11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 31da177e4SLinus Torvalds * (C) William Irwin, 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 { 13090481622SDavid Gibson return subpool_inode(vma->vm_file->f_dentry->d_inode); 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 32208fba699SMel Gorman return 1UL << (hstate->order + PAGE_SHIFT); 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); 630aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 6310edaecfaSAneesh Kumar K.V /* remove the page from active list */ 6320edaecfaSAneesh Kumar K.V list_del(&page->lru); 633a5516438SAndi Kleen update_and_free_page(h, page); 634a5516438SAndi Kleen h->surplus_huge_pages--; 635a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 6367893d1d5SAdam Litke } else { 637a5516438SAndi Kleen enqueue_huge_page(h, page); 6387893d1d5SAdam Litke } 63927a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 64090481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 64127a85ef1SDavid Gibson } 64227a85ef1SDavid Gibson 643a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 644b7ba30c6SAndi Kleen { 6450edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 646b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 647b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 6489dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 649a5516438SAndi Kleen h->nr_huge_pages++; 650a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 651b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 652b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 653b7ba30c6SAndi Kleen } 654b7ba30c6SAndi Kleen 65520a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 65620a0307cSWu Fengguang { 65720a0307cSWu Fengguang int i; 65820a0307cSWu Fengguang int nr_pages = 1 << order; 65920a0307cSWu Fengguang struct page *p = page + 1; 66020a0307cSWu Fengguang 66120a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 66220a0307cSWu Fengguang set_compound_order(page, order); 66320a0307cSWu Fengguang __SetPageHead(page); 66420a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 66520a0307cSWu Fengguang __SetPageTail(p); 66658a84aa9SYouquan Song set_page_count(p, 0); 66720a0307cSWu Fengguang p->first_page = page; 66820a0307cSWu Fengguang } 66920a0307cSWu Fengguang } 67020a0307cSWu Fengguang 67120a0307cSWu Fengguang int PageHuge(struct page *page) 67220a0307cSWu Fengguang { 67320a0307cSWu Fengguang compound_page_dtor *dtor; 67420a0307cSWu Fengguang 67520a0307cSWu Fengguang if (!PageCompound(page)) 67620a0307cSWu Fengguang return 0; 67720a0307cSWu Fengguang 67820a0307cSWu Fengguang page = compound_head(page); 67920a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 68020a0307cSWu Fengguang 68120a0307cSWu Fengguang return dtor == free_huge_page; 68220a0307cSWu Fengguang } 68343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 68443131e14SNaoya Horiguchi 685a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6861da177e4SLinus Torvalds { 6871da177e4SLinus Torvalds struct page *page; 688f96efd58SJoe Jin 689aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 690aa888a74SAndi Kleen return NULL; 691aa888a74SAndi Kleen 6926484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 693551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 694551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 695a5516438SAndi Kleen huge_page_order(h)); 6961da177e4SLinus Torvalds if (page) { 6977f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 698caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6997b8ee84dSHarvey Harrison return NULL; 7007f2e9525SGerald Schaefer } 701a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 7021da177e4SLinus Torvalds } 70363b4613cSNishanth Aravamudan 70463b4613cSNishanth Aravamudan return page; 70563b4613cSNishanth Aravamudan } 70663b4613cSNishanth Aravamudan 7075ced66c9SAndi Kleen /* 7086ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 7096ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 7106ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7116ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 7126ae11b27SLee Schermerhorn * node for alloc or free. 7139a76db09SLee Schermerhorn */ 7146ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7159a76db09SLee Schermerhorn { 7166ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 7179a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 7186ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 7199a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 7209a76db09SLee Schermerhorn 7219a76db09SLee Schermerhorn return nid; 7229a76db09SLee Schermerhorn } 7239a76db09SLee Schermerhorn 7246ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 7255ced66c9SAndi Kleen { 7266ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 7276ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 7289a76db09SLee Schermerhorn return nid; 7295ced66c9SAndi Kleen } 7305ced66c9SAndi Kleen 7316ae11b27SLee Schermerhorn /* 7326ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 7336ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 7346ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 7356ae11b27SLee Schermerhorn * mask. 7366ae11b27SLee Schermerhorn */ 7376ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 7386ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 7396ae11b27SLee Schermerhorn { 7406ae11b27SLee Schermerhorn int nid; 7416ae11b27SLee Schermerhorn 7426ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7436ae11b27SLee Schermerhorn 7446ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 7456ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 7466ae11b27SLee Schermerhorn 7476ae11b27SLee Schermerhorn return nid; 7486ae11b27SLee Schermerhorn } 7496ae11b27SLee Schermerhorn 7506ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 75163b4613cSNishanth Aravamudan { 75263b4613cSNishanth Aravamudan struct page *page; 75363b4613cSNishanth Aravamudan int start_nid; 75463b4613cSNishanth Aravamudan int next_nid; 75563b4613cSNishanth Aravamudan int ret = 0; 75663b4613cSNishanth Aravamudan 7576ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 758e8c5c824SLee Schermerhorn next_nid = start_nid; 75963b4613cSNishanth Aravamudan 76063b4613cSNishanth Aravamudan do { 761e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 7629a76db09SLee Schermerhorn if (page) { 76363b4613cSNishanth Aravamudan ret = 1; 7649a76db09SLee Schermerhorn break; 7659a76db09SLee Schermerhorn } 7666ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 7679a76db09SLee Schermerhorn } while (next_nid != start_nid); 76863b4613cSNishanth Aravamudan 7693b116300SAdam Litke if (ret) 7703b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7713b116300SAdam Litke else 7723b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7733b116300SAdam Litke 77463b4613cSNishanth Aravamudan return ret; 7751da177e4SLinus Torvalds } 7761da177e4SLinus Torvalds 777e8c5c824SLee Schermerhorn /* 7786ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7796ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7806ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7816ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 782e8c5c824SLee Schermerhorn */ 7836ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 784e8c5c824SLee Schermerhorn { 7856ae11b27SLee Schermerhorn int nid; 7869a76db09SLee Schermerhorn 7876ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7886ae11b27SLee Schermerhorn 7896ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7906ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7916ae11b27SLee Schermerhorn 7929a76db09SLee Schermerhorn return nid; 793e8c5c824SLee Schermerhorn } 794e8c5c824SLee Schermerhorn 795e8c5c824SLee Schermerhorn /* 796e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 797e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 798e8c5c824SLee Schermerhorn * balanced over allowed nodes. 799e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 800e8c5c824SLee Schermerhorn */ 8016ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 8026ae11b27SLee Schermerhorn bool acct_surplus) 803e8c5c824SLee Schermerhorn { 804e8c5c824SLee Schermerhorn int start_nid; 805e8c5c824SLee Schermerhorn int next_nid; 806e8c5c824SLee Schermerhorn int ret = 0; 807e8c5c824SLee Schermerhorn 8086ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 809e8c5c824SLee Schermerhorn next_nid = start_nid; 810e8c5c824SLee Schermerhorn 811e8c5c824SLee Schermerhorn do { 812685f3457SLee Schermerhorn /* 813685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 814685f3457SLee Schermerhorn * nodes with surplus pages. 815685f3457SLee Schermerhorn */ 816685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 817685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 818e8c5c824SLee Schermerhorn struct page *page = 819e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 820e8c5c824SLee Schermerhorn struct page, lru); 821e8c5c824SLee Schermerhorn list_del(&page->lru); 822e8c5c824SLee Schermerhorn h->free_huge_pages--; 823e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 824685f3457SLee Schermerhorn if (acct_surplus) { 825685f3457SLee Schermerhorn h->surplus_huge_pages--; 826685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 827685f3457SLee Schermerhorn } 828e8c5c824SLee Schermerhorn update_and_free_page(h, page); 829e8c5c824SLee Schermerhorn ret = 1; 8309a76db09SLee Schermerhorn break; 831e8c5c824SLee Schermerhorn } 8326ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 8339a76db09SLee Schermerhorn } while (next_nid != start_nid); 834e8c5c824SLee Schermerhorn 835e8c5c824SLee Schermerhorn return ret; 836e8c5c824SLee Schermerhorn } 837e8c5c824SLee Schermerhorn 838bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 8397893d1d5SAdam Litke { 8407893d1d5SAdam Litke struct page *page; 841bf50bab2SNaoya Horiguchi unsigned int r_nid; 8427893d1d5SAdam Litke 843aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 844aa888a74SAndi Kleen return NULL; 845aa888a74SAndi Kleen 846d1c3fb1fSNishanth Aravamudan /* 847d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 848d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 849d1c3fb1fSNishanth Aravamudan * overcommit 850d1c3fb1fSNishanth Aravamudan * 851d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 852d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 853d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 854d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 855d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 856d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 857d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 858d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 859d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 860d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 861d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 862d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 863d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 864d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 865d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 866d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 867d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 868d1c3fb1fSNishanth Aravamudan */ 869d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 870a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 871d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 872d1c3fb1fSNishanth Aravamudan return NULL; 873d1c3fb1fSNishanth Aravamudan } else { 874a5516438SAndi Kleen h->nr_huge_pages++; 875a5516438SAndi Kleen h->surplus_huge_pages++; 876d1c3fb1fSNishanth Aravamudan } 877d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 878d1c3fb1fSNishanth Aravamudan 879bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 880551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 881551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 882a5516438SAndi Kleen huge_page_order(h)); 883bf50bab2SNaoya Horiguchi else 884bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 885bf50bab2SNaoya Horiguchi htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 886bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 887d1c3fb1fSNishanth Aravamudan 888caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 889caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 890ea5768c7SHillf Danton page = NULL; 891caff3a2cSGerald Schaefer } 892caff3a2cSGerald Schaefer 8937893d1d5SAdam Litke spin_lock(&hugetlb_lock); 894d1c3fb1fSNishanth Aravamudan if (page) { 8950edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 896bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 897d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 8989dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 899d1c3fb1fSNishanth Aravamudan /* 900d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 901d1c3fb1fSNishanth Aravamudan */ 902bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 903bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 9043b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 905d1c3fb1fSNishanth Aravamudan } else { 906a5516438SAndi Kleen h->nr_huge_pages--; 907a5516438SAndi Kleen h->surplus_huge_pages--; 9083b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 9097893d1d5SAdam Litke } 910d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 9117893d1d5SAdam Litke 9127893d1d5SAdam Litke return page; 9137893d1d5SAdam Litke } 9147893d1d5SAdam Litke 915e4e574b7SAdam Litke /* 916bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 917bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 918bf50bab2SNaoya Horiguchi * address of error page. 919bf50bab2SNaoya Horiguchi */ 920bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 921bf50bab2SNaoya Horiguchi { 922bf50bab2SNaoya Horiguchi struct page *page; 923bf50bab2SNaoya Horiguchi 924bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 925bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 926bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 927bf50bab2SNaoya Horiguchi 928bf50bab2SNaoya Horiguchi if (!page) 929bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 930bf50bab2SNaoya Horiguchi 931bf50bab2SNaoya Horiguchi return page; 932bf50bab2SNaoya Horiguchi } 933bf50bab2SNaoya Horiguchi 934bf50bab2SNaoya Horiguchi /* 93525985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 936e4e574b7SAdam Litke * of size 'delta'. 937e4e574b7SAdam Litke */ 938a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 939e4e574b7SAdam Litke { 940e4e574b7SAdam Litke struct list_head surplus_list; 941e4e574b7SAdam Litke struct page *page, *tmp; 942e4e574b7SAdam Litke int ret, i; 943e4e574b7SAdam Litke int needed, allocated; 94428073b02SHillf Danton bool alloc_ok = true; 945e4e574b7SAdam Litke 946a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 947ac09b3a1SAdam Litke if (needed <= 0) { 948a5516438SAndi Kleen h->resv_huge_pages += delta; 949e4e574b7SAdam Litke return 0; 950ac09b3a1SAdam Litke } 951e4e574b7SAdam Litke 952e4e574b7SAdam Litke allocated = 0; 953e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 954e4e574b7SAdam Litke 955e4e574b7SAdam Litke ret = -ENOMEM; 956e4e574b7SAdam Litke retry: 957e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 958e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 959bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 96028073b02SHillf Danton if (!page) { 96128073b02SHillf Danton alloc_ok = false; 96228073b02SHillf Danton break; 96328073b02SHillf Danton } 964e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 965e4e574b7SAdam Litke } 96628073b02SHillf Danton allocated += i; 967e4e574b7SAdam Litke 968e4e574b7SAdam Litke /* 969e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 970e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 971e4e574b7SAdam Litke */ 972e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 973a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 974a5516438SAndi Kleen (h->free_huge_pages + allocated); 97528073b02SHillf Danton if (needed > 0) { 97628073b02SHillf Danton if (alloc_ok) 977e4e574b7SAdam Litke goto retry; 97828073b02SHillf Danton /* 97928073b02SHillf Danton * We were not able to allocate enough pages to 98028073b02SHillf Danton * satisfy the entire reservation so we free what 98128073b02SHillf Danton * we've allocated so far. 98228073b02SHillf Danton */ 98328073b02SHillf Danton goto free; 98428073b02SHillf Danton } 985e4e574b7SAdam Litke /* 986e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 98725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 988e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 989ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 990ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 991ac09b3a1SAdam Litke * before they are reserved. 992e4e574b7SAdam Litke */ 993e4e574b7SAdam Litke needed += allocated; 994a5516438SAndi Kleen h->resv_huge_pages += delta; 995e4e574b7SAdam Litke ret = 0; 996a9869b83SNaoya Horiguchi 99719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 99819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 99919fc3f0aSAdam Litke if ((--needed) < 0) 100019fc3f0aSAdam Litke break; 1001a9869b83SNaoya Horiguchi /* 1002a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1003a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1004a9869b83SNaoya Horiguchi */ 1005a9869b83SNaoya Horiguchi put_page_testzero(page); 1006a9869b83SNaoya Horiguchi VM_BUG_ON(page_count(page)); 1007a5516438SAndi Kleen enqueue_huge_page(h, page); 100819fc3f0aSAdam Litke } 100928073b02SHillf Danton free: 1010b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 101119fc3f0aSAdam Litke 101219fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 101319fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 1014e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1015a9869b83SNaoya Horiguchi put_page(page); 1016a9869b83SNaoya Horiguchi } 1017af767cbdSAdam Litke } 101819fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1019e4e574b7SAdam Litke 1020e4e574b7SAdam Litke return ret; 1021e4e574b7SAdam Litke } 1022e4e574b7SAdam Litke 1023e4e574b7SAdam Litke /* 1024e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1025e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1026e4e574b7SAdam Litke * never used. 1027685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1028e4e574b7SAdam Litke */ 1029a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1030a5516438SAndi Kleen unsigned long unused_resv_pages) 1031e4e574b7SAdam Litke { 1032e4e574b7SAdam Litke unsigned long nr_pages; 1033e4e574b7SAdam Litke 1034ac09b3a1SAdam Litke /* Uncommit the reservation */ 1035a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1036ac09b3a1SAdam Litke 1037aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1038aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1039aa888a74SAndi Kleen return; 1040aa888a74SAndi Kleen 1041a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1042e4e574b7SAdam Litke 1043685f3457SLee Schermerhorn /* 1044685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 10459b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 10469b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 10479b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 10489b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 10499b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1050685f3457SLee Schermerhorn */ 1051685f3457SLee Schermerhorn while (nr_pages--) { 10529b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 1053685f3457SLee Schermerhorn break; 1054e4e574b7SAdam Litke } 1055e4e574b7SAdam Litke } 1056e4e574b7SAdam Litke 1057c37f9fb1SAndy Whitcroft /* 1058c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1059c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 106090481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 106190481622SDavid Gibson * can occur. Where any new reservation would be required the 106290481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 106390481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 106490481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 106590481622SDavid Gibson * failure. 1066c37f9fb1SAndy Whitcroft */ 1067e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1068a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1069c37f9fb1SAndy Whitcroft { 1070c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1071c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1072c37f9fb1SAndy Whitcroft 1073f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1074a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1075c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 1076c37f9fb1SAndy Whitcroft idx, idx + 1); 1077c37f9fb1SAndy Whitcroft 107884afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1079c37f9fb1SAndy Whitcroft return 1; 1080c37f9fb1SAndy Whitcroft 108184afd99bSAndy Whitcroft } else { 1082e2f17d94SRoel Kluin long err; 1083a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 108484afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 108584afd99bSAndy Whitcroft 108684afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 108784afd99bSAndy Whitcroft if (err < 0) 108884afd99bSAndy Whitcroft return err; 1089c37f9fb1SAndy Whitcroft return 0; 1090c37f9fb1SAndy Whitcroft } 109184afd99bSAndy Whitcroft } 1092a5516438SAndi Kleen static void vma_commit_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 region_add(&inode->i_mapping->private_list, idx, idx + 1); 110184afd99bSAndy Whitcroft 110284afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1103a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 110484afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 110584afd99bSAndy Whitcroft 110684afd99bSAndy Whitcroft /* Mark this page used in the map. */ 110784afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1108c37f9fb1SAndy Whitcroft } 1109c37f9fb1SAndy Whitcroft } 1110c37f9fb1SAndy Whitcroft 1111348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 111204f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1113348ea204SAdam Litke { 111490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1115a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1116348ea204SAdam Litke struct page *page; 1117e2f17d94SRoel Kluin long chg; 11182fc39cecSAdam Litke 1119a1e78772SMel Gorman /* 112090481622SDavid Gibson * Processes that did not create the mapping will have no 112190481622SDavid Gibson * reserves and will not have accounted against subpool 112290481622SDavid Gibson * limit. Check that the subpool limit can be made before 112390481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 112490481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 112590481622SDavid Gibson * mapping overlaps. 1126a1e78772SMel Gorman */ 1127a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1128c37f9fb1SAndy Whitcroft if (chg < 0) 112976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1130c37f9fb1SAndy Whitcroft if (chg) 113190481622SDavid Gibson if (hugepage_subpool_get_pages(spool, chg)) 113276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 113390d8b7e6SAdam Litke 1134a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1135a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1136a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1137a1e78772SMel Gorman 1138a1e78772SMel Gorman if (!page) { 1139bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1140a1e78772SMel Gorman if (!page) { 114190481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 114276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 1143a1e78772SMel Gorman } 1144a1e78772SMel Gorman } 1145a1e78772SMel Gorman 114690481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1147a1e78772SMel Gorman 1148a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1149c37f9fb1SAndy Whitcroft 11507893d1d5SAdam Litke return page; 1151b45b5bd6SDavid Gibson } 1152b45b5bd6SDavid Gibson 115391f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1154aa888a74SAndi Kleen { 1155aa888a74SAndi Kleen struct huge_bootmem_page *m; 11569b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1157aa888a74SAndi Kleen 1158aa888a74SAndi Kleen while (nr_nodes) { 1159aa888a74SAndi Kleen void *addr; 1160aa888a74SAndi Kleen 1161aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 11626ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 11639b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1164aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1165aa888a74SAndi Kleen 1166aa888a74SAndi Kleen if (addr) { 1167aa888a74SAndi Kleen /* 1168aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1169aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1170aa888a74SAndi Kleen * puts them into the mem_map). 1171aa888a74SAndi Kleen */ 1172aa888a74SAndi Kleen m = addr; 1173aa888a74SAndi Kleen goto found; 1174aa888a74SAndi Kleen } 1175aa888a74SAndi Kleen nr_nodes--; 1176aa888a74SAndi Kleen } 1177aa888a74SAndi Kleen return 0; 1178aa888a74SAndi Kleen 1179aa888a74SAndi Kleen found: 1180aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1181aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1182aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1183aa888a74SAndi Kleen m->hstate = h; 1184aa888a74SAndi Kleen return 1; 1185aa888a74SAndi Kleen } 1186aa888a74SAndi Kleen 118718229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 118818229df5SAndy Whitcroft { 118918229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 119018229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 119118229df5SAndy Whitcroft else 119218229df5SAndy Whitcroft prep_compound_page(page, order); 119318229df5SAndy Whitcroft } 119418229df5SAndy Whitcroft 1195aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1196aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1197aa888a74SAndi Kleen { 1198aa888a74SAndi Kleen struct huge_bootmem_page *m; 1199aa888a74SAndi Kleen 1200aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1201aa888a74SAndi Kleen struct hstate *h = m->hstate; 1202ee8f248dSBecky Bruce struct page *page; 1203ee8f248dSBecky Bruce 1204ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1205ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 1206ee8f248dSBecky Bruce free_bootmem_late((unsigned long)m, 1207ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1208ee8f248dSBecky Bruce #else 1209ee8f248dSBecky Bruce page = virt_to_page(m); 1210ee8f248dSBecky Bruce #endif 1211aa888a74SAndi Kleen __ClearPageReserved(page); 1212aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 121318229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1214aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1215b0320c7bSRafael Aquini /* 1216b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1217b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1218b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1219b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1220b0320c7bSRafael Aquini */ 1221b0320c7bSRafael Aquini if (h->order > (MAX_ORDER - 1)) 1222b0320c7bSRafael Aquini totalram_pages += 1 << h->order; 1223aa888a74SAndi Kleen } 1224aa888a74SAndi Kleen } 1225aa888a74SAndi Kleen 12268faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 12271da177e4SLinus Torvalds { 12281da177e4SLinus Torvalds unsigned long i; 12291da177e4SLinus Torvalds 1230e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1231aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1232aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1233aa888a74SAndi Kleen break; 12349b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 12359b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 12361da177e4SLinus Torvalds break; 12371da177e4SLinus Torvalds } 12388faa8b07SAndi Kleen h->max_huge_pages = i; 1239e5ff2159SAndi Kleen } 1240e5ff2159SAndi Kleen 1241e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1242e5ff2159SAndi Kleen { 1243e5ff2159SAndi Kleen struct hstate *h; 1244e5ff2159SAndi Kleen 1245e5ff2159SAndi Kleen for_each_hstate(h) { 12468faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 12478faa8b07SAndi Kleen if (h->order < MAX_ORDER) 12488faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1249e5ff2159SAndi Kleen } 1250e5ff2159SAndi Kleen } 1251e5ff2159SAndi Kleen 12524abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 12534abd32dbSAndi Kleen { 12544abd32dbSAndi Kleen if (n >= (1UL << 30)) 12554abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 12564abd32dbSAndi Kleen else if (n >= (1UL << 20)) 12574abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 12584abd32dbSAndi Kleen else 12594abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 12604abd32dbSAndi Kleen return buf; 12614abd32dbSAndi Kleen } 12624abd32dbSAndi Kleen 1263e5ff2159SAndi Kleen static void __init report_hugepages(void) 1264e5ff2159SAndi Kleen { 1265e5ff2159SAndi Kleen struct hstate *h; 1266e5ff2159SAndi Kleen 1267e5ff2159SAndi Kleen for_each_hstate(h) { 12684abd32dbSAndi Kleen char buf[32]; 12694abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 12704abd32dbSAndi Kleen "pre-allocated %ld pages\n", 12714abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 12724abd32dbSAndi Kleen h->free_huge_pages); 1273e5ff2159SAndi Kleen } 1274e5ff2159SAndi Kleen } 1275e5ff2159SAndi Kleen 12761da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 12776ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 12786ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12791da177e4SLinus Torvalds { 12804415cc8dSChristoph Lameter int i; 12814415cc8dSChristoph Lameter 1282aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1283aa888a74SAndi Kleen return; 1284aa888a74SAndi Kleen 12856ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 12861da177e4SLinus Torvalds struct page *page, *next; 1287a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1288a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1289a5516438SAndi Kleen if (count >= h->nr_huge_pages) 12906b0c880dSAdam Litke return; 12911da177e4SLinus Torvalds if (PageHighMem(page)) 12921da177e4SLinus Torvalds continue; 12931da177e4SLinus Torvalds list_del(&page->lru); 1294e5ff2159SAndi Kleen update_and_free_page(h, page); 1295a5516438SAndi Kleen h->free_huge_pages--; 1296a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 12971da177e4SLinus Torvalds } 12981da177e4SLinus Torvalds } 12991da177e4SLinus Torvalds } 13001da177e4SLinus Torvalds #else 13016ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 13026ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13031da177e4SLinus Torvalds { 13041da177e4SLinus Torvalds } 13051da177e4SLinus Torvalds #endif 13061da177e4SLinus Torvalds 130720a0307cSWu Fengguang /* 130820a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 130920a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 131020a0307cSWu Fengguang * Returns 1 if an adjustment was made. 131120a0307cSWu Fengguang */ 13126ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 13136ae11b27SLee Schermerhorn int delta) 131420a0307cSWu Fengguang { 1315e8c5c824SLee Schermerhorn int start_nid, next_nid; 131620a0307cSWu Fengguang int ret = 0; 131720a0307cSWu Fengguang 131820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 131920a0307cSWu Fengguang 1320e8c5c824SLee Schermerhorn if (delta < 0) 13216ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1322e8c5c824SLee Schermerhorn else 13236ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1324e8c5c824SLee Schermerhorn next_nid = start_nid; 1325e8c5c824SLee Schermerhorn 1326e8c5c824SLee Schermerhorn do { 1327e8c5c824SLee Schermerhorn int nid = next_nid; 1328e8c5c824SLee Schermerhorn if (delta < 0) { 1329e8c5c824SLee Schermerhorn /* 1330e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1331e8c5c824SLee Schermerhorn */ 13329a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 13336ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 13346ae11b27SLee Schermerhorn nodes_allowed); 133520a0307cSWu Fengguang continue; 1336e8c5c824SLee Schermerhorn } 13379a76db09SLee Schermerhorn } 1338e8c5c824SLee Schermerhorn if (delta > 0) { 1339e8c5c824SLee Schermerhorn /* 1340e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1341e8c5c824SLee Schermerhorn */ 1342e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 13439a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 13446ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 13456ae11b27SLee Schermerhorn nodes_allowed); 134620a0307cSWu Fengguang continue; 1347e8c5c824SLee Schermerhorn } 13489a76db09SLee Schermerhorn } 134920a0307cSWu Fengguang 135020a0307cSWu Fengguang h->surplus_huge_pages += delta; 135120a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 135220a0307cSWu Fengguang ret = 1; 135320a0307cSWu Fengguang break; 1354e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 135520a0307cSWu Fengguang 135620a0307cSWu Fengguang return ret; 135720a0307cSWu Fengguang } 135820a0307cSWu Fengguang 1359a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 13606ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 13616ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13621da177e4SLinus Torvalds { 13637893d1d5SAdam Litke unsigned long min_count, ret; 13641da177e4SLinus Torvalds 1365aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1366aa888a74SAndi Kleen return h->max_huge_pages; 1367aa888a74SAndi Kleen 13687893d1d5SAdam Litke /* 13697893d1d5SAdam Litke * Increase the pool size 13707893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 13717893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1372d1c3fb1fSNishanth Aravamudan * 1373d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1374d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1375d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1376d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1377d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 13787893d1d5SAdam Litke */ 13791da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1380a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 13816ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 13827893d1d5SAdam Litke break; 13837893d1d5SAdam Litke } 13847893d1d5SAdam Litke 1385a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 13867893d1d5SAdam Litke /* 13877893d1d5SAdam Litke * If this allocation races such that we no longer need the 13887893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 13897893d1d5SAdam Litke * and reducing the surplus. 13907893d1d5SAdam Litke */ 13917893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 13926ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 13937893d1d5SAdam Litke spin_lock(&hugetlb_lock); 13947893d1d5SAdam Litke if (!ret) 13957893d1d5SAdam Litke goto out; 13967893d1d5SAdam Litke 1397536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1398536240f2SMel Gorman if (signal_pending(current)) 1399536240f2SMel Gorman goto out; 14007893d1d5SAdam Litke } 14017893d1d5SAdam Litke 14027893d1d5SAdam Litke /* 14037893d1d5SAdam Litke * Decrease the pool size 14047893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 14057893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 14067893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 14077893d1d5SAdam Litke * to the desired size as pages become free. 1408d1c3fb1fSNishanth Aravamudan * 1409d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1410d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1411d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1412d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1413d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1414d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1415d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 14167893d1d5SAdam Litke */ 1417a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 14186b0c880dSAdam Litke min_count = max(count, min_count); 14196ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1420a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 14216ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 14221da177e4SLinus Torvalds break; 14231da177e4SLinus Torvalds } 1424a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 14256ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 14267893d1d5SAdam Litke break; 14277893d1d5SAdam Litke } 14287893d1d5SAdam Litke out: 1429a5516438SAndi Kleen ret = persistent_huge_pages(h); 14301da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 14317893d1d5SAdam Litke return ret; 14321da177e4SLinus Torvalds } 14331da177e4SLinus Torvalds 1434a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1435a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1436a3437870SNishanth Aravamudan 1437a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1438a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1439a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1440a3437870SNishanth Aravamudan 1441a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1442a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1443a3437870SNishanth Aravamudan 14449a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 14459a305230SLee Schermerhorn 14469a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1447a3437870SNishanth Aravamudan { 1448a3437870SNishanth Aravamudan int i; 14499a305230SLee Schermerhorn 1450a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 14519a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 14529a305230SLee Schermerhorn if (nidp) 14539a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1454a3437870SNishanth Aravamudan return &hstates[i]; 14559a305230SLee Schermerhorn } 14569a305230SLee Schermerhorn 14579a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1458a3437870SNishanth Aravamudan } 1459a3437870SNishanth Aravamudan 146006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1461a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1462a3437870SNishanth Aravamudan { 14639a305230SLee Schermerhorn struct hstate *h; 14649a305230SLee Schermerhorn unsigned long nr_huge_pages; 14659a305230SLee Schermerhorn int nid; 14669a305230SLee Schermerhorn 14679a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14689a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14699a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 14709a305230SLee Schermerhorn else 14719a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 14729a305230SLee Schermerhorn 14739a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1474a3437870SNishanth Aravamudan } 1475adbe8726SEric B Munson 147606808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 147706808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 147806808b08SLee Schermerhorn const char *buf, size_t len) 1479a3437870SNishanth Aravamudan { 1480a3437870SNishanth Aravamudan int err; 14819a305230SLee Schermerhorn int nid; 148206808b08SLee Schermerhorn unsigned long count; 14839a305230SLee Schermerhorn struct hstate *h; 1484bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1485a3437870SNishanth Aravamudan 148606808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 148773ae31e5SEric B Munson if (err) 1488adbe8726SEric B Munson goto out; 1489a3437870SNishanth Aravamudan 14909a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1491adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1492adbe8726SEric B Munson err = -EINVAL; 1493adbe8726SEric B Munson goto out; 1494adbe8726SEric B Munson } 1495adbe8726SEric B Munson 14969a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 14979a305230SLee Schermerhorn /* 14989a305230SLee Schermerhorn * global hstate attribute 14999a305230SLee Schermerhorn */ 15009a305230SLee Schermerhorn if (!(obey_mempolicy && 15019a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 150206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 15039a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 150406808b08SLee Schermerhorn } 15059a305230SLee Schermerhorn } else if (nodes_allowed) { 15069a305230SLee Schermerhorn /* 15079a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 15089a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 15099a305230SLee Schermerhorn */ 15109a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 15119a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 15129a305230SLee Schermerhorn } else 15139a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 15149a305230SLee Schermerhorn 151506808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1516a3437870SNishanth Aravamudan 15179b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 151806808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 151906808b08SLee Schermerhorn 152006808b08SLee Schermerhorn return len; 1521adbe8726SEric B Munson out: 1522adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1523adbe8726SEric B Munson return err; 152406808b08SLee Schermerhorn } 152506808b08SLee Schermerhorn 152606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 152706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 152806808b08SLee Schermerhorn { 152906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 153006808b08SLee Schermerhorn } 153106808b08SLee Schermerhorn 153206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 153306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 153406808b08SLee Schermerhorn { 153506808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1536a3437870SNishanth Aravamudan } 1537a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1538a3437870SNishanth Aravamudan 153906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 154006808b08SLee Schermerhorn 154106808b08SLee Schermerhorn /* 154206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 154306808b08SLee Schermerhorn * huge page alloc/free. 154406808b08SLee Schermerhorn */ 154506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 154606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 154706808b08SLee Schermerhorn { 154806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 154906808b08SLee Schermerhorn } 155006808b08SLee Schermerhorn 155106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 155206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 155306808b08SLee Schermerhorn { 155406808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 155506808b08SLee Schermerhorn } 155606808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 155706808b08SLee Schermerhorn #endif 155806808b08SLee Schermerhorn 155906808b08SLee Schermerhorn 1560a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1561a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1562a3437870SNishanth Aravamudan { 15639a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1564a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1565a3437870SNishanth Aravamudan } 1566adbe8726SEric B Munson 1567a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1568a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1569a3437870SNishanth Aravamudan { 1570a3437870SNishanth Aravamudan int err; 1571a3437870SNishanth Aravamudan unsigned long input; 15729a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1573a3437870SNishanth Aravamudan 1574adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1575adbe8726SEric B Munson return -EINVAL; 1576adbe8726SEric B Munson 1577a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1578a3437870SNishanth Aravamudan if (err) 157973ae31e5SEric B Munson return err; 1580a3437870SNishanth Aravamudan 1581a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1582a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1583a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1584a3437870SNishanth Aravamudan 1585a3437870SNishanth Aravamudan return count; 1586a3437870SNishanth Aravamudan } 1587a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1588a3437870SNishanth Aravamudan 1589a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1590a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1591a3437870SNishanth Aravamudan { 15929a305230SLee Schermerhorn struct hstate *h; 15939a305230SLee Schermerhorn unsigned long free_huge_pages; 15949a305230SLee Schermerhorn int nid; 15959a305230SLee Schermerhorn 15969a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15979a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15989a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 15999a305230SLee Schermerhorn else 16009a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 16019a305230SLee Schermerhorn 16029a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1603a3437870SNishanth Aravamudan } 1604a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1605a3437870SNishanth Aravamudan 1606a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1607a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1608a3437870SNishanth Aravamudan { 16099a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1610a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1611a3437870SNishanth Aravamudan } 1612a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1613a3437870SNishanth Aravamudan 1614a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1615a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1616a3437870SNishanth Aravamudan { 16179a305230SLee Schermerhorn struct hstate *h; 16189a305230SLee Schermerhorn unsigned long surplus_huge_pages; 16199a305230SLee Schermerhorn int nid; 16209a305230SLee Schermerhorn 16219a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16229a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16239a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 16249a305230SLee Schermerhorn else 16259a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 16269a305230SLee Schermerhorn 16279a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1628a3437870SNishanth Aravamudan } 1629a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1630a3437870SNishanth Aravamudan 1631a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1632a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1633a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1634a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1635a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1636a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 163706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 163806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 163906808b08SLee Schermerhorn #endif 1640a3437870SNishanth Aravamudan NULL, 1641a3437870SNishanth Aravamudan }; 1642a3437870SNishanth Aravamudan 1643a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1644a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1645a3437870SNishanth Aravamudan }; 1646a3437870SNishanth Aravamudan 1647094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 16489a305230SLee Schermerhorn struct kobject **hstate_kobjs, 16499a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1650a3437870SNishanth Aravamudan { 1651a3437870SNishanth Aravamudan int retval; 1652972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1653a3437870SNishanth Aravamudan 16549a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 16559a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1656a3437870SNishanth Aravamudan return -ENOMEM; 1657a3437870SNishanth Aravamudan 16589a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1659a3437870SNishanth Aravamudan if (retval) 16609a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1661a3437870SNishanth Aravamudan 1662a3437870SNishanth Aravamudan return retval; 1663a3437870SNishanth Aravamudan } 1664a3437870SNishanth Aravamudan 1665a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1666a3437870SNishanth Aravamudan { 1667a3437870SNishanth Aravamudan struct hstate *h; 1668a3437870SNishanth Aravamudan int err; 1669a3437870SNishanth Aravamudan 1670a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1671a3437870SNishanth Aravamudan if (!hugepages_kobj) 1672a3437870SNishanth Aravamudan return; 1673a3437870SNishanth Aravamudan 1674a3437870SNishanth Aravamudan for_each_hstate(h) { 16759a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 16769a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1677a3437870SNishanth Aravamudan if (err) 1678a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1679a3437870SNishanth Aravamudan h->name); 1680a3437870SNishanth Aravamudan } 1681a3437870SNishanth Aravamudan } 1682a3437870SNishanth Aravamudan 16839a305230SLee Schermerhorn #ifdef CONFIG_NUMA 16849a305230SLee Schermerhorn 16859a305230SLee Schermerhorn /* 16869a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 168710fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 168810fbcf4cSKay Sievers * index of a node device or _hstate == node id. 168910fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 16909a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 16919a305230SLee Schermerhorn */ 16929a305230SLee Schermerhorn struct node_hstate { 16939a305230SLee Schermerhorn struct kobject *hugepages_kobj; 16949a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 16959a305230SLee Schermerhorn }; 16969a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 16979a305230SLee Schermerhorn 16989a305230SLee Schermerhorn /* 169910fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 17009a305230SLee Schermerhorn */ 17019a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 17029a305230SLee Schermerhorn &nr_hugepages_attr.attr, 17039a305230SLee Schermerhorn &free_hugepages_attr.attr, 17049a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 17059a305230SLee Schermerhorn NULL, 17069a305230SLee Schermerhorn }; 17079a305230SLee Schermerhorn 17089a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 17099a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 17109a305230SLee Schermerhorn }; 17119a305230SLee Schermerhorn 17129a305230SLee Schermerhorn /* 171310fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 17149a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 17159a305230SLee Schermerhorn */ 17169a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17179a305230SLee Schermerhorn { 17189a305230SLee Schermerhorn int nid; 17199a305230SLee Schermerhorn 17209a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 17219a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 17229a305230SLee Schermerhorn int i; 17239a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 17249a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 17259a305230SLee Schermerhorn if (nidp) 17269a305230SLee Schermerhorn *nidp = nid; 17279a305230SLee Schermerhorn return &hstates[i]; 17289a305230SLee Schermerhorn } 17299a305230SLee Schermerhorn } 17309a305230SLee Schermerhorn 17319a305230SLee Schermerhorn BUG(); 17329a305230SLee Schermerhorn return NULL; 17339a305230SLee Schermerhorn } 17349a305230SLee Schermerhorn 17359a305230SLee Schermerhorn /* 173610fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 17379a305230SLee Schermerhorn * No-op if no hstate attributes attached. 17389a305230SLee Schermerhorn */ 17399a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 17409a305230SLee Schermerhorn { 17419a305230SLee Schermerhorn struct hstate *h; 174210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 17439a305230SLee Schermerhorn 17449a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 17459b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 17469a305230SLee Schermerhorn 1747972dc4deSAneesh Kumar K.V for_each_hstate(h) { 1748972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 1749972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 1750972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 1751972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 1752972dc4deSAneesh Kumar K.V } 17539a305230SLee Schermerhorn } 17549a305230SLee Schermerhorn 17559a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 17569a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 17579a305230SLee Schermerhorn } 17589a305230SLee Schermerhorn 17599a305230SLee Schermerhorn /* 176010fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 17619a305230SLee Schermerhorn * that have them. 17629a305230SLee Schermerhorn */ 17639a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 17649a305230SLee Schermerhorn { 17659a305230SLee Schermerhorn int nid; 17669a305230SLee Schermerhorn 17679a305230SLee Schermerhorn /* 176810fbcf4cSKay Sievers * disable node device registrations. 17699a305230SLee Schermerhorn */ 17709a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 17719a305230SLee Schermerhorn 17729a305230SLee Schermerhorn /* 17739a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 17749a305230SLee Schermerhorn */ 17759a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 17769a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 17779a305230SLee Schermerhorn } 17789a305230SLee Schermerhorn 17799a305230SLee Schermerhorn /* 178010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 17819a305230SLee Schermerhorn * No-op if attributes already registered. 17829a305230SLee Schermerhorn */ 17839a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 17849a305230SLee Schermerhorn { 17859a305230SLee Schermerhorn struct hstate *h; 178610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 17879a305230SLee Schermerhorn int err; 17889a305230SLee Schermerhorn 17899a305230SLee Schermerhorn if (nhs->hugepages_kobj) 17909a305230SLee Schermerhorn return; /* already allocated */ 17919a305230SLee Schermerhorn 17929a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 179310fbcf4cSKay Sievers &node->dev.kobj); 17949a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 17959a305230SLee Schermerhorn return; 17969a305230SLee Schermerhorn 17979a305230SLee Schermerhorn for_each_hstate(h) { 17989a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 17999a305230SLee Schermerhorn nhs->hstate_kobjs, 18009a305230SLee Schermerhorn &per_node_hstate_attr_group); 18019a305230SLee Schermerhorn if (err) { 18029a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 18039a305230SLee Schermerhorn " for node %d\n", 180410fbcf4cSKay Sievers h->name, node->dev.id); 18059a305230SLee Schermerhorn hugetlb_unregister_node(node); 18069a305230SLee Schermerhorn break; 18079a305230SLee Schermerhorn } 18089a305230SLee Schermerhorn } 18099a305230SLee Schermerhorn } 18109a305230SLee Schermerhorn 18119a305230SLee Schermerhorn /* 18129b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 181310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 181410fbcf4cSKay Sievers * registered their associated device by this time. 18159a305230SLee Schermerhorn */ 18169a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 18179a305230SLee Schermerhorn { 18189a305230SLee Schermerhorn int nid; 18199a305230SLee Schermerhorn 18209b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 18219a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 182210fbcf4cSKay Sievers if (node->dev.id == nid) 18239a305230SLee Schermerhorn hugetlb_register_node(node); 18249a305230SLee Schermerhorn } 18259a305230SLee Schermerhorn 18269a305230SLee Schermerhorn /* 182710fbcf4cSKay Sievers * Let the node device driver know we're here so it can 18289a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 18299a305230SLee Schermerhorn */ 18309a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 18319a305230SLee Schermerhorn hugetlb_unregister_node); 18329a305230SLee Schermerhorn } 18339a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 18349a305230SLee Schermerhorn 18359a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 18369a305230SLee Schermerhorn { 18379a305230SLee Schermerhorn BUG(); 18389a305230SLee Schermerhorn if (nidp) 18399a305230SLee Schermerhorn *nidp = -1; 18409a305230SLee Schermerhorn return NULL; 18419a305230SLee Schermerhorn } 18429a305230SLee Schermerhorn 18439a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 18449a305230SLee Schermerhorn 18459a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 18469a305230SLee Schermerhorn 18479a305230SLee Schermerhorn #endif 18489a305230SLee Schermerhorn 1849a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1850a3437870SNishanth Aravamudan { 1851a3437870SNishanth Aravamudan struct hstate *h; 1852a3437870SNishanth Aravamudan 18539a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 18549a305230SLee Schermerhorn 1855a3437870SNishanth Aravamudan for_each_hstate(h) { 1856972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 1857a3437870SNishanth Aravamudan } 1858a3437870SNishanth Aravamudan 1859a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1860a3437870SNishanth Aravamudan } 1861a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1862a3437870SNishanth Aravamudan 1863a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1864a3437870SNishanth Aravamudan { 18650ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 18660ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 18670ef89d25SBenjamin Herrenschmidt * there is no such support 18680ef89d25SBenjamin Herrenschmidt */ 18690ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 18700ef89d25SBenjamin Herrenschmidt return 0; 1871a3437870SNishanth Aravamudan 1872e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1873e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1874e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1875a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1876a3437870SNishanth Aravamudan } 1877972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1878e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1879e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1880a3437870SNishanth Aravamudan 1881a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1882a3437870SNishanth Aravamudan 1883aa888a74SAndi Kleen gather_bootmem_prealloc(); 1884aa888a74SAndi Kleen 1885a3437870SNishanth Aravamudan report_hugepages(); 1886a3437870SNishanth Aravamudan 1887a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1888a3437870SNishanth Aravamudan 18899a305230SLee Schermerhorn hugetlb_register_all_nodes(); 18909a305230SLee Schermerhorn 1891a3437870SNishanth Aravamudan return 0; 1892a3437870SNishanth Aravamudan } 1893a3437870SNishanth Aravamudan module_init(hugetlb_init); 1894a3437870SNishanth Aravamudan 1895a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1896a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1897a3437870SNishanth Aravamudan { 1898a3437870SNishanth Aravamudan struct hstate *h; 18998faa8b07SAndi Kleen unsigned long i; 19008faa8b07SAndi Kleen 1901a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1902a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1903a3437870SNishanth Aravamudan return; 1904a3437870SNishanth Aravamudan } 190547d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 1906a3437870SNishanth Aravamudan BUG_ON(order == 0); 190747d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 1908a3437870SNishanth Aravamudan h->order = order; 1909a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 19108faa8b07SAndi Kleen h->nr_huge_pages = 0; 19118faa8b07SAndi Kleen h->free_huge_pages = 0; 19128faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 19138faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 19140edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 19159b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 19169b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1917a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1918a3437870SNishanth Aravamudan huge_page_size(h)/1024); 19198faa8b07SAndi Kleen 1920a3437870SNishanth Aravamudan parsed_hstate = h; 1921a3437870SNishanth Aravamudan } 1922a3437870SNishanth Aravamudan 1923e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1924a3437870SNishanth Aravamudan { 1925a3437870SNishanth Aravamudan unsigned long *mhp; 19268faa8b07SAndi Kleen static unsigned long *last_mhp; 1927a3437870SNishanth Aravamudan 1928a3437870SNishanth Aravamudan /* 192947d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 1930a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1931a3437870SNishanth Aravamudan */ 193247d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 1933a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1934a3437870SNishanth Aravamudan else 1935a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1936a3437870SNishanth Aravamudan 19378faa8b07SAndi Kleen if (mhp == last_mhp) { 19388faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 19398faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 19408faa8b07SAndi Kleen return 1; 19418faa8b07SAndi Kleen } 19428faa8b07SAndi Kleen 1943a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1944a3437870SNishanth Aravamudan *mhp = 0; 1945a3437870SNishanth Aravamudan 19468faa8b07SAndi Kleen /* 19478faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 19488faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 19498faa8b07SAndi Kleen * use the bootmem allocator. 19508faa8b07SAndi Kleen */ 195147d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 19528faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 19538faa8b07SAndi Kleen 19548faa8b07SAndi Kleen last_mhp = mhp; 19558faa8b07SAndi Kleen 1956a3437870SNishanth Aravamudan return 1; 1957a3437870SNishanth Aravamudan } 1958e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1959e11bfbfcSNick Piggin 1960e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1961e11bfbfcSNick Piggin { 1962e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1963e11bfbfcSNick Piggin return 1; 1964e11bfbfcSNick Piggin } 1965e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1966a3437870SNishanth Aravamudan 19678a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 19688a213460SNishanth Aravamudan { 19698a213460SNishanth Aravamudan int node; 19708a213460SNishanth Aravamudan unsigned int nr = 0; 19718a213460SNishanth Aravamudan 19728a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 19738a213460SNishanth Aravamudan nr += array[node]; 19748a213460SNishanth Aravamudan 19758a213460SNishanth Aravamudan return nr; 19768a213460SNishanth Aravamudan } 19778a213460SNishanth Aravamudan 19788a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 197906808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 198006808b08SLee Schermerhorn struct ctl_table *table, int write, 198106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 19821da177e4SLinus Torvalds { 1983e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1984e5ff2159SAndi Kleen unsigned long tmp; 198508d4a246SMichal Hocko int ret; 1986e5ff2159SAndi Kleen 1987e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1988e5ff2159SAndi Kleen 1989adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 1990adbe8726SEric B Munson return -EINVAL; 1991adbe8726SEric B Munson 1992e5ff2159SAndi Kleen table->data = &tmp; 1993e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 199408d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 199508d4a246SMichal Hocko if (ret) 199608d4a246SMichal Hocko goto out; 1997e5ff2159SAndi Kleen 199806808b08SLee Schermerhorn if (write) { 1999bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2000bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 200106808b08SLee Schermerhorn if (!(obey_mempolicy && 200206808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 200306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 200406808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 200506808b08SLee Schermerhorn } 200606808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 200706808b08SLee Schermerhorn 200806808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 200906808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 201006808b08SLee Schermerhorn } 201108d4a246SMichal Hocko out: 201208d4a246SMichal Hocko return ret; 20131da177e4SLinus Torvalds } 2014396faf03SMel Gorman 201506808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 201606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 201706808b08SLee Schermerhorn { 201806808b08SLee Schermerhorn 201906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 202006808b08SLee Schermerhorn buffer, length, ppos); 202106808b08SLee Schermerhorn } 202206808b08SLee Schermerhorn 202306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 202406808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 202506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 202606808b08SLee Schermerhorn { 202706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 202806808b08SLee Schermerhorn buffer, length, ppos); 202906808b08SLee Schermerhorn } 203006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 203106808b08SLee Schermerhorn 2032396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 20338d65af78SAlexey Dobriyan void __user *buffer, 2034396faf03SMel Gorman size_t *length, loff_t *ppos) 2035396faf03SMel Gorman { 20368d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 2037396faf03SMel Gorman if (hugepages_treat_as_movable) 2038396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 2039396faf03SMel Gorman else 2040396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 2041396faf03SMel Gorman return 0; 2042396faf03SMel Gorman } 2043396faf03SMel Gorman 2044a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 20458d65af78SAlexey Dobriyan void __user *buffer, 2046a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2047a3d0c6aaSNishanth Aravamudan { 2048a5516438SAndi Kleen struct hstate *h = &default_hstate; 2049e5ff2159SAndi Kleen unsigned long tmp; 205008d4a246SMichal Hocko int ret; 2051e5ff2159SAndi Kleen 2052e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2053e5ff2159SAndi Kleen 2054adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2055adbe8726SEric B Munson return -EINVAL; 2056adbe8726SEric B Munson 2057e5ff2159SAndi Kleen table->data = &tmp; 2058e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 205908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 206008d4a246SMichal Hocko if (ret) 206108d4a246SMichal Hocko goto out; 2062e5ff2159SAndi Kleen 2063e5ff2159SAndi Kleen if (write) { 2064064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2065e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2066a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2067e5ff2159SAndi Kleen } 206808d4a246SMichal Hocko out: 206908d4a246SMichal Hocko return ret; 2070a3d0c6aaSNishanth Aravamudan } 2071a3d0c6aaSNishanth Aravamudan 20721da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 20731da177e4SLinus Torvalds 2074e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 20751da177e4SLinus Torvalds { 2076a5516438SAndi Kleen struct hstate *h = &default_hstate; 2077e1759c21SAlexey Dobriyan seq_printf(m, 20781da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 20791da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2080b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 20817893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 20824f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2083a5516438SAndi Kleen h->nr_huge_pages, 2084a5516438SAndi Kleen h->free_huge_pages, 2085a5516438SAndi Kleen h->resv_huge_pages, 2086a5516438SAndi Kleen h->surplus_huge_pages, 2087a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 20881da177e4SLinus Torvalds } 20891da177e4SLinus Torvalds 20901da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 20911da177e4SLinus Torvalds { 2092a5516438SAndi Kleen struct hstate *h = &default_hstate; 20931da177e4SLinus Torvalds return sprintf(buf, 20941da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2095a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2096a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2097a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2098a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2099a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 21001da177e4SLinus Torvalds } 21011da177e4SLinus Torvalds 21021da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 21031da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 21041da177e4SLinus Torvalds { 2105a5516438SAndi Kleen struct hstate *h = &default_hstate; 2106a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 21071da177e4SLinus Torvalds } 21081da177e4SLinus Torvalds 2109a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2110fc1b8a73SMel Gorman { 2111fc1b8a73SMel Gorman int ret = -ENOMEM; 2112fc1b8a73SMel Gorman 2113fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2114fc1b8a73SMel Gorman /* 2115fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2116fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2117fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2118fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2119fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2120fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2121fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2122fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2123fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2124fc1b8a73SMel Gorman * 2125fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2126fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2127fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2128fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2129fc1b8a73SMel Gorman * semantics that cpuset has. 2130fc1b8a73SMel Gorman */ 2131fc1b8a73SMel Gorman if (delta > 0) { 2132a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2133fc1b8a73SMel Gorman goto out; 2134fc1b8a73SMel Gorman 2135a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2136a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2137fc1b8a73SMel Gorman goto out; 2138fc1b8a73SMel Gorman } 2139fc1b8a73SMel Gorman } 2140fc1b8a73SMel Gorman 2141fc1b8a73SMel Gorman ret = 0; 2142fc1b8a73SMel Gorman if (delta < 0) 2143a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2144fc1b8a73SMel Gorman 2145fc1b8a73SMel Gorman out: 2146fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2147fc1b8a73SMel Gorman return ret; 2148fc1b8a73SMel Gorman } 2149fc1b8a73SMel Gorman 215084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 215184afd99bSAndy Whitcroft { 215284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 215384afd99bSAndy Whitcroft 215484afd99bSAndy Whitcroft /* 215584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 215684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 215784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 215825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 215984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 216084afd99bSAndy Whitcroft * new reference here without additional locking. 216184afd99bSAndy Whitcroft */ 216284afd99bSAndy Whitcroft if (reservations) 216384afd99bSAndy Whitcroft kref_get(&reservations->refs); 216484afd99bSAndy Whitcroft } 216584afd99bSAndy Whitcroft 2166c50ac050SDave Hansen static void resv_map_put(struct vm_area_struct *vma) 2167c50ac050SDave Hansen { 2168c50ac050SDave Hansen struct resv_map *reservations = vma_resv_map(vma); 2169c50ac050SDave Hansen 2170c50ac050SDave Hansen if (!reservations) 2171c50ac050SDave Hansen return; 2172c50ac050SDave Hansen kref_put(&reservations->refs, resv_map_release); 2173c50ac050SDave Hansen } 2174c50ac050SDave Hansen 2175a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2176a1e78772SMel Gorman { 2177a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 217884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 217990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 218084afd99bSAndy Whitcroft unsigned long reserve; 218184afd99bSAndy Whitcroft unsigned long start; 218284afd99bSAndy Whitcroft unsigned long end; 218384afd99bSAndy Whitcroft 218484afd99bSAndy Whitcroft if (reservations) { 2185a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2186a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 218784afd99bSAndy Whitcroft 218884afd99bSAndy Whitcroft reserve = (end - start) - 218984afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 219084afd99bSAndy Whitcroft 2191c50ac050SDave Hansen resv_map_put(vma); 219284afd99bSAndy Whitcroft 21937251ff78SAdam Litke if (reserve) { 2194a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 219590481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 21967251ff78SAdam Litke } 2197a1e78772SMel Gorman } 219884afd99bSAndy Whitcroft } 2199a1e78772SMel Gorman 22001da177e4SLinus Torvalds /* 22011da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 22021da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 22031da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 22041da177e4SLinus Torvalds * this far. 22051da177e4SLinus Torvalds */ 2206d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 22071da177e4SLinus Torvalds { 22081da177e4SLinus Torvalds BUG(); 2209d0217ac0SNick Piggin return 0; 22101da177e4SLinus Torvalds } 22111da177e4SLinus Torvalds 2212f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2213d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 221484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2215a1e78772SMel Gorman .close = hugetlb_vm_op_close, 22161da177e4SLinus Torvalds }; 22171da177e4SLinus Torvalds 22181e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 22191e8f889bSDavid Gibson int writable) 222063551ae0SDavid Gibson { 222163551ae0SDavid Gibson pte_t entry; 222263551ae0SDavid Gibson 22231e8f889bSDavid Gibson if (writable) { 222463551ae0SDavid Gibson entry = 222563551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 222663551ae0SDavid Gibson } else { 22277f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 222863551ae0SDavid Gibson } 222963551ae0SDavid Gibson entry = pte_mkyoung(entry); 223063551ae0SDavid Gibson entry = pte_mkhuge(entry); 2231d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 223263551ae0SDavid Gibson 223363551ae0SDavid Gibson return entry; 223463551ae0SDavid Gibson } 223563551ae0SDavid Gibson 22361e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 22371e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 22381e8f889bSDavid Gibson { 22391e8f889bSDavid Gibson pte_t entry; 22401e8f889bSDavid Gibson 22417f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 224232f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 22434b3073e1SRussell King update_mmu_cache(vma, address, ptep); 22441e8f889bSDavid Gibson } 22451e8f889bSDavid Gibson 22461e8f889bSDavid Gibson 224763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 224863551ae0SDavid Gibson struct vm_area_struct *vma) 224963551ae0SDavid Gibson { 225063551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 225163551ae0SDavid Gibson struct page *ptepage; 22521c59827dSHugh Dickins unsigned long addr; 22531e8f889bSDavid Gibson int cow; 2254a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2255a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 22561e8f889bSDavid Gibson 22571e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 225863551ae0SDavid Gibson 2259a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2260c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2261c74df32cSHugh Dickins if (!src_pte) 2262c74df32cSHugh Dickins continue; 2263a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 226463551ae0SDavid Gibson if (!dst_pte) 226563551ae0SDavid Gibson goto nomem; 2266c5c99429SLarry Woodman 2267c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2268c5c99429SLarry Woodman if (dst_pte == src_pte) 2269c5c99429SLarry Woodman continue; 2270c5c99429SLarry Woodman 2271c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 227246478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 22737f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 22741e8f889bSDavid Gibson if (cow) 22757f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 22767f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 227763551ae0SDavid Gibson ptepage = pte_page(entry); 227863551ae0SDavid Gibson get_page(ptepage); 22790fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 228063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 22811c59827dSHugh Dickins } 22821c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2283c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 228463551ae0SDavid Gibson } 228563551ae0SDavid Gibson return 0; 228663551ae0SDavid Gibson 228763551ae0SDavid Gibson nomem: 228863551ae0SDavid Gibson return -ENOMEM; 228963551ae0SDavid Gibson } 229063551ae0SDavid Gibson 2291290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2292290408d4SNaoya Horiguchi { 2293290408d4SNaoya Horiguchi swp_entry_t swp; 2294290408d4SNaoya Horiguchi 2295290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2296290408d4SNaoya Horiguchi return 0; 2297290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 229832f84528SChris Forbes if (non_swap_entry(swp) && is_migration_entry(swp)) 2299290408d4SNaoya Horiguchi return 1; 230032f84528SChris Forbes else 2301290408d4SNaoya Horiguchi return 0; 2302290408d4SNaoya Horiguchi } 2303290408d4SNaoya Horiguchi 2304fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2305fd6a03edSNaoya Horiguchi { 2306fd6a03edSNaoya Horiguchi swp_entry_t swp; 2307fd6a03edSNaoya Horiguchi 2308fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2309fd6a03edSNaoya Horiguchi return 0; 2310fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 231132f84528SChris Forbes if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2312fd6a03edSNaoya Horiguchi return 1; 231332f84528SChris Forbes else 2314fd6a03edSNaoya Horiguchi return 0; 2315fd6a03edSNaoya Horiguchi } 2316fd6a03edSNaoya Horiguchi 231724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 231824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 231924669e58SAneesh Kumar K.V struct page *ref_page) 232063551ae0SDavid Gibson { 232124669e58SAneesh Kumar K.V int force_flush = 0; 232263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 232363551ae0SDavid Gibson unsigned long address; 2324c7546f8fSDavid Gibson pte_t *ptep; 232563551ae0SDavid Gibson pte_t pte; 232663551ae0SDavid Gibson struct page *page; 2327a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2328a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2329a5516438SAndi Kleen 233063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2331a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2332a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 233363551ae0SDavid Gibson 233424669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 2335cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 233624669e58SAneesh Kumar K.V again: 2337508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2338a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2339c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2340c7546f8fSDavid Gibson if (!ptep) 2341c7546f8fSDavid Gibson continue; 2342c7546f8fSDavid Gibson 234339dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 234439dde65cSChen, Kenneth W continue; 234539dde65cSChen, Kenneth W 23466629326bSHillf Danton pte = huge_ptep_get(ptep); 23476629326bSHillf Danton if (huge_pte_none(pte)) 23486629326bSHillf Danton continue; 23496629326bSHillf Danton 23506629326bSHillf Danton /* 23516629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 23526629326bSHillf Danton */ 23536629326bSHillf Danton if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 23546629326bSHillf Danton continue; 23556629326bSHillf Danton 23566629326bSHillf Danton page = pte_page(pte); 235704f2cbe3SMel Gorman /* 235804f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 235904f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 236004f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 236104f2cbe3SMel Gorman */ 236204f2cbe3SMel Gorman if (ref_page) { 236304f2cbe3SMel Gorman if (page != ref_page) 236404f2cbe3SMel Gorman continue; 236504f2cbe3SMel Gorman 236604f2cbe3SMel Gorman /* 236704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 236804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 236904f2cbe3SMel Gorman * looking like data was lost 237004f2cbe3SMel Gorman */ 237104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 237204f2cbe3SMel Gorman } 237304f2cbe3SMel Gorman 2374c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 237524669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 23766649a386SKen Chen if (pte_dirty(pte)) 23776649a386SKen Chen set_page_dirty(page); 23789e81130bSHillf Danton 237924669e58SAneesh Kumar K.V page_remove_rmap(page); 238024669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 238124669e58SAneesh Kumar K.V if (force_flush) 238224669e58SAneesh Kumar K.V break; 23839e81130bSHillf Danton /* Bail out after unmapping reference page if supplied */ 23849e81130bSHillf Danton if (ref_page) 23859e81130bSHillf Danton break; 238663551ae0SDavid Gibson } 2387cd2934a3SAl Viro spin_unlock(&mm->page_table_lock); 238824669e58SAneesh Kumar K.V /* 238924669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 239024669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 239124669e58SAneesh Kumar K.V * and page-free while holding it. 239224669e58SAneesh Kumar K.V */ 239324669e58SAneesh Kumar K.V if (force_flush) { 239424669e58SAneesh Kumar K.V force_flush = 0; 239524669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 239624669e58SAneesh Kumar K.V if (address < end && !ref_page) 239724669e58SAneesh Kumar K.V goto again; 2398fe1668aeSChen, Kenneth W } 239924669e58SAneesh Kumar K.V mmu_notifier_invalidate_range_end(mm, start, end); 240024669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 24011da177e4SLinus Torvalds } 240263551ae0SDavid Gibson 2403502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 240404f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2405502717f4SChen, Kenneth W { 240624669e58SAneesh Kumar K.V struct mm_struct *mm; 240724669e58SAneesh Kumar K.V struct mmu_gather tlb; 240824669e58SAneesh Kumar K.V 240924669e58SAneesh Kumar K.V mm = vma->vm_mm; 241024669e58SAneesh Kumar K.V 241124669e58SAneesh Kumar K.V tlb_gather_mmu(&tlb, mm, 0); 241224669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 241324669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2414502717f4SChen, Kenneth W } 2415502717f4SChen, Kenneth W 241604f2cbe3SMel Gorman /* 241704f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 241804f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 241904f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 242004f2cbe3SMel Gorman * same region. 242104f2cbe3SMel Gorman */ 24222a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 24232a4b3dedSHarvey Harrison struct page *page, unsigned long address) 242404f2cbe3SMel Gorman { 24257526674dSAdam Litke struct hstate *h = hstate_vma(vma); 242604f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 242704f2cbe3SMel Gorman struct address_space *mapping; 242804f2cbe3SMel Gorman struct prio_tree_iter iter; 242904f2cbe3SMel Gorman pgoff_t pgoff; 243004f2cbe3SMel Gorman 243104f2cbe3SMel Gorman /* 243204f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 243304f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 243404f2cbe3SMel Gorman */ 24357526674dSAdam Litke address = address & huge_page_mask(h); 24360c176d52SHillf Danton pgoff = vma_hugecache_offset(h, vma, address); 243790481622SDavid Gibson mapping = vma->vm_file->f_dentry->d_inode->i_mapping; 243804f2cbe3SMel Gorman 24394eb2b1dcSMel Gorman /* 24404eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 24414eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 24424eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 24434eb2b1dcSMel Gorman */ 24443d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 244504f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 244604f2cbe3SMel Gorman /* Do not unmap the current VMA */ 244704f2cbe3SMel Gorman if (iter_vma == vma) 244804f2cbe3SMel Gorman continue; 244904f2cbe3SMel Gorman 245004f2cbe3SMel Gorman /* 245104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 245204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 245304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 245404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 245504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 245604f2cbe3SMel Gorman */ 245704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 245824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 245924669e58SAneesh Kumar K.V address + huge_page_size(h), page); 246004f2cbe3SMel Gorman } 24613d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 246204f2cbe3SMel Gorman 246304f2cbe3SMel Gorman return 1; 246404f2cbe3SMel Gorman } 246504f2cbe3SMel Gorman 24660fe6e20bSNaoya Horiguchi /* 24670fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2468ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2469ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2470ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 24710fe6e20bSNaoya Horiguchi */ 24721e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 247304f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 247404f2cbe3SMel Gorman struct page *pagecache_page) 24751e8f889bSDavid Gibson { 2476a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 24771e8f889bSDavid Gibson struct page *old_page, *new_page; 247879ac6ba4SDavid Gibson int avoidcopy; 247904f2cbe3SMel Gorman int outside_reserve = 0; 24801e8f889bSDavid Gibson 24811e8f889bSDavid Gibson old_page = pte_page(pte); 24821e8f889bSDavid Gibson 248304f2cbe3SMel Gorman retry_avoidcopy: 24841e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 24851e8f889bSDavid Gibson * and just make the page writable */ 24860fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 24871e8f889bSDavid Gibson if (avoidcopy) { 24880fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 24890fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 24901e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 249183c54070SNick Piggin return 0; 24921e8f889bSDavid Gibson } 24931e8f889bSDavid Gibson 249404f2cbe3SMel Gorman /* 249504f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 249604f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 249704f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 249804f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 249904f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 250004f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 250104f2cbe3SMel Gorman * of the full address range. 250204f2cbe3SMel Gorman */ 2503f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 250404f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 250504f2cbe3SMel Gorman old_page != pagecache_page) 250604f2cbe3SMel Gorman outside_reserve = 1; 250704f2cbe3SMel Gorman 25081e8f889bSDavid Gibson page_cache_get(old_page); 2509b76c8cfbSLarry Woodman 2510b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2511b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 251204f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 25131e8f889bSDavid Gibson 25142fc39cecSAdam Litke if (IS_ERR(new_page)) { 251576dcee75SAneesh Kumar K.V long err = PTR_ERR(new_page); 25161e8f889bSDavid Gibson page_cache_release(old_page); 251704f2cbe3SMel Gorman 251804f2cbe3SMel Gorman /* 251904f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 252004f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 252104f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 252204f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 252304f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 252404f2cbe3SMel Gorman */ 252504f2cbe3SMel Gorman if (outside_reserve) { 252604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 252704f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 252804f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2529b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2530a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2531a734bcc8SHillf Danton if (likely(pte_same(huge_ptep_get(ptep), pte))) 253204f2cbe3SMel Gorman goto retry_avoidcopy; 2533a734bcc8SHillf Danton /* 2534a734bcc8SHillf Danton * race occurs while re-acquiring page_table_lock, and 2535a734bcc8SHillf Danton * our job is done. 2536a734bcc8SHillf Danton */ 2537a734bcc8SHillf Danton return 0; 253804f2cbe3SMel Gorman } 253904f2cbe3SMel Gorman WARN_ON_ONCE(1); 254004f2cbe3SMel Gorman } 254104f2cbe3SMel Gorman 2542b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2543b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 254476dcee75SAneesh Kumar K.V if (err == -ENOMEM) 254576dcee75SAneesh Kumar K.V return VM_FAULT_OOM; 254676dcee75SAneesh Kumar K.V else 254776dcee75SAneesh Kumar K.V return VM_FAULT_SIGBUS; 25481e8f889bSDavid Gibson } 25491e8f889bSDavid Gibson 25500fe6e20bSNaoya Horiguchi /* 25510fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 25520fe6e20bSNaoya Horiguchi * anon_vma prepared. 25530fe6e20bSNaoya Horiguchi */ 255444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2555ea4039a3SHillf Danton page_cache_release(new_page); 2556ea4039a3SHillf Danton page_cache_release(old_page); 255744e2aa93SDean Nelson /* Caller expects lock to be held */ 255844e2aa93SDean Nelson spin_lock(&mm->page_table_lock); 25590fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 256044e2aa93SDean Nelson } 25610fe6e20bSNaoya Horiguchi 256247ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 256347ad8475SAndrea Arcangeli pages_per_huge_page(h)); 25640ed361deSNick Piggin __SetPageUptodate(new_page); 25651e8f889bSDavid Gibson 2566b76c8cfbSLarry Woodman /* 2567b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2568b76c8cfbSLarry Woodman * before the page tables are altered 2569b76c8cfbSLarry Woodman */ 2570b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2571a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 25727f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 25731e8f889bSDavid Gibson /* Break COW */ 25743edd4fc9SDoug Doan mmu_notifier_invalidate_range_start(mm, 25753edd4fc9SDoug Doan address & huge_page_mask(h), 25763edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 25778fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 25781e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 25791e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 25800fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2581cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 25821e8f889bSDavid Gibson /* Make the old page be freed below */ 25831e8f889bSDavid Gibson new_page = old_page; 25843edd4fc9SDoug Doan mmu_notifier_invalidate_range_end(mm, 25853edd4fc9SDoug Doan address & huge_page_mask(h), 25863edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 25871e8f889bSDavid Gibson } 25881e8f889bSDavid Gibson page_cache_release(new_page); 25891e8f889bSDavid Gibson page_cache_release(old_page); 259083c54070SNick Piggin return 0; 25911e8f889bSDavid Gibson } 25921e8f889bSDavid Gibson 259304f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2594a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2595a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 259604f2cbe3SMel Gorman { 259704f2cbe3SMel Gorman struct address_space *mapping; 2598e7c4b0bfSAndy Whitcroft pgoff_t idx; 259904f2cbe3SMel Gorman 260004f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2601a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 260204f2cbe3SMel Gorman 260304f2cbe3SMel Gorman return find_lock_page(mapping, idx); 260404f2cbe3SMel Gorman } 260504f2cbe3SMel Gorman 26063ae77f43SHugh Dickins /* 26073ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 26083ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 26093ae77f43SHugh Dickins */ 26103ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 26112a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 26122a15efc9SHugh Dickins { 26132a15efc9SHugh Dickins struct address_space *mapping; 26142a15efc9SHugh Dickins pgoff_t idx; 26152a15efc9SHugh Dickins struct page *page; 26162a15efc9SHugh Dickins 26172a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 26182a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 26192a15efc9SHugh Dickins 26202a15efc9SHugh Dickins page = find_get_page(mapping, idx); 26212a15efc9SHugh Dickins if (page) 26222a15efc9SHugh Dickins put_page(page); 26232a15efc9SHugh Dickins return page != NULL; 26242a15efc9SHugh Dickins } 26252a15efc9SHugh Dickins 2626a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2627788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2628ac9b9c66SHugh Dickins { 2629a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2630ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2631409eb8c2SHillf Danton int anon_rmap = 0; 2632e7c4b0bfSAndy Whitcroft pgoff_t idx; 26334c887265SAdam Litke unsigned long size; 26344c887265SAdam Litke struct page *page; 26354c887265SAdam Litke struct address_space *mapping; 26361e8f889bSDavid Gibson pte_t new_pte; 26374c887265SAdam Litke 263804f2cbe3SMel Gorman /* 263904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 264004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 264125985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 264204f2cbe3SMel Gorman */ 264304f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 264404f2cbe3SMel Gorman printk(KERN_WARNING 264504f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 264604f2cbe3SMel Gorman current->pid); 264704f2cbe3SMel Gorman return ret; 264804f2cbe3SMel Gorman } 264904f2cbe3SMel Gorman 26504c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2651a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 26524c887265SAdam Litke 26534c887265SAdam Litke /* 26544c887265SAdam Litke * Use page lock to guard against racing truncation 26554c887265SAdam Litke * before we get page_table_lock. 26564c887265SAdam Litke */ 26576bda666aSChristoph Lameter retry: 26586bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 26596bda666aSChristoph Lameter if (!page) { 2660a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2661ebed4bfcSHugh Dickins if (idx >= size) 2662ebed4bfcSHugh Dickins goto out; 266304f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 26642fc39cecSAdam Litke if (IS_ERR(page)) { 266576dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 266676dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 266776dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 266876dcee75SAneesh Kumar K.V else 266976dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 26706bda666aSChristoph Lameter goto out; 26716bda666aSChristoph Lameter } 267247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 26730ed361deSNick Piggin __SetPageUptodate(page); 2674ac9b9c66SHugh Dickins 2675f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 26766bda666aSChristoph Lameter int err; 267745c682a6SKen Chen struct inode *inode = mapping->host; 26786bda666aSChristoph Lameter 26796bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 26806bda666aSChristoph Lameter if (err) { 26816bda666aSChristoph Lameter put_page(page); 26826bda666aSChristoph Lameter if (err == -EEXIST) 26836bda666aSChristoph Lameter goto retry; 26846bda666aSChristoph Lameter goto out; 26856bda666aSChristoph Lameter } 268645c682a6SKen Chen 268745c682a6SKen Chen spin_lock(&inode->i_lock); 2688a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 268945c682a6SKen Chen spin_unlock(&inode->i_lock); 269023be7468SMel Gorman } else { 26916bda666aSChristoph Lameter lock_page(page); 26920fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 26930fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 26940fe6e20bSNaoya Horiguchi goto backout_unlocked; 269523be7468SMel Gorman } 2696409eb8c2SHillf Danton anon_rmap = 1; 26970fe6e20bSNaoya Horiguchi } 26980fe6e20bSNaoya Horiguchi } else { 269957303d80SAndy Whitcroft /* 2700998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2701998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2702998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2703fd6a03edSNaoya Horiguchi */ 2704fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2705aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2706972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2707fd6a03edSNaoya Horiguchi goto backout_unlocked; 27086bda666aSChristoph Lameter } 2709998b4382SNaoya Horiguchi } 27101e8f889bSDavid Gibson 271157303d80SAndy Whitcroft /* 271257303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 271357303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 271457303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 271557303d80SAndy Whitcroft * the spinlock. 271657303d80SAndy Whitcroft */ 2717788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 27182b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 27192b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 27202b26736cSAndy Whitcroft goto backout_unlocked; 27212b26736cSAndy Whitcroft } 272257303d80SAndy Whitcroft 2723ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2724a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 27254c887265SAdam Litke if (idx >= size) 27264c887265SAdam Litke goto backout; 27274c887265SAdam Litke 272883c54070SNick Piggin ret = 0; 27297f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 27304c887265SAdam Litke goto backout; 27314c887265SAdam Litke 2732409eb8c2SHillf Danton if (anon_rmap) 2733409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 2734409eb8c2SHillf Danton else 2735409eb8c2SHillf Danton page_dup_rmap(page); 27361e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 27371e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 27381e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 27391e8f889bSDavid Gibson 2740788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 27411e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 274204f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 27431e8f889bSDavid Gibson } 27441e8f889bSDavid Gibson 2745ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 27464c887265SAdam Litke unlock_page(page); 27474c887265SAdam Litke out: 2748ac9b9c66SHugh Dickins return ret; 27494c887265SAdam Litke 27504c887265SAdam Litke backout: 27514c887265SAdam Litke spin_unlock(&mm->page_table_lock); 27522b26736cSAndy Whitcroft backout_unlocked: 27534c887265SAdam Litke unlock_page(page); 27544c887265SAdam Litke put_page(page); 27554c887265SAdam Litke goto out; 2756ac9b9c66SHugh Dickins } 2757ac9b9c66SHugh Dickins 275886e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2759788c7df4SHugh Dickins unsigned long address, unsigned int flags) 276086e5216fSAdam Litke { 276186e5216fSAdam Litke pte_t *ptep; 276286e5216fSAdam Litke pte_t entry; 27631e8f889bSDavid Gibson int ret; 27640fe6e20bSNaoya Horiguchi struct page *page = NULL; 276557303d80SAndy Whitcroft struct page *pagecache_page = NULL; 27663935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2767a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 276886e5216fSAdam Litke 27691e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 27701e16a539SKAMEZAWA Hiroyuki 2771fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2772fd6a03edSNaoya Horiguchi if (ptep) { 2773fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2774290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 2775290408d4SNaoya Horiguchi migration_entry_wait(mm, (pmd_t *)ptep, address); 2776290408d4SNaoya Horiguchi return 0; 2777290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2778aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2779972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2780fd6a03edSNaoya Horiguchi } 2781fd6a03edSNaoya Horiguchi 2782a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 278386e5216fSAdam Litke if (!ptep) 278486e5216fSAdam Litke return VM_FAULT_OOM; 278586e5216fSAdam Litke 27863935baa9SDavid Gibson /* 27873935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 27883935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 27893935baa9SDavid Gibson * the same page in the page cache. 27903935baa9SDavid Gibson */ 27913935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 27927f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 27937f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2794788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2795b4d1d99fSDavid Gibson goto out_mutex; 27963935baa9SDavid Gibson } 279786e5216fSAdam Litke 279883c54070SNick Piggin ret = 0; 27991e8f889bSDavid Gibson 280057303d80SAndy Whitcroft /* 280157303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 280257303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 280357303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 280457303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 280557303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 280657303d80SAndy Whitcroft * consumed. 280757303d80SAndy Whitcroft */ 2808788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 28092b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28102b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2811b4d1d99fSDavid Gibson goto out_mutex; 28122b26736cSAndy Whitcroft } 281357303d80SAndy Whitcroft 2814f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 281557303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 281657303d80SAndy Whitcroft vma, address); 281757303d80SAndy Whitcroft } 281857303d80SAndy Whitcroft 281956c9cfb1SNaoya Horiguchi /* 282056c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 282156c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 282256c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 282356c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 282456c9cfb1SNaoya Horiguchi * so no worry about deadlock. 282556c9cfb1SNaoya Horiguchi */ 28260fe6e20bSNaoya Horiguchi page = pte_page(entry); 282766aebce7SChris Metcalf get_page(page); 282856c9cfb1SNaoya Horiguchi if (page != pagecache_page) 28290fe6e20bSNaoya Horiguchi lock_page(page); 28300fe6e20bSNaoya Horiguchi 28311e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 28321e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2833b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2834b4d1d99fSDavid Gibson goto out_page_table_lock; 2835b4d1d99fSDavid Gibson 2836b4d1d99fSDavid Gibson 2837788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2838b4d1d99fSDavid Gibson if (!pte_write(entry)) { 283957303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 284057303d80SAndy Whitcroft pagecache_page); 2841b4d1d99fSDavid Gibson goto out_page_table_lock; 2842b4d1d99fSDavid Gibson } 2843b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2844b4d1d99fSDavid Gibson } 2845b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2846788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2847788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 28484b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2849b4d1d99fSDavid Gibson 2850b4d1d99fSDavid Gibson out_page_table_lock: 28511e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 285257303d80SAndy Whitcroft 285357303d80SAndy Whitcroft if (pagecache_page) { 285457303d80SAndy Whitcroft unlock_page(pagecache_page); 285557303d80SAndy Whitcroft put_page(pagecache_page); 285657303d80SAndy Whitcroft } 28571f64d69cSDean Nelson if (page != pagecache_page) 285856c9cfb1SNaoya Horiguchi unlock_page(page); 285966aebce7SChris Metcalf put_page(page); 286057303d80SAndy Whitcroft 2861b4d1d99fSDavid Gibson out_mutex: 28623935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 28631e8f889bSDavid Gibson 28641e8f889bSDavid Gibson return ret; 286586e5216fSAdam Litke } 286686e5216fSAdam Litke 2867ceb86879SAndi Kleen /* Can be overriden by architectures */ 2868ceb86879SAndi Kleen __attribute__((weak)) struct page * 2869ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2870ceb86879SAndi Kleen pud_t *pud, int write) 2871ceb86879SAndi Kleen { 2872ceb86879SAndi Kleen BUG(); 2873ceb86879SAndi Kleen return NULL; 2874ceb86879SAndi Kleen } 2875ceb86879SAndi Kleen 287663551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 287763551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 28785b23dbe8SAdam Litke unsigned long *position, int *length, int i, 28792a15efc9SHugh Dickins unsigned int flags) 288063551ae0SDavid Gibson { 2881d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2882d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 288363551ae0SDavid Gibson int remainder = *length; 2884a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 288563551ae0SDavid Gibson 28861c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 288763551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 288863551ae0SDavid Gibson pte_t *pte; 28892a15efc9SHugh Dickins int absent; 289063551ae0SDavid Gibson struct page *page; 289163551ae0SDavid Gibson 28924c887265SAdam Litke /* 28934c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 28942a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 28954c887265SAdam Litke * first, for the page indexing below to work. 28964c887265SAdam Litke */ 2897a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 28982a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 289963551ae0SDavid Gibson 29002a15efc9SHugh Dickins /* 29012a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 29023ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 29033ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 29043ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 29053ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 29062a15efc9SHugh Dickins */ 29073ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 29083ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 29092a15efc9SHugh Dickins remainder = 0; 29102a15efc9SHugh Dickins break; 29112a15efc9SHugh Dickins } 29122a15efc9SHugh Dickins 29132a15efc9SHugh Dickins if (absent || 29142a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 29154c887265SAdam Litke int ret; 29164c887265SAdam Litke 29174c887265SAdam Litke spin_unlock(&mm->page_table_lock); 29182a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 29192a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 29204c887265SAdam Litke spin_lock(&mm->page_table_lock); 2921a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 29224c887265SAdam Litke continue; 29234c887265SAdam Litke 29241c59827dSHugh Dickins remainder = 0; 29251c59827dSHugh Dickins break; 29261c59827dSHugh Dickins } 292763551ae0SDavid Gibson 2928a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 29297f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2930d5d4b0aaSChen, Kenneth W same_page: 2931d6692183SChen, Kenneth W if (pages) { 293269d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 29334b2e38adSKOSAKI Motohiro get_page(pages[i]); 2934d6692183SChen, Kenneth W } 293563551ae0SDavid Gibson 293663551ae0SDavid Gibson if (vmas) 293763551ae0SDavid Gibson vmas[i] = vma; 293863551ae0SDavid Gibson 293963551ae0SDavid Gibson vaddr += PAGE_SIZE; 2940d5d4b0aaSChen, Kenneth W ++pfn_offset; 294163551ae0SDavid Gibson --remainder; 294263551ae0SDavid Gibson ++i; 2943d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2944a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2945d5d4b0aaSChen, Kenneth W /* 2946d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2947d5d4b0aaSChen, Kenneth W * of this compound page. 2948d5d4b0aaSChen, Kenneth W */ 2949d5d4b0aaSChen, Kenneth W goto same_page; 2950d5d4b0aaSChen, Kenneth W } 295163551ae0SDavid Gibson } 29521c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 295363551ae0SDavid Gibson *length = remainder; 295463551ae0SDavid Gibson *position = vaddr; 295563551ae0SDavid Gibson 29562a15efc9SHugh Dickins return i ? i : -EFAULT; 295763551ae0SDavid Gibson } 29588f860591SZhang, Yanmin 29598f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 29608f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 29618f860591SZhang, Yanmin { 29628f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 29638f860591SZhang, Yanmin unsigned long start = address; 29648f860591SZhang, Yanmin pte_t *ptep; 29658f860591SZhang, Yanmin pte_t pte; 2966a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 29678f860591SZhang, Yanmin 29688f860591SZhang, Yanmin BUG_ON(address >= end); 29698f860591SZhang, Yanmin flush_cache_range(vma, address, end); 29708f860591SZhang, Yanmin 29713d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 29728f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2973a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 29748f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 29758f860591SZhang, Yanmin if (!ptep) 29768f860591SZhang, Yanmin continue; 297739dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 297839dde65cSChen, Kenneth W continue; 29797f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 29808f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 29818f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 29828f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 29838f860591SZhang, Yanmin } 29848f860591SZhang, Yanmin } 29858f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 29863d48ae45SPeter Zijlstra mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 29878f860591SZhang, Yanmin 29888f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 29898f860591SZhang, Yanmin } 29908f860591SZhang, Yanmin 2991a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2992a1e78772SMel Gorman long from, long to, 29935a6fe125SMel Gorman struct vm_area_struct *vma, 2994ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 2995e4e574b7SAdam Litke { 299617c9d12eSMel Gorman long ret, chg; 2997a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 299890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 2999e4e574b7SAdam Litke 3000a1e78772SMel Gorman /* 300117c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 300217c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 300390481622SDavid Gibson * without using reserves 300417c9d12eSMel Gorman */ 3005ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 300617c9d12eSMel Gorman return 0; 300717c9d12eSMel Gorman 300817c9d12eSMel Gorman /* 3009a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3010a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3011a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3012a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3013a1e78772SMel Gorman */ 3014f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 3015e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 30165a6fe125SMel Gorman else { 30175a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 30185a6fe125SMel Gorman if (!resv_map) 30195a6fe125SMel Gorman return -ENOMEM; 30205a6fe125SMel Gorman 302117c9d12eSMel Gorman chg = to - from; 302217c9d12eSMel Gorman 30235a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 30245a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 30255a6fe125SMel Gorman } 30265a6fe125SMel Gorman 3027c50ac050SDave Hansen if (chg < 0) { 3028c50ac050SDave Hansen ret = chg; 3029c50ac050SDave Hansen goto out_err; 3030c50ac050SDave Hansen } 303117c9d12eSMel Gorman 303290481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3033c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3034c50ac050SDave Hansen ret = -ENOSPC; 3035c50ac050SDave Hansen goto out_err; 3036c50ac050SDave Hansen } 303717c9d12eSMel Gorman 303817c9d12eSMel Gorman /* 303917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 304090481622SDavid Gibson * Hand the pages back to the subpool if there are not 304117c9d12eSMel Gorman */ 304217c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 304317c9d12eSMel Gorman if (ret < 0) { 304490481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3045c50ac050SDave Hansen goto out_err; 304617c9d12eSMel Gorman } 304717c9d12eSMel Gorman 304817c9d12eSMel Gorman /* 304917c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 305017c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 305117c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 305217c9d12eSMel Gorman * the reservation was and the page cache tells how much of 305317c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 305417c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 305517c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 305617c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 305717c9d12eSMel Gorman * else has to be done for private mappings here 305817c9d12eSMel Gorman */ 3059f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 306017c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 3061a43a8c39SChen, Kenneth W return 0; 3062c50ac050SDave Hansen out_err: 30634523e145SDave Hansen if (vma) 3064c50ac050SDave Hansen resv_map_put(vma); 3065c50ac050SDave Hansen return ret; 3066a43a8c39SChen, Kenneth W } 3067a43a8c39SChen, Kenneth W 3068a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3069a43a8c39SChen, Kenneth W { 3070a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 3071a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 307290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 307345c682a6SKen Chen 307445c682a6SKen Chen spin_lock(&inode->i_lock); 3075e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 307645c682a6SKen Chen spin_unlock(&inode->i_lock); 307745c682a6SKen Chen 307890481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3079a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3080a43a8c39SChen, Kenneth W } 308193f70f90SNaoya Horiguchi 3082d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3083d5bd9106SAndi Kleen 30846de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 30856de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 30866de2b1aaSNaoya Horiguchi { 30876de2b1aaSNaoya Horiguchi struct page *page; 30886de2b1aaSNaoya Horiguchi struct page *tmp; 30896de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 30906de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 30916de2b1aaSNaoya Horiguchi 30926de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 30936de2b1aaSNaoya Horiguchi if (page == hpage) 30946de2b1aaSNaoya Horiguchi return 1; 30956de2b1aaSNaoya Horiguchi return 0; 30966de2b1aaSNaoya Horiguchi } 30976de2b1aaSNaoya Horiguchi 309893f70f90SNaoya Horiguchi /* 309993f70f90SNaoya Horiguchi * This function is called from memory failure code. 310093f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 310193f70f90SNaoya Horiguchi */ 31026de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 310393f70f90SNaoya Horiguchi { 310493f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 310593f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 31066de2b1aaSNaoya Horiguchi int ret = -EBUSY; 310793f70f90SNaoya Horiguchi 310893f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 31096de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 311093f70f90SNaoya Horiguchi list_del(&hpage->lru); 31118c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 311293f70f90SNaoya Horiguchi h->free_huge_pages--; 311393f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 31146de2b1aaSNaoya Horiguchi ret = 0; 311593f70f90SNaoya Horiguchi } 31166de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 31176de2b1aaSNaoya Horiguchi return ret; 31186de2b1aaSNaoya Horiguchi } 31196de2b1aaSNaoya Horiguchi #endif 3120