11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24d6606683SLinus Torvalds 2563551ae0SDavid Gibson #include <asm/page.h> 2663551ae0SDavid Gibson #include <asm/pgtable.h> 2724669e58SAneesh Kumar K.V #include <asm/tlb.h> 2863551ae0SDavid Gibson 2924669e58SAneesh Kumar K.V #include <linux/io.h> 3063551ae0SDavid Gibson #include <linux/hugetlb.h> 319dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 329a305230SLee Schermerhorn #include <linux/node.h> 337835e98bSNick Piggin #include "internal.h" 341da177e4SLinus Torvalds 351da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 36396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 37396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 38a5516438SAndi Kleen 39c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 40e5ff2159SAndi Kleen unsigned int default_hstate_idx; 41e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 42e5ff2159SAndi Kleen 4353ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4453ba51d2SJon Tollefson 45e5ff2159SAndi Kleen /* for command line parsing */ 46e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 47e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 48e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 49e5ff2159SAndi Kleen 503935baa9SDavid Gibson /* 513935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 523935baa9SDavid Gibson */ 53c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 540bd0f9fbSEric Paris 5590481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 5690481622SDavid Gibson { 5790481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 5890481622SDavid Gibson 5990481622SDavid Gibson spin_unlock(&spool->lock); 6090481622SDavid Gibson 6190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 6290481622SDavid Gibson * remain, free the subpool the subpool remain */ 6390481622SDavid Gibson if (free) 6490481622SDavid Gibson kfree(spool); 6590481622SDavid Gibson } 6690481622SDavid Gibson 6790481622SDavid Gibson struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 6890481622SDavid Gibson { 6990481622SDavid Gibson struct hugepage_subpool *spool; 7090481622SDavid Gibson 7190481622SDavid Gibson spool = kmalloc(sizeof(*spool), GFP_KERNEL); 7290481622SDavid Gibson if (!spool) 7390481622SDavid Gibson return NULL; 7490481622SDavid Gibson 7590481622SDavid Gibson spin_lock_init(&spool->lock); 7690481622SDavid Gibson spool->count = 1; 7790481622SDavid Gibson spool->max_hpages = nr_blocks; 7890481622SDavid Gibson spool->used_hpages = 0; 7990481622SDavid Gibson 8090481622SDavid Gibson return spool; 8190481622SDavid Gibson } 8290481622SDavid Gibson 8390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 8490481622SDavid Gibson { 8590481622SDavid Gibson spin_lock(&spool->lock); 8690481622SDavid Gibson BUG_ON(!spool->count); 8790481622SDavid Gibson spool->count--; 8890481622SDavid Gibson unlock_or_release_subpool(spool); 8990481622SDavid Gibson } 9090481622SDavid Gibson 9190481622SDavid Gibson static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 9290481622SDavid Gibson long delta) 9390481622SDavid Gibson { 9490481622SDavid Gibson int ret = 0; 9590481622SDavid Gibson 9690481622SDavid Gibson if (!spool) 9790481622SDavid Gibson return 0; 9890481622SDavid Gibson 9990481622SDavid Gibson spin_lock(&spool->lock); 10090481622SDavid Gibson if ((spool->used_hpages + delta) <= spool->max_hpages) { 10190481622SDavid Gibson spool->used_hpages += delta; 10290481622SDavid Gibson } else { 10390481622SDavid Gibson ret = -ENOMEM; 10490481622SDavid Gibson } 10590481622SDavid Gibson spin_unlock(&spool->lock); 10690481622SDavid Gibson 10790481622SDavid Gibson return ret; 10890481622SDavid Gibson } 10990481622SDavid Gibson 11090481622SDavid Gibson static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 11190481622SDavid Gibson long delta) 11290481622SDavid Gibson { 11390481622SDavid Gibson if (!spool) 11490481622SDavid Gibson return; 11590481622SDavid Gibson 11690481622SDavid Gibson spin_lock(&spool->lock); 11790481622SDavid Gibson spool->used_hpages -= delta; 11890481622SDavid Gibson /* If hugetlbfs_put_super couldn't free spool due to 11990481622SDavid Gibson * an outstanding quota reference, free it now. */ 12090481622SDavid Gibson unlock_or_release_subpool(spool); 12190481622SDavid Gibson } 12290481622SDavid Gibson 12390481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 12490481622SDavid Gibson { 12590481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 12690481622SDavid Gibson } 12790481622SDavid Gibson 12890481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 12990481622SDavid Gibson { 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); 6306d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 6316d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 632aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 6330edaecfaSAneesh Kumar K.V /* remove the page from active list */ 6340edaecfaSAneesh Kumar K.V list_del(&page->lru); 635a5516438SAndi Kleen update_and_free_page(h, page); 636a5516438SAndi Kleen h->surplus_huge_pages--; 637a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 6387893d1d5SAdam Litke } else { 6395d3a551cSWill Deacon arch_clear_hugepage_flags(page); 640a5516438SAndi Kleen enqueue_huge_page(h, page); 6417893d1d5SAdam Litke } 64227a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 64390481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 64427a85ef1SDavid Gibson } 64527a85ef1SDavid Gibson 646a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 647b7ba30c6SAndi Kleen { 6480edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 649b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 650b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 6519dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 652a5516438SAndi Kleen h->nr_huge_pages++; 653a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 654b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 655b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 656b7ba30c6SAndi Kleen } 657b7ba30c6SAndi Kleen 65820a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 65920a0307cSWu Fengguang { 66020a0307cSWu Fengguang int i; 66120a0307cSWu Fengguang int nr_pages = 1 << order; 66220a0307cSWu Fengguang struct page *p = page + 1; 66320a0307cSWu Fengguang 66420a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 66520a0307cSWu Fengguang set_compound_order(page, order); 66620a0307cSWu Fengguang __SetPageHead(page); 66720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 66820a0307cSWu Fengguang __SetPageTail(p); 66958a84aa9SYouquan Song set_page_count(p, 0); 67020a0307cSWu Fengguang p->first_page = page; 67120a0307cSWu Fengguang } 67220a0307cSWu Fengguang } 67320a0307cSWu Fengguang 6747795912cSAndrew Morton /* 6757795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 6767795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 6777795912cSAndrew Morton * details. 6787795912cSAndrew Morton */ 67920a0307cSWu Fengguang int PageHuge(struct page *page) 68020a0307cSWu Fengguang { 68120a0307cSWu Fengguang compound_page_dtor *dtor; 68220a0307cSWu Fengguang 68320a0307cSWu Fengguang if (!PageCompound(page)) 68420a0307cSWu Fengguang return 0; 68520a0307cSWu Fengguang 68620a0307cSWu Fengguang page = compound_head(page); 68720a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 68820a0307cSWu Fengguang 68920a0307cSWu Fengguang return dtor == free_huge_page; 69020a0307cSWu Fengguang } 69143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 69243131e14SNaoya Horiguchi 693a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6941da177e4SLinus Torvalds { 6951da177e4SLinus Torvalds struct page *page; 696f96efd58SJoe Jin 697aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 698aa888a74SAndi Kleen return NULL; 699aa888a74SAndi Kleen 7006484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 701551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 702551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 703a5516438SAndi Kleen huge_page_order(h)); 7041da177e4SLinus Torvalds if (page) { 7057f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 706caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 7077b8ee84dSHarvey Harrison return NULL; 7087f2e9525SGerald Schaefer } 709a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 7101da177e4SLinus Torvalds } 71163b4613cSNishanth Aravamudan 71263b4613cSNishanth Aravamudan return page; 71363b4613cSNishanth Aravamudan } 71463b4613cSNishanth Aravamudan 7155ced66c9SAndi Kleen /* 7166ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 7176ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 7186ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7196ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 7206ae11b27SLee Schermerhorn * node for alloc or free. 7219a76db09SLee Schermerhorn */ 7226ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7239a76db09SLee Schermerhorn { 7246ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 7259a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 7266ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 7279a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 7289a76db09SLee Schermerhorn 7299a76db09SLee Schermerhorn return nid; 7309a76db09SLee Schermerhorn } 7319a76db09SLee Schermerhorn 7326ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 7335ced66c9SAndi Kleen { 7346ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 7356ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 7369a76db09SLee Schermerhorn return nid; 7375ced66c9SAndi Kleen } 7385ced66c9SAndi Kleen 7396ae11b27SLee Schermerhorn /* 7406ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 7416ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 7426ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 7436ae11b27SLee Schermerhorn * mask. 7446ae11b27SLee Schermerhorn */ 7456ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 7466ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 7476ae11b27SLee Schermerhorn { 7486ae11b27SLee Schermerhorn int nid; 7496ae11b27SLee Schermerhorn 7506ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7516ae11b27SLee Schermerhorn 7526ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 7536ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 7546ae11b27SLee Schermerhorn 7556ae11b27SLee Schermerhorn return nid; 7566ae11b27SLee Schermerhorn } 7576ae11b27SLee Schermerhorn 7586ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 75963b4613cSNishanth Aravamudan { 76063b4613cSNishanth Aravamudan struct page *page; 76163b4613cSNishanth Aravamudan int start_nid; 76263b4613cSNishanth Aravamudan int next_nid; 76363b4613cSNishanth Aravamudan int ret = 0; 76463b4613cSNishanth Aravamudan 7656ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 766e8c5c824SLee Schermerhorn next_nid = start_nid; 76763b4613cSNishanth Aravamudan 76863b4613cSNishanth Aravamudan do { 769e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 7709a76db09SLee Schermerhorn if (page) { 77163b4613cSNishanth Aravamudan ret = 1; 7729a76db09SLee Schermerhorn break; 7739a76db09SLee Schermerhorn } 7746ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 7759a76db09SLee Schermerhorn } while (next_nid != start_nid); 77663b4613cSNishanth Aravamudan 7773b116300SAdam Litke if (ret) 7783b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7793b116300SAdam Litke else 7803b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7813b116300SAdam Litke 78263b4613cSNishanth Aravamudan return ret; 7831da177e4SLinus Torvalds } 7841da177e4SLinus Torvalds 785e8c5c824SLee Schermerhorn /* 7866ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7876ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7886ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7896ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 790e8c5c824SLee Schermerhorn */ 7916ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 792e8c5c824SLee Schermerhorn { 7936ae11b27SLee Schermerhorn int nid; 7949a76db09SLee Schermerhorn 7956ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7966ae11b27SLee Schermerhorn 7976ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7986ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7996ae11b27SLee Schermerhorn 8009a76db09SLee Schermerhorn return nid; 801e8c5c824SLee Schermerhorn } 802e8c5c824SLee Schermerhorn 803e8c5c824SLee Schermerhorn /* 804e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 805e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 806e8c5c824SLee Schermerhorn * balanced over allowed nodes. 807e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 808e8c5c824SLee Schermerhorn */ 8096ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 8106ae11b27SLee Schermerhorn bool acct_surplus) 811e8c5c824SLee Schermerhorn { 812e8c5c824SLee Schermerhorn int start_nid; 813e8c5c824SLee Schermerhorn int next_nid; 814e8c5c824SLee Schermerhorn int ret = 0; 815e8c5c824SLee Schermerhorn 8166ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 817e8c5c824SLee Schermerhorn next_nid = start_nid; 818e8c5c824SLee Schermerhorn 819e8c5c824SLee Schermerhorn do { 820685f3457SLee Schermerhorn /* 821685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 822685f3457SLee Schermerhorn * nodes with surplus pages. 823685f3457SLee Schermerhorn */ 824685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 825685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 826e8c5c824SLee Schermerhorn struct page *page = 827e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 828e8c5c824SLee Schermerhorn struct page, lru); 829e8c5c824SLee Schermerhorn list_del(&page->lru); 830e8c5c824SLee Schermerhorn h->free_huge_pages--; 831e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 832685f3457SLee Schermerhorn if (acct_surplus) { 833685f3457SLee Schermerhorn h->surplus_huge_pages--; 834685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 835685f3457SLee Schermerhorn } 836e8c5c824SLee Schermerhorn update_and_free_page(h, page); 837e8c5c824SLee Schermerhorn ret = 1; 8389a76db09SLee Schermerhorn break; 839e8c5c824SLee Schermerhorn } 8406ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 8419a76db09SLee Schermerhorn } while (next_nid != start_nid); 842e8c5c824SLee Schermerhorn 843e8c5c824SLee Schermerhorn return ret; 844e8c5c824SLee Schermerhorn } 845e8c5c824SLee Schermerhorn 846bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 8477893d1d5SAdam Litke { 8487893d1d5SAdam Litke struct page *page; 849bf50bab2SNaoya Horiguchi unsigned int r_nid; 8507893d1d5SAdam Litke 851aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 852aa888a74SAndi Kleen return NULL; 853aa888a74SAndi Kleen 854d1c3fb1fSNishanth Aravamudan /* 855d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 856d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 857d1c3fb1fSNishanth Aravamudan * overcommit 858d1c3fb1fSNishanth Aravamudan * 859d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 860d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 861d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 862d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 863d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 864d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 865d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 866d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 867d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 868d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 869d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 870d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 871d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 872d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 873d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 874d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 875d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 876d1c3fb1fSNishanth Aravamudan */ 877d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 878a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 879d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 880d1c3fb1fSNishanth Aravamudan return NULL; 881d1c3fb1fSNishanth Aravamudan } else { 882a5516438SAndi Kleen h->nr_huge_pages++; 883a5516438SAndi Kleen h->surplus_huge_pages++; 884d1c3fb1fSNishanth Aravamudan } 885d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 886d1c3fb1fSNishanth Aravamudan 887bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 888551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 889551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 890a5516438SAndi Kleen huge_page_order(h)); 891bf50bab2SNaoya Horiguchi else 892bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 893bf50bab2SNaoya Horiguchi htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 894bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 895d1c3fb1fSNishanth Aravamudan 896caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 897caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 898ea5768c7SHillf Danton page = NULL; 899caff3a2cSGerald Schaefer } 900caff3a2cSGerald Schaefer 9017893d1d5SAdam Litke spin_lock(&hugetlb_lock); 902d1c3fb1fSNishanth Aravamudan if (page) { 9030edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 904bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 905d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 9069dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 907d1c3fb1fSNishanth Aravamudan /* 908d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 909d1c3fb1fSNishanth Aravamudan */ 910bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 911bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 9123b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 913d1c3fb1fSNishanth Aravamudan } else { 914a5516438SAndi Kleen h->nr_huge_pages--; 915a5516438SAndi Kleen h->surplus_huge_pages--; 9163b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 9177893d1d5SAdam Litke } 918d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 9197893d1d5SAdam Litke 9207893d1d5SAdam Litke return page; 9217893d1d5SAdam Litke } 9227893d1d5SAdam Litke 923e4e574b7SAdam Litke /* 924bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 925bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 926bf50bab2SNaoya Horiguchi * address of error page. 927bf50bab2SNaoya Horiguchi */ 928bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 929bf50bab2SNaoya Horiguchi { 930bf50bab2SNaoya Horiguchi struct page *page; 931bf50bab2SNaoya Horiguchi 932bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 933bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 934bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 935bf50bab2SNaoya Horiguchi 93694ae8ba7SAneesh Kumar K.V if (!page) 937bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 938bf50bab2SNaoya Horiguchi 939bf50bab2SNaoya Horiguchi return page; 940bf50bab2SNaoya Horiguchi } 941bf50bab2SNaoya Horiguchi 942bf50bab2SNaoya Horiguchi /* 94325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 944e4e574b7SAdam Litke * of size 'delta'. 945e4e574b7SAdam Litke */ 946a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 947e4e574b7SAdam Litke { 948e4e574b7SAdam Litke struct list_head surplus_list; 949e4e574b7SAdam Litke struct page *page, *tmp; 950e4e574b7SAdam Litke int ret, i; 951e4e574b7SAdam Litke int needed, allocated; 95228073b02SHillf Danton bool alloc_ok = true; 953e4e574b7SAdam Litke 954a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 955ac09b3a1SAdam Litke if (needed <= 0) { 956a5516438SAndi Kleen h->resv_huge_pages += delta; 957e4e574b7SAdam Litke return 0; 958ac09b3a1SAdam Litke } 959e4e574b7SAdam Litke 960e4e574b7SAdam Litke allocated = 0; 961e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 962e4e574b7SAdam Litke 963e4e574b7SAdam Litke ret = -ENOMEM; 964e4e574b7SAdam Litke retry: 965e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 966e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 967bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 96828073b02SHillf Danton if (!page) { 96928073b02SHillf Danton alloc_ok = false; 97028073b02SHillf Danton break; 97128073b02SHillf Danton } 972e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 973e4e574b7SAdam Litke } 97428073b02SHillf Danton allocated += i; 975e4e574b7SAdam Litke 976e4e574b7SAdam Litke /* 977e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 978e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 979e4e574b7SAdam Litke */ 980e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 981a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 982a5516438SAndi Kleen (h->free_huge_pages + allocated); 98328073b02SHillf Danton if (needed > 0) { 98428073b02SHillf Danton if (alloc_ok) 985e4e574b7SAdam Litke goto retry; 98628073b02SHillf Danton /* 98728073b02SHillf Danton * We were not able to allocate enough pages to 98828073b02SHillf Danton * satisfy the entire reservation so we free what 98928073b02SHillf Danton * we've allocated so far. 99028073b02SHillf Danton */ 99128073b02SHillf Danton goto free; 99228073b02SHillf Danton } 993e4e574b7SAdam Litke /* 994e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 99525985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 996e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 997ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 998ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 999ac09b3a1SAdam Litke * before they are reserved. 1000e4e574b7SAdam Litke */ 1001e4e574b7SAdam Litke needed += allocated; 1002a5516438SAndi Kleen h->resv_huge_pages += delta; 1003e4e574b7SAdam Litke ret = 0; 1004a9869b83SNaoya Horiguchi 100519fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 100619fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 100719fc3f0aSAdam Litke if ((--needed) < 0) 100819fc3f0aSAdam Litke break; 1009a9869b83SNaoya Horiguchi /* 1010a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1011a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1012a9869b83SNaoya Horiguchi */ 1013a9869b83SNaoya Horiguchi put_page_testzero(page); 1014a9869b83SNaoya Horiguchi VM_BUG_ON(page_count(page)); 1015a5516438SAndi Kleen enqueue_huge_page(h, page); 101619fc3f0aSAdam Litke } 101728073b02SHillf Danton free: 1018b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 101919fc3f0aSAdam Litke 102019fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 102119fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 1022e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1023a9869b83SNaoya Horiguchi put_page(page); 1024a9869b83SNaoya Horiguchi } 1025af767cbdSAdam Litke } 102619fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1027e4e574b7SAdam Litke 1028e4e574b7SAdam Litke return ret; 1029e4e574b7SAdam Litke } 1030e4e574b7SAdam Litke 1031e4e574b7SAdam Litke /* 1032e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1033e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1034e4e574b7SAdam Litke * never used. 1035685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1036e4e574b7SAdam Litke */ 1037a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1038a5516438SAndi Kleen unsigned long unused_resv_pages) 1039e4e574b7SAdam Litke { 1040e4e574b7SAdam Litke unsigned long nr_pages; 1041e4e574b7SAdam Litke 1042ac09b3a1SAdam Litke /* Uncommit the reservation */ 1043a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1044ac09b3a1SAdam Litke 1045aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1046aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1047aa888a74SAndi Kleen return; 1048aa888a74SAndi Kleen 1049a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1050e4e574b7SAdam Litke 1051685f3457SLee Schermerhorn /* 1052685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 10539b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 10549b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 10559b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 10569b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 10579b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1058685f3457SLee Schermerhorn */ 1059685f3457SLee Schermerhorn while (nr_pages--) { 10609b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 1061685f3457SLee Schermerhorn break; 1062e4e574b7SAdam Litke } 1063e4e574b7SAdam Litke } 1064e4e574b7SAdam Litke 1065c37f9fb1SAndy Whitcroft /* 1066c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1067c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 106890481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 106990481622SDavid Gibson * can occur. Where any new reservation would be required the 107090481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 107190481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 107290481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 107390481622SDavid Gibson * failure. 1074c37f9fb1SAndy Whitcroft */ 1075e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1076a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1077c37f9fb1SAndy Whitcroft { 1078c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1079c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1080c37f9fb1SAndy Whitcroft 1081f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1082a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1083c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 1084c37f9fb1SAndy Whitcroft idx, idx + 1); 1085c37f9fb1SAndy Whitcroft 108684afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1087c37f9fb1SAndy Whitcroft return 1; 1088c37f9fb1SAndy Whitcroft 108984afd99bSAndy Whitcroft } else { 1090e2f17d94SRoel Kluin long err; 1091a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 109284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 109384afd99bSAndy Whitcroft 109484afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 109584afd99bSAndy Whitcroft if (err < 0) 109684afd99bSAndy Whitcroft return err; 1097c37f9fb1SAndy Whitcroft return 0; 1098c37f9fb1SAndy Whitcroft } 109984afd99bSAndy Whitcroft } 1100a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1101a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1102c37f9fb1SAndy Whitcroft { 1103c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1104c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1105c37f9fb1SAndy Whitcroft 1106f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1107a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1108c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 110984afd99bSAndy Whitcroft 111084afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1111a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 111284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 111384afd99bSAndy Whitcroft 111484afd99bSAndy Whitcroft /* Mark this page used in the map. */ 111584afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1116c37f9fb1SAndy Whitcroft } 1117c37f9fb1SAndy Whitcroft } 1118c37f9fb1SAndy Whitcroft 1119348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 112004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1121348ea204SAdam Litke { 112290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1123a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1124348ea204SAdam Litke struct page *page; 1125e2f17d94SRoel Kluin long chg; 11266d76dcf4SAneesh Kumar K.V int ret, idx; 11276d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 11282fc39cecSAdam Litke 11296d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1130a1e78772SMel Gorman /* 113190481622SDavid Gibson * Processes that did not create the mapping will have no 113290481622SDavid Gibson * reserves and will not have accounted against subpool 113390481622SDavid Gibson * limit. Check that the subpool limit can be made before 113490481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 113590481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 113690481622SDavid Gibson * mapping overlaps. 1137a1e78772SMel Gorman */ 1138a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1139c37f9fb1SAndy Whitcroft if (chg < 0) 114076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1141c37f9fb1SAndy Whitcroft if (chg) 114290481622SDavid Gibson if (hugepage_subpool_get_pages(spool, chg)) 114376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 114490d8b7e6SAdam Litke 11456d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 11466d76dcf4SAneesh Kumar K.V if (ret) { 11476d76dcf4SAneesh Kumar K.V hugepage_subpool_put_pages(spool, chg); 11486d76dcf4SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 11496d76dcf4SAneesh Kumar K.V } 1150a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1151a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 115294ae8ba7SAneesh Kumar K.V if (page) { 115394ae8ba7SAneesh Kumar K.V /* update page cgroup details */ 115494ae8ba7SAneesh Kumar K.V hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), 115594ae8ba7SAneesh Kumar K.V h_cg, page); 1156a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 115794ae8ba7SAneesh Kumar K.V } else { 115894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1159bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1160a1e78772SMel Gorman if (!page) { 11616d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_cgroup(idx, 11626d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), 11636d76dcf4SAneesh Kumar K.V h_cg); 116490481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 116576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 1166a1e78772SMel Gorman } 116779dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 116894ae8ba7SAneesh Kumar K.V hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), 116994ae8ba7SAneesh Kumar K.V h_cg, page); 117079dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 117179dbb236SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1172a1e78772SMel Gorman } 1173a1e78772SMel Gorman 117490481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1175a1e78772SMel Gorman 1176a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 11777893d1d5SAdam Litke return page; 1178b45b5bd6SDavid Gibson } 1179b45b5bd6SDavid Gibson 118091f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1181aa888a74SAndi Kleen { 1182aa888a74SAndi Kleen struct huge_bootmem_page *m; 11839b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1184aa888a74SAndi Kleen 1185aa888a74SAndi Kleen while (nr_nodes) { 1186aa888a74SAndi Kleen void *addr; 1187aa888a74SAndi Kleen 1188aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 11896ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 11909b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1191aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1192aa888a74SAndi Kleen 1193aa888a74SAndi Kleen if (addr) { 1194aa888a74SAndi Kleen /* 1195aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1196aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1197aa888a74SAndi Kleen * puts them into the mem_map). 1198aa888a74SAndi Kleen */ 1199aa888a74SAndi Kleen m = addr; 1200aa888a74SAndi Kleen goto found; 1201aa888a74SAndi Kleen } 1202aa888a74SAndi Kleen nr_nodes--; 1203aa888a74SAndi Kleen } 1204aa888a74SAndi Kleen return 0; 1205aa888a74SAndi Kleen 1206aa888a74SAndi Kleen found: 1207aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1208aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1209aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1210aa888a74SAndi Kleen m->hstate = h; 1211aa888a74SAndi Kleen return 1; 1212aa888a74SAndi Kleen } 1213aa888a74SAndi Kleen 121418229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 121518229df5SAndy Whitcroft { 121618229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 121718229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 121818229df5SAndy Whitcroft else 121918229df5SAndy Whitcroft prep_compound_page(page, order); 122018229df5SAndy Whitcroft } 122118229df5SAndy Whitcroft 1222aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1223aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1224aa888a74SAndi Kleen { 1225aa888a74SAndi Kleen struct huge_bootmem_page *m; 1226aa888a74SAndi Kleen 1227aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1228aa888a74SAndi Kleen struct hstate *h = m->hstate; 1229ee8f248dSBecky Bruce struct page *page; 1230ee8f248dSBecky Bruce 1231ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1232ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 1233ee8f248dSBecky Bruce free_bootmem_late((unsigned long)m, 1234ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1235ee8f248dSBecky Bruce #else 1236ee8f248dSBecky Bruce page = virt_to_page(m); 1237ee8f248dSBecky Bruce #endif 1238aa888a74SAndi Kleen __ClearPageReserved(page); 1239aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 124018229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1241aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1242b0320c7bSRafael Aquini /* 1243b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1244b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1245b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1246b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1247b0320c7bSRafael Aquini */ 1248b0320c7bSRafael Aquini if (h->order > (MAX_ORDER - 1)) 1249b0320c7bSRafael Aquini totalram_pages += 1 << h->order; 1250aa888a74SAndi Kleen } 1251aa888a74SAndi Kleen } 1252aa888a74SAndi Kleen 12538faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 12541da177e4SLinus Torvalds { 12551da177e4SLinus Torvalds unsigned long i; 12561da177e4SLinus Torvalds 1257e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1258aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1259aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1260aa888a74SAndi Kleen break; 12619b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 12629b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 12631da177e4SLinus Torvalds break; 12641da177e4SLinus Torvalds } 12658faa8b07SAndi Kleen h->max_huge_pages = i; 1266e5ff2159SAndi Kleen } 1267e5ff2159SAndi Kleen 1268e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1269e5ff2159SAndi Kleen { 1270e5ff2159SAndi Kleen struct hstate *h; 1271e5ff2159SAndi Kleen 1272e5ff2159SAndi Kleen for_each_hstate(h) { 12738faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 12748faa8b07SAndi Kleen if (h->order < MAX_ORDER) 12758faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1276e5ff2159SAndi Kleen } 1277e5ff2159SAndi Kleen } 1278e5ff2159SAndi Kleen 12794abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 12804abd32dbSAndi Kleen { 12814abd32dbSAndi Kleen if (n >= (1UL << 30)) 12824abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 12834abd32dbSAndi Kleen else if (n >= (1UL << 20)) 12844abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 12854abd32dbSAndi Kleen else 12864abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 12874abd32dbSAndi Kleen return buf; 12884abd32dbSAndi Kleen } 12894abd32dbSAndi Kleen 1290e5ff2159SAndi Kleen static void __init report_hugepages(void) 1291e5ff2159SAndi Kleen { 1292e5ff2159SAndi Kleen struct hstate *h; 1293e5ff2159SAndi Kleen 1294e5ff2159SAndi Kleen for_each_hstate(h) { 12954abd32dbSAndi Kleen char buf[32]; 12964abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 12974abd32dbSAndi Kleen "pre-allocated %ld pages\n", 12984abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 12994abd32dbSAndi Kleen h->free_huge_pages); 1300e5ff2159SAndi Kleen } 1301e5ff2159SAndi Kleen } 1302e5ff2159SAndi Kleen 13031da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 13046ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 13056ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13061da177e4SLinus Torvalds { 13074415cc8dSChristoph Lameter int i; 13084415cc8dSChristoph Lameter 1309aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1310aa888a74SAndi Kleen return; 1311aa888a74SAndi Kleen 13126ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 13131da177e4SLinus Torvalds struct page *page, *next; 1314a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1315a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1316a5516438SAndi Kleen if (count >= h->nr_huge_pages) 13176b0c880dSAdam Litke return; 13181da177e4SLinus Torvalds if (PageHighMem(page)) 13191da177e4SLinus Torvalds continue; 13201da177e4SLinus Torvalds list_del(&page->lru); 1321e5ff2159SAndi Kleen update_and_free_page(h, page); 1322a5516438SAndi Kleen h->free_huge_pages--; 1323a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 13241da177e4SLinus Torvalds } 13251da177e4SLinus Torvalds } 13261da177e4SLinus Torvalds } 13271da177e4SLinus Torvalds #else 13286ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 13296ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13301da177e4SLinus Torvalds { 13311da177e4SLinus Torvalds } 13321da177e4SLinus Torvalds #endif 13331da177e4SLinus Torvalds 133420a0307cSWu Fengguang /* 133520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 133620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 133720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 133820a0307cSWu Fengguang */ 13396ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 13406ae11b27SLee Schermerhorn int delta) 134120a0307cSWu Fengguang { 1342e8c5c824SLee Schermerhorn int start_nid, next_nid; 134320a0307cSWu Fengguang int ret = 0; 134420a0307cSWu Fengguang 134520a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 134620a0307cSWu Fengguang 1347e8c5c824SLee Schermerhorn if (delta < 0) 13486ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1349e8c5c824SLee Schermerhorn else 13506ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1351e8c5c824SLee Schermerhorn next_nid = start_nid; 1352e8c5c824SLee Schermerhorn 1353e8c5c824SLee Schermerhorn do { 1354e8c5c824SLee Schermerhorn int nid = next_nid; 1355e8c5c824SLee Schermerhorn if (delta < 0) { 1356e8c5c824SLee Schermerhorn /* 1357e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1358e8c5c824SLee Schermerhorn */ 13599a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 13606ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 13616ae11b27SLee Schermerhorn nodes_allowed); 136220a0307cSWu Fengguang continue; 1363e8c5c824SLee Schermerhorn } 13649a76db09SLee Schermerhorn } 1365e8c5c824SLee Schermerhorn if (delta > 0) { 1366e8c5c824SLee Schermerhorn /* 1367e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1368e8c5c824SLee Schermerhorn */ 1369e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 13709a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 13716ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 13726ae11b27SLee Schermerhorn nodes_allowed); 137320a0307cSWu Fengguang continue; 1374e8c5c824SLee Schermerhorn } 13759a76db09SLee Schermerhorn } 137620a0307cSWu Fengguang 137720a0307cSWu Fengguang h->surplus_huge_pages += delta; 137820a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 137920a0307cSWu Fengguang ret = 1; 138020a0307cSWu Fengguang break; 1381e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 138220a0307cSWu Fengguang 138320a0307cSWu Fengguang return ret; 138420a0307cSWu Fengguang } 138520a0307cSWu Fengguang 1386a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 13876ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 13886ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13891da177e4SLinus Torvalds { 13907893d1d5SAdam Litke unsigned long min_count, ret; 13911da177e4SLinus Torvalds 1392aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1393aa888a74SAndi Kleen return h->max_huge_pages; 1394aa888a74SAndi Kleen 13957893d1d5SAdam Litke /* 13967893d1d5SAdam Litke * Increase the pool size 13977893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 13987893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1399d1c3fb1fSNishanth Aravamudan * 1400d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1401d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1402d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1403d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1404d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 14057893d1d5SAdam Litke */ 14061da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1407a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 14086ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 14097893d1d5SAdam Litke break; 14107893d1d5SAdam Litke } 14117893d1d5SAdam Litke 1412a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 14137893d1d5SAdam Litke /* 14147893d1d5SAdam Litke * If this allocation races such that we no longer need the 14157893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 14167893d1d5SAdam Litke * and reducing the surplus. 14177893d1d5SAdam Litke */ 14187893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 14196ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 14207893d1d5SAdam Litke spin_lock(&hugetlb_lock); 14217893d1d5SAdam Litke if (!ret) 14227893d1d5SAdam Litke goto out; 14237893d1d5SAdam Litke 1424536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1425536240f2SMel Gorman if (signal_pending(current)) 1426536240f2SMel Gorman goto out; 14277893d1d5SAdam Litke } 14287893d1d5SAdam Litke 14297893d1d5SAdam Litke /* 14307893d1d5SAdam Litke * Decrease the pool size 14317893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 14327893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 14337893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 14347893d1d5SAdam Litke * to the desired size as pages become free. 1435d1c3fb1fSNishanth Aravamudan * 1436d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1437d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1438d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1439d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1440d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1441d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1442d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 14437893d1d5SAdam Litke */ 1444a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 14456b0c880dSAdam Litke min_count = max(count, min_count); 14466ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1447a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 14486ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 14491da177e4SLinus Torvalds break; 14501da177e4SLinus Torvalds } 1451a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 14526ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 14537893d1d5SAdam Litke break; 14547893d1d5SAdam Litke } 14557893d1d5SAdam Litke out: 1456a5516438SAndi Kleen ret = persistent_huge_pages(h); 14571da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 14587893d1d5SAdam Litke return ret; 14591da177e4SLinus Torvalds } 14601da177e4SLinus Torvalds 1461a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1462a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1463a3437870SNishanth Aravamudan 1464a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1465a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1466a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1467a3437870SNishanth Aravamudan 1468a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1469a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1470a3437870SNishanth Aravamudan 14719a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 14729a305230SLee Schermerhorn 14739a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1474a3437870SNishanth Aravamudan { 1475a3437870SNishanth Aravamudan int i; 14769a305230SLee Schermerhorn 1477a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 14789a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 14799a305230SLee Schermerhorn if (nidp) 14809a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1481a3437870SNishanth Aravamudan return &hstates[i]; 14829a305230SLee Schermerhorn } 14839a305230SLee Schermerhorn 14849a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1485a3437870SNishanth Aravamudan } 1486a3437870SNishanth Aravamudan 148706808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1488a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1489a3437870SNishanth Aravamudan { 14909a305230SLee Schermerhorn struct hstate *h; 14919a305230SLee Schermerhorn unsigned long nr_huge_pages; 14929a305230SLee Schermerhorn int nid; 14939a305230SLee Schermerhorn 14949a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14959a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14969a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 14979a305230SLee Schermerhorn else 14989a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 14999a305230SLee Schermerhorn 15009a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1501a3437870SNishanth Aravamudan } 1502adbe8726SEric B Munson 150306808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 150406808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 150506808b08SLee Schermerhorn const char *buf, size_t len) 1506a3437870SNishanth Aravamudan { 1507a3437870SNishanth Aravamudan int err; 15089a305230SLee Schermerhorn int nid; 150906808b08SLee Schermerhorn unsigned long count; 15109a305230SLee Schermerhorn struct hstate *h; 1511bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1512a3437870SNishanth Aravamudan 151306808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 151473ae31e5SEric B Munson if (err) 1515adbe8726SEric B Munson goto out; 1516a3437870SNishanth Aravamudan 15179a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1518adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1519adbe8726SEric B Munson err = -EINVAL; 1520adbe8726SEric B Munson goto out; 1521adbe8726SEric B Munson } 1522adbe8726SEric B Munson 15239a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 15249a305230SLee Schermerhorn /* 15259a305230SLee Schermerhorn * global hstate attribute 15269a305230SLee Schermerhorn */ 15279a305230SLee Schermerhorn if (!(obey_mempolicy && 15289a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 152906808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 15309a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 153106808b08SLee Schermerhorn } 15329a305230SLee Schermerhorn } else if (nodes_allowed) { 15339a305230SLee Schermerhorn /* 15349a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 15359a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 15369a305230SLee Schermerhorn */ 15379a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 15389a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 15399a305230SLee Schermerhorn } else 15409a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 15419a305230SLee Schermerhorn 154206808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1543a3437870SNishanth Aravamudan 15449b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 154506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 154606808b08SLee Schermerhorn 154706808b08SLee Schermerhorn return len; 1548adbe8726SEric B Munson out: 1549adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1550adbe8726SEric B Munson return err; 155106808b08SLee Schermerhorn } 155206808b08SLee Schermerhorn 155306808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 155406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 155506808b08SLee Schermerhorn { 155606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 155706808b08SLee Schermerhorn } 155806808b08SLee Schermerhorn 155906808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 156006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 156106808b08SLee Schermerhorn { 156206808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1563a3437870SNishanth Aravamudan } 1564a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1565a3437870SNishanth Aravamudan 156606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 156706808b08SLee Schermerhorn 156806808b08SLee Schermerhorn /* 156906808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 157006808b08SLee Schermerhorn * huge page alloc/free. 157106808b08SLee Schermerhorn */ 157206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 157306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 157406808b08SLee Schermerhorn { 157506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 157606808b08SLee Schermerhorn } 157706808b08SLee Schermerhorn 157806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 157906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 158006808b08SLee Schermerhorn { 158106808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 158206808b08SLee Schermerhorn } 158306808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 158406808b08SLee Schermerhorn #endif 158506808b08SLee Schermerhorn 158606808b08SLee Schermerhorn 1587a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1588a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1589a3437870SNishanth Aravamudan { 15909a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1591a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1592a3437870SNishanth Aravamudan } 1593adbe8726SEric B Munson 1594a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1595a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1596a3437870SNishanth Aravamudan { 1597a3437870SNishanth Aravamudan int err; 1598a3437870SNishanth Aravamudan unsigned long input; 15999a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1600a3437870SNishanth Aravamudan 1601adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1602adbe8726SEric B Munson return -EINVAL; 1603adbe8726SEric B Munson 1604a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1605a3437870SNishanth Aravamudan if (err) 160673ae31e5SEric B Munson return err; 1607a3437870SNishanth Aravamudan 1608a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1609a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1610a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1611a3437870SNishanth Aravamudan 1612a3437870SNishanth Aravamudan return count; 1613a3437870SNishanth Aravamudan } 1614a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1615a3437870SNishanth Aravamudan 1616a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1617a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1618a3437870SNishanth Aravamudan { 16199a305230SLee Schermerhorn struct hstate *h; 16209a305230SLee Schermerhorn unsigned long free_huge_pages; 16219a305230SLee Schermerhorn int nid; 16229a305230SLee Schermerhorn 16239a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16249a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16259a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 16269a305230SLee Schermerhorn else 16279a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 16289a305230SLee Schermerhorn 16299a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1630a3437870SNishanth Aravamudan } 1631a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1632a3437870SNishanth Aravamudan 1633a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1634a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1635a3437870SNishanth Aravamudan { 16369a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1637a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1638a3437870SNishanth Aravamudan } 1639a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1640a3437870SNishanth Aravamudan 1641a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1642a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1643a3437870SNishanth Aravamudan { 16449a305230SLee Schermerhorn struct hstate *h; 16459a305230SLee Schermerhorn unsigned long surplus_huge_pages; 16469a305230SLee Schermerhorn int nid; 16479a305230SLee Schermerhorn 16489a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16499a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16509a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 16519a305230SLee Schermerhorn else 16529a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 16539a305230SLee Schermerhorn 16549a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1655a3437870SNishanth Aravamudan } 1656a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1657a3437870SNishanth Aravamudan 1658a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1659a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1660a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1661a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1662a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1663a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 166406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 166506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 166606808b08SLee Schermerhorn #endif 1667a3437870SNishanth Aravamudan NULL, 1668a3437870SNishanth Aravamudan }; 1669a3437870SNishanth Aravamudan 1670a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1671a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1672a3437870SNishanth Aravamudan }; 1673a3437870SNishanth Aravamudan 1674094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 16759a305230SLee Schermerhorn struct kobject **hstate_kobjs, 16769a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1677a3437870SNishanth Aravamudan { 1678a3437870SNishanth Aravamudan int retval; 1679972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1680a3437870SNishanth Aravamudan 16819a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 16829a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1683a3437870SNishanth Aravamudan return -ENOMEM; 1684a3437870SNishanth Aravamudan 16859a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1686a3437870SNishanth Aravamudan if (retval) 16879a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1688a3437870SNishanth Aravamudan 1689a3437870SNishanth Aravamudan return retval; 1690a3437870SNishanth Aravamudan } 1691a3437870SNishanth Aravamudan 1692a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1693a3437870SNishanth Aravamudan { 1694a3437870SNishanth Aravamudan struct hstate *h; 1695a3437870SNishanth Aravamudan int err; 1696a3437870SNishanth Aravamudan 1697a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1698a3437870SNishanth Aravamudan if (!hugepages_kobj) 1699a3437870SNishanth Aravamudan return; 1700a3437870SNishanth Aravamudan 1701a3437870SNishanth Aravamudan for_each_hstate(h) { 17029a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 17039a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1704a3437870SNishanth Aravamudan if (err) 1705a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1706a3437870SNishanth Aravamudan h->name); 1707a3437870SNishanth Aravamudan } 1708a3437870SNishanth Aravamudan } 1709a3437870SNishanth Aravamudan 17109a305230SLee Schermerhorn #ifdef CONFIG_NUMA 17119a305230SLee Schermerhorn 17129a305230SLee Schermerhorn /* 17139a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 171410fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 171510fbcf4cSKay Sievers * index of a node device or _hstate == node id. 171610fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 17179a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 17189a305230SLee Schermerhorn */ 17199a305230SLee Schermerhorn struct node_hstate { 17209a305230SLee Schermerhorn struct kobject *hugepages_kobj; 17219a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 17229a305230SLee Schermerhorn }; 17239a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 17249a305230SLee Schermerhorn 17259a305230SLee Schermerhorn /* 172610fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 17279a305230SLee Schermerhorn */ 17289a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 17299a305230SLee Schermerhorn &nr_hugepages_attr.attr, 17309a305230SLee Schermerhorn &free_hugepages_attr.attr, 17319a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 17329a305230SLee Schermerhorn NULL, 17339a305230SLee Schermerhorn }; 17349a305230SLee Schermerhorn 17359a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 17369a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 17379a305230SLee Schermerhorn }; 17389a305230SLee Schermerhorn 17399a305230SLee Schermerhorn /* 174010fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 17419a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 17429a305230SLee Schermerhorn */ 17439a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17449a305230SLee Schermerhorn { 17459a305230SLee Schermerhorn int nid; 17469a305230SLee Schermerhorn 17479a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 17489a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 17499a305230SLee Schermerhorn int i; 17509a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 17519a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 17529a305230SLee Schermerhorn if (nidp) 17539a305230SLee Schermerhorn *nidp = nid; 17549a305230SLee Schermerhorn return &hstates[i]; 17559a305230SLee Schermerhorn } 17569a305230SLee Schermerhorn } 17579a305230SLee Schermerhorn 17589a305230SLee Schermerhorn BUG(); 17599a305230SLee Schermerhorn return NULL; 17609a305230SLee Schermerhorn } 17619a305230SLee Schermerhorn 17629a305230SLee Schermerhorn /* 176310fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 17649a305230SLee Schermerhorn * No-op if no hstate attributes attached. 17659a305230SLee Schermerhorn */ 17669a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 17679a305230SLee Schermerhorn { 17689a305230SLee Schermerhorn struct hstate *h; 176910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 17709a305230SLee Schermerhorn 17719a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 17729b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 17739a305230SLee Schermerhorn 1774972dc4deSAneesh Kumar K.V for_each_hstate(h) { 1775972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 1776972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 1777972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 1778972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 1779972dc4deSAneesh Kumar K.V } 17809a305230SLee Schermerhorn } 17819a305230SLee Schermerhorn 17829a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 17839a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 17849a305230SLee Schermerhorn } 17859a305230SLee Schermerhorn 17869a305230SLee Schermerhorn /* 178710fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 17889a305230SLee Schermerhorn * that have them. 17899a305230SLee Schermerhorn */ 17909a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 17919a305230SLee Schermerhorn { 17929a305230SLee Schermerhorn int nid; 17939a305230SLee Schermerhorn 17949a305230SLee Schermerhorn /* 179510fbcf4cSKay Sievers * disable node device registrations. 17969a305230SLee Schermerhorn */ 17979a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 17989a305230SLee Schermerhorn 17999a305230SLee Schermerhorn /* 18009a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 18019a305230SLee Schermerhorn */ 18029a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 18039a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 18049a305230SLee Schermerhorn } 18059a305230SLee Schermerhorn 18069a305230SLee Schermerhorn /* 180710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 18089a305230SLee Schermerhorn * No-op if attributes already registered. 18099a305230SLee Schermerhorn */ 18109a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 18119a305230SLee Schermerhorn { 18129a305230SLee Schermerhorn struct hstate *h; 181310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 18149a305230SLee Schermerhorn int err; 18159a305230SLee Schermerhorn 18169a305230SLee Schermerhorn if (nhs->hugepages_kobj) 18179a305230SLee Schermerhorn return; /* already allocated */ 18189a305230SLee Schermerhorn 18199a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 182010fbcf4cSKay Sievers &node->dev.kobj); 18219a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 18229a305230SLee Schermerhorn return; 18239a305230SLee Schermerhorn 18249a305230SLee Schermerhorn for_each_hstate(h) { 18259a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 18269a305230SLee Schermerhorn nhs->hstate_kobjs, 18279a305230SLee Schermerhorn &per_node_hstate_attr_group); 18289a305230SLee Schermerhorn if (err) { 18299a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 18309a305230SLee Schermerhorn " for node %d\n", 183110fbcf4cSKay Sievers h->name, node->dev.id); 18329a305230SLee Schermerhorn hugetlb_unregister_node(node); 18339a305230SLee Schermerhorn break; 18349a305230SLee Schermerhorn } 18359a305230SLee Schermerhorn } 18369a305230SLee Schermerhorn } 18379a305230SLee Schermerhorn 18389a305230SLee Schermerhorn /* 18399b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 184010fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 184110fbcf4cSKay Sievers * registered their associated device by this time. 18429a305230SLee Schermerhorn */ 18439a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 18449a305230SLee Schermerhorn { 18459a305230SLee Schermerhorn int nid; 18469a305230SLee Schermerhorn 18479b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 18489a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 184910fbcf4cSKay Sievers if (node->dev.id == nid) 18509a305230SLee Schermerhorn hugetlb_register_node(node); 18519a305230SLee Schermerhorn } 18529a305230SLee Schermerhorn 18539a305230SLee Schermerhorn /* 185410fbcf4cSKay Sievers * Let the node device driver know we're here so it can 18559a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 18569a305230SLee Schermerhorn */ 18579a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 18589a305230SLee Schermerhorn hugetlb_unregister_node); 18599a305230SLee Schermerhorn } 18609a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 18619a305230SLee Schermerhorn 18629a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 18639a305230SLee Schermerhorn { 18649a305230SLee Schermerhorn BUG(); 18659a305230SLee Schermerhorn if (nidp) 18669a305230SLee Schermerhorn *nidp = -1; 18679a305230SLee Schermerhorn return NULL; 18689a305230SLee Schermerhorn } 18699a305230SLee Schermerhorn 18709a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 18719a305230SLee Schermerhorn 18729a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 18739a305230SLee Schermerhorn 18749a305230SLee Schermerhorn #endif 18759a305230SLee Schermerhorn 1876a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1877a3437870SNishanth Aravamudan { 1878a3437870SNishanth Aravamudan struct hstate *h; 1879a3437870SNishanth Aravamudan 18809a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 18819a305230SLee Schermerhorn 1882a3437870SNishanth Aravamudan for_each_hstate(h) { 1883972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 1884a3437870SNishanth Aravamudan } 1885a3437870SNishanth Aravamudan 1886a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1887a3437870SNishanth Aravamudan } 1888a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1889a3437870SNishanth Aravamudan 1890a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1891a3437870SNishanth Aravamudan { 18920ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 18930ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 18940ef89d25SBenjamin Herrenschmidt * there is no such support 18950ef89d25SBenjamin Herrenschmidt */ 18960ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 18970ef89d25SBenjamin Herrenschmidt return 0; 1898a3437870SNishanth Aravamudan 1899e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1900e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1901e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1902a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1903a3437870SNishanth Aravamudan } 1904972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1905e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1906e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1907a3437870SNishanth Aravamudan 1908a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1909a3437870SNishanth Aravamudan 1910aa888a74SAndi Kleen gather_bootmem_prealloc(); 1911aa888a74SAndi Kleen 1912a3437870SNishanth Aravamudan report_hugepages(); 1913a3437870SNishanth Aravamudan 1914a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1915a3437870SNishanth Aravamudan 19169a305230SLee Schermerhorn hugetlb_register_all_nodes(); 19179a305230SLee Schermerhorn 1918a3437870SNishanth Aravamudan return 0; 1919a3437870SNishanth Aravamudan } 1920a3437870SNishanth Aravamudan module_init(hugetlb_init); 1921a3437870SNishanth Aravamudan 1922a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1923a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1924a3437870SNishanth Aravamudan { 1925a3437870SNishanth Aravamudan struct hstate *h; 19268faa8b07SAndi Kleen unsigned long i; 19278faa8b07SAndi Kleen 1928a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1929a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1930a3437870SNishanth Aravamudan return; 1931a3437870SNishanth Aravamudan } 193247d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 1933a3437870SNishanth Aravamudan BUG_ON(order == 0); 193447d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 1935a3437870SNishanth Aravamudan h->order = order; 1936a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 19378faa8b07SAndi Kleen h->nr_huge_pages = 0; 19388faa8b07SAndi Kleen h->free_huge_pages = 0; 19398faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 19408faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 19410edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 19429b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 19439b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1944a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1945a3437870SNishanth Aravamudan huge_page_size(h)/1024); 1946abb8206cSAneesh Kumar K.V /* 1947abb8206cSAneesh Kumar K.V * Add cgroup control files only if the huge page consists 1948abb8206cSAneesh Kumar K.V * of more than two normal pages. This is because we use 1949abb8206cSAneesh Kumar K.V * page[2].lru.next for storing cgoup details. 1950abb8206cSAneesh Kumar K.V */ 1951abb8206cSAneesh Kumar K.V if (order >= HUGETLB_CGROUP_MIN_ORDER) 1952abb8206cSAneesh Kumar K.V hugetlb_cgroup_file_init(hugetlb_max_hstate - 1); 19538faa8b07SAndi Kleen 1954a3437870SNishanth Aravamudan parsed_hstate = h; 1955a3437870SNishanth Aravamudan } 1956a3437870SNishanth Aravamudan 1957e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1958a3437870SNishanth Aravamudan { 1959a3437870SNishanth Aravamudan unsigned long *mhp; 19608faa8b07SAndi Kleen static unsigned long *last_mhp; 1961a3437870SNishanth Aravamudan 1962a3437870SNishanth Aravamudan /* 196347d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 1964a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1965a3437870SNishanth Aravamudan */ 196647d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 1967a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1968a3437870SNishanth Aravamudan else 1969a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1970a3437870SNishanth Aravamudan 19718faa8b07SAndi Kleen if (mhp == last_mhp) { 19728faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 19738faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 19748faa8b07SAndi Kleen return 1; 19758faa8b07SAndi Kleen } 19768faa8b07SAndi Kleen 1977a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1978a3437870SNishanth Aravamudan *mhp = 0; 1979a3437870SNishanth Aravamudan 19808faa8b07SAndi Kleen /* 19818faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 19828faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 19838faa8b07SAndi Kleen * use the bootmem allocator. 19848faa8b07SAndi Kleen */ 198547d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 19868faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 19878faa8b07SAndi Kleen 19888faa8b07SAndi Kleen last_mhp = mhp; 19898faa8b07SAndi Kleen 1990a3437870SNishanth Aravamudan return 1; 1991a3437870SNishanth Aravamudan } 1992e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1993e11bfbfcSNick Piggin 1994e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1995e11bfbfcSNick Piggin { 1996e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1997e11bfbfcSNick Piggin return 1; 1998e11bfbfcSNick Piggin } 1999e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2000a3437870SNishanth Aravamudan 20018a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 20028a213460SNishanth Aravamudan { 20038a213460SNishanth Aravamudan int node; 20048a213460SNishanth Aravamudan unsigned int nr = 0; 20058a213460SNishanth Aravamudan 20068a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 20078a213460SNishanth Aravamudan nr += array[node]; 20088a213460SNishanth Aravamudan 20098a213460SNishanth Aravamudan return nr; 20108a213460SNishanth Aravamudan } 20118a213460SNishanth Aravamudan 20128a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 201306808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 201406808b08SLee Schermerhorn struct ctl_table *table, int write, 201506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 20161da177e4SLinus Torvalds { 2017e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2018e5ff2159SAndi Kleen unsigned long tmp; 201908d4a246SMichal Hocko int ret; 2020e5ff2159SAndi Kleen 2021e5ff2159SAndi Kleen tmp = h->max_huge_pages; 2022e5ff2159SAndi Kleen 2023adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2024adbe8726SEric B Munson return -EINVAL; 2025adbe8726SEric B Munson 2026e5ff2159SAndi Kleen table->data = &tmp; 2027e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 202808d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 202908d4a246SMichal Hocko if (ret) 203008d4a246SMichal Hocko goto out; 2031e5ff2159SAndi Kleen 203206808b08SLee Schermerhorn if (write) { 2033bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2034bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 203506808b08SLee Schermerhorn if (!(obey_mempolicy && 203606808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 203706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 203806808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 203906808b08SLee Schermerhorn } 204006808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 204106808b08SLee Schermerhorn 204206808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 204306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 204406808b08SLee Schermerhorn } 204508d4a246SMichal Hocko out: 204608d4a246SMichal Hocko return ret; 20471da177e4SLinus Torvalds } 2048396faf03SMel Gorman 204906808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 205006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 205106808b08SLee Schermerhorn { 205206808b08SLee Schermerhorn 205306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 205406808b08SLee Schermerhorn buffer, length, ppos); 205506808b08SLee Schermerhorn } 205606808b08SLee Schermerhorn 205706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 205806808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 205906808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 206006808b08SLee Schermerhorn { 206106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 206206808b08SLee Schermerhorn buffer, length, ppos); 206306808b08SLee Schermerhorn } 206406808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 206506808b08SLee Schermerhorn 2066396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 20678d65af78SAlexey Dobriyan void __user *buffer, 2068396faf03SMel Gorman size_t *length, loff_t *ppos) 2069396faf03SMel Gorman { 20708d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 2071396faf03SMel Gorman if (hugepages_treat_as_movable) 2072396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 2073396faf03SMel Gorman else 2074396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 2075396faf03SMel Gorman return 0; 2076396faf03SMel Gorman } 2077396faf03SMel Gorman 2078a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 20798d65af78SAlexey Dobriyan void __user *buffer, 2080a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2081a3d0c6aaSNishanth Aravamudan { 2082a5516438SAndi Kleen struct hstate *h = &default_hstate; 2083e5ff2159SAndi Kleen unsigned long tmp; 208408d4a246SMichal Hocko int ret; 2085e5ff2159SAndi Kleen 2086e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2087e5ff2159SAndi Kleen 2088adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2089adbe8726SEric B Munson return -EINVAL; 2090adbe8726SEric B Munson 2091e5ff2159SAndi Kleen table->data = &tmp; 2092e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 209308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 209408d4a246SMichal Hocko if (ret) 209508d4a246SMichal Hocko goto out; 2096e5ff2159SAndi Kleen 2097e5ff2159SAndi Kleen if (write) { 2098064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2099e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2100a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2101e5ff2159SAndi Kleen } 210208d4a246SMichal Hocko out: 210308d4a246SMichal Hocko return ret; 2104a3d0c6aaSNishanth Aravamudan } 2105a3d0c6aaSNishanth Aravamudan 21061da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 21071da177e4SLinus Torvalds 2108e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 21091da177e4SLinus Torvalds { 2110a5516438SAndi Kleen struct hstate *h = &default_hstate; 2111e1759c21SAlexey Dobriyan seq_printf(m, 21121da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 21131da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2114b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 21157893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 21164f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2117a5516438SAndi Kleen h->nr_huge_pages, 2118a5516438SAndi Kleen h->free_huge_pages, 2119a5516438SAndi Kleen h->resv_huge_pages, 2120a5516438SAndi Kleen h->surplus_huge_pages, 2121a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 21221da177e4SLinus Torvalds } 21231da177e4SLinus Torvalds 21241da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 21251da177e4SLinus Torvalds { 2126a5516438SAndi Kleen struct hstate *h = &default_hstate; 21271da177e4SLinus Torvalds return sprintf(buf, 21281da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2129a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2130a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2131a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2132a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2133a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 21341da177e4SLinus Torvalds } 21351da177e4SLinus Torvalds 21361da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 21371da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 21381da177e4SLinus Torvalds { 2139a5516438SAndi Kleen struct hstate *h = &default_hstate; 2140a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 21411da177e4SLinus Torvalds } 21421da177e4SLinus Torvalds 2143a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2144fc1b8a73SMel Gorman { 2145fc1b8a73SMel Gorman int ret = -ENOMEM; 2146fc1b8a73SMel Gorman 2147fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2148fc1b8a73SMel Gorman /* 2149fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2150fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2151fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2152fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2153fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2154fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2155fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2156fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2157fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2158fc1b8a73SMel Gorman * 2159fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2160fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2161fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2162fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2163fc1b8a73SMel Gorman * semantics that cpuset has. 2164fc1b8a73SMel Gorman */ 2165fc1b8a73SMel Gorman if (delta > 0) { 2166a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2167fc1b8a73SMel Gorman goto out; 2168fc1b8a73SMel Gorman 2169a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2170a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2171fc1b8a73SMel Gorman goto out; 2172fc1b8a73SMel Gorman } 2173fc1b8a73SMel Gorman } 2174fc1b8a73SMel Gorman 2175fc1b8a73SMel Gorman ret = 0; 2176fc1b8a73SMel Gorman if (delta < 0) 2177a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2178fc1b8a73SMel Gorman 2179fc1b8a73SMel Gorman out: 2180fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2181fc1b8a73SMel Gorman return ret; 2182fc1b8a73SMel Gorman } 2183fc1b8a73SMel Gorman 218484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 218584afd99bSAndy Whitcroft { 218684afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 218784afd99bSAndy Whitcroft 218884afd99bSAndy Whitcroft /* 218984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 219084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 219184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 219225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 219384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 219484afd99bSAndy Whitcroft * new reference here without additional locking. 219584afd99bSAndy Whitcroft */ 219684afd99bSAndy Whitcroft if (reservations) 219784afd99bSAndy Whitcroft kref_get(&reservations->refs); 219884afd99bSAndy Whitcroft } 219984afd99bSAndy Whitcroft 2200c50ac050SDave Hansen static void resv_map_put(struct vm_area_struct *vma) 2201c50ac050SDave Hansen { 2202c50ac050SDave Hansen struct resv_map *reservations = vma_resv_map(vma); 2203c50ac050SDave Hansen 2204c50ac050SDave Hansen if (!reservations) 2205c50ac050SDave Hansen return; 2206c50ac050SDave Hansen kref_put(&reservations->refs, resv_map_release); 2207c50ac050SDave Hansen } 2208c50ac050SDave Hansen 2209a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2210a1e78772SMel Gorman { 2211a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 221284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 221390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 221484afd99bSAndy Whitcroft unsigned long reserve; 221584afd99bSAndy Whitcroft unsigned long start; 221684afd99bSAndy Whitcroft unsigned long end; 221784afd99bSAndy Whitcroft 221884afd99bSAndy Whitcroft if (reservations) { 2219a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2220a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 222184afd99bSAndy Whitcroft 222284afd99bSAndy Whitcroft reserve = (end - start) - 222384afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 222484afd99bSAndy Whitcroft 2225c50ac050SDave Hansen resv_map_put(vma); 222684afd99bSAndy Whitcroft 22277251ff78SAdam Litke if (reserve) { 2228a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 222990481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 22307251ff78SAdam Litke } 2231a1e78772SMel Gorman } 223284afd99bSAndy Whitcroft } 2233a1e78772SMel Gorman 22341da177e4SLinus Torvalds /* 22351da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 22361da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 22371da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 22381da177e4SLinus Torvalds * this far. 22391da177e4SLinus Torvalds */ 2240d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 22411da177e4SLinus Torvalds { 22421da177e4SLinus Torvalds BUG(); 2243d0217ac0SNick Piggin return 0; 22441da177e4SLinus Torvalds } 22451da177e4SLinus Torvalds 2246f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2247d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 224884afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2249a1e78772SMel Gorman .close = hugetlb_vm_op_close, 22501da177e4SLinus Torvalds }; 22511da177e4SLinus Torvalds 22521e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 22531e8f889bSDavid Gibson int writable) 225463551ae0SDavid Gibson { 225563551ae0SDavid Gibson pte_t entry; 225663551ae0SDavid Gibson 22571e8f889bSDavid Gibson if (writable) { 225863551ae0SDavid Gibson entry = 225963551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 226063551ae0SDavid Gibson } else { 22617f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 226263551ae0SDavid Gibson } 226363551ae0SDavid Gibson entry = pte_mkyoung(entry); 226463551ae0SDavid Gibson entry = pte_mkhuge(entry); 2265d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 226663551ae0SDavid Gibson 226763551ae0SDavid Gibson return entry; 226863551ae0SDavid Gibson } 226963551ae0SDavid Gibson 22701e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 22711e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 22721e8f889bSDavid Gibson { 22731e8f889bSDavid Gibson pte_t entry; 22741e8f889bSDavid Gibson 22757f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 227632f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 22774b3073e1SRussell King update_mmu_cache(vma, address, ptep); 22781e8f889bSDavid Gibson } 22791e8f889bSDavid Gibson 22801e8f889bSDavid Gibson 228163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 228263551ae0SDavid Gibson struct vm_area_struct *vma) 228363551ae0SDavid Gibson { 228463551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 228563551ae0SDavid Gibson struct page *ptepage; 22861c59827dSHugh Dickins unsigned long addr; 22871e8f889bSDavid Gibson int cow; 2288a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2289a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 22901e8f889bSDavid Gibson 22911e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 229263551ae0SDavid Gibson 2293a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2294c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2295c74df32cSHugh Dickins if (!src_pte) 2296c74df32cSHugh Dickins continue; 2297a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 229863551ae0SDavid Gibson if (!dst_pte) 229963551ae0SDavid Gibson goto nomem; 2300c5c99429SLarry Woodman 2301c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2302c5c99429SLarry Woodman if (dst_pte == src_pte) 2303c5c99429SLarry Woodman continue; 2304c5c99429SLarry Woodman 2305c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 230646478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 23077f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 23081e8f889bSDavid Gibson if (cow) 23097f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 23107f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 231163551ae0SDavid Gibson ptepage = pte_page(entry); 231263551ae0SDavid Gibson get_page(ptepage); 23130fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 231463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 23151c59827dSHugh Dickins } 23161c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2317c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 231863551ae0SDavid Gibson } 231963551ae0SDavid Gibson return 0; 232063551ae0SDavid Gibson 232163551ae0SDavid Gibson nomem: 232263551ae0SDavid Gibson return -ENOMEM; 232363551ae0SDavid Gibson } 232463551ae0SDavid Gibson 2325290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2326290408d4SNaoya Horiguchi { 2327290408d4SNaoya Horiguchi swp_entry_t swp; 2328290408d4SNaoya Horiguchi 2329290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2330290408d4SNaoya Horiguchi return 0; 2331290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 233232f84528SChris Forbes if (non_swap_entry(swp) && is_migration_entry(swp)) 2333290408d4SNaoya Horiguchi return 1; 233432f84528SChris Forbes else 2335290408d4SNaoya Horiguchi return 0; 2336290408d4SNaoya Horiguchi } 2337290408d4SNaoya Horiguchi 2338fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2339fd6a03edSNaoya Horiguchi { 2340fd6a03edSNaoya Horiguchi swp_entry_t swp; 2341fd6a03edSNaoya Horiguchi 2342fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2343fd6a03edSNaoya Horiguchi return 0; 2344fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 234532f84528SChris Forbes if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2346fd6a03edSNaoya Horiguchi return 1; 234732f84528SChris Forbes else 2348fd6a03edSNaoya Horiguchi return 0; 2349fd6a03edSNaoya Horiguchi } 2350fd6a03edSNaoya Horiguchi 235124669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 235224669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 235324669e58SAneesh Kumar K.V struct page *ref_page) 235463551ae0SDavid Gibson { 235524669e58SAneesh Kumar K.V int force_flush = 0; 235663551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 235763551ae0SDavid Gibson unsigned long address; 2358c7546f8fSDavid Gibson pte_t *ptep; 235963551ae0SDavid Gibson pte_t pte; 236063551ae0SDavid Gibson struct page *page; 2361a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2362a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 23632ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 23642ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 2365a5516438SAndi Kleen 236663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2367a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2368a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 236963551ae0SDavid Gibson 237024669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 23712ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 237224669e58SAneesh Kumar K.V again: 2373508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2374a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2375c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2376c7546f8fSDavid Gibson if (!ptep) 2377c7546f8fSDavid Gibson continue; 2378c7546f8fSDavid Gibson 237939dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 238039dde65cSChen, Kenneth W continue; 238139dde65cSChen, Kenneth W 23826629326bSHillf Danton pte = huge_ptep_get(ptep); 23836629326bSHillf Danton if (huge_pte_none(pte)) 23846629326bSHillf Danton continue; 23856629326bSHillf Danton 23866629326bSHillf Danton /* 23876629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 23886629326bSHillf Danton */ 23896629326bSHillf Danton if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 23906629326bSHillf Danton continue; 23916629326bSHillf Danton 23926629326bSHillf Danton page = pte_page(pte); 239304f2cbe3SMel Gorman /* 239404f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 239504f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 239604f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 239704f2cbe3SMel Gorman */ 239804f2cbe3SMel Gorman if (ref_page) { 239904f2cbe3SMel Gorman if (page != ref_page) 240004f2cbe3SMel Gorman continue; 240104f2cbe3SMel Gorman 240204f2cbe3SMel Gorman /* 240304f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 240404f2cbe3SMel Gorman * future faults in this VMA will fail rather than 240504f2cbe3SMel Gorman * looking like data was lost 240604f2cbe3SMel Gorman */ 240704f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 240804f2cbe3SMel Gorman } 240904f2cbe3SMel Gorman 2410c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 241124669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 24126649a386SKen Chen if (pte_dirty(pte)) 24136649a386SKen Chen set_page_dirty(page); 24149e81130bSHillf Danton 241524669e58SAneesh Kumar K.V page_remove_rmap(page); 241624669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 241724669e58SAneesh Kumar K.V if (force_flush) 241824669e58SAneesh Kumar K.V break; 24199e81130bSHillf Danton /* Bail out after unmapping reference page if supplied */ 24209e81130bSHillf Danton if (ref_page) 24219e81130bSHillf Danton break; 242263551ae0SDavid Gibson } 2423cd2934a3SAl Viro spin_unlock(&mm->page_table_lock); 242424669e58SAneesh Kumar K.V /* 242524669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 242624669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 242724669e58SAneesh Kumar K.V * and page-free while holding it. 242824669e58SAneesh Kumar K.V */ 242924669e58SAneesh Kumar K.V if (force_flush) { 243024669e58SAneesh Kumar K.V force_flush = 0; 243124669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 243224669e58SAneesh Kumar K.V if (address < end && !ref_page) 243324669e58SAneesh Kumar K.V goto again; 2434fe1668aeSChen, Kenneth W } 24352ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 243624669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 24371da177e4SLinus Torvalds } 243863551ae0SDavid Gibson 2439d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2440d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2441d833352aSMel Gorman unsigned long end, struct page *ref_page) 2442d833352aSMel Gorman { 2443d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2444d833352aSMel Gorman 2445d833352aSMel Gorman /* 2446d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2447d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2448d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2449d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2450d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2451d833352aSMel Gorman * is to clear it before releasing the i_mmap_mutex. This works 2452d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2453d833352aSMel Gorman * destroyed and the i_mmap_mutex is held. 2454d833352aSMel Gorman */ 2455d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2456d833352aSMel Gorman } 2457d833352aSMel Gorman 2458502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 245904f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2460502717f4SChen, Kenneth W { 246124669e58SAneesh Kumar K.V struct mm_struct *mm; 246224669e58SAneesh Kumar K.V struct mmu_gather tlb; 246324669e58SAneesh Kumar K.V 246424669e58SAneesh Kumar K.V mm = vma->vm_mm; 246524669e58SAneesh Kumar K.V 246624669e58SAneesh Kumar K.V tlb_gather_mmu(&tlb, mm, 0); 246724669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 246824669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2469502717f4SChen, Kenneth W } 2470502717f4SChen, Kenneth W 247104f2cbe3SMel Gorman /* 247204f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 247304f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 247404f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 247504f2cbe3SMel Gorman * same region. 247604f2cbe3SMel Gorman */ 24772a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 24782a4b3dedSHarvey Harrison struct page *page, unsigned long address) 247904f2cbe3SMel Gorman { 24807526674dSAdam Litke struct hstate *h = hstate_vma(vma); 248104f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 248204f2cbe3SMel Gorman struct address_space *mapping; 248304f2cbe3SMel Gorman pgoff_t pgoff; 248404f2cbe3SMel Gorman 248504f2cbe3SMel Gorman /* 248604f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 248704f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 248804f2cbe3SMel Gorman */ 24897526674dSAdam Litke address = address & huge_page_mask(h); 249036e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 249136e4f20aSMichal Hocko vma->vm_pgoff; 249290481622SDavid Gibson mapping = vma->vm_file->f_dentry->d_inode->i_mapping; 249304f2cbe3SMel Gorman 24944eb2b1dcSMel Gorman /* 24954eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 24964eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 24974eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 24984eb2b1dcSMel Gorman */ 24993d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 25006b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 250104f2cbe3SMel Gorman /* Do not unmap the current VMA */ 250204f2cbe3SMel Gorman if (iter_vma == vma) 250304f2cbe3SMel Gorman continue; 250404f2cbe3SMel Gorman 250504f2cbe3SMel Gorman /* 250604f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 250704f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 250804f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 250904f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 251004f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 251104f2cbe3SMel Gorman */ 251204f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 251324669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 251424669e58SAneesh Kumar K.V address + huge_page_size(h), page); 251504f2cbe3SMel Gorman } 25163d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 251704f2cbe3SMel Gorman 251804f2cbe3SMel Gorman return 1; 251904f2cbe3SMel Gorman } 252004f2cbe3SMel Gorman 25210fe6e20bSNaoya Horiguchi /* 25220fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2523ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2524ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2525ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 25260fe6e20bSNaoya Horiguchi */ 25271e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 252804f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 252904f2cbe3SMel Gorman struct page *pagecache_page) 25301e8f889bSDavid Gibson { 2531a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 25321e8f889bSDavid Gibson struct page *old_page, *new_page; 253379ac6ba4SDavid Gibson int avoidcopy; 253404f2cbe3SMel Gorman int outside_reserve = 0; 25352ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 25362ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 25371e8f889bSDavid Gibson 25381e8f889bSDavid Gibson old_page = pte_page(pte); 25391e8f889bSDavid Gibson 254004f2cbe3SMel Gorman retry_avoidcopy: 25411e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 25421e8f889bSDavid Gibson * and just make the page writable */ 25430fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 25441e8f889bSDavid Gibson if (avoidcopy) { 25450fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 25460fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 25471e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 254883c54070SNick Piggin return 0; 25491e8f889bSDavid Gibson } 25501e8f889bSDavid Gibson 255104f2cbe3SMel Gorman /* 255204f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 255304f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 255404f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 255504f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 255604f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 255704f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 255804f2cbe3SMel Gorman * of the full address range. 255904f2cbe3SMel Gorman */ 2560f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 256104f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 256204f2cbe3SMel Gorman old_page != pagecache_page) 256304f2cbe3SMel Gorman outside_reserve = 1; 256404f2cbe3SMel Gorman 25651e8f889bSDavid Gibson page_cache_get(old_page); 2566b76c8cfbSLarry Woodman 2567b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2568b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 256904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 25701e8f889bSDavid Gibson 25712fc39cecSAdam Litke if (IS_ERR(new_page)) { 257276dcee75SAneesh Kumar K.V long err = PTR_ERR(new_page); 25731e8f889bSDavid Gibson page_cache_release(old_page); 257404f2cbe3SMel Gorman 257504f2cbe3SMel Gorman /* 257604f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 257704f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 257804f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 257904f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 258004f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 258104f2cbe3SMel Gorman */ 258204f2cbe3SMel Gorman if (outside_reserve) { 258304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 258404f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 258504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2586b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2587a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2588a734bcc8SHillf Danton if (likely(pte_same(huge_ptep_get(ptep), pte))) 258904f2cbe3SMel Gorman goto retry_avoidcopy; 2590a734bcc8SHillf Danton /* 2591a734bcc8SHillf Danton * race occurs while re-acquiring page_table_lock, and 2592a734bcc8SHillf Danton * our job is done. 2593a734bcc8SHillf Danton */ 2594a734bcc8SHillf Danton return 0; 259504f2cbe3SMel Gorman } 259604f2cbe3SMel Gorman WARN_ON_ONCE(1); 259704f2cbe3SMel Gorman } 259804f2cbe3SMel Gorman 2599b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2600b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 260176dcee75SAneesh Kumar K.V if (err == -ENOMEM) 260276dcee75SAneesh Kumar K.V return VM_FAULT_OOM; 260376dcee75SAneesh Kumar K.V else 260476dcee75SAneesh Kumar K.V return VM_FAULT_SIGBUS; 26051e8f889bSDavid Gibson } 26061e8f889bSDavid Gibson 26070fe6e20bSNaoya Horiguchi /* 26080fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 26090fe6e20bSNaoya Horiguchi * anon_vma prepared. 26100fe6e20bSNaoya Horiguchi */ 261144e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2612ea4039a3SHillf Danton page_cache_release(new_page); 2613ea4039a3SHillf Danton page_cache_release(old_page); 261444e2aa93SDean Nelson /* Caller expects lock to be held */ 261544e2aa93SDean Nelson spin_lock(&mm->page_table_lock); 26160fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 261744e2aa93SDean Nelson } 26180fe6e20bSNaoya Horiguchi 261947ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 262047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 26210ed361deSNick Piggin __SetPageUptodate(new_page); 26221e8f889bSDavid Gibson 26232ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 26242ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 26252ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2626b76c8cfbSLarry Woodman /* 2627b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2628b76c8cfbSLarry Woodman * before the page tables are altered 2629b76c8cfbSLarry Woodman */ 2630b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2631a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 26327f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 26331e8f889bSDavid Gibson /* Break COW */ 26348fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 26351e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 26361e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 26370fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2638cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 26391e8f889bSDavid Gibson /* Make the old page be freed below */ 26401e8f889bSDavid Gibson new_page = old_page; 26411e8f889bSDavid Gibson } 26422ec74c3eSSagi Grimberg spin_unlock(&mm->page_table_lock); 26432ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 26442ec74c3eSSagi Grimberg /* Caller expects lock to be held */ 26452ec74c3eSSagi Grimberg spin_lock(&mm->page_table_lock); 26461e8f889bSDavid Gibson page_cache_release(new_page); 26471e8f889bSDavid Gibson page_cache_release(old_page); 264883c54070SNick Piggin return 0; 26491e8f889bSDavid Gibson } 26501e8f889bSDavid Gibson 265104f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2652a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2653a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 265404f2cbe3SMel Gorman { 265504f2cbe3SMel Gorman struct address_space *mapping; 2656e7c4b0bfSAndy Whitcroft pgoff_t idx; 265704f2cbe3SMel Gorman 265804f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2659a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 266004f2cbe3SMel Gorman 266104f2cbe3SMel Gorman return find_lock_page(mapping, idx); 266204f2cbe3SMel Gorman } 266304f2cbe3SMel Gorman 26643ae77f43SHugh Dickins /* 26653ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 26663ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 26673ae77f43SHugh Dickins */ 26683ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 26692a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 26702a15efc9SHugh Dickins { 26712a15efc9SHugh Dickins struct address_space *mapping; 26722a15efc9SHugh Dickins pgoff_t idx; 26732a15efc9SHugh Dickins struct page *page; 26742a15efc9SHugh Dickins 26752a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 26762a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 26772a15efc9SHugh Dickins 26782a15efc9SHugh Dickins page = find_get_page(mapping, idx); 26792a15efc9SHugh Dickins if (page) 26802a15efc9SHugh Dickins put_page(page); 26812a15efc9SHugh Dickins return page != NULL; 26822a15efc9SHugh Dickins } 26832a15efc9SHugh Dickins 2684a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2685788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2686ac9b9c66SHugh Dickins { 2687a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2688ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2689409eb8c2SHillf Danton int anon_rmap = 0; 2690e7c4b0bfSAndy Whitcroft pgoff_t idx; 26914c887265SAdam Litke unsigned long size; 26924c887265SAdam Litke struct page *page; 26934c887265SAdam Litke struct address_space *mapping; 26941e8f889bSDavid Gibson pte_t new_pte; 26954c887265SAdam Litke 269604f2cbe3SMel Gorman /* 269704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 269804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 269925985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 270004f2cbe3SMel Gorman */ 270104f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 270204f2cbe3SMel Gorman printk(KERN_WARNING 270304f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 270404f2cbe3SMel Gorman current->pid); 270504f2cbe3SMel Gorman return ret; 270604f2cbe3SMel Gorman } 270704f2cbe3SMel Gorman 27084c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2709a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 27104c887265SAdam Litke 27114c887265SAdam Litke /* 27124c887265SAdam Litke * Use page lock to guard against racing truncation 27134c887265SAdam Litke * before we get page_table_lock. 27144c887265SAdam Litke */ 27156bda666aSChristoph Lameter retry: 27166bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 27176bda666aSChristoph Lameter if (!page) { 2718a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2719ebed4bfcSHugh Dickins if (idx >= size) 2720ebed4bfcSHugh Dickins goto out; 272104f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 27222fc39cecSAdam Litke if (IS_ERR(page)) { 272376dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 272476dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 272576dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 272676dcee75SAneesh Kumar K.V else 272776dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 27286bda666aSChristoph Lameter goto out; 27296bda666aSChristoph Lameter } 273047ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 27310ed361deSNick Piggin __SetPageUptodate(page); 2732ac9b9c66SHugh Dickins 2733f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 27346bda666aSChristoph Lameter int err; 273545c682a6SKen Chen struct inode *inode = mapping->host; 27366bda666aSChristoph Lameter 27376bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 27386bda666aSChristoph Lameter if (err) { 27396bda666aSChristoph Lameter put_page(page); 27406bda666aSChristoph Lameter if (err == -EEXIST) 27416bda666aSChristoph Lameter goto retry; 27426bda666aSChristoph Lameter goto out; 27436bda666aSChristoph Lameter } 274445c682a6SKen Chen 274545c682a6SKen Chen spin_lock(&inode->i_lock); 2746a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 274745c682a6SKen Chen spin_unlock(&inode->i_lock); 274823be7468SMel Gorman } else { 27496bda666aSChristoph Lameter lock_page(page); 27500fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 27510fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 27520fe6e20bSNaoya Horiguchi goto backout_unlocked; 275323be7468SMel Gorman } 2754409eb8c2SHillf Danton anon_rmap = 1; 27550fe6e20bSNaoya Horiguchi } 27560fe6e20bSNaoya Horiguchi } else { 275757303d80SAndy Whitcroft /* 2758998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2759998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2760998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2761fd6a03edSNaoya Horiguchi */ 2762fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2763aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2764972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2765fd6a03edSNaoya Horiguchi goto backout_unlocked; 27666bda666aSChristoph Lameter } 2767998b4382SNaoya Horiguchi } 27681e8f889bSDavid Gibson 276957303d80SAndy Whitcroft /* 277057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 277157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 277257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 277357303d80SAndy Whitcroft * the spinlock. 277457303d80SAndy Whitcroft */ 2775788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 27762b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 27772b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 27782b26736cSAndy Whitcroft goto backout_unlocked; 27792b26736cSAndy Whitcroft } 278057303d80SAndy Whitcroft 2781ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2782a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 27834c887265SAdam Litke if (idx >= size) 27844c887265SAdam Litke goto backout; 27854c887265SAdam Litke 278683c54070SNick Piggin ret = 0; 27877f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 27884c887265SAdam Litke goto backout; 27894c887265SAdam Litke 2790409eb8c2SHillf Danton if (anon_rmap) 2791409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 2792409eb8c2SHillf Danton else 2793409eb8c2SHillf Danton page_dup_rmap(page); 27941e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 27951e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 27961e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 27971e8f889bSDavid Gibson 2798788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 27991e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 280004f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 28011e8f889bSDavid Gibson } 28021e8f889bSDavid Gibson 2803ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 28044c887265SAdam Litke unlock_page(page); 28054c887265SAdam Litke out: 2806ac9b9c66SHugh Dickins return ret; 28074c887265SAdam Litke 28084c887265SAdam Litke backout: 28094c887265SAdam Litke spin_unlock(&mm->page_table_lock); 28102b26736cSAndy Whitcroft backout_unlocked: 28114c887265SAdam Litke unlock_page(page); 28124c887265SAdam Litke put_page(page); 28134c887265SAdam Litke goto out; 2814ac9b9c66SHugh Dickins } 2815ac9b9c66SHugh Dickins 281686e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2817788c7df4SHugh Dickins unsigned long address, unsigned int flags) 281886e5216fSAdam Litke { 281986e5216fSAdam Litke pte_t *ptep; 282086e5216fSAdam Litke pte_t entry; 28211e8f889bSDavid Gibson int ret; 28220fe6e20bSNaoya Horiguchi struct page *page = NULL; 282357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 28243935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2825a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 282686e5216fSAdam Litke 28271e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 28281e16a539SKAMEZAWA Hiroyuki 2829fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2830fd6a03edSNaoya Horiguchi if (ptep) { 2831fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2832290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 2833290408d4SNaoya Horiguchi migration_entry_wait(mm, (pmd_t *)ptep, address); 2834290408d4SNaoya Horiguchi return 0; 2835290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2836aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2837972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2838fd6a03edSNaoya Horiguchi } 2839fd6a03edSNaoya Horiguchi 2840a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 284186e5216fSAdam Litke if (!ptep) 284286e5216fSAdam Litke return VM_FAULT_OOM; 284386e5216fSAdam Litke 28443935baa9SDavid Gibson /* 28453935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 28463935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 28473935baa9SDavid Gibson * the same page in the page cache. 28483935baa9SDavid Gibson */ 28493935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 28507f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 28517f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2852788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2853b4d1d99fSDavid Gibson goto out_mutex; 28543935baa9SDavid Gibson } 285586e5216fSAdam Litke 285683c54070SNick Piggin ret = 0; 28571e8f889bSDavid Gibson 285857303d80SAndy Whitcroft /* 285957303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 286057303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 286157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 286257303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 286357303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 286457303d80SAndy Whitcroft * consumed. 286557303d80SAndy Whitcroft */ 2866788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 28672b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28682b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2869b4d1d99fSDavid Gibson goto out_mutex; 28702b26736cSAndy Whitcroft } 287157303d80SAndy Whitcroft 2872f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 287357303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 287457303d80SAndy Whitcroft vma, address); 287557303d80SAndy Whitcroft } 287657303d80SAndy Whitcroft 287756c9cfb1SNaoya Horiguchi /* 287856c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 287956c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 288056c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 288156c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 288256c9cfb1SNaoya Horiguchi * so no worry about deadlock. 288356c9cfb1SNaoya Horiguchi */ 28840fe6e20bSNaoya Horiguchi page = pte_page(entry); 288566aebce7SChris Metcalf get_page(page); 288656c9cfb1SNaoya Horiguchi if (page != pagecache_page) 28870fe6e20bSNaoya Horiguchi lock_page(page); 28880fe6e20bSNaoya Horiguchi 28891e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 28901e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2891b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2892b4d1d99fSDavid Gibson goto out_page_table_lock; 2893b4d1d99fSDavid Gibson 2894b4d1d99fSDavid Gibson 2895788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2896b4d1d99fSDavid Gibson if (!pte_write(entry)) { 289757303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 289857303d80SAndy Whitcroft pagecache_page); 2899b4d1d99fSDavid Gibson goto out_page_table_lock; 2900b4d1d99fSDavid Gibson } 2901b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2902b4d1d99fSDavid Gibson } 2903b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2904788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2905788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 29064b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2907b4d1d99fSDavid Gibson 2908b4d1d99fSDavid Gibson out_page_table_lock: 29091e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 291057303d80SAndy Whitcroft 291157303d80SAndy Whitcroft if (pagecache_page) { 291257303d80SAndy Whitcroft unlock_page(pagecache_page); 291357303d80SAndy Whitcroft put_page(pagecache_page); 291457303d80SAndy Whitcroft } 29151f64d69cSDean Nelson if (page != pagecache_page) 291656c9cfb1SNaoya Horiguchi unlock_page(page); 291766aebce7SChris Metcalf put_page(page); 291857303d80SAndy Whitcroft 2919b4d1d99fSDavid Gibson out_mutex: 29203935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 29211e8f889bSDavid Gibson 29221e8f889bSDavid Gibson return ret; 292386e5216fSAdam Litke } 292486e5216fSAdam Litke 2925ceb86879SAndi Kleen /* Can be overriden by architectures */ 2926ceb86879SAndi Kleen __attribute__((weak)) struct page * 2927ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2928ceb86879SAndi Kleen pud_t *pud, int write) 2929ceb86879SAndi Kleen { 2930ceb86879SAndi Kleen BUG(); 2931ceb86879SAndi Kleen return NULL; 2932ceb86879SAndi Kleen } 2933ceb86879SAndi Kleen 293463551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 293563551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 29365b23dbe8SAdam Litke unsigned long *position, int *length, int i, 29372a15efc9SHugh Dickins unsigned int flags) 293863551ae0SDavid Gibson { 2939d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2940d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 294163551ae0SDavid Gibson int remainder = *length; 2942a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 294363551ae0SDavid Gibson 29441c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 294563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 294663551ae0SDavid Gibson pte_t *pte; 29472a15efc9SHugh Dickins int absent; 294863551ae0SDavid Gibson struct page *page; 294963551ae0SDavid Gibson 29504c887265SAdam Litke /* 29514c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 29522a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 29534c887265SAdam Litke * first, for the page indexing below to work. 29544c887265SAdam Litke */ 2955a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 29562a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 295763551ae0SDavid Gibson 29582a15efc9SHugh Dickins /* 29592a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 29603ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 29613ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 29623ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 29633ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 29642a15efc9SHugh Dickins */ 29653ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 29663ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 29672a15efc9SHugh Dickins remainder = 0; 29682a15efc9SHugh Dickins break; 29692a15efc9SHugh Dickins } 29702a15efc9SHugh Dickins 29712a15efc9SHugh Dickins if (absent || 29722a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 29734c887265SAdam Litke int ret; 29744c887265SAdam Litke 29754c887265SAdam Litke spin_unlock(&mm->page_table_lock); 29762a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 29772a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 29784c887265SAdam Litke spin_lock(&mm->page_table_lock); 2979a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 29804c887265SAdam Litke continue; 29814c887265SAdam Litke 29821c59827dSHugh Dickins remainder = 0; 29831c59827dSHugh Dickins break; 29841c59827dSHugh Dickins } 298563551ae0SDavid Gibson 2986a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 29877f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2988d5d4b0aaSChen, Kenneth W same_page: 2989d6692183SChen, Kenneth W if (pages) { 299069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 29914b2e38adSKOSAKI Motohiro get_page(pages[i]); 2992d6692183SChen, Kenneth W } 299363551ae0SDavid Gibson 299463551ae0SDavid Gibson if (vmas) 299563551ae0SDavid Gibson vmas[i] = vma; 299663551ae0SDavid Gibson 299763551ae0SDavid Gibson vaddr += PAGE_SIZE; 2998d5d4b0aaSChen, Kenneth W ++pfn_offset; 299963551ae0SDavid Gibson --remainder; 300063551ae0SDavid Gibson ++i; 3001d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3002a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3003d5d4b0aaSChen, Kenneth W /* 3004d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3005d5d4b0aaSChen, Kenneth W * of this compound page. 3006d5d4b0aaSChen, Kenneth W */ 3007d5d4b0aaSChen, Kenneth W goto same_page; 3008d5d4b0aaSChen, Kenneth W } 300963551ae0SDavid Gibson } 30101c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 301163551ae0SDavid Gibson *length = remainder; 301263551ae0SDavid Gibson *position = vaddr; 301363551ae0SDavid Gibson 30142a15efc9SHugh Dickins return i ? i : -EFAULT; 301563551ae0SDavid Gibson } 30168f860591SZhang, Yanmin 30178f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 30188f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 30198f860591SZhang, Yanmin { 30208f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 30218f860591SZhang, Yanmin unsigned long start = address; 30228f860591SZhang, Yanmin pte_t *ptep; 30238f860591SZhang, Yanmin pte_t pte; 3024a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 30258f860591SZhang, Yanmin 30268f860591SZhang, Yanmin BUG_ON(address >= end); 30278f860591SZhang, Yanmin flush_cache_range(vma, address, end); 30288f860591SZhang, Yanmin 30293d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 30308f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 3031a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 30328f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 30338f860591SZhang, Yanmin if (!ptep) 30348f860591SZhang, Yanmin continue; 303539dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 303639dde65cSChen, Kenneth W continue; 30377f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 30388f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 30398f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 30408f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 30418f860591SZhang, Yanmin } 30428f860591SZhang, Yanmin } 30438f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 3044d833352aSMel Gorman /* 3045d833352aSMel Gorman * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3046d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3047d833352aSMel Gorman * once we release i_mmap_mutex, another task can do the final put_page 3048d833352aSMel Gorman * and that page table be reused and filled with junk. 3049d833352aSMel Gorman */ 30508f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 3051d833352aSMel Gorman mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 30528f860591SZhang, Yanmin } 30538f860591SZhang, Yanmin 3054a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3055a1e78772SMel Gorman long from, long to, 30565a6fe125SMel Gorman struct vm_area_struct *vma, 3057ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3058e4e574b7SAdam Litke { 305917c9d12eSMel Gorman long ret, chg; 3060a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 306190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 3062e4e574b7SAdam Litke 3063a1e78772SMel Gorman /* 306417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 306517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 306690481622SDavid Gibson * without using reserves 306717c9d12eSMel Gorman */ 3068ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 306917c9d12eSMel Gorman return 0; 307017c9d12eSMel Gorman 307117c9d12eSMel Gorman /* 3072a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3073a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3074a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3075a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3076a1e78772SMel Gorman */ 3077f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 3078e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 30795a6fe125SMel Gorman else { 30805a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 30815a6fe125SMel Gorman if (!resv_map) 30825a6fe125SMel Gorman return -ENOMEM; 30835a6fe125SMel Gorman 308417c9d12eSMel Gorman chg = to - from; 308517c9d12eSMel Gorman 30865a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 30875a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 30885a6fe125SMel Gorman } 30895a6fe125SMel Gorman 3090c50ac050SDave Hansen if (chg < 0) { 3091c50ac050SDave Hansen ret = chg; 3092c50ac050SDave Hansen goto out_err; 3093c50ac050SDave Hansen } 309417c9d12eSMel Gorman 309590481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3096c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3097c50ac050SDave Hansen ret = -ENOSPC; 3098c50ac050SDave Hansen goto out_err; 3099c50ac050SDave Hansen } 310017c9d12eSMel Gorman 310117c9d12eSMel Gorman /* 310217c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 310390481622SDavid Gibson * Hand the pages back to the subpool if there are not 310417c9d12eSMel Gorman */ 310517c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 310617c9d12eSMel Gorman if (ret < 0) { 310790481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3108c50ac050SDave Hansen goto out_err; 310917c9d12eSMel Gorman } 311017c9d12eSMel Gorman 311117c9d12eSMel Gorman /* 311217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 311317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 311417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 311517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 311617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 311717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 311817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 311917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 312017c9d12eSMel Gorman * else has to be done for private mappings here 312117c9d12eSMel Gorman */ 3122f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 312317c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 3124a43a8c39SChen, Kenneth W return 0; 3125c50ac050SDave Hansen out_err: 31264523e145SDave Hansen if (vma) 3127c50ac050SDave Hansen resv_map_put(vma); 3128c50ac050SDave Hansen return ret; 3129a43a8c39SChen, Kenneth W } 3130a43a8c39SChen, Kenneth W 3131a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3132a43a8c39SChen, Kenneth W { 3133a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 3134a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 313590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 313645c682a6SKen Chen 313745c682a6SKen Chen spin_lock(&inode->i_lock); 3138e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 313945c682a6SKen Chen spin_unlock(&inode->i_lock); 314045c682a6SKen Chen 314190481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3142a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3143a43a8c39SChen, Kenneth W } 314493f70f90SNaoya Horiguchi 3145d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3146d5bd9106SAndi Kleen 31476de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 31486de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 31496de2b1aaSNaoya Horiguchi { 31506de2b1aaSNaoya Horiguchi struct page *page; 31516de2b1aaSNaoya Horiguchi struct page *tmp; 31526de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 31536de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 31546de2b1aaSNaoya Horiguchi 31556de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 31566de2b1aaSNaoya Horiguchi if (page == hpage) 31576de2b1aaSNaoya Horiguchi return 1; 31586de2b1aaSNaoya Horiguchi return 0; 31596de2b1aaSNaoya Horiguchi } 31606de2b1aaSNaoya Horiguchi 316193f70f90SNaoya Horiguchi /* 316293f70f90SNaoya Horiguchi * This function is called from memory failure code. 316393f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 316493f70f90SNaoya Horiguchi */ 31656de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 316693f70f90SNaoya Horiguchi { 316793f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 316893f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 31696de2b1aaSNaoya Horiguchi int ret = -EBUSY; 317093f70f90SNaoya Horiguchi 317193f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 31726de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 317393f70f90SNaoya Horiguchi list_del(&hpage->lru); 31748c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 317593f70f90SNaoya Horiguchi h->free_huge_pages--; 317693f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 31776de2b1aaSNaoya Horiguchi ret = 0; 317893f70f90SNaoya Horiguchi } 31796de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 31806de2b1aaSNaoya Horiguchi return ret; 31816de2b1aaSNaoya Horiguchi } 31826de2b1aaSNaoya Horiguchi #endif 3183