230SLee Schermerhorn /* 177810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 17799a305230SLee Schermerhorn * No-op if no hstate attributes attached. 17809a305230SLee Schermerhorn */ 17813cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 17829a305230SLee Schermerhorn { 17839a305230SLee Schermerhorn struct hstate *h; 178410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 17859a305230SLee Schermerhorn 17869a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 17879b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 17889a305230SLee Schermerhorn 1789972dc4deSAneesh Kumar K.V for_each_hstate(h) { 1790972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 1791972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 1792972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 1793972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 1794972dc4deSAneesh Kumar K.V } 17959a305230SLee Schermerhorn } 17969a305230SLee Schermerhorn 17979a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 17989a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 17999a305230SLee Schermerhorn } 18009a305230SLee Schermerhorn 18019a305230SLee Schermerhorn /* 180210fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 18039a305230SLee Schermerhorn * that have them. 18049a305230SLee Schermerhorn */ 18059a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 18069a305230SLee Schermerhorn { 18079a305230SLee Schermerhorn int nid; 18089a305230SLee Schermerhorn 18099a305230SLee Schermerhorn /* 181010fbcf4cSKay Sievers * disable node device registrations. 18119a305230SLee Schermerhorn */ 18129a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 18139a305230SLee Schermerhorn 18149a305230SLee Schermerhorn /* 18159a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 18169a305230SLee Schermerhorn */ 18179a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 18188732794bSWen Congyang hugetlb_unregister_node(node_devices[nid]); 18199a305230SLee Schermerhorn } 18209a305230SLee Schermerhorn 18219a305230SLee Schermerhorn /* 182210fbcf4cSKay Sievers * Register hstate attributes for a single node device. 18239a305230SLee Schermerhorn * No-op if attributes already registered. 18249a305230SLee Schermerhorn */ 18253cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 18269a305230SLee Schermerhorn { 18279a305230SLee Schermerhorn struct hstate *h; 182810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 18299a305230SLee Schermerhorn int err; 18309a305230SLee Schermerhorn 18319a305230SLee Schermerhorn if (nhs->hugepages_kobj) 18329a305230SLee Schermerhorn return; /* already allocated */ 18339a305230SLee Schermerhorn 18349a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 183510fbcf4cSKay Sievers &node->dev.kobj); 18369a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 18379a305230SLee Schermerhorn return; 18389a305230SLee Schermerhorn 18399a305230SLee Schermerhorn for_each_hstate(h) { 18409a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 18419a305230SLee Schermerhorn nhs->hstate_kobjs, 18429a305230SLee Schermerhorn &per_node_hstate_attr_group); 18439a305230SLee Schermerhorn if (err) { 1844ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 184510fbcf4cSKay Sievers h->name, node->dev.id); 18469a305230SLee Schermerhorn hugetlb_unregister_node(node); 18479a305230SLee Schermerhorn break; 18489a305230SLee Schermerhorn } 18499a305230SLee Schermerhorn } 18509a305230SLee Schermerhorn } 18519a305230SLee Schermerhorn 18529a305230SLee Schermerhorn /* 18539b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 185410fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 185510fbcf4cSKay Sievers * registered their associated device by this time. 18569a305230SLee Schermerhorn */ 18579a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 18589a305230SLee Schermerhorn { 18599a305230SLee Schermerhorn int nid; 18609a305230SLee Schermerhorn 18618cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 18628732794bSWen Congyang struct node *node = node_devices[nid]; 186310fbcf4cSKay Sievers if (node->dev.id == nid) 18649a305230SLee Schermerhorn hugetlb_register_node(node); 18659a305230SLee Schermerhorn } 18669a305230SLee Schermerhorn 18679a305230SLee Schermerhorn /* 186810fbcf4cSKay Sievers * Let the node device driver know we're here so it can 18699a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 18709a305230SLee Schermerhorn */ 18719a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 18729a305230SLee Schermerhorn hugetlb_unregister_node); 18739a305230SLee Schermerhorn } 18749a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 18759a305230SLee Schermerhorn 18769a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 18779a305230SLee Schermerhorn { 18789a305230SLee Schermerhorn BUG(); 18799a305230SLee Schermerhorn if (nidp) 18809a305230SLee Schermerhorn *nidp = -1; 18819a305230SLee Schermerhorn return NULL; 18829a305230SLee Schermerhorn } 18839a305230SLee Schermerhorn 18849a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 18859a305230SLee Schermerhorn 18869a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 18879a305230SLee Schermerhorn 18889a305230SLee Schermerhorn #endif 18899a305230SLee Schermerhorn 1890a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1891a3437870SNishanth Aravamudan { 1892a3437870SNishanth Aravamudan struct hstate *h; 1893a3437870SNishanth Aravamudan 18949a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 18959a305230SLee Schermerhorn 1896a3437870SNishanth Aravamudan for_each_hstate(h) { 1897972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 1898a3437870SNishanth Aravamudan } 1899a3437870SNishanth Aravamudan 1900a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1901a3437870SNishanth Aravamudan } 1902a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1903a3437870SNishanth Aravamudan 1904a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1905a3437870SNishanth Aravamudan { 19060ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 19070ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 19080ef89d25SBenjamin Herrenschmidt * there is no such support 19090ef89d25SBenjamin Herrenschmidt */ 19100ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 19110ef89d25SBenjamin Herrenschmidt return 0; 1912a3437870SNishanth Aravamudan 1913e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1914e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1915e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1916a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1917a3437870SNishanth Aravamudan } 1918972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1919e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1920e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1921a3437870SNishanth Aravamudan 1922a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1923aa888a74SAndi Kleen gather_bootmem_prealloc(); 1924a3437870SNishanth Aravamudan report_hugepages(); 1925a3437870SNishanth Aravamudan 1926a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 19279a305230SLee Schermerhorn hugetlb_register_all_nodes(); 19287179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 19299a305230SLee Schermerhorn 1930a3437870SNishanth Aravamudan return 0; 1931a3437870SNishanth Aravamudan } 1932a3437870SNishanth Aravamudan module_init(hugetlb_init); 1933a3437870SNishanth Aravamudan 1934a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1935a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1936a3437870SNishanth Aravamudan { 1937a3437870SNishanth Aravamudan struct hstate *h; 19388faa8b07SAndi Kleen unsigned long i; 19398faa8b07SAndi Kleen 1940a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1941ffb22af5SAndrew Morton pr_warning("hugepagesz= specified twice, ignoring\n"); 1942a3437870SNishanth Aravamudan return; 1943a3437870SNishanth Aravamudan } 194447d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 1945a3437870SNishanth Aravamudan BUG_ON(order == 0); 194647d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 1947a3437870SNishanth Aravamudan h->order = order; 1948a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 19498faa8b07SAndi Kleen h->nr_huge_pages = 0; 19508faa8b07SAndi Kleen h->free_huge_pages = 0; 19518faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 19528faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 19530edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 19548cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 19558cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 1956a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1957a3437870SNishanth Aravamudan huge_page_size(h)/1024); 19588faa8b07SAndi Kleen 1959a3437870SNishanth Aravamudan parsed_hstate = h; 1960a3437870SNishanth Aravamudan } 1961a3437870SNishanth Aravamudan 1962e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1963a3437870SNishanth Aravamudan { 1964a3437870SNishanth Aravamudan unsigned long *mhp; 19658faa8b07SAndi Kleen static unsigned long *last_mhp; 1966a3437870SNishanth Aravamudan 1967a3437870SNishanth Aravamudan /* 196847d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 1969a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1970a3437870SNishanth Aravamudan */ 197147d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 1972a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1973a3437870SNishanth Aravamudan else 1974a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1975a3437870SNishanth Aravamudan 19768faa8b07SAndi Kleen if (mhp == last_mhp) { 1977ffb22af5SAndrew Morton pr_warning("hugepages= specified twice without " 19788faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 19798faa8b07SAndi Kleen return 1; 19808faa8b07SAndi Kleen } 19818faa8b07SAndi Kleen 1982a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1983a3437870SNishanth Aravamudan *mhp = 0; 1984a3437870SNishanth Aravamudan 19858faa8b07SAndi Kleen /* 19868faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 19878faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 19888faa8b07SAndi Kleen * use the bootmem allocator. 19898faa8b07SAndi Kleen */ 199047d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 19918faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 19928faa8b07SAndi Kleen 19938faa8b07SAndi Kleen last_mhp = mhp; 19948faa8b07SAndi Kleen 1995a3437870SNishanth Aravamudan return 1; 1996a3437870SNishanth Aravamudan } 1997e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1998e11bfbfcSNick Piggin 1999e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2000e11bfbfcSNick Piggin { 2001e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2002e11bfbfcSNick Piggin return 1; 2003e11bfbfcSNick Piggin } 2004e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2005a3437870SNishanth Aravamudan 20068a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 20078a213460SNishanth Aravamudan { 20088a213460SNishanth Aravamudan int node; 20098a213460SNishanth Aravamudan unsigned int nr = 0; 20108a213460SNishanth Aravamudan 20118a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 20128a213460SNishanth Aravamudan nr += array[node]; 20138a213460SNishanth Aravamudan 20148a213460SNishanth Aravamudan return nr; 20158a213460SNishanth Aravamudan } 20168a213460SNishanth Aravamudan 20178a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 201806808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 201906808b08SLee Schermerhorn struct ctl_table *table, int write, 202006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 20211da177e4SLinus Torvalds { 2022e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2023e5ff2159SAndi Kleen unsigned long tmp; 202408d4a246SMichal Hocko int ret; 2025e5ff2159SAndi Kleen 2026e5ff2159SAndi Kleen tmp = h->max_huge_pages; 2027e5ff2159SAndi Kleen 2028adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2029adbe8726SEric B Munson return -EINVAL; 2030adbe8726SEric B Munson 2031e5ff2159SAndi Kleen table->data = &tmp; 2032e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 203308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 203408d4a246SMichal Hocko if (ret) 203508d4a246SMichal Hocko goto out; 2036e5ff2159SAndi Kleen 203706808b08SLee Schermerhorn if (write) { 2038bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2039bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 204006808b08SLee Schermerhorn if (!(obey_mempolicy && 204106808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 204206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 20438cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 204406808b08SLee Schermerhorn } 204506808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 204606808b08SLee Schermerhorn 20478cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 204806808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 204906808b08SLee Schermerhorn } 205008d4a246SMichal Hocko out: 205108d4a246SMichal Hocko return ret; 20521da177e4SLinus Torvalds } 2053396faf03SMel Gorman 205406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 205506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 205606808b08SLee Schermerhorn { 205706808b08SLee Schermerhorn 205806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 205906808b08SLee Schermerhorn buffer, length, ppos); 206006808b08SLee Schermerhorn } 206106808b08SLee Schermerhorn 206206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 206306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 206406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 206506808b08SLee Schermerhorn { 206606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 206706808b08SLee Schermerhorn buffer, length, ppos); 206806808b08SLee Schermerhorn } 206906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 207006808b08SLee Schermerhorn 2071396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 20728d65af78SAlexey Dobriyan void __user *buffer, 2073396faf03SMel Gorman size_t *length, loff_t *ppos) 2074396faf03SMel Gorman { 20758d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 2076396faf03SMel Gorman if (hugepages_treat_as_movable) 2077396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 2078396faf03SMel Gorman else 2079396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 2080396faf03SMel Gorman return 0; 2081396faf03SMel Gorman } 2082396faf03SMel Gorman 2083a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 20848d65af78SAlexey Dobriyan void __user *buffer, 2085a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2086a3d0c6aaSNishanth Aravamudan { 2087a5516438SAndi Kleen struct hstate *h = &default_hstate; 2088e5ff2159SAndi Kleen unsigned long tmp; 208908d4a246SMichal Hocko int ret; 2090e5ff2159SAndi Kleen 2091e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2092e5ff2159SAndi Kleen 2093adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2094adbe8726SEric B Munson return -EINVAL; 2095adbe8726SEric B Munson 2096e5ff2159SAndi Kleen table->data = &tmp; 2097e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 209808d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 209908d4a246SMichal Hocko if (ret) 210008d4a246SMichal Hocko goto out; 2101e5ff2159SAndi Kleen 2102e5ff2159SAndi Kleen if (write) { 2103064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2104e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2105a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2106e5ff2159SAndi Kleen } 210708d4a246SMichal Hocko out: 210808d4a246SMichal Hocko return ret; 2109a3d0c6aaSNishanth Aravamudan } 2110a3d0c6aaSNishanth Aravamudan 21111da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 21121da177e4SLinus Torvalds 2113e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 21141da177e4SLinus Torvalds { 2115a5516438SAndi Kleen struct hstate *h = &default_hstate; 2116e1759c21SAlexey Dobriyan seq_printf(m, 21171da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 21181da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2119b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 21207893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 21214f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2122a5516438SAndi Kleen h->nr_huge_pages, 2123a5516438SAndi Kleen h->free_huge_pages, 2124a5516438SAndi Kleen h->resv_huge_pages, 2125a5516438SAndi Kleen h->surplus_huge_pages, 2126a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 21271da177e4SLinus Torvalds } 21281da177e4SLinus Torvalds 21291da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 21301da177e4SLinus Torvalds { 2131a5516438SAndi Kleen struct hstate *h = &default_hstate; 21321da177e4SLinus Torvalds return sprintf(buf, 21331da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2134a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2135a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2136a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2137a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2138a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 21391da177e4SLinus Torvalds } 21401da177e4SLinus Torvalds 2141949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2142949f7ec5SDavid Rientjes { 2143949f7ec5SDavid Rientjes struct hstate *h; 2144949f7ec5SDavid Rientjes int nid; 2145949f7ec5SDavid Rientjes 2146949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2147949f7ec5SDavid Rientjes for_each_hstate(h) 2148949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2149949f7ec5SDavid Rientjes nid, 2150949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2151949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2152949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2153949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2154949f7ec5SDavid Rientjes } 2155949f7ec5SDavid Rientjes 21561da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 21571da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 21581da177e4SLinus Torvalds { 2159d0028588SWanpeng Li struct hstate *h; 2160d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2161d0028588SWanpeng Li 2162d0028588SWanpeng Li for_each_hstate(h) 2163d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2164d0028588SWanpeng Li return nr_total_pages; 21651da177e4SLinus Torvalds } 21661da177e4SLinus Torvalds 2167a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2168fc1b8a73SMel Gorman { 2169fc1b8a73SMel Gorman int ret = -ENOMEM; 2170fc1b8a73SMel Gorman 2171fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2172fc1b8a73SMel Gorman /* 2173fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2174fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2175fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2176fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2177fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2178fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2179fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2180fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2181fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2182fc1b8a73SMel Gorman * 2183fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2184fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2185fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2186fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2187fc1b8a73SMel Gorman * semantics that cpuset has. 2188fc1b8a73SMel Gorman */ 2189fc1b8a73SMel Gorman if (delta > 0) { 2190a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2191fc1b8a73SMel Gorman goto out; 2192fc1b8a73SMel Gorman 2193a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2194a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2195fc1b8a73SMel Gorman goto out; 2196fc1b8a73SMel Gorman } 2197fc1b8a73SMel Gorman } 2198fc1b8a73SMel Gorman 2199fc1b8a73SMel Gorman ret = 0; 2200fc1b8a73SMel Gorman if (delta < 0) 2201a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2202fc1b8a73SMel Gorman 2203fc1b8a73SMel Gorman out: 2204fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2205fc1b8a73SMel Gorman return ret; 2206fc1b8a73SMel Gorman } 2207fc1b8a73SMel Gorman 220884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 220984afd99bSAndy Whitcroft { 221084afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 221184afd99bSAndy Whitcroft 221284afd99bSAndy Whitcroft /* 221384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 221484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 221584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 221625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 221784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 221884afd99bSAndy Whitcroft * new reference here without additional locking. 221984afd99bSAndy Whitcroft */ 222084afd99bSAndy Whitcroft if (reservations) 222184afd99bSAndy Whitcroft kref_get(&reservations->refs); 222284afd99bSAndy Whitcroft } 222384afd99bSAndy Whitcroft 2224c50ac050SDave Hansen static void resv_map_put(struct vm_area_struct *vma) 2225c50ac050SDave Hansen { 2226c50ac050SDave Hansen struct resv_map *reservations = vma_resv_map(vma); 2227c50ac050SDave Hansen 2228c50ac050SDave Hansen if (!reservations) 2229c50ac050SDave Hansen return; 2230c50ac050SDave Hansen kref_put(&reservations->refs, resv_map_release); 2231c50ac050SDave Hansen } 2232c50ac050SDave Hansen 2233a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2234a1e78772SMel Gorman { 2235a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 223684afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 223790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 223884afd99bSAndy Whitcroft unsigned long reserve; 223984afd99bSAndy Whitcroft unsigned long start; 224084afd99bSAndy Whitcroft unsigned long end; 224184afd99bSAndy Whitcroft 224284afd99bSAndy Whitcroft if (reservations) { 2243a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2244a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 224584afd99bSAndy Whitcroft 224684afd99bSAndy Whitcroft reserve = (end - start) - 224784afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 224884afd99bSAndy Whitcroft 2249c50ac050SDave Hansen resv_map_put(vma); 225084afd99bSAndy Whitcroft 22517251ff78SAdam Litke if (reserve) { 2252a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 225390481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 22547251ff78SAdam Litke } 2255a1e78772SMel Gorman } 225684afd99bSAndy Whitcroft } 2257a1e78772SMel Gorman 22581da177e4SLinus Torvalds /* 22591da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 22601da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 22611da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 22621da177e4SLinus Torvalds * this far. 22631da177e4SLinus Torvalds */ 2264d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 22651da177e4SLinus Torvalds { 22661da177e4SLinus Torvalds BUG(); 2267d0217ac0SNick Piggin return 0; 22681da177e4SLinus Torvalds } 22691da177e4SLinus Torvalds 2270f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2271d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 227284afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2273a1e78772SMel Gorman .close = hugetlb_vm_op_close, 22741da177e4SLinus Torvalds }; 22751da177e4SLinus Torvalds 22761e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 22771e8f889bSDavid Gibson int writable) 227863551ae0SDavid Gibson { 227963551ae0SDavid Gibson pte_t entry; 228063551ae0SDavid Gibson 22811e8f889bSDavid Gibson if (writable) { 2282106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2283106c992aSGerald Schaefer vma->vm_page_prot))); 228463551ae0SDavid Gibson } else { 2285106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 2286106c992aSGerald Schaefer vma->vm_page_prot)); 228763551ae0SDavid Gibson } 228863551ae0SDavid Gibson entry = pte_mkyoung(entry); 228963551ae0SDavid Gibson entry = pte_mkhuge(entry); 2290d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 229163551ae0SDavid Gibson 229263551ae0SDavid Gibson return entry; 229363551ae0SDavid Gibson } 229463551ae0SDavid Gibson 22951e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 22961e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 22971e8f889bSDavid Gibson { 22981e8f889bSDavid Gibson pte_t entry; 22991e8f889bSDavid Gibson 2300106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 230132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 23024b3073e1SRussell King update_mmu_cache(vma, address, ptep); 23031e8f889bSDavid Gibson } 23041e8f889bSDavid Gibson 23051e8f889bSDavid Gibson 230663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 230763551ae0SDavid Gibson struct vm_area_struct *vma) 230863551ae0SDavid Gibson { 230963551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 231063551ae0SDavid Gibson struct page *ptepage; 23111c59827dSHugh Dickins unsigned long addr; 23121e8f889bSDavid Gibson int cow; 2313a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2314a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 23151e8f889bSDavid Gibson 23161e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 231763551ae0SDavid Gibson 2318a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2319c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2320c74df32cSHugh Dickins if (!src_pte) 2321c74df32cSHugh Dickins continue; 2322a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 232363551ae0SDavid Gibson if (!dst_pte) 232463551ae0SDavid Gibson goto nomem; 2325c5c99429SLarry Woodman 2326c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2327c5c99429SLarry Woodman if (dst_pte == src_pte) 2328c5c99429SLarry Woodman continue; 2329c5c99429SLarry Woodman 2330c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 233146478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 23327f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 23331e8f889bSDavid Gibson if (cow) 23347f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 23357f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 233663551ae0SDavid Gibson ptepage = pte_page(entry); 233763551ae0SDavid Gibson get_page(ptepage); 23380fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 233963551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 23401c59827dSHugh Dickins } 23411c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2342c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 234363551ae0SDavid Gibson } 234463551ae0SDavid Gibson return 0; 234563551ae0SDavid Gibson 234663551ae0SDavid Gibson nomem: 234763551ae0SDavid Gibson return -ENOMEM; 234863551ae0SDavid Gibson } 234963551ae0SDavid Gibson 2350290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2351290408d4SNaoya Horiguchi { 2352290408d4SNaoya Horiguchi swp_entry_t swp; 2353290408d4SNaoya Horiguchi 2354290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2355290408d4SNaoya Horiguchi return 0; 2356290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 235732f84528SChris Forbes if (non_swap_entry(swp) && is_migration_entry(swp)) 2358290408d4SNaoya Horiguchi return 1; 235932f84528SChris Forbes else 2360290408d4SNaoya Horiguchi return 0; 2361290408d4SNaoya Horiguchi } 2362290408d4SNaoya Horiguchi 2363fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2364fd6a03edSNaoya Horiguchi { 2365fd6a03edSNaoya Horiguchi swp_entry_t swp; 2366fd6a03edSNaoya Horiguchi 2367fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2368fd6a03edSNaoya Horiguchi return 0; 2369fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 237032f84528SChris Forbes if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2371fd6a03edSNaoya Horiguchi return 1; 237232f84528SChris Forbes else 2373fd6a03edSNaoya Horiguchi return 0; 2374fd6a03edSNaoya Horiguchi } 2375fd6a03edSNaoya Horiguchi 237624669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 237724669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 237824669e58SAneesh Kumar K.V struct page *ref_page) 237963551ae0SDavid Gibson { 238024669e58SAneesh Kumar K.V int force_flush = 0; 238163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 238263551ae0SDavid Gibson unsigned long address; 2383c7546f8fSDavid Gibson pte_t *ptep; 238463551ae0SDavid Gibson pte_t pte; 238563551ae0SDavid Gibson struct page *page; 2386a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2387a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 23882ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 23892ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 2390a5516438SAndi Kleen 239163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2392a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2393a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 239463551ae0SDavid Gibson 239524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 23962ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 239724669e58SAneesh Kumar K.V again: 2398508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2399a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2400c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2401c7546f8fSDavid Gibson if (!ptep) 2402c7546f8fSDavid Gibson continue; 2403c7546f8fSDavid Gibson 240439dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 240539dde65cSChen, Kenneth W continue; 240639dde65cSChen, Kenneth W 24076629326bSHillf Danton pte = huge_ptep_get(ptep); 24086629326bSHillf Danton if (huge_pte_none(pte)) 24096629326bSHillf Danton continue; 24106629326bSHillf Danton 24116629326bSHillf Danton /* 24126629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 24136629326bSHillf Danton */ 24148c4894c6SNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 2415106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 24166629326bSHillf Danton continue; 24178c4894c6SNaoya Horiguchi } 24186629326bSHillf Danton 24196629326bSHillf Danton page = pte_page(pte); 242004f2cbe3SMel Gorman /* 242104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 242204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 242304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 242404f2cbe3SMel Gorman */ 242504f2cbe3SMel Gorman if (ref_page) { 242604f2cbe3SMel Gorman if (page != ref_page) 242704f2cbe3SMel Gorman continue; 242804f2cbe3SMel Gorman 242904f2cbe3SMel Gorman /* 243004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 243104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 243204f2cbe3SMel Gorman * looking like data was lost 243304f2cbe3SMel Gorman */ 243404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 243504f2cbe3SMel Gorman } 243604f2cbe3SMel Gorman 2437c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 243824669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 2439106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 24406649a386SKen Chen set_page_dirty(page); 24419e81130bSHillf Danton 244224669e58SAneesh Kumar K.V page_remove_rmap(page); 244324669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 244424669e58SAneesh Kumar K.V if (force_flush) 244524669e58SAneesh Kumar K.V break; 24469e81130bSHillf Danton /* Bail out after unmapping reference page if supplied */ 24479e81130bSHillf Danton if (ref_page) 24489e81130bSHillf Danton break; 244963551ae0SDavid Gibson } 2450cd2934a3SAl Viro spin_unlock(&mm->page_table_lock); 245124669e58SAneesh Kumar K.V /* 245224669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 245324669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 245424669e58SAneesh Kumar K.V * and page-free while holding it. 245524669e58SAneesh Kumar K.V */ 245624669e58SAneesh Kumar K.V if (force_flush) { 245724669e58SAneesh Kumar K.V force_flush = 0; 245824669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 245924669e58SAneesh Kumar K.V if (address < end && !ref_page) 246024669e58SAneesh Kumar K.V goto again; 2461fe1668aeSChen, Kenneth W } 24622ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 246324669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 24641da177e4SLinus Torvalds } 246563551ae0SDavid Gibson 2466d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2467d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2468d833352aSMel Gorman unsigned long end, struct page *ref_page) 2469d833352aSMel Gorman { 2470d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2471d833352aSMel Gorman 2472d833352aSMel Gorman /* 2473d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2474d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2475d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2476d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2477d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2478d833352aSMel Gorman * is to clear it before releasing the i_mmap_mutex. This works 2479d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2480d833352aSMel Gorman * destroyed and the i_mmap_mutex is held. 2481d833352aSMel Gorman */ 2482d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2483d833352aSMel Gorman } 2484d833352aSMel Gorman 2485502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 248604f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2487502717f4SChen, Kenneth W { 248824669e58SAneesh Kumar K.V struct mm_struct *mm; 248924669e58SAneesh Kumar K.V struct mmu_gather tlb; 249024669e58SAneesh Kumar K.V 249124669e58SAneesh Kumar K.V mm = vma->vm_mm; 249224669e58SAneesh Kumar K.V 24932b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 249424669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 249524669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2496502717f4SChen, Kenneth W } 2497502717f4SChen, Kenneth W 249804f2cbe3SMel Gorman /* 249904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 250004f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 250104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 250204f2cbe3SMel Gorman * same region. 250304f2cbe3SMel Gorman */ 25042a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 25052a4b3dedSHarvey Harrison struct page *page, unsigned long address) 250604f2cbe3SMel Gorman { 25077526674dSAdam Litke struct hstate *h = hstate_vma(vma); 250804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 250904f2cbe3SMel Gorman struct address_space *mapping; 251004f2cbe3SMel Gorman pgoff_t pgoff; 251104f2cbe3SMel Gorman 251204f2cbe3SMel Gorman /* 251304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 251404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 251504f2cbe3SMel Gorman */ 25167526674dSAdam Litke address = address & huge_page_mask(h); 251736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 251836e4f20aSMichal Hocko vma->vm_pgoff; 2519496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 252004f2cbe3SMel Gorman 25214eb2b1dcSMel Gorman /* 25224eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 25234eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 25244eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 25254eb2b1dcSMel Gorman */ 25263d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 25276b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 252804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 252904f2cbe3SMel Gorman if (iter_vma == vma) 253004f2cbe3SMel Gorman continue; 253104f2cbe3SMel Gorman 253204f2cbe3SMel Gorman /* 253304f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 253404f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 253504f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 253604f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 253704f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 253804f2cbe3SMel Gorman */ 253904f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 254024669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 254124669e58SAneesh Kumar K.V address + huge_page_size(h), page); 254204f2cbe3SMel Gorman } 25433d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 254404f2cbe3SMel Gorman 254504f2cbe3SMel Gorman return 1; 254604f2cbe3SMel Gorman } 254704f2cbe3SMel Gorman 25480fe6e20bSNaoya Horiguchi /* 25490fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2550ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2551ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2552ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 25530fe6e20bSNaoya Horiguchi */ 25541e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 255504f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 255604f2cbe3SMel Gorman struct page *pagecache_page) 25571e8f889bSDavid Gibson { 2558a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 25591e8f889bSDavid Gibson struct page *old_page, *new_page; 256079ac6ba4SDavid Gibson int avoidcopy; 256104f2cbe3SMel Gorman int outside_reserve = 0; 25622ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 25632ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 25641e8f889bSDavid Gibson 25651e8f889bSDavid Gibson old_page = pte_page(pte); 25661e8f889bSDavid Gibson 256704f2cbe3SMel Gorman retry_avoidcopy: 25681e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 25691e8f889bSDavid Gibson * and just make the page writable */ 25700fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 25711e8f889bSDavid Gibson if (avoidcopy) { 25720fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 25730fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 25741e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 257583c54070SNick Piggin return 0; 25761e8f889bSDavid Gibson } 25771e8f889bSDavid Gibson 257804f2cbe3SMel Gorman /* 257904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 258004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 258104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 258204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 258304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 258404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 258504f2cbe3SMel Gorman * of the full address range. 258604f2cbe3SMel Gorman */ 2587f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 258804f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 258904f2cbe3SMel Gorman old_page != pagecache_page) 259004f2cbe3SMel Gorman outside_reserve = 1; 259104f2cbe3SMel Gorman 25921e8f889bSDavid Gibson page_cache_get(old_page); 2593b76c8cfbSLarry Woodman 2594b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2595b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 259604f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 25971e8f889bSDavid Gibson 25982fc39cecSAdam Litke if (IS_ERR(new_page)) { 259976dcee75SAneesh Kumar K.V long err = PTR_ERR(new_page); 26001e8f889bSDavid Gibson page_cache_release(old_page); 260104f2cbe3SMel Gorman 260204f2cbe3SMel Gorman /* 260304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 260404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 260504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 260604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 260704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 260804f2cbe3SMel Gorman */ 260904f2cbe3SMel Gorman if (outside_reserve) { 261004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 261104f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 261204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2613b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2614a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2615a734bcc8SHillf Danton if (likely(pte_same(huge_ptep_get(ptep), pte))) 261604f2cbe3SMel Gorman goto retry_avoidcopy; 2617a734bcc8SHillf Danton /* 2618a734bcc8SHillf Danton * race occurs while re-acquiring page_table_lock, and 2619a734bcc8SHillf Danton * our job is done. 2620a734bcc8SHillf Danton */ 2621a734bcc8SHillf Danton return 0; 262204f2cbe3SMel Gorman } 262304f2cbe3SMel Gorman WARN_ON_ONCE(1); 262404f2cbe3SMel Gorman } 262504f2cbe3SMel Gorman 2626b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2627b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 262876dcee75SAneesh Kumar K.V if (err == -ENOMEM) 262976dcee75SAneesh Kumar K.V return VM_FAULT_OOM; 263076dcee75SAneesh Kumar K.V else 263176dcee75SAneesh Kumar K.V return VM_FAULT_SIGBUS; 26321e8f889bSDavid Gibson } 26331e8f889bSDavid Gibson 26340fe6e20bSNaoya Horiguchi /* 26350fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 26360fe6e20bSNaoya Horiguchi * anon_vma prepared. 26370fe6e20bSNaoya Horiguchi */ 263844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2639ea4039a3SHillf Danton page_cache_release(new_page); 2640ea4039a3SHillf Danton page_cache_release(old_page); 264144e2aa93SDean Nelson /* Caller expects lock to be held */ 264244e2aa93SDean Nelson spin_lock(&mm->page_table_lock); 26430fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 264444e2aa93SDean Nelson } 26450fe6e20bSNaoya Horiguchi 264647ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 264747ad8475SAndrea Arcangeli pages_per_huge_page(h)); 26480ed361deSNick Piggin __SetPageUptodate(new_page); 26491e8f889bSDavid Gibson 26502ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 26512ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 26522ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2653b76c8cfbSLarry Woodman /* 2654b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2655b76c8cfbSLarry Woodman * before the page tables are altered 2656b76c8cfbSLarry Woodman */ 2657b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2658a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 26597f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 26601e8f889bSDavid Gibson /* Break COW */ 26618fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 26621e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 26631e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 26640fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2665cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 26661e8f889bSDavid Gibson /* Make the old page be freed below */ 26671e8f889bSDavid Gibson new_page = old_page; 26681e8f889bSDavid Gibson } 26692ec74c3eSSagi Grimberg spin_unlock(&mm->page_table_lock); 26702ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 26712ec74c3eSSagi Grimberg /* Caller expects lock to be held */ 26722ec74c3eSSagi Grimberg spin_lock(&mm->page_table_lock); 26731e8f889bSDavid Gibson page_cache_release(new_page); 26741e8f889bSDavid Gibson page_cache_release(old_page); 267583c54070SNick Piggin return 0; 26761e8f889bSDavid Gibson } 26771e8f889bSDavid Gibson 267804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2679a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2680a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 268104f2cbe3SMel Gorman { 268204f2cbe3SMel Gorman struct address_space *mapping; 2683e7c4b0bfSAndy Whitcroft pgoff_t idx; 268404f2cbe3SMel Gorman 268504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2686a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 268704f2cbe3SMel Gorman 268804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 268904f2cbe3SMel Gorman } 269004f2cbe3SMel Gorman 26913ae77f43SHugh Dickins /* 26923ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 26933ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 26943ae77f43SHugh Dickins */ 26953ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 26962a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 26972a15efc9SHugh Dickins { 26982a15efc9SHugh Dickins struct address_space *mapping; 26992a15efc9SHugh Dickins pgoff_t idx; 27002a15efc9SHugh Dickins struct page *page; 27012a15efc9SHugh Dickins 27022a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 27032a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 27042a15efc9SHugh Dickins 27052a15efc9SHugh Dickins page = find_get_page(mapping, idx); 27062a15efc9SHugh Dickins if (page) 27072a15efc9SHugh Dickins put_page(page); 27082a15efc9SHugh Dickins return page != NULL; 27092a15efc9SHugh Dickins } 27102a15efc9SHugh Dickins 2711a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2712788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2713ac9b9c66SHugh Dickins { 2714a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2715ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2716409eb8c2SHillf Danton int anon_rmap = 0; 2717e7c4b0bfSAndy Whitcroft pgoff_t idx; 27184c887265SAdam Litke unsigned long size; 27194c887265SAdam Litke struct page *page; 27204c887265SAdam Litke struct address_space *mapping; 27211e8f889bSDavid Gibson pte_t new_pte; 27224c887265SAdam Litke 272304f2cbe3SMel Gorman /* 272404f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 272504f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 272625985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 272704f2cbe3SMel Gorman */ 272804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2729ffb22af5SAndrew Morton pr_warning("PID %d killed due to inadequate hugepage pool\n", 273004f2cbe3SMel Gorman current->pid); 273104f2cbe3SMel Gorman return ret; 273204f2cbe3SMel Gorman } 273304f2cbe3SMel Gorman 27344c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2735a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 27364c887265SAdam Litke 27374c887265SAdam Litke /* 27384c887265SAdam Litke * Use page lock to guard against racing truncation 27394c887265SAdam Litke * before we get page_table_lock. 27404c887265SAdam Litke */ 27416bda666aSChristoph Lameter retry: 27426bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 27436bda666aSChristoph Lameter if (!page) { 2744a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2745ebed4bfcSHugh Dickins if (idx >= size) 2746ebed4bfcSHugh Dickins goto out; 274704f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 27482fc39cecSAdam Litke if (IS_ERR(page)) { 274976dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 275076dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 275176dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 275276dcee75SAneesh Kumar K.V else 275376dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 27546bda666aSChristoph Lameter goto out; 27556bda666aSChristoph Lameter } 275647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 27570ed361deSNick Piggin __SetPageUptodate(page); 2758ac9b9c66SHugh Dickins 2759f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 27606bda666aSChristoph Lameter int err; 276145c682a6SKen Chen struct inode *inode = mapping->host; 27626bda666aSChristoph Lameter 27636bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 27646bda666aSChristoph Lameter if (err) { 27656bda666aSChristoph Lameter put_page(page); 27666bda666aSChristoph Lameter if (err == -EEXIST) 27676bda666aSChristoph Lameter goto retry; 27686bda666aSChristoph Lameter goto out; 27696bda666aSChristoph Lameter } 277045c682a6SKen Chen 277145c682a6SKen Chen spin_lock(&inode->i_lock); 2772a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 277345c682a6SKen Chen spin_unlock(&inode->i_lock); 277423be7468SMel Gorman } else { 27756bda666aSChristoph Lameter lock_page(page); 27760fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 27770fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 27780fe6e20bSNaoya Horiguchi goto backout_unlocked; 277923be7468SMel Gorman } 2780409eb8c2SHillf Danton anon_rmap = 1; 27810fe6e20bSNaoya Horiguchi } 27820fe6e20bSNaoya Horiguchi } else { 278357303d80SAndy Whitcroft /* 2784998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2785998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2786998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2787fd6a03edSNaoya Horiguchi */ 2788fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2789aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2790972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2791fd6a03edSNaoya Horiguchi goto backout_unlocked; 27926bda666aSChristoph Lameter } 2793998b4382SNaoya Horiguchi } 27941e8f889bSDavid Gibson 279557303d80SAndy Whitcroft /* 279657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 279757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 279857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 279957303d80SAndy Whitcroft * the spinlock. 280057303d80SAndy Whitcroft */ 2801788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 28022b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28032b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 28042b26736cSAndy Whitcroft goto backout_unlocked; 28052b26736cSAndy Whitcroft } 280657303d80SAndy Whitcroft 2807ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2808a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 28094c887265SAdam Litke if (idx >= size) 28104c887265SAdam Litke goto backout; 28114c887265SAdam Litke 281283c54070SNick Piggin ret = 0; 28137f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 28144c887265SAdam Litke goto backout; 28154c887265SAdam Litke 2816409eb8c2SHillf Danton if (anon_rmap) 2817409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 2818409eb8c2SHillf Danton else 2819409eb8c2SHillf Danton page_dup_rmap(page); 28201e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 28211e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 28221e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 28231e8f889bSDavid Gibson 2824788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 28251e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 282604f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 28271e8f889bSDavid Gibson } 28281e8f889bSDavid Gibson 2829ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 28304c887265SAdam Litke unlock_page(page); 28314c887265SAdam Litke out: 2832ac9b9c66SHugh Dickins return ret; 28334c887265SAdam Litke 28344c887265SAdam Litke backout: 28354c887265SAdam Litke spin_unlock(&mm->page_table_lock); 28362b26736cSAndy Whitcroft backout_unlocked: 28374c887265SAdam Litke unlock_page(page); 28384c887265SAdam Litke put_page(page); 28394c887265SAdam Litke goto out; 2840ac9b9c66SHugh Dickins } 2841ac9b9c66SHugh Dickins 284286e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2843788c7df4SHugh Dickins unsigned long address, unsigned int flags) 284486e5216fSAdam Litke { 284586e5216fSAdam Litke pte_t *ptep; 284686e5216fSAdam Litke pte_t entry; 28471e8f889bSDavid Gibson int ret; 28480fe6e20bSNaoya Horiguchi struct page *page = NULL; 284957303d80SAndy Whitcroft struct page *pagecache_page = NULL; 28503935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2851a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 285286e5216fSAdam Litke 28531e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 28541e16a539SKAMEZAWA Hiroyuki 2855fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2856fd6a03edSNaoya Horiguchi if (ptep) { 2857fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2858290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 285930dad309SNaoya Horiguchi migration_entry_wait_huge(mm, ptep); 2860290408d4SNaoya Horiguchi return 0; 2861290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2862aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2863972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2864fd6a03edSNaoya Horiguchi } 2865fd6a03edSNaoya Horiguchi 2866a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 286786e5216fSAdam Litke if (!ptep) 286886e5216fSAdam Litke return VM_FAULT_OOM; 286986e5216fSAdam Litke 28703935baa9SDavid Gibson /* 28713935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 28723935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 28733935baa9SDavid Gibson * the same page in the page cache. 28743935baa9SDavid Gibson */ 28753935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 28767f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 28777f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2878788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2879b4d1d99fSDavid Gibson goto out_mutex; 28803935baa9SDavid Gibson } 288186e5216fSAdam Litke 288283c54070SNick Piggin ret = 0; 28831e8f889bSDavid Gibson 288457303d80SAndy Whitcroft /* 288557303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 288657303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 288757303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 288857303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 288957303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 289057303d80SAndy Whitcroft * consumed. 289157303d80SAndy Whitcroft */ 2892106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 28932b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28942b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2895b4d1d99fSDavid Gibson goto out_mutex; 28962b26736cSAndy Whitcroft } 289757303d80SAndy Whitcroft 2898f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 289957303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 290057303d80SAndy Whitcroft vma, address); 290157303d80SAndy Whitcroft } 290257303d80SAndy Whitcroft 290356c9cfb1SNaoya Horiguchi /* 290456c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 290556c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 290656c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 290756c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 290856c9cfb1SNaoya Horiguchi * so no worry about deadlock. 290956c9cfb1SNaoya Horiguchi */ 29100fe6e20bSNaoya Horiguchi page = pte_page(entry); 291166aebce7SChris Metcalf get_page(page); 291256c9cfb1SNaoya Horiguchi if (page != pagecache_page) 29130fe6e20bSNaoya Horiguchi lock_page(page); 29140fe6e20bSNaoya Horiguchi 29151e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 29161e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2917b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2918b4d1d99fSDavid Gibson goto out_page_table_lock; 2919b4d1d99fSDavid Gibson 2920b4d1d99fSDavid Gibson 2921788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2922106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 292357303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 292457303d80SAndy Whitcroft pagecache_page); 2925b4d1d99fSDavid Gibson goto out_page_table_lock; 2926b4d1d99fSDavid Gibson } 2927106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 2928b4d1d99fSDavid Gibson } 2929b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2930788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2931788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 29324b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2933b4d1d99fSDavid Gibson 2934b4d1d99fSDavid Gibson out_page_table_lock: 29351e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 293657303d80SAndy Whitcroft 293757303d80SAndy Whitcroft if (pagecache_page) { 293857303d80SAndy Whitcroft unlock_page(pagecache_page); 293957303d80SAndy Whitcroft put_page(pagecache_page); 294057303d80SAndy Whitcroft } 29411f64d69cSDean Nelson if (page != pagecache_page) 294256c9cfb1SNaoya Horiguchi unlock_page(page); 294366aebce7SChris Metcalf put_page(page); 294457303d80SAndy Whitcroft 2945b4d1d99fSDavid Gibson out_mutex: 29463935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 29471e8f889bSDavid Gibson 29481e8f889bSDavid Gibson return ret; 294986e5216fSAdam Litke } 295086e5216fSAdam Litke 295128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 295263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 295328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 295428a35716SMichel Lespinasse long i, unsigned int flags) 295563551ae0SDavid Gibson { 2956d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2957d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 295828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 2959a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 296063551ae0SDavid Gibson 29611c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 296263551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 296363551ae0SDavid Gibson pte_t *pte; 29642a15efc9SHugh Dickins int absent; 296563551ae0SDavid Gibson struct page *page; 296663551ae0SDavid Gibson 29674c887265SAdam Litke /* 29684c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 29692a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 29704c887265SAdam Litke * first, for the page indexing below to work. 29714c887265SAdam Litke */ 2972a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 29732a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 297463551ae0SDavid Gibson 29752a15efc9SHugh Dickins /* 29762a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 29773ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 29783ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 29793ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 29803ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 29812a15efc9SHugh Dickins */ 29823ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 29833ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 29842a15efc9SHugh Dickins remainder = 0; 29852a15efc9SHugh Dickins break; 29862a15efc9SHugh Dickins } 29872a15efc9SHugh Dickins 29889cc3a5bdSNaoya Horiguchi /* 29899cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 29909cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 29919cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 29929cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 29939cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 29949cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 29959cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 29969cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 29979cc3a5bdSNaoya Horiguchi */ 29989cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 2999106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3000106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 30014c887265SAdam Litke int ret; 30024c887265SAdam Litke 30034c887265SAdam Litke spin_unlock(&mm->page_table_lock); 30042a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 30052a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 30064c887265SAdam Litke spin_lock(&mm->page_table_lock); 3007a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 30084c887265SAdam Litke continue; 30094c887265SAdam Litke 30101c59827dSHugh Dickins remainder = 0; 30111c59827dSHugh Dickins break; 30121c59827dSHugh Dickins } 301363551ae0SDavid Gibson 3014a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 30157f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3016d5d4b0aaSChen, Kenneth W same_page: 3017d6692183SChen, Kenneth W if (pages) { 301869d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 30194b2e38adSKOSAKI Motohiro get_page(pages[i]); 3020d6692183SChen, Kenneth W } 302163551ae0SDavid Gibson 302263551ae0SDavid Gibson if (vmas) 302363551ae0SDavid Gibson vmas[i] = vma; 302463551ae0SDavid Gibson 302563551ae0SDavid Gibson vaddr += PAGE_SIZE; 3026d5d4b0aaSChen, Kenneth W ++pfn_offset; 302763551ae0SDavid Gibson --remainder; 302863551ae0SDavid Gibson ++i; 3029d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3030a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3031d5d4b0aaSChen, Kenneth W /* 3032d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3033d5d4b0aaSChen, Kenneth W * of this compound page. 3034d5d4b0aaSChen, Kenneth W */ 3035d5d4b0aaSChen, Kenneth W goto same_page; 3036d5d4b0aaSChen, Kenneth W } 303763551ae0SDavid Gibson } 30381c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 303928a35716SMichel Lespinasse *nr_pages = remainder; 304063551ae0SDavid Gibson *position = vaddr; 304163551ae0SDavid Gibson 30422a15efc9SHugh Dickins return i ? i : -EFAULT; 304363551ae0SDavid Gibson } 30448f860591SZhang, Yanmin 30457da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 30468f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 30478f860591SZhang, Yanmin { 30488f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 30498f860591SZhang, Yanmin unsigned long start = address; 30508f860591SZhang, Yanmin pte_t *ptep; 30518f860591SZhang, Yanmin pte_t pte; 3052a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 30537da4d641SPeter Zijlstra unsigned long pages = 0; 30548f860591SZhang, Yanmin 30558f860591SZhang, Yanmin BUG_ON(address >= end); 30568f860591SZhang, Yanmin flush_cache_range(vma, address, end); 30578f860591SZhang, Yanmin 30583d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 30598f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 3060a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 30618f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 30628f860591SZhang, Yanmin if (!ptep) 30638f860591SZhang, Yanmin continue; 30647da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 30657da4d641SPeter Zijlstra pages++; 306639dde65cSChen, Kenneth W continue; 30677da4d641SPeter Zijlstra } 30687f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 30698f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3070106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3071be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 30728f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 30737da4d641SPeter Zijlstra pages++; 30748f860591SZhang, Yanmin } 30758f860591SZhang, Yanmin } 30768f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 3077d833352aSMel Gorman /* 3078d833352aSMel Gorman * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3079d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3080d833352aSMel Gorman * once we release i_mmap_mutex, another task can do the final put_page 3081d833352aSMel Gorman * and that page table be reused and filled with junk. 3082d833352aSMel Gorman */ 30838f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 3084d833352aSMel Gorman mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 30857da4d641SPeter Zijlstra 30867da4d641SPeter Zijlstra return pages << h->order; 30878f860591SZhang, Yanmin } 30888f860591SZhang, Yanmin 3089a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3090a1e78772SMel Gorman long from, long to, 30915a6fe125SMel Gorman struct vm_area_struct *vma, 3092ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3093e4e574b7SAdam Litke { 309417c9d12eSMel Gorman long ret, chg; 3095a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 309690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 3097e4e574b7SAdam Litke 3098a1e78772SMel Gorman /* 309917c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 310017c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 310190481622SDavid Gibson * without using reserves 310217c9d12eSMel Gorman */ 3103ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 310417c9d12eSMel Gorman return 0; 310517c9d12eSMel Gorman 310617c9d12eSMel Gorman /* 3107a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3108a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3109a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3110a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3111a1e78772SMel Gorman */ 3112f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 3113e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 31145a6fe125SMel Gorman else { 31155a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 31165a6fe125SMel Gorman if (!resv_map) 31175a6fe125SMel Gorman return -ENOMEM; 31185a6fe125SMel Gorman 311917c9d12eSMel Gorman chg = to - from; 312017c9d12eSMel Gorman 31215a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 31225a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 31235a6fe125SMel Gorman } 31245a6fe125SMel Gorman 3125c50ac050SDave Hansen if (chg < 0) { 3126c50ac050SDave Hansen ret = chg; 3127c50ac050SDave Hansen goto out_err; 3128c50ac050SDave Hansen } 312917c9d12eSMel Gorman 313090481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3131c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3132c50ac050SDave Hansen ret = -ENOSPC; 3133c50ac050SDave Hansen goto out_err; 3134c50ac050SDave Hansen } 313517c9d12eSMel Gorman 313617c9d12eSMel Gorman /* 313717c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 313890481622SDavid Gibson * Hand the pages back to the subpool if there are not 313917c9d12eSMel Gorman */ 314017c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 314117c9d12eSMel Gorman if (ret < 0) { 314290481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3143c50ac050SDave Hansen goto out_err; 314417c9d12eSMel Gorman } 314517c9d12eSMel Gorman 314617c9d12eSMel Gorman /* 314717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 314817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 314917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 315017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 315117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 315217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 315317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 315417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 315517c9d12eSMel Gorman * else has to be done for private mappings here 315617c9d12eSMel Gorman */ 3157f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 315817c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 3159a43a8c39SChen, Kenneth W return 0; 3160c50ac050SDave Hansen out_err: 31614523e145SDave Hansen if (vma) 3162c50ac050SDave Hansen resv_map_put(vma); 3163c50ac050SDave Hansen return ret; 3164a43a8c39SChen, Kenneth W } 3165a43a8c39SChen, Kenneth W 3166a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3167a43a8c39SChen, Kenneth W { 3168a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 3169a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 317090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 317145c682a6SKen Chen 317245c682a6SKen Chen spin_lock(&inode->i_lock); 3173e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 317445c682a6SKen Chen spin_unlock(&inode->i_lock); 317545c682a6SKen Chen 317690481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3177a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3178a43a8c39SChen, Kenneth W } 317993f70f90SNaoya Horiguchi 31803212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 31813212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 31823212b535SSteve Capper struct vm_area_struct *vma, 31833212b535SSteve Capper unsigned long addr, pgoff_t idx) 31843212b535SSteve Capper { 31853212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 31863212b535SSteve Capper svma->vm_start; 31873212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 31883212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 31893212b535SSteve Capper 31903212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 31913212b535SSteve Capper unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 31923212b535SSteve Capper unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 31933212b535SSteve Capper 31943212b535SSteve Capper /* 31953212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 31963212b535SSteve Capper * page table page. 31973212b535SSteve Capper */ 31983212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 31993212b535SSteve Capper vm_flags != svm_flags || 32003212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 32013212b535SSteve Capper return 0; 32023212b535SSteve Capper 32033212b535SSteve Capper return saddr; 32043212b535SSteve Capper } 32053212b535SSteve Capper 32063212b535SSteve Capper static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 32073212b535SSteve Capper { 32083212b535SSteve Capper unsigned long base = addr & PUD_MASK; 32093212b535SSteve Capper unsigned long end = base + PUD_SIZE; 32103212b535SSteve Capper 32113212b535SSteve Capper /* 32123212b535SSteve Capper * check on proper vm_flags and page table alignment 32133212b535SSteve Capper */ 32143212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 32153212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 32163212b535SSteve Capper return 1; 32173212b535SSteve Capper return 0; 32183212b535SSteve Capper } 32193212b535SSteve Capper 32203212b535SSteve Capper /* 32213212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 32223212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 32233212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 32243212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 32253212b535SSteve Capper * pud has to be populated inside the same i_mmap_mutex section - otherwise 32263212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 32273212b535SSteve Capper * bad pmd for sharing. 32283212b535SSteve Capper */ 32293212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 32303212b535SSteve Capper { 32313212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 32323212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 32333212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 32343212b535SSteve Capper vma->vm_pgoff; 32353212b535SSteve Capper struct vm_area_struct *svma; 32363212b535SSteve Capper unsigned long saddr; 32373212b535SSteve Capper pte_t *spte = NULL; 32383212b535SSteve Capper pte_t *pte; 32393212b535SSteve Capper 32403212b535SSteve Capper if (!vma_shareable(vma, addr)) 32413212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 32423212b535SSteve Capper 32433212b535SSteve Capper mutex_lock(&mapping->i_mmap_mutex); 32443212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 32453212b535SSteve Capper if (svma == vma) 32463212b535SSteve Capper continue; 32473212b535SSteve Capper 32483212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 32493212b535SSteve Capper if (saddr) { 32503212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 32513212b535SSteve Capper if (spte) { 32523212b535SSteve Capper get_page(virt_to_page(spte)); 32533212b535SSteve Capper break; 32543212b535SSteve Capper } 32553212b535SSteve Capper } 32563212b535SSteve Capper } 32573212b535SSteve Capper 32583212b535SSteve Capper if (!spte) 32593212b535SSteve Capper goto out; 32603212b535SSteve Capper 32613212b535SSteve Capper spin_lock(&mm->page_table_lock); 32623212b535SSteve Capper if (pud_none(*pud)) 32633212b535SSteve Capper pud_populate(mm, pud, 32643212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 32653212b535SSteve Capper else 32663212b535SSteve Capper put_page(virt_to_page(spte)); 32673212b535SSteve Capper spin_unlock(&mm->page_table_lock); 32683212b535SSteve Capper out: 32693212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 32703212b535SSteve Capper mutex_unlock(&mapping->i_mmap_mutex); 32713212b535SSteve Capper return pte; 32723212b535SSteve Capper } 32733212b535SSteve Capper 32743212b535SSteve Capper /* 32753212b535SSteve Capper * unmap huge page backed by shared pte. 32763212b535SSteve Capper * 32773212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 32783212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 32793212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 32803212b535SSteve Capper * 32813212b535SSteve Capper * called with vma->vm_mm->page_table_lock held. 32823212b535SSteve Capper * 32833212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 32843212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 32853212b535SSteve Capper */ 32863212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 32873212b535SSteve Capper { 32883212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 32893212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 32903212b535SSteve Capper 32913212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 32923212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 32933212b535SSteve Capper return 0; 32943212b535SSteve Capper 32953212b535SSteve Capper pud_clear(pud); 32963212b535SSteve Capper put_page(virt_to_page(ptep)); 32973212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 32983212b535SSteve Capper return 1; 32993212b535SSteve Capper } 33009e5fc74cSSteve Capper #define want_pmd_share() (1) 33019e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 33029e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 33039e5fc74cSSteve Capper { 33049e5fc74cSSteve Capper return NULL; 33059e5fc74cSSteve Capper } 33069e5fc74cSSteve Capper #define want_pmd_share() (0) 33073212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 33083212b535SSteve Capper 33099e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 33109e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 33119e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 33129e5fc74cSSteve Capper { 33139e5fc74cSSteve Capper pgd_t *pgd; 33149e5fc74cSSteve Capper pud_t *pud; 33159e5fc74cSSteve Capper pte_t *pte = NULL; 33169e5fc74cSSteve Capper 33179e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 33189e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 33199e5fc74cSSteve Capper if (pud) { 33209e5fc74cSSteve Capper if (sz == PUD_SIZE) { 33219e5fc74cSSteve Capper pte = (pte_t *)pud; 33229e5fc74cSSteve Capper } else { 33239e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 33249e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 33259e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 33269e5fc74cSSteve Capper else 33279e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 33289e5fc74cSSteve Capper } 33299e5fc74cSSteve Capper } 33309e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 33319e5fc74cSSteve Capper 33329e5fc74cSSteve Capper return pte; 33339e5fc74cSSteve Capper } 33349e5fc74cSSteve Capper 33359e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 33369e5fc74cSSteve Capper { 33379e5fc74cSSteve Capper pgd_t *pgd; 33389e5fc74cSSteve Capper pud_t *pud; 33399e5fc74cSSteve Capper pmd_t *pmd = NULL; 33409e5fc74cSSteve Capper 33419e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 33429e5fc74cSSteve Capper if (pgd_present(*pgd)) { 33439e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 33449e5fc74cSSteve Capper if (pud_present(*pud)) { 33459e5fc74cSSteve Capper if (pud_huge(*pud)) 33469e5fc74cSSteve Capper return (pte_t *)pud; 33479e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 33489e5fc74cSSteve Capper } 33499e5fc74cSSteve Capper } 33509e5fc74cSSteve Capper return (pte_t *) pmd; 33519e5fc74cSSteve Capper } 33529e5fc74cSSteve Capper 33539e5fc74cSSteve Capper struct page * 33549e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 33559e5fc74cSSteve Capper pmd_t *pmd, int write) 33569e5fc74cSSteve Capper { 33579e5fc74cSSteve Capper struct page *page; 33589e5fc74cSSteve Capper 33599e5fc74cSSteve Capper page = pte_page(*(pte_t *)pmd); 33609e5fc74cSSteve Capper if (page) 33619e5fc74cSSteve Capper page += ((address & ~PMD_MASK) >> PAGE_SHIFT); 33629e5fc74cSSteve Capper return page; 33639e5fc74cSSteve Capper } 33649e5fc74cSSteve Capper 33659e5fc74cSSteve Capper struct page * 33669e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 33679e5fc74cSSteve Capper pud_t *pud, int write) 33689e5fc74cSSteve Capper { 33699e5fc74cSSteve Capper struct page *page; 33709e5fc74cSSteve Capper 33719e5fc74cSSteve Capper page = pte_page(*(pte_t *)pud); 33729e5fc74cSSteve Capper if (page) 33739e5fc74cSSteve Capper page += ((address & ~PUD_MASK) >> PAGE_SHIFT); 33749e5fc74cSSteve Capper return page; 33759e5fc74cSSteve Capper } 33769e5fc74cSSteve Capper 33779e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 33789e5fc74cSSteve Capper 33799e5fc74cSSteve Capper /* Can be overriden by architectures */ 33809e5fc74cSSteve Capper __attribute__((weak)) struct page * 33819e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 33829e5fc74cSSteve Capper pud_t *pud, int write) 33839e5fc74cSSteve Capper { 33849e5fc74cSSteve Capper BUG(); 33859e5fc74cSSteve Capper return NULL; 33869e5fc74cSSteve Capper } 33879e5fc74cSSteve Capper 33889e5fc74cSSteve Capper #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 33899e5fc74cSSteve Capper 3390d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3391d5bd9106SAndi Kleen 33926de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 33936de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 33946de2b1aaSNaoya Horiguchi { 33956de2b1aaSNaoya Horiguchi struct page *page; 33966de2b1aaSNaoya Horiguchi struct page *tmp; 33976de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 33986de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 33996de2b1aaSNaoya Horiguchi 34006de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 34016de2b1aaSNaoya Horiguchi if (page == hpage) 34026de2b1aaSNaoya Horiguchi return 1; 34036de2b1aaSNaoya Horiguchi return 0; 34046de2b1aaSNaoya Horiguchi } 34056de2b1aaSNaoya Horiguchi 340693f70f90SNaoya Horiguchi /* 340793f70f90SNaoya Horiguchi * This function is called from memory failure code. 340893f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 340993f70f90SNaoya Horiguchi */ 34106de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 341193f70f90SNaoya Horiguchi { 341293f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 341393f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 34146de2b1aaSNaoya Horiguchi int ret = -EBUSY; 341593f70f90SNaoya Horiguchi 341693f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 34176de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 341856f2fb14SNaoya Horiguchi /* 341956f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 342056f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 342156f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 342256f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 342356f2fb14SNaoya Horiguchi */ 342456f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 34258c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 342693f70f90SNaoya Horiguchi h->free_huge_pages--; 342793f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 34286de2b1aaSNaoya Horiguchi ret = 0; 342993f70f90SNaoya Horiguchi } 34306de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 34316de2b1aaSNaoya Horiguchi return ret; 34326de2b1aaSNaoya Horiguchi } 34336de2b1aaSNaoya Horiguchi #endif 3434