11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 31da177e4SLinus Torvalds * (C) William Irwin, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24d6606683SLinus Torvalds 2563551ae0SDavid Gibson #include <asm/page.h> 2663551ae0SDavid Gibson #include <asm/pgtable.h> 2724669e58SAneesh Kumar K.V #include <asm/tlb.h> 2863551ae0SDavid Gibson 2924669e58SAneesh Kumar K.V #include <linux/io.h> 3063551ae0SDavid Gibson #include <linux/hugetlb.h> 319dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 329a305230SLee Schermerhorn #include <linux/node.h> 33abb8206cSAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 347835e98bSNick Piggin #include "internal.h" 351da177e4SLinus Torvalds 361da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 37396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 38396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 39a5516438SAndi Kleen 40c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 41e5ff2159SAndi Kleen unsigned int default_hstate_idx; 42e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 43e5ff2159SAndi Kleen 4453ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4553ba51d2SJon Tollefson 46e5ff2159SAndi Kleen /* for command line parsing */ 47e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 48e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 49e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 50e5ff2159SAndi Kleen 513935baa9SDavid Gibson /* 523935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 533935baa9SDavid Gibson */ 54c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 550bd0f9fbSEric Paris 5690481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 5790481622SDavid Gibson { 5890481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 5990481622SDavid Gibson 6090481622SDavid Gibson spin_unlock(&spool->lock); 6190481622SDavid Gibson 6290481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 6390481622SDavid Gibson * remain, free the subpool the subpool remain */ 6490481622SDavid Gibson if (free) 6590481622SDavid Gibson kfree(spool); 6690481622SDavid Gibson } 6790481622SDavid Gibson 6890481622SDavid Gibson struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 6990481622SDavid Gibson { 7090481622SDavid Gibson struct hugepage_subpool *spool; 7190481622SDavid Gibson 7290481622SDavid Gibson spool = kmalloc(sizeof(*spool), GFP_KERNEL); 7390481622SDavid Gibson if (!spool) 7490481622SDavid Gibson return NULL; 7590481622SDavid Gibson 7690481622SDavid Gibson spin_lock_init(&spool->lock); 7790481622SDavid Gibson spool->count = 1; 7890481622SDavid Gibson spool->max_hpages = nr_blocks; 7990481622SDavid Gibson spool->used_hpages = 0; 8090481622SDavid Gibson 8190481622SDavid Gibson return spool; 8290481622SDavid Gibson } 8390481622SDavid Gibson 8490481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 8590481622SDavid Gibson { 8690481622SDavid Gibson spin_lock(&spool->lock); 8790481622SDavid Gibson BUG_ON(!spool->count); 8890481622SDavid Gibson spool->count--; 8990481622SDavid Gibson unlock_or_release_subpool(spool); 9090481622SDavid Gibson } 9190481622SDavid Gibson 9290481622SDavid Gibson static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 9390481622SDavid Gibson long delta) 9490481622SDavid Gibson { 9590481622SDavid Gibson int ret = 0; 9690481622SDavid Gibson 9790481622SDavid Gibson if (!spool) 9890481622SDavid Gibson return 0; 9990481622SDavid Gibson 10090481622SDavid Gibson spin_lock(&spool->lock); 10190481622SDavid Gibson if ((spool->used_hpages + delta) <= spool->max_hpages) { 10290481622SDavid Gibson spool->used_hpages += delta; 10390481622SDavid Gibson } else { 10490481622SDavid Gibson ret = -ENOMEM; 10590481622SDavid Gibson } 10690481622SDavid Gibson spin_unlock(&spool->lock); 10790481622SDavid Gibson 10890481622SDavid Gibson return ret; 10990481622SDavid Gibson } 11090481622SDavid Gibson 11190481622SDavid Gibson static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 11290481622SDavid Gibson long delta) 11390481622SDavid Gibson { 11490481622SDavid Gibson if (!spool) 11590481622SDavid Gibson return; 11690481622SDavid Gibson 11790481622SDavid Gibson spin_lock(&spool->lock); 11890481622SDavid Gibson spool->used_hpages -= delta; 11990481622SDavid Gibson /* If hugetlbfs_put_super couldn't free spool due to 12090481622SDavid Gibson * an outstanding quota reference, free it now. */ 12190481622SDavid Gibson unlock_or_release_subpool(spool); 12290481622SDavid Gibson } 12390481622SDavid Gibson 12490481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 12590481622SDavid Gibson { 12690481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 12790481622SDavid Gibson } 12890481622SDavid Gibson 12990481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 13090481622SDavid Gibson { 13190481622SDavid Gibson return subpool_inode(vma->vm_file->f_dentry->d_inode); 13290481622SDavid Gibson } 13390481622SDavid Gibson 134e7c4b0bfSAndy Whitcroft /* 13596822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 13696822904SAndy Whitcroft * across the pages in a mapping. 13784afd99bSAndy Whitcroft * 13884afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 13984afd99bSAndy Whitcroft * and the hugetlb_instantion_mutex. To access or modify a region the caller 14084afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 14184afd99bSAndy Whitcroft * the hugetlb_instantiation mutex: 14284afd99bSAndy Whitcroft * 14384afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 14484afd99bSAndy Whitcroft * or 14584afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 14684afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 14796822904SAndy Whitcroft */ 14896822904SAndy Whitcroft struct file_region { 14996822904SAndy Whitcroft struct list_head link; 15096822904SAndy Whitcroft long from; 15196822904SAndy Whitcroft long to; 15296822904SAndy Whitcroft }; 15396822904SAndy Whitcroft 15496822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 15596822904SAndy Whitcroft { 15696822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 15796822904SAndy Whitcroft 15896822904SAndy Whitcroft /* Locate the region we are either in or before. */ 15996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16096822904SAndy Whitcroft if (f <= rg->to) 16196822904SAndy Whitcroft break; 16296822904SAndy Whitcroft 16396822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 16496822904SAndy Whitcroft if (f > rg->from) 16596822904SAndy Whitcroft f = rg->from; 16696822904SAndy Whitcroft 16796822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 16896822904SAndy Whitcroft nrg = rg; 16996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 17096822904SAndy Whitcroft if (&rg->link == head) 17196822904SAndy Whitcroft break; 17296822904SAndy Whitcroft if (rg->from > t) 17396822904SAndy Whitcroft break; 17496822904SAndy Whitcroft 17596822904SAndy Whitcroft /* If this area reaches higher then extend our area to 17696822904SAndy Whitcroft * include it completely. If this is not the first area 17796822904SAndy Whitcroft * which we intend to reuse, free it. */ 17896822904SAndy Whitcroft if (rg->to > t) 17996822904SAndy Whitcroft t = rg->to; 18096822904SAndy Whitcroft if (rg != nrg) { 18196822904SAndy Whitcroft list_del(&rg->link); 18296822904SAndy Whitcroft kfree(rg); 18396822904SAndy Whitcroft } 18496822904SAndy Whitcroft } 18596822904SAndy Whitcroft nrg->from = f; 18696822904SAndy Whitcroft nrg->to = t; 18796822904SAndy Whitcroft return 0; 18896822904SAndy Whitcroft } 18996822904SAndy Whitcroft 19096822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 19196822904SAndy Whitcroft { 19296822904SAndy Whitcroft struct file_region *rg, *nrg; 19396822904SAndy Whitcroft long chg = 0; 19496822904SAndy Whitcroft 19596822904SAndy Whitcroft /* Locate the region we are before or in. */ 19696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 19796822904SAndy Whitcroft if (f <= rg->to) 19896822904SAndy Whitcroft break; 19996822904SAndy Whitcroft 20096822904SAndy Whitcroft /* If we are below the current region then a new region is required. 20196822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 20296822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 20396822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 20496822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 20596822904SAndy Whitcroft if (!nrg) 20696822904SAndy Whitcroft return -ENOMEM; 20796822904SAndy Whitcroft nrg->from = f; 20896822904SAndy Whitcroft nrg->to = f; 20996822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 21096822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 21196822904SAndy Whitcroft 21296822904SAndy Whitcroft return t - f; 21396822904SAndy Whitcroft } 21496822904SAndy Whitcroft 21596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 21696822904SAndy Whitcroft if (f > rg->from) 21796822904SAndy Whitcroft f = rg->from; 21896822904SAndy Whitcroft chg = t - f; 21996822904SAndy Whitcroft 22096822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 22196822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 22296822904SAndy Whitcroft if (&rg->link == head) 22396822904SAndy Whitcroft break; 22496822904SAndy Whitcroft if (rg->from > t) 22596822904SAndy Whitcroft return chg; 22696822904SAndy Whitcroft 22725985edcSLucas De Marchi /* We overlap with this area, if it extends further than 22896822904SAndy Whitcroft * us then we must extend ourselves. Account for its 22996822904SAndy Whitcroft * existing reservation. */ 23096822904SAndy Whitcroft if (rg->to > t) { 23196822904SAndy Whitcroft chg += rg->to - t; 23296822904SAndy Whitcroft t = rg->to; 23396822904SAndy Whitcroft } 23496822904SAndy Whitcroft chg -= rg->to - rg->from; 23596822904SAndy Whitcroft } 23696822904SAndy Whitcroft return chg; 23796822904SAndy Whitcroft } 23896822904SAndy Whitcroft 23996822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 24096822904SAndy Whitcroft { 24196822904SAndy Whitcroft struct file_region *rg, *trg; 24296822904SAndy Whitcroft long chg = 0; 24396822904SAndy Whitcroft 24496822904SAndy Whitcroft /* Locate the region we are either in or before. */ 24596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 24696822904SAndy Whitcroft if (end <= rg->to) 24796822904SAndy Whitcroft break; 24896822904SAndy Whitcroft if (&rg->link == head) 24996822904SAndy Whitcroft return 0; 25096822904SAndy Whitcroft 25196822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 25296822904SAndy Whitcroft if (end > rg->from) { 25396822904SAndy Whitcroft chg = rg->to - end; 25496822904SAndy Whitcroft rg->to = end; 25596822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 25696822904SAndy Whitcroft } 25796822904SAndy Whitcroft 25896822904SAndy Whitcroft /* Drop any remaining regions. */ 25996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 26096822904SAndy Whitcroft if (&rg->link == head) 26196822904SAndy Whitcroft break; 26296822904SAndy Whitcroft chg += rg->to - rg->from; 26396822904SAndy Whitcroft list_del(&rg->link); 26496822904SAndy Whitcroft kfree(rg); 26596822904SAndy Whitcroft } 26696822904SAndy Whitcroft return chg; 26796822904SAndy Whitcroft } 26896822904SAndy Whitcroft 26984afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 27084afd99bSAndy Whitcroft { 27184afd99bSAndy Whitcroft struct file_region *rg; 27284afd99bSAndy Whitcroft long chg = 0; 27384afd99bSAndy Whitcroft 27484afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 27584afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 276f2135a4aSWang Sheng-Hui long seg_from; 277f2135a4aSWang Sheng-Hui long seg_to; 27884afd99bSAndy Whitcroft 27984afd99bSAndy Whitcroft if (rg->to <= f) 28084afd99bSAndy Whitcroft continue; 28184afd99bSAndy Whitcroft if (rg->from >= t) 28284afd99bSAndy Whitcroft break; 28384afd99bSAndy Whitcroft 28484afd99bSAndy Whitcroft seg_from = max(rg->from, f); 28584afd99bSAndy Whitcroft seg_to = min(rg->to, t); 28684afd99bSAndy Whitcroft 28784afd99bSAndy Whitcroft chg += seg_to - seg_from; 28884afd99bSAndy Whitcroft } 28984afd99bSAndy Whitcroft 29084afd99bSAndy Whitcroft return chg; 29184afd99bSAndy Whitcroft } 29284afd99bSAndy Whitcroft 29396822904SAndy Whitcroft /* 294e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 295e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 296e7c4b0bfSAndy Whitcroft */ 297a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 298a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 299e7c4b0bfSAndy Whitcroft { 300a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 301a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 302e7c4b0bfSAndy Whitcroft } 303e7c4b0bfSAndy Whitcroft 3040fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 3050fe6e20bSNaoya Horiguchi unsigned long address) 3060fe6e20bSNaoya Horiguchi { 3070fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 3080fe6e20bSNaoya Horiguchi } 3090fe6e20bSNaoya Horiguchi 31084afd99bSAndy Whitcroft /* 31108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 31208fba699SMel Gorman * cases this will be same size as used by the page table entries. 31308fba699SMel Gorman */ 31408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 31508fba699SMel Gorman { 31608fba699SMel Gorman struct hstate *hstate; 31708fba699SMel Gorman 31808fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 31908fba699SMel Gorman return PAGE_SIZE; 32008fba699SMel Gorman 32108fba699SMel Gorman hstate = hstate_vma(vma); 32208fba699SMel Gorman 32308fba699SMel Gorman return 1UL << (hstate->order + PAGE_SHIFT); 32408fba699SMel Gorman } 325f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 32608fba699SMel Gorman 32708fba699SMel Gorman /* 3283340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 3293340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 3303340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 3313340289dSMel Gorman * function is required. 3323340289dSMel Gorman */ 3333340289dSMel Gorman #ifndef vma_mmu_pagesize 3343340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 3353340289dSMel Gorman { 3363340289dSMel Gorman return vma_kernel_pagesize(vma); 3373340289dSMel Gorman } 3383340289dSMel Gorman #endif 3393340289dSMel Gorman 3403340289dSMel Gorman /* 34184afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 34284afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 34384afd99bSAndy Whitcroft * alignment. 34484afd99bSAndy Whitcroft */ 34584afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 34684afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 34704f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 34884afd99bSAndy Whitcroft 349a1e78772SMel Gorman /* 350a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 351a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 352a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 353a1e78772SMel Gorman * 354a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 355a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 356a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 357a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 35884afd99bSAndy Whitcroft * 35984afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 36084afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 36184afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 36284afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 36384afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 36484afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 36584afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 36684afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 367a1e78772SMel Gorman */ 368e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 369e7c4b0bfSAndy Whitcroft { 370e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 371e7c4b0bfSAndy Whitcroft } 372e7c4b0bfSAndy Whitcroft 373e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 374e7c4b0bfSAndy Whitcroft unsigned long value) 375e7c4b0bfSAndy Whitcroft { 376e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 377e7c4b0bfSAndy Whitcroft } 378e7c4b0bfSAndy Whitcroft 37984afd99bSAndy Whitcroft struct resv_map { 38084afd99bSAndy Whitcroft struct kref refs; 38184afd99bSAndy Whitcroft struct list_head regions; 38284afd99bSAndy Whitcroft }; 38384afd99bSAndy Whitcroft 3842a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 38584afd99bSAndy Whitcroft { 38684afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 38784afd99bSAndy Whitcroft if (!resv_map) 38884afd99bSAndy Whitcroft return NULL; 38984afd99bSAndy Whitcroft 39084afd99bSAndy Whitcroft kref_init(&resv_map->refs); 39184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 39284afd99bSAndy Whitcroft 39384afd99bSAndy Whitcroft return resv_map; 39484afd99bSAndy Whitcroft } 39584afd99bSAndy Whitcroft 3962a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 39784afd99bSAndy Whitcroft { 39884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 39984afd99bSAndy Whitcroft 40084afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 40184afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 40284afd99bSAndy Whitcroft kfree(resv_map); 40384afd99bSAndy Whitcroft } 40484afd99bSAndy Whitcroft 40584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 406a1e78772SMel Gorman { 407a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 408f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 40984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 41084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 4112a4b3dedSHarvey Harrison return NULL; 412a1e78772SMel Gorman } 413a1e78772SMel Gorman 41484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 415a1e78772SMel Gorman { 416a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 417f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 418a1e78772SMel Gorman 41984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 42084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 42104f2cbe3SMel Gorman } 42204f2cbe3SMel Gorman 42304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 42404f2cbe3SMel Gorman { 42504f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 426f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 427e7c4b0bfSAndy Whitcroft 428e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 42904f2cbe3SMel Gorman } 43004f2cbe3SMel Gorman 43104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 43204f2cbe3SMel Gorman { 43304f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 434e7c4b0bfSAndy Whitcroft 435e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 436a1e78772SMel Gorman } 437a1e78772SMel Gorman 438a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 439a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 440a5516438SAndi Kleen struct vm_area_struct *vma) 441a1e78772SMel Gorman { 442c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 443c37f9fb1SAndy Whitcroft return; 444c37f9fb1SAndy Whitcroft 445f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 446a1e78772SMel Gorman /* Shared mappings always use reserves */ 447a5516438SAndi Kleen h->resv_huge_pages--; 44884afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 449a1e78772SMel Gorman /* 450a1e78772SMel Gorman * Only the process that called mmap() has reserves for 451a1e78772SMel Gorman * private mappings. 452a1e78772SMel Gorman */ 453a5516438SAndi Kleen h->resv_huge_pages--; 454a1e78772SMel Gorman } 455a1e78772SMel Gorman } 456a1e78772SMel Gorman 45704f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 458a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 459a1e78772SMel Gorman { 460a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 461f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 462a1e78772SMel Gorman vma->vm_private_data = (void *)0; 463a1e78772SMel Gorman } 464a1e78772SMel Gorman 465a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 4667f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 467a1e78772SMel Gorman { 468f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 469a1e78772SMel Gorman return 1; 4707f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4717f09ca51SMel Gorman return 1; 4727f09ca51SMel Gorman return 0; 473a1e78772SMel Gorman } 474a1e78772SMel Gorman 4750ebabb41SNaoya Horiguchi static void copy_gigantic_page(struct page *dst, struct page *src) 4760ebabb41SNaoya Horiguchi { 4770ebabb41SNaoya Horiguchi int i; 4780ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4790ebabb41SNaoya Horiguchi struct page *dst_base = dst; 4800ebabb41SNaoya Horiguchi struct page *src_base = src; 4810ebabb41SNaoya Horiguchi 4820ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); ) { 4830ebabb41SNaoya Horiguchi cond_resched(); 4840ebabb41SNaoya Horiguchi copy_highpage(dst, src); 4850ebabb41SNaoya Horiguchi 4860ebabb41SNaoya Horiguchi i++; 4870ebabb41SNaoya Horiguchi dst = mem_map_next(dst, dst_base, i); 4880ebabb41SNaoya Horiguchi src = mem_map_next(src, src_base, i); 4890ebabb41SNaoya Horiguchi } 4900ebabb41SNaoya Horiguchi } 4910ebabb41SNaoya Horiguchi 4920ebabb41SNaoya Horiguchi void copy_huge_page(struct page *dst, struct page *src) 4930ebabb41SNaoya Horiguchi { 4940ebabb41SNaoya Horiguchi int i; 4950ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4960ebabb41SNaoya Horiguchi 4970ebabb41SNaoya Horiguchi if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 4980ebabb41SNaoya Horiguchi copy_gigantic_page(dst, src); 4990ebabb41SNaoya Horiguchi return; 5000ebabb41SNaoya Horiguchi } 5010ebabb41SNaoya Horiguchi 5020ebabb41SNaoya Horiguchi might_sleep(); 5030ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); i++) { 5040ebabb41SNaoya Horiguchi cond_resched(); 5050ebabb41SNaoya Horiguchi copy_highpage(dst + i, src + i); 5060ebabb41SNaoya Horiguchi } 5070ebabb41SNaoya Horiguchi } 5080ebabb41SNaoya Horiguchi 509a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 5101da177e4SLinus Torvalds { 5111da177e4SLinus Torvalds int nid = page_to_nid(page); 5120edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 513a5516438SAndi Kleen h->free_huge_pages++; 514a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 5151da177e4SLinus Torvalds } 5161da177e4SLinus Torvalds 517bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 518bf50bab2SNaoya Horiguchi { 519bf50bab2SNaoya Horiguchi struct page *page; 520bf50bab2SNaoya Horiguchi 521bf50bab2SNaoya Horiguchi if (list_empty(&h->hugepage_freelists[nid])) 522bf50bab2SNaoya Horiguchi return NULL; 523bf50bab2SNaoya Horiguchi page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); 5240edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 525a9869b83SNaoya Horiguchi set_page_refcounted(page); 526bf50bab2SNaoya Horiguchi h->free_huge_pages--; 527bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 528bf50bab2SNaoya Horiguchi return page; 529bf50bab2SNaoya Horiguchi } 530bf50bab2SNaoya Horiguchi 531a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 532a5516438SAndi Kleen struct vm_area_struct *vma, 53304f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 5341da177e4SLinus Torvalds { 535b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 536480eccf9SLee Schermerhorn struct mempolicy *mpol; 53719770b32SMel Gorman nodemask_t *nodemask; 538c0ff7453SMiao Xie struct zonelist *zonelist; 539dd1a239fSMel Gorman struct zone *zone; 540dd1a239fSMel Gorman struct zoneref *z; 541cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 5421da177e4SLinus Torvalds 543cc9a6c87SMel Gorman retry_cpuset: 544cc9a6c87SMel Gorman cpuset_mems_cookie = get_mems_allowed(); 545c0ff7453SMiao Xie zonelist = huge_zonelist(vma, address, 546c0ff7453SMiao Xie htlb_alloc_mask, &mpol, &nodemask); 547a1e78772SMel Gorman /* 548a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 549a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 550a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 551a1e78772SMel Gorman */ 5527f09ca51SMel Gorman if (!vma_has_reserves(vma) && 553a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 554c0ff7453SMiao Xie goto err; 555a1e78772SMel Gorman 55604f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 557a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 5586eab04a8SJustin P. Mattock goto err; 55904f2cbe3SMel Gorman 56019770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 56119770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 562bf50bab2SNaoya Horiguchi if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { 563bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 564bf50bab2SNaoya Horiguchi if (page) { 56504f2cbe3SMel Gorman if (!avoid_reserve) 566a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 5675ab3ee7bSKen Chen break; 5681da177e4SLinus Torvalds } 5693abf7afdSAndrew Morton } 570bf50bab2SNaoya Horiguchi } 571cc9a6c87SMel Gorman 572cc9a6c87SMel Gorman mpol_cond_put(mpol); 573cc9a6c87SMel Gorman if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) 574cc9a6c87SMel Gorman goto retry_cpuset; 575cc9a6c87SMel Gorman return page; 576cc9a6c87SMel Gorman 577c0ff7453SMiao Xie err: 57852cd3b07SLee Schermerhorn mpol_cond_put(mpol); 579cc9a6c87SMel Gorman return NULL; 5801da177e4SLinus Torvalds } 5811da177e4SLinus Torvalds 582a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5836af2acb6SAdam Litke { 5846af2acb6SAdam Litke int i; 585a5516438SAndi Kleen 58618229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 58718229df5SAndy Whitcroft 588a5516438SAndi Kleen h->nr_huge_pages--; 589a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 590a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 59132f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 59232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 59332f84528SChris Forbes 1 << PG_active | 1 << PG_reserved | 5946af2acb6SAdam Litke 1 << PG_private | 1 << PG_writeback); 5956af2acb6SAdam Litke } 5969dd540e2SAneesh Kumar K.V VM_BUG_ON(hugetlb_cgroup_from_page(page)); 5976af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5986af2acb6SAdam Litke set_page_refcounted(page); 5997f2e9525SGerald Schaefer arch_release_hugepage(page); 600a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 6016af2acb6SAdam Litke } 6026af2acb6SAdam Litke 603e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 604e5ff2159SAndi Kleen { 605e5ff2159SAndi Kleen struct hstate *h; 606e5ff2159SAndi Kleen 607e5ff2159SAndi Kleen for_each_hstate(h) { 608e5ff2159SAndi Kleen if (huge_page_size(h) == size) 609e5ff2159SAndi Kleen return h; 610e5ff2159SAndi Kleen } 611e5ff2159SAndi Kleen return NULL; 612e5ff2159SAndi Kleen } 613e5ff2159SAndi Kleen 61427a85ef1SDavid Gibson static void free_huge_page(struct page *page) 61527a85ef1SDavid Gibson { 616a5516438SAndi Kleen /* 617a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 618a5516438SAndi Kleen * compound page destructor. 619a5516438SAndi Kleen */ 620e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 6217893d1d5SAdam Litke int nid = page_to_nid(page); 62290481622SDavid Gibson struct hugepage_subpool *spool = 62390481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 62427a85ef1SDavid Gibson 625e5df70abSAndy Whitcroft set_page_private(page, 0); 62623be7468SMel Gorman page->mapping = NULL; 6277893d1d5SAdam Litke BUG_ON(page_count(page)); 6280fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 62927a85ef1SDavid Gibson 63027a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 6316d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 6326d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 633aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 6340edaecfaSAneesh Kumar K.V /* remove the page from active list */ 6350edaecfaSAneesh Kumar K.V list_del(&page->lru); 636a5516438SAndi Kleen update_and_free_page(h, page); 637a5516438SAndi Kleen h->surplus_huge_pages--; 638a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 6397893d1d5SAdam Litke } else { 6405d3a551cSWill Deacon arch_clear_hugepage_flags(page); 641a5516438SAndi Kleen enqueue_huge_page(h, page); 6427893d1d5SAdam Litke } 64327a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 64490481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 64527a85ef1SDavid Gibson } 64627a85ef1SDavid Gibson 647a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 648b7ba30c6SAndi Kleen { 6490edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 650b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 651b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 6529dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 653a5516438SAndi Kleen h->nr_huge_pages++; 654a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 655b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 656b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 657b7ba30c6SAndi Kleen } 658b7ba30c6SAndi Kleen 65920a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 66020a0307cSWu Fengguang { 66120a0307cSWu Fengguang int i; 66220a0307cSWu Fengguang int nr_pages = 1 << order; 66320a0307cSWu Fengguang struct page *p = page + 1; 66420a0307cSWu Fengguang 66520a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 66620a0307cSWu Fengguang set_compound_order(page, order); 66720a0307cSWu Fengguang __SetPageHead(page); 66820a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 66920a0307cSWu Fengguang __SetPageTail(p); 67058a84aa9SYouquan Song set_page_count(p, 0); 67120a0307cSWu Fengguang p->first_page = page; 67220a0307cSWu Fengguang } 67320a0307cSWu Fengguang } 67420a0307cSWu Fengguang 67520a0307cSWu Fengguang int PageHuge(struct page *page) 67620a0307cSWu Fengguang { 67720a0307cSWu Fengguang compound_page_dtor *dtor; 67820a0307cSWu Fengguang 67920a0307cSWu Fengguang if (!PageCompound(page)) 68020a0307cSWu Fengguang return 0; 68120a0307cSWu Fengguang 68220a0307cSWu Fengguang page = compound_head(page); 68320a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 68420a0307cSWu Fengguang 68520a0307cSWu Fengguang return dtor == free_huge_page; 68620a0307cSWu Fengguang } 68743131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 68843131e14SNaoya Horiguchi 689a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6901da177e4SLinus Torvalds { 6911da177e4SLinus Torvalds struct page *page; 692f96efd58SJoe Jin 693aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 694aa888a74SAndi Kleen return NULL; 695aa888a74SAndi Kleen 6966484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 697551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 698551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 699a5516438SAndi Kleen huge_page_order(h)); 7001da177e4SLinus Torvalds if (page) { 7017f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 702caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 7037b8ee84dSHarvey Harrison return NULL; 7047f2e9525SGerald Schaefer } 705a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 7061da177e4SLinus Torvalds } 70763b4613cSNishanth Aravamudan 70863b4613cSNishanth Aravamudan return page; 70963b4613cSNishanth Aravamudan } 71063b4613cSNishanth Aravamudan 7115ced66c9SAndi Kleen /* 7126ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 7136ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 7146ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7156ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 7166ae11b27SLee Schermerhorn * node for alloc or free. 7179a76db09SLee Schermerhorn */ 7186ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7199a76db09SLee Schermerhorn { 7206ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 7219a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 7226ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 7239a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 7249a76db09SLee Schermerhorn 7259a76db09SLee Schermerhorn return nid; 7269a76db09SLee Schermerhorn } 7279a76db09SLee Schermerhorn 7286ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 7295ced66c9SAndi Kleen { 7306ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 7316ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 7329a76db09SLee Schermerhorn return nid; 7335ced66c9SAndi Kleen } 7345ced66c9SAndi Kleen 7356ae11b27SLee Schermerhorn /* 7366ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 7376ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 7386ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 7396ae11b27SLee Schermerhorn * mask. 7406ae11b27SLee Schermerhorn */ 7416ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 7426ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 7436ae11b27SLee Schermerhorn { 7446ae11b27SLee Schermerhorn int nid; 7456ae11b27SLee Schermerhorn 7466ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7476ae11b27SLee Schermerhorn 7486ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 7496ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 7506ae11b27SLee Schermerhorn 7516ae11b27SLee Schermerhorn return nid; 7526ae11b27SLee Schermerhorn } 7536ae11b27SLee Schermerhorn 7546ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 75563b4613cSNishanth Aravamudan { 75663b4613cSNishanth Aravamudan struct page *page; 75763b4613cSNishanth Aravamudan int start_nid; 75863b4613cSNishanth Aravamudan int next_nid; 75963b4613cSNishanth Aravamudan int ret = 0; 76063b4613cSNishanth Aravamudan 7616ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 762e8c5c824SLee Schermerhorn next_nid = start_nid; 76363b4613cSNishanth Aravamudan 76463b4613cSNishanth Aravamudan do { 765e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 7669a76db09SLee Schermerhorn if (page) { 76763b4613cSNishanth Aravamudan ret = 1; 7689a76db09SLee Schermerhorn break; 7699a76db09SLee Schermerhorn } 7706ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 7719a76db09SLee Schermerhorn } while (next_nid != start_nid); 77263b4613cSNishanth Aravamudan 7733b116300SAdam Litke if (ret) 7743b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7753b116300SAdam Litke else 7763b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7773b116300SAdam Litke 77863b4613cSNishanth Aravamudan return ret; 7791da177e4SLinus Torvalds } 7801da177e4SLinus Torvalds 781e8c5c824SLee Schermerhorn /* 7826ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7836ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7846ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7856ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 786e8c5c824SLee Schermerhorn */ 7876ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 788e8c5c824SLee Schermerhorn { 7896ae11b27SLee Schermerhorn int nid; 7909a76db09SLee Schermerhorn 7916ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7926ae11b27SLee Schermerhorn 7936ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7946ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7956ae11b27SLee Schermerhorn 7969a76db09SLee Schermerhorn return nid; 797e8c5c824SLee Schermerhorn } 798e8c5c824SLee Schermerhorn 799e8c5c824SLee Schermerhorn /* 800e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 801e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 802e8c5c824SLee Schermerhorn * balanced over allowed nodes. 803e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 804e8c5c824SLee Schermerhorn */ 8056ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 8066ae11b27SLee Schermerhorn bool acct_surplus) 807e8c5c824SLee Schermerhorn { 808e8c5c824SLee Schermerhorn int start_nid; 809e8c5c824SLee Schermerhorn int next_nid; 810e8c5c824SLee Schermerhorn int ret = 0; 811e8c5c824SLee Schermerhorn 8126ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 813e8c5c824SLee Schermerhorn next_nid = start_nid; 814e8c5c824SLee Schermerhorn 815e8c5c824SLee Schermerhorn do { 816685f3457SLee Schermerhorn /* 817685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 818685f3457SLee Schermerhorn * nodes with surplus pages. 819685f3457SLee Schermerhorn */ 820685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 821685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 822e8c5c824SLee Schermerhorn struct page *page = 823e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 824e8c5c824SLee Schermerhorn struct page, lru); 825e8c5c824SLee Schermerhorn list_del(&page->lru); 826e8c5c824SLee Schermerhorn h->free_huge_pages--; 827e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 828685f3457SLee Schermerhorn if (acct_surplus) { 829685f3457SLee Schermerhorn h->surplus_huge_pages--; 830685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 831685f3457SLee Schermerhorn } 832e8c5c824SLee Schermerhorn update_and_free_page(h, page); 833e8c5c824SLee Schermerhorn ret = 1; 8349a76db09SLee Schermerhorn break; 835e8c5c824SLee Schermerhorn } 8366ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 8379a76db09SLee Schermerhorn } while (next_nid != start_nid); 838e8c5c824SLee Schermerhorn 839e8c5c824SLee Schermerhorn return ret; 840e8c5c824SLee Schermerhorn } 841e8c5c824SLee Schermerhorn 842bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 8437893d1d5SAdam Litke { 8447893d1d5SAdam Litke struct page *page; 845bf50bab2SNaoya Horiguchi unsigned int r_nid; 8467893d1d5SAdam Litke 847aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 848aa888a74SAndi Kleen return NULL; 849aa888a74SAndi Kleen 850d1c3fb1fSNishanth Aravamudan /* 851d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 852d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 853d1c3fb1fSNishanth Aravamudan * overcommit 854d1c3fb1fSNishanth Aravamudan * 855d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 856d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 857d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 858d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 859d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 860d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 861d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 862d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 863d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 864d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 865d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 866d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 867d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 868d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 869d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 870d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 871d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 872d1c3fb1fSNishanth Aravamudan */ 873d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 874a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 875d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 876d1c3fb1fSNishanth Aravamudan return NULL; 877d1c3fb1fSNishanth Aravamudan } else { 878a5516438SAndi Kleen h->nr_huge_pages++; 879a5516438SAndi Kleen h->surplus_huge_pages++; 880d1c3fb1fSNishanth Aravamudan } 881d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 882d1c3fb1fSNishanth Aravamudan 883bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 884551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 885551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 886a5516438SAndi Kleen huge_page_order(h)); 887bf50bab2SNaoya Horiguchi else 888bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 889bf50bab2SNaoya Horiguchi htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 890bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 891d1c3fb1fSNishanth Aravamudan 892caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 893caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 894ea5768c7SHillf Danton page = NULL; 895caff3a2cSGerald Schaefer } 896caff3a2cSGerald Schaefer 8977893d1d5SAdam Litke spin_lock(&hugetlb_lock); 898d1c3fb1fSNishanth Aravamudan if (page) { 8990edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 900bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 901d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 9029dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 903d1c3fb1fSNishanth Aravamudan /* 904d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 905d1c3fb1fSNishanth Aravamudan */ 906bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 907bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 9083b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 909d1c3fb1fSNishanth Aravamudan } else { 910a5516438SAndi Kleen h->nr_huge_pages--; 911a5516438SAndi Kleen h->surplus_huge_pages--; 9123b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 9137893d1d5SAdam Litke } 914d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 9157893d1d5SAdam Litke 9167893d1d5SAdam Litke return page; 9177893d1d5SAdam Litke } 9187893d1d5SAdam Litke 919e4e574b7SAdam Litke /* 920bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 921bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 922bf50bab2SNaoya Horiguchi * address of error page. 923bf50bab2SNaoya Horiguchi */ 924bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 925bf50bab2SNaoya Horiguchi { 926bf50bab2SNaoya Horiguchi struct page *page; 927bf50bab2SNaoya Horiguchi 928bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 929bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 930bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 931bf50bab2SNaoya Horiguchi 93294ae8ba7SAneesh Kumar K.V if (!page) 933bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 934bf50bab2SNaoya Horiguchi 935bf50bab2SNaoya Horiguchi return page; 936bf50bab2SNaoya Horiguchi } 937bf50bab2SNaoya Horiguchi 938bf50bab2SNaoya Horiguchi /* 93925985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 940e4e574b7SAdam Litke * of size 'delta'. 941e4e574b7SAdam Litke */ 942a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 943e4e574b7SAdam Litke { 944e4e574b7SAdam Litke struct list_head surplus_list; 945e4e574b7SAdam Litke struct page *page, *tmp; 946e4e574b7SAdam Litke int ret, i; 947e4e574b7SAdam Litke int needed, allocated; 94828073b02SHillf Danton bool alloc_ok = true; 949e4e574b7SAdam Litke 950a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 951ac09b3a1SAdam Litke if (needed <= 0) { 952a5516438SAndi Kleen h->resv_huge_pages += delta; 953e4e574b7SAdam Litke return 0; 954ac09b3a1SAdam Litke } 955e4e574b7SAdam Litke 956e4e574b7SAdam Litke allocated = 0; 957e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 958e4e574b7SAdam Litke 959e4e574b7SAdam Litke ret = -ENOMEM; 960e4e574b7SAdam Litke retry: 961e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 962e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 963bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 96428073b02SHillf Danton if (!page) { 96528073b02SHillf Danton alloc_ok = false; 96628073b02SHillf Danton break; 96728073b02SHillf Danton } 968e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 969e4e574b7SAdam Litke } 97028073b02SHillf Danton allocated += i; 971e4e574b7SAdam Litke 972e4e574b7SAdam Litke /* 973e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 974e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 975e4e574b7SAdam Litke */ 976e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 977a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 978a5516438SAndi Kleen (h->free_huge_pages + allocated); 97928073b02SHillf Danton if (needed > 0) { 98028073b02SHillf Danton if (alloc_ok) 981e4e574b7SAdam Litke goto retry; 98228073b02SHillf Danton /* 98328073b02SHillf Danton * We were not able to allocate enough pages to 98428073b02SHillf Danton * satisfy the entire reservation so we free what 98528073b02SHillf Danton * we've allocated so far. 98628073b02SHillf Danton */ 98728073b02SHillf Danton goto free; 98828073b02SHillf Danton } 989e4e574b7SAdam Litke /* 990e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 99125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 992e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 993ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 994ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 995ac09b3a1SAdam Litke * before they are reserved. 996e4e574b7SAdam Litke */ 997e4e574b7SAdam Litke needed += allocated; 998a5516438SAndi Kleen h->resv_huge_pages += delta; 999e4e574b7SAdam Litke ret = 0; 1000a9869b83SNaoya Horiguchi 100119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 100219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 100319fc3f0aSAdam Litke if ((--needed) < 0) 100419fc3f0aSAdam Litke break; 1005a9869b83SNaoya Horiguchi /* 1006a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1007a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1008a9869b83SNaoya Horiguchi */ 1009a9869b83SNaoya Horiguchi put_page_testzero(page); 1010a9869b83SNaoya Horiguchi VM_BUG_ON(page_count(page)); 1011a5516438SAndi Kleen enqueue_huge_page(h, page); 101219fc3f0aSAdam Litke } 101328073b02SHillf Danton free: 1014b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 101519fc3f0aSAdam Litke 101619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 101719fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 1018e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1019a9869b83SNaoya Horiguchi put_page(page); 1020a9869b83SNaoya Horiguchi } 1021af767cbdSAdam Litke } 102219fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1023e4e574b7SAdam Litke 1024e4e574b7SAdam Litke return ret; 1025e4e574b7SAdam Litke } 1026e4e574b7SAdam Litke 1027e4e574b7SAdam Litke /* 1028e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1029e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1030e4e574b7SAdam Litke * never used. 1031685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1032e4e574b7SAdam Litke */ 1033a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1034a5516438SAndi Kleen unsigned long unused_resv_pages) 1035e4e574b7SAdam Litke { 1036e4e574b7SAdam Litke unsigned long nr_pages; 1037e4e574b7SAdam Litke 1038ac09b3a1SAdam Litke /* Uncommit the reservation */ 1039a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1040ac09b3a1SAdam Litke 1041aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1042aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1043aa888a74SAndi Kleen return; 1044aa888a74SAndi Kleen 1045a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1046e4e574b7SAdam Litke 1047685f3457SLee Schermerhorn /* 1048685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 10499b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 10509b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 10519b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 10529b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 10539b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1054685f3457SLee Schermerhorn */ 1055685f3457SLee Schermerhorn while (nr_pages--) { 10569b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 1057685f3457SLee Schermerhorn break; 1058e4e574b7SAdam Litke } 1059e4e574b7SAdam Litke } 1060e4e574b7SAdam Litke 1061c37f9fb1SAndy Whitcroft /* 1062c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1063c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 106490481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 106590481622SDavid Gibson * can occur. Where any new reservation would be required the 106690481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 106790481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 106890481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 106990481622SDavid Gibson * failure. 1070c37f9fb1SAndy Whitcroft */ 1071e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1072a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1073c37f9fb1SAndy Whitcroft { 1074c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1075c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1076c37f9fb1SAndy Whitcroft 1077f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1078a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1079c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 1080c37f9fb1SAndy Whitcroft idx, idx + 1); 1081c37f9fb1SAndy Whitcroft 108284afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1083c37f9fb1SAndy Whitcroft return 1; 1084c37f9fb1SAndy Whitcroft 108584afd99bSAndy Whitcroft } else { 1086e2f17d94SRoel Kluin long err; 1087a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 108884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 108984afd99bSAndy Whitcroft 109084afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 109184afd99bSAndy Whitcroft if (err < 0) 109284afd99bSAndy Whitcroft return err; 1093c37f9fb1SAndy Whitcroft return 0; 1094c37f9fb1SAndy Whitcroft } 109584afd99bSAndy Whitcroft } 1096a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1097a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1098c37f9fb1SAndy Whitcroft { 1099c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1100c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1101c37f9fb1SAndy Whitcroft 1102f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1103a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1104c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 110584afd99bSAndy Whitcroft 110684afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1107a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 110884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 110984afd99bSAndy Whitcroft 111084afd99bSAndy Whitcroft /* Mark this page used in the map. */ 111184afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1112c37f9fb1SAndy Whitcroft } 1113c37f9fb1SAndy Whitcroft } 1114c37f9fb1SAndy Whitcroft 1115348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 111604f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1117348ea204SAdam Litke { 111890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1119a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1120348ea204SAdam Litke struct page *page; 1121e2f17d94SRoel Kluin long chg; 11226d76dcf4SAneesh Kumar K.V int ret, idx; 11236d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 11242fc39cecSAdam Litke 11256d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1126a1e78772SMel Gorman /* 112790481622SDavid Gibson * Processes that did not create the mapping will have no 112890481622SDavid Gibson * reserves and will not have accounted against subpool 112990481622SDavid Gibson * limit. Check that the subpool limit can be made before 113090481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 113190481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 113290481622SDavid Gibson * mapping overlaps. 1133a1e78772SMel Gorman */ 1134a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1135c37f9fb1SAndy Whitcroft if (chg < 0) 113676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1137c37f9fb1SAndy Whitcroft if (chg) 113890481622SDavid Gibson if (hugepage_subpool_get_pages(spool, chg)) 113976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 114090d8b7e6SAdam Litke 11416d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 11426d76dcf4SAneesh Kumar K.V if (ret) { 11436d76dcf4SAneesh Kumar K.V hugepage_subpool_put_pages(spool, chg); 11446d76dcf4SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 11456d76dcf4SAneesh Kumar K.V } 1146a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1147a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 114894ae8ba7SAneesh Kumar K.V if (page) { 114994ae8ba7SAneesh Kumar K.V /* update page cgroup details */ 115094ae8ba7SAneesh Kumar K.V hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), 115194ae8ba7SAneesh Kumar K.V h_cg, page); 1152a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 115394ae8ba7SAneesh Kumar K.V } else { 115494ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1155bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1156a1e78772SMel Gorman if (!page) { 11576d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_cgroup(idx, 11586d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), 11596d76dcf4SAneesh Kumar K.V h_cg); 116090481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 116176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 1162a1e78772SMel Gorman } 116379dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 116494ae8ba7SAneesh Kumar K.V hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), 116594ae8ba7SAneesh Kumar K.V h_cg, page); 116679dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 116779dbb236SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1168a1e78772SMel Gorman } 1169a1e78772SMel Gorman 117090481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1171a1e78772SMel Gorman 1172a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 11737893d1d5SAdam Litke return page; 1174b45b5bd6SDavid Gibson } 1175b45b5bd6SDavid Gibson 117691f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1177aa888a74SAndi Kleen { 1178aa888a74SAndi Kleen struct huge_bootmem_page *m; 11799b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1180aa888a74SAndi Kleen 1181aa888a74SAndi Kleen while (nr_nodes) { 1182aa888a74SAndi Kleen void *addr; 1183aa888a74SAndi Kleen 1184aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 11856ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 11869b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1187aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1188aa888a74SAndi Kleen 1189aa888a74SAndi Kleen if (addr) { 1190aa888a74SAndi Kleen /* 1191aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1192aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1193aa888a74SAndi Kleen * puts them into the mem_map). 1194aa888a74SAndi Kleen */ 1195aa888a74SAndi Kleen m = addr; 1196aa888a74SAndi Kleen goto found; 1197aa888a74SAndi Kleen } 1198aa888a74SAndi Kleen nr_nodes--; 1199aa888a74SAndi Kleen } 1200aa888a74SAndi Kleen return 0; 1201aa888a74SAndi Kleen 1202aa888a74SAndi Kleen found: 1203aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1204aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1205aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1206aa888a74SAndi Kleen m->hstate = h; 1207aa888a74SAndi Kleen return 1; 1208aa888a74SAndi Kleen } 1209aa888a74SAndi Kleen 121018229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 121118229df5SAndy Whitcroft { 121218229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 121318229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 121418229df5SAndy Whitcroft else 121518229df5SAndy Whitcroft prep_compound_page(page, order); 121618229df5SAndy Whitcroft } 121718229df5SAndy Whitcroft 1218aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1219aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1220aa888a74SAndi Kleen { 1221aa888a74SAndi Kleen struct huge_bootmem_page *m; 1222aa888a74SAndi Kleen 1223aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1224aa888a74SAndi Kleen struct hstate *h = m->hstate; 1225ee8f248dSBecky Bruce struct page *page; 1226ee8f248dSBecky Bruce 1227ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1228ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 1229ee8f248dSBecky Bruce free_bootmem_late((unsigned long)m, 1230ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1231ee8f248dSBecky Bruce #else 1232ee8f248dSBecky Bruce page = virt_to_page(m); 1233ee8f248dSBecky Bruce #endif 1234aa888a74SAndi Kleen __ClearPageReserved(page); 1235aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 123618229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1237aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1238b0320c7bSRafael Aquini /* 1239b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1240b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1241b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1242b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1243b0320c7bSRafael Aquini */ 1244b0320c7bSRafael Aquini if (h->order > (MAX_ORDER - 1)) 1245b0320c7bSRafael Aquini totalram_pages += 1 << h->order; 1246aa888a74SAndi Kleen } 1247aa888a74SAndi Kleen } 1248aa888a74SAndi Kleen 12498faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 12501da177e4SLinus Torvalds { 12511da177e4SLinus Torvalds unsigned long i; 12521da177e4SLinus Torvalds 1253e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1254aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1255aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1256aa888a74SAndi Kleen break; 12579b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 12589b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 12591da177e4SLinus Torvalds break; 12601da177e4SLinus Torvalds } 12618faa8b07SAndi Kleen h->max_huge_pages = i; 1262e5ff2159SAndi Kleen } 1263e5ff2159SAndi Kleen 1264e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1265e5ff2159SAndi Kleen { 1266e5ff2159SAndi Kleen struct hstate *h; 1267e5ff2159SAndi Kleen 1268e5ff2159SAndi Kleen for_each_hstate(h) { 12698faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 12708faa8b07SAndi Kleen if (h->order < MAX_ORDER) 12718faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1272e5ff2159SAndi Kleen } 1273e5ff2159SAndi Kleen } 1274e5ff2159SAndi Kleen 12754abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 12764abd32dbSAndi Kleen { 12774abd32dbSAndi Kleen if (n >= (1UL << 30)) 12784abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 12794abd32dbSAndi Kleen else if (n >= (1UL << 20)) 12804abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 12814abd32dbSAndi Kleen else 12824abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 12834abd32dbSAndi Kleen return buf; 12844abd32dbSAndi Kleen } 12854abd32dbSAndi Kleen 1286e5ff2159SAndi Kleen static void __init report_hugepages(void) 1287e5ff2159SAndi Kleen { 1288e5ff2159SAndi Kleen struct hstate *h; 1289e5ff2159SAndi Kleen 1290e5ff2159SAndi Kleen for_each_hstate(h) { 12914abd32dbSAndi Kleen char buf[32]; 12924abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 12934abd32dbSAndi Kleen "pre-allocated %ld pages\n", 12944abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 12954abd32dbSAndi Kleen h->free_huge_pages); 1296e5ff2159SAndi Kleen } 1297e5ff2159SAndi Kleen } 1298e5ff2159SAndi Kleen 12991da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 13006ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 13016ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13021da177e4SLinus Torvalds { 13034415cc8dSChristoph Lameter int i; 13044415cc8dSChristoph Lameter 1305aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1306aa888a74SAndi Kleen return; 1307aa888a74SAndi Kleen 13086ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 13091da177e4SLinus Torvalds struct page *page, *next; 1310a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1311a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1312a5516438SAndi Kleen if (count >= h->nr_huge_pages) 13136b0c880dSAdam Litke return; 13141da177e4SLinus Torvalds if (PageHighMem(page)) 13151da177e4SLinus Torvalds continue; 13161da177e4SLinus Torvalds list_del(&page->lru); 1317e5ff2159SAndi Kleen update_and_free_page(h, page); 1318a5516438SAndi Kleen h->free_huge_pages--; 1319a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 13201da177e4SLinus Torvalds } 13211da177e4SLinus Torvalds } 13221da177e4SLinus Torvalds } 13231da177e4SLinus Torvalds #else 13246ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 13256ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13261da177e4SLinus Torvalds { 13271da177e4SLinus Torvalds } 13281da177e4SLinus Torvalds #endif 13291da177e4SLinus Torvalds 133020a0307cSWu Fengguang /* 133120a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 133220a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 133320a0307cSWu Fengguang * Returns 1 if an adjustment was made. 133420a0307cSWu Fengguang */ 13356ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 13366ae11b27SLee Schermerhorn int delta) 133720a0307cSWu Fengguang { 1338e8c5c824SLee Schermerhorn int start_nid, next_nid; 133920a0307cSWu Fengguang int ret = 0; 134020a0307cSWu Fengguang 134120a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 134220a0307cSWu Fengguang 1343e8c5c824SLee Schermerhorn if (delta < 0) 13446ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1345e8c5c824SLee Schermerhorn else 13466ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1347e8c5c824SLee Schermerhorn next_nid = start_nid; 1348e8c5c824SLee Schermerhorn 1349e8c5c824SLee Schermerhorn do { 1350e8c5c824SLee Schermerhorn int nid = next_nid; 1351e8c5c824SLee Schermerhorn if (delta < 0) { 1352e8c5c824SLee Schermerhorn /* 1353e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1354e8c5c824SLee Schermerhorn */ 13559a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 13566ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 13576ae11b27SLee Schermerhorn nodes_allowed); 135820a0307cSWu Fengguang continue; 1359e8c5c824SLee Schermerhorn } 13609a76db09SLee Schermerhorn } 1361e8c5c824SLee Schermerhorn if (delta > 0) { 1362e8c5c824SLee Schermerhorn /* 1363e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1364e8c5c824SLee Schermerhorn */ 1365e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 13669a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 13676ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 13686ae11b27SLee Schermerhorn nodes_allowed); 136920a0307cSWu Fengguang continue; 1370e8c5c824SLee Schermerhorn } 13719a76db09SLee Schermerhorn } 137220a0307cSWu Fengguang 137320a0307cSWu Fengguang h->surplus_huge_pages += delta; 137420a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 137520a0307cSWu Fengguang ret = 1; 137620a0307cSWu Fengguang break; 1377e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 137820a0307cSWu Fengguang 137920a0307cSWu Fengguang return ret; 138020a0307cSWu Fengguang } 138120a0307cSWu Fengguang 1382a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 13836ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 13846ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 13851da177e4SLinus Torvalds { 13867893d1d5SAdam Litke unsigned long min_count, ret; 13871da177e4SLinus Torvalds 1388aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1389aa888a74SAndi Kleen return h->max_huge_pages; 1390aa888a74SAndi Kleen 13917893d1d5SAdam Litke /* 13927893d1d5SAdam Litke * Increase the pool size 13937893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 13947893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1395d1c3fb1fSNishanth Aravamudan * 1396d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1397d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1398d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1399d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1400d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 14017893d1d5SAdam Litke */ 14021da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1403a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 14046ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 14057893d1d5SAdam Litke break; 14067893d1d5SAdam Litke } 14077893d1d5SAdam Litke 1408a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 14097893d1d5SAdam Litke /* 14107893d1d5SAdam Litke * If this allocation races such that we no longer need the 14117893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 14127893d1d5SAdam Litke * and reducing the surplus. 14137893d1d5SAdam Litke */ 14147893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 14156ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 14167893d1d5SAdam Litke spin_lock(&hugetlb_lock); 14177893d1d5SAdam Litke if (!ret) 14187893d1d5SAdam Litke goto out; 14197893d1d5SAdam Litke 1420536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1421536240f2SMel Gorman if (signal_pending(current)) 1422536240f2SMel Gorman goto out; 14237893d1d5SAdam Litke } 14247893d1d5SAdam Litke 14257893d1d5SAdam Litke /* 14267893d1d5SAdam Litke * Decrease the pool size 14277893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 14287893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 14297893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 14307893d1d5SAdam Litke * to the desired size as pages become free. 1431d1c3fb1fSNishanth Aravamudan * 1432d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1433d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1434d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1435d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1436d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1437d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1438d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 14397893d1d5SAdam Litke */ 1440a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 14416b0c880dSAdam Litke min_count = max(count, min_count); 14426ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1443a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 14446ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 14451da177e4SLinus Torvalds break; 14461da177e4SLinus Torvalds } 1447a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 14486ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 14497893d1d5SAdam Litke break; 14507893d1d5SAdam Litke } 14517893d1d5SAdam Litke out: 1452a5516438SAndi Kleen ret = persistent_huge_pages(h); 14531da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 14547893d1d5SAdam Litke return ret; 14551da177e4SLinus Torvalds } 14561da177e4SLinus Torvalds 1457a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1458a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1459a3437870SNishanth Aravamudan 1460a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1461a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1462a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1463a3437870SNishanth Aravamudan 1464a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1465a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1466a3437870SNishanth Aravamudan 14679a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 14689a305230SLee Schermerhorn 14699a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1470a3437870SNishanth Aravamudan { 1471a3437870SNishanth Aravamudan int i; 14729a305230SLee Schermerhorn 1473a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 14749a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 14759a305230SLee Schermerhorn if (nidp) 14769a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1477a3437870SNishanth Aravamudan return &hstates[i]; 14789a305230SLee Schermerhorn } 14799a305230SLee Schermerhorn 14809a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1481a3437870SNishanth Aravamudan } 1482a3437870SNishanth Aravamudan 148306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1484a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1485a3437870SNishanth Aravamudan { 14869a305230SLee Schermerhorn struct hstate *h; 14879a305230SLee Schermerhorn unsigned long nr_huge_pages; 14889a305230SLee Schermerhorn int nid; 14899a305230SLee Schermerhorn 14909a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14919a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14929a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 14939a305230SLee Schermerhorn else 14949a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 14959a305230SLee Schermerhorn 14969a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1497a3437870SNishanth Aravamudan } 1498adbe8726SEric B Munson 149906808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 150006808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 150106808b08SLee Schermerhorn const char *buf, size_t len) 1502a3437870SNishanth Aravamudan { 1503a3437870SNishanth Aravamudan int err; 15049a305230SLee Schermerhorn int nid; 150506808b08SLee Schermerhorn unsigned long count; 15069a305230SLee Schermerhorn struct hstate *h; 1507bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1508a3437870SNishanth Aravamudan 150906808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 151073ae31e5SEric B Munson if (err) 1511adbe8726SEric B Munson goto out; 1512a3437870SNishanth Aravamudan 15139a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1514adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1515adbe8726SEric B Munson err = -EINVAL; 1516adbe8726SEric B Munson goto out; 1517adbe8726SEric B Munson } 1518adbe8726SEric B Munson 15199a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 15209a305230SLee Schermerhorn /* 15219a305230SLee Schermerhorn * global hstate attribute 15229a305230SLee Schermerhorn */ 15239a305230SLee Schermerhorn if (!(obey_mempolicy && 15249a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 152506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 15269a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 152706808b08SLee Schermerhorn } 15289a305230SLee Schermerhorn } else if (nodes_allowed) { 15299a305230SLee Schermerhorn /* 15309a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 15319a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 15329a305230SLee Schermerhorn */ 15339a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 15349a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 15359a305230SLee Schermerhorn } else 15369a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 15379a305230SLee Schermerhorn 153806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1539a3437870SNishanth Aravamudan 15409b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 154106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 154206808b08SLee Schermerhorn 154306808b08SLee Schermerhorn return len; 1544adbe8726SEric B Munson out: 1545adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1546adbe8726SEric B Munson return err; 154706808b08SLee Schermerhorn } 154806808b08SLee Schermerhorn 154906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 155006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 155106808b08SLee Schermerhorn { 155206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 155306808b08SLee Schermerhorn } 155406808b08SLee Schermerhorn 155506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 155606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 155706808b08SLee Schermerhorn { 155806808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1559a3437870SNishanth Aravamudan } 1560a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1561a3437870SNishanth Aravamudan 156206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 156306808b08SLee Schermerhorn 156406808b08SLee Schermerhorn /* 156506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 156606808b08SLee Schermerhorn * huge page alloc/free. 156706808b08SLee Schermerhorn */ 156806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 156906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 157006808b08SLee Schermerhorn { 157106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 157206808b08SLee Schermerhorn } 157306808b08SLee Schermerhorn 157406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 157506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 157606808b08SLee Schermerhorn { 157706808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 157806808b08SLee Schermerhorn } 157906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 158006808b08SLee Schermerhorn #endif 158106808b08SLee Schermerhorn 158206808b08SLee Schermerhorn 1583a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1584a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1585a3437870SNishanth Aravamudan { 15869a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1587a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1588a3437870SNishanth Aravamudan } 1589adbe8726SEric B Munson 1590a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1591a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1592a3437870SNishanth Aravamudan { 1593a3437870SNishanth Aravamudan int err; 1594a3437870SNishanth Aravamudan unsigned long input; 15959a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1596a3437870SNishanth Aravamudan 1597adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1598adbe8726SEric B Munson return -EINVAL; 1599adbe8726SEric B Munson 1600a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1601a3437870SNishanth Aravamudan if (err) 160273ae31e5SEric B Munson return err; 1603a3437870SNishanth Aravamudan 1604a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1605a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1606a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1607a3437870SNishanth Aravamudan 1608a3437870SNishanth Aravamudan return count; 1609a3437870SNishanth Aravamudan } 1610a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1611a3437870SNishanth Aravamudan 1612a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1613a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1614a3437870SNishanth Aravamudan { 16159a305230SLee Schermerhorn struct hstate *h; 16169a305230SLee Schermerhorn unsigned long free_huge_pages; 16179a305230SLee Schermerhorn int nid; 16189a305230SLee Schermerhorn 16199a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16209a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16219a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 16229a305230SLee Schermerhorn else 16239a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 16249a305230SLee Schermerhorn 16259a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1626a3437870SNishanth Aravamudan } 1627a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1628a3437870SNishanth Aravamudan 1629a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1630a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1631a3437870SNishanth Aravamudan { 16329a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1633a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1634a3437870SNishanth Aravamudan } 1635a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1636a3437870SNishanth Aravamudan 1637a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1638a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1639a3437870SNishanth Aravamudan { 16409a305230SLee Schermerhorn struct hstate *h; 16419a305230SLee Schermerhorn unsigned long surplus_huge_pages; 16429a305230SLee Schermerhorn int nid; 16439a305230SLee Schermerhorn 16449a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 16459a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 16469a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 16479a305230SLee Schermerhorn else 16489a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 16499a305230SLee Schermerhorn 16509a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1651a3437870SNishanth Aravamudan } 1652a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1653a3437870SNishanth Aravamudan 1654a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1655a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1656a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1657a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1658a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1659a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 166006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 166106808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 166206808b08SLee Schermerhorn #endif 1663a3437870SNishanth Aravamudan NULL, 1664a3437870SNishanth Aravamudan }; 1665a3437870SNishanth Aravamudan 1666a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1667a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1668a3437870SNishanth Aravamudan }; 1669a3437870SNishanth Aravamudan 1670094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 16719a305230SLee Schermerhorn struct kobject **hstate_kobjs, 16729a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1673a3437870SNishanth Aravamudan { 1674a3437870SNishanth Aravamudan int retval; 1675972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1676a3437870SNishanth Aravamudan 16779a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 16789a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1679a3437870SNishanth Aravamudan return -ENOMEM; 1680a3437870SNishanth Aravamudan 16819a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1682a3437870SNishanth Aravamudan if (retval) 16839a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1684a3437870SNishanth Aravamudan 1685a3437870SNishanth Aravamudan return retval; 1686a3437870SNishanth Aravamudan } 1687a3437870SNishanth Aravamudan 1688a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1689a3437870SNishanth Aravamudan { 1690a3437870SNishanth Aravamudan struct hstate *h; 1691a3437870SNishanth Aravamudan int err; 1692a3437870SNishanth Aravamudan 1693a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1694a3437870SNishanth Aravamudan if (!hugepages_kobj) 1695a3437870SNishanth Aravamudan return; 1696a3437870SNishanth Aravamudan 1697a3437870SNishanth Aravamudan for_each_hstate(h) { 16989a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 16999a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1700a3437870SNishanth Aravamudan if (err) 1701a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1702a3437870SNishanth Aravamudan h->name); 1703a3437870SNishanth Aravamudan } 1704a3437870SNishanth Aravamudan } 1705a3437870SNishanth Aravamudan 17069a305230SLee Schermerhorn #ifdef CONFIG_NUMA 17079a305230SLee Schermerhorn 17089a305230SLee Schermerhorn /* 17099a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 171010fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 171110fbcf4cSKay Sievers * index of a node device or _hstate == node id. 171210fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 17139a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 17149a305230SLee Schermerhorn */ 17159a305230SLee Schermerhorn struct node_hstate { 17169a305230SLee Schermerhorn struct kobject *hugepages_kobj; 17179a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 17189a305230SLee Schermerhorn }; 17199a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 17209a305230SLee Schermerhorn 17219a305230SLee Schermerhorn /* 172210fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 17239a305230SLee Schermerhorn */ 17249a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 17259a305230SLee Schermerhorn &nr_hugepages_attr.attr, 17269a305230SLee Schermerhorn &free_hugepages_attr.attr, 17279a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 17289a305230SLee Schermerhorn NULL, 17299a305230SLee Schermerhorn }; 17309a305230SLee Schermerhorn 17319a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 17329a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 17339a305230SLee Schermerhorn }; 17349a305230SLee Schermerhorn 17359a305230SLee Schermerhorn /* 173610fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 17379a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 17389a305230SLee Schermerhorn */ 17399a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17409a305230SLee Schermerhorn { 17419a305230SLee Schermerhorn int nid; 17429a305230SLee Schermerhorn 17439a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 17449a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 17459a305230SLee Schermerhorn int i; 17469a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 17479a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 17489a305230SLee Schermerhorn if (nidp) 17499a305230SLee Schermerhorn *nidp = nid; 17509a305230SLee Schermerhorn return &hstates[i]; 17519a305230SLee Schermerhorn } 17529a305230SLee Schermerhorn } 17539a305230SLee Schermerhorn 17549a305230SLee Schermerhorn BUG(); 17559a305230SLee Schermerhorn return NULL; 17569a305230SLee Schermerhorn } 17579a305230SLee Schermerhorn 17589a305230SLee Schermerhorn /* 175910fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 17609a305230SLee Schermerhorn * No-op if no hstate attributes attached. 17619a305230SLee Schermerhorn */ 17629a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 17639a305230SLee Schermerhorn { 17649a305230SLee Schermerhorn struct hstate *h; 176510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 17669a305230SLee Schermerhorn 17679a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 17689b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 17699a305230SLee Schermerhorn 1770972dc4deSAneesh Kumar K.V for_each_hstate(h) { 1771972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 1772972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 1773972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 1774972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 1775972dc4deSAneesh Kumar K.V } 17769a305230SLee Schermerhorn } 17779a305230SLee Schermerhorn 17789a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 17799a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 17809a305230SLee Schermerhorn } 17819a305230SLee Schermerhorn 17829a305230SLee Schermerhorn /* 178310fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 17849a305230SLee Schermerhorn * that have them. 17859a305230SLee Schermerhorn */ 17869a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 17879a305230SLee Schermerhorn { 17889a305230SLee Schermerhorn int nid; 17899a305230SLee Schermerhorn 17909a305230SLee Schermerhorn /* 179110fbcf4cSKay Sievers * disable node device registrations. 17929a305230SLee Schermerhorn */ 17939a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 17949a305230SLee Schermerhorn 17959a305230SLee Schermerhorn /* 17969a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 17979a305230SLee Schermerhorn */ 17989a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 17999a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 18009a305230SLee Schermerhorn } 18019a305230SLee Schermerhorn 18029a305230SLee Schermerhorn /* 180310fbcf4cSKay Sievers * Register hstate attributes for a single node device. 18049a305230SLee Schermerhorn * No-op if attributes already registered. 18059a305230SLee Schermerhorn */ 18069a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 18079a305230SLee Schermerhorn { 18089a305230SLee Schermerhorn struct hstate *h; 180910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 18109a305230SLee Schermerhorn int err; 18119a305230SLee Schermerhorn 18129a305230SLee Schermerhorn if (nhs->hugepages_kobj) 18139a305230SLee Schermerhorn return; /* already allocated */ 18149a305230SLee Schermerhorn 18159a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 181610fbcf4cSKay Sievers &node->dev.kobj); 18179a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 18189a305230SLee Schermerhorn return; 18199a305230SLee Schermerhorn 18209a305230SLee Schermerhorn for_each_hstate(h) { 18219a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 18229a305230SLee Schermerhorn nhs->hstate_kobjs, 18239a305230SLee Schermerhorn &per_node_hstate_attr_group); 18249a305230SLee Schermerhorn if (err) { 18259a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 18269a305230SLee Schermerhorn " for node %d\n", 182710fbcf4cSKay Sievers h->name, node->dev.id); 18289a305230SLee Schermerhorn hugetlb_unregister_node(node); 18299a305230SLee Schermerhorn break; 18309a305230SLee Schermerhorn } 18319a305230SLee Schermerhorn } 18329a305230SLee Schermerhorn } 18339a305230SLee Schermerhorn 18349a305230SLee Schermerhorn /* 18359b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 183610fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 183710fbcf4cSKay Sievers * registered their associated device by this time. 18389a305230SLee Schermerhorn */ 18399a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 18409a305230SLee Schermerhorn { 18419a305230SLee Schermerhorn int nid; 18429a305230SLee Schermerhorn 18439b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 18449a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 184510fbcf4cSKay Sievers if (node->dev.id == nid) 18469a305230SLee Schermerhorn hugetlb_register_node(node); 18479a305230SLee Schermerhorn } 18489a305230SLee Schermerhorn 18499a305230SLee Schermerhorn /* 185010fbcf4cSKay Sievers * Let the node device driver know we're here so it can 18519a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 18529a305230SLee Schermerhorn */ 18539a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 18549a305230SLee Schermerhorn hugetlb_unregister_node); 18559a305230SLee Schermerhorn } 18569a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 18579a305230SLee Schermerhorn 18589a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 18599a305230SLee Schermerhorn { 18609a305230SLee Schermerhorn BUG(); 18619a305230SLee Schermerhorn if (nidp) 18629a305230SLee Schermerhorn *nidp = -1; 18639a305230SLee Schermerhorn return NULL; 18649a305230SLee Schermerhorn } 18659a305230SLee Schermerhorn 18669a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 18679a305230SLee Schermerhorn 18689a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 18699a305230SLee Schermerhorn 18709a305230SLee Schermerhorn #endif 18719a305230SLee Schermerhorn 1872a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1873a3437870SNishanth Aravamudan { 1874a3437870SNishanth Aravamudan struct hstate *h; 1875a3437870SNishanth Aravamudan 18769a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 18779a305230SLee Schermerhorn 1878a3437870SNishanth Aravamudan for_each_hstate(h) { 1879972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 1880a3437870SNishanth Aravamudan } 1881a3437870SNishanth Aravamudan 1882a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1883a3437870SNishanth Aravamudan } 1884a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1885a3437870SNishanth Aravamudan 1886a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1887a3437870SNishanth Aravamudan { 18880ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 18890ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 18900ef89d25SBenjamin Herrenschmidt * there is no such support 18910ef89d25SBenjamin Herrenschmidt */ 18920ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 18930ef89d25SBenjamin Herrenschmidt return 0; 1894a3437870SNishanth Aravamudan 1895e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1896e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1897e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1898a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1899a3437870SNishanth Aravamudan } 1900972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1901e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1902e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1903a3437870SNishanth Aravamudan 1904a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1905a3437870SNishanth Aravamudan 1906aa888a74SAndi Kleen gather_bootmem_prealloc(); 1907aa888a74SAndi Kleen 1908a3437870SNishanth Aravamudan report_hugepages(); 1909a3437870SNishanth Aravamudan 1910a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1911a3437870SNishanth Aravamudan 19129a305230SLee Schermerhorn hugetlb_register_all_nodes(); 19139a305230SLee Schermerhorn 1914a3437870SNishanth Aravamudan return 0; 1915a3437870SNishanth Aravamudan } 1916a3437870SNishanth Aravamudan module_init(hugetlb_init); 1917a3437870SNishanth Aravamudan 1918a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1919a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1920a3437870SNishanth Aravamudan { 1921a3437870SNishanth Aravamudan struct hstate *h; 19228faa8b07SAndi Kleen unsigned long i; 19238faa8b07SAndi Kleen 1924a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1925a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1926a3437870SNishanth Aravamudan return; 1927a3437870SNishanth Aravamudan } 192847d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 1929a3437870SNishanth Aravamudan BUG_ON(order == 0); 193047d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 1931a3437870SNishanth Aravamudan h->order = order; 1932a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 19338faa8b07SAndi Kleen h->nr_huge_pages = 0; 19348faa8b07SAndi Kleen h->free_huge_pages = 0; 19358faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 19368faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 19370edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 19389b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 19399b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1940a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1941a3437870SNishanth Aravamudan huge_page_size(h)/1024); 1942abb8206cSAneesh Kumar K.V /* 1943abb8206cSAneesh Kumar K.V * Add cgroup control files only if the huge page consists 1944abb8206cSAneesh Kumar K.V * of more than two normal pages. This is because we use 1945abb8206cSAneesh Kumar K.V * page[2].lru.next for storing cgoup details. 1946abb8206cSAneesh Kumar K.V */ 1947abb8206cSAneesh Kumar K.V if (order >= HUGETLB_CGROUP_MIN_ORDER) 1948abb8206cSAneesh Kumar K.V hugetlb_cgroup_file_init(hugetlb_max_hstate - 1); 19498faa8b07SAndi Kleen 1950a3437870SNishanth Aravamudan parsed_hstate = h; 1951a3437870SNishanth Aravamudan } 1952a3437870SNishanth Aravamudan 1953e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1954a3437870SNishanth Aravamudan { 1955a3437870SNishanth Aravamudan unsigned long *mhp; 19568faa8b07SAndi Kleen static unsigned long *last_mhp; 1957a3437870SNishanth Aravamudan 1958a3437870SNishanth Aravamudan /* 195947d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 1960a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1961a3437870SNishanth Aravamudan */ 196247d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 1963a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1964a3437870SNishanth Aravamudan else 1965a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1966a3437870SNishanth Aravamudan 19678faa8b07SAndi Kleen if (mhp == last_mhp) { 19688faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 19698faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 19708faa8b07SAndi Kleen return 1; 19718faa8b07SAndi Kleen } 19728faa8b07SAndi Kleen 1973a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1974a3437870SNishanth Aravamudan *mhp = 0; 1975a3437870SNishanth Aravamudan 19768faa8b07SAndi Kleen /* 19778faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 19788faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 19798faa8b07SAndi Kleen * use the bootmem allocator. 19808faa8b07SAndi Kleen */ 198147d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 19828faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 19838faa8b07SAndi Kleen 19848faa8b07SAndi Kleen last_mhp = mhp; 19858faa8b07SAndi Kleen 1986a3437870SNishanth Aravamudan return 1; 1987a3437870SNishanth Aravamudan } 1988e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1989e11bfbfcSNick Piggin 1990e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1991e11bfbfcSNick Piggin { 1992e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1993e11bfbfcSNick Piggin return 1; 1994e11bfbfcSNick Piggin } 1995e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1996a3437870SNishanth Aravamudan 19978a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 19988a213460SNishanth Aravamudan { 19998a213460SNishanth Aravamudan int node; 20008a213460SNishanth Aravamudan unsigned int nr = 0; 20018a213460SNishanth Aravamudan 20028a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 20038a213460SNishanth Aravamudan nr += array[node]; 20048a213460SNishanth Aravamudan 20058a213460SNishanth Aravamudan return nr; 20068a213460SNishanth Aravamudan } 20078a213460SNishanth Aravamudan 20088a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 200906808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 201006808b08SLee Schermerhorn struct ctl_table *table, int write, 201106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 20121da177e4SLinus Torvalds { 2013e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2014e5ff2159SAndi Kleen unsigned long tmp; 201508d4a246SMichal Hocko int ret; 2016e5ff2159SAndi Kleen 2017e5ff2159SAndi Kleen tmp = h->max_huge_pages; 2018e5ff2159SAndi Kleen 2019adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2020adbe8726SEric B Munson return -EINVAL; 2021adbe8726SEric B Munson 2022e5ff2159SAndi Kleen table->data = &tmp; 2023e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 202408d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 202508d4a246SMichal Hocko if (ret) 202608d4a246SMichal Hocko goto out; 2027e5ff2159SAndi Kleen 202806808b08SLee Schermerhorn if (write) { 2029bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2030bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 203106808b08SLee Schermerhorn if (!(obey_mempolicy && 203206808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 203306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 203406808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 203506808b08SLee Schermerhorn } 203606808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 203706808b08SLee Schermerhorn 203806808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 203906808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 204006808b08SLee Schermerhorn } 204108d4a246SMichal Hocko out: 204208d4a246SMichal Hocko return ret; 20431da177e4SLinus Torvalds } 2044396faf03SMel Gorman 204506808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 204606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 204706808b08SLee Schermerhorn { 204806808b08SLee Schermerhorn 204906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 205006808b08SLee Schermerhorn buffer, length, ppos); 205106808b08SLee Schermerhorn } 205206808b08SLee Schermerhorn 205306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 205406808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 205506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 205606808b08SLee Schermerhorn { 205706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 205806808b08SLee Schermerhorn buffer, length, ppos); 205906808b08SLee Schermerhorn } 206006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 206106808b08SLee Schermerhorn 2062396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 20638d65af78SAlexey Dobriyan void __user *buffer, 2064396faf03SMel Gorman size_t *length, loff_t *ppos) 2065396faf03SMel Gorman { 20668d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 2067396faf03SMel Gorman if (hugepages_treat_as_movable) 2068396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 2069396faf03SMel Gorman else 2070396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 2071396faf03SMel Gorman return 0; 2072396faf03SMel Gorman } 2073396faf03SMel Gorman 2074a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 20758d65af78SAlexey Dobriyan void __user *buffer, 2076a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2077a3d0c6aaSNishanth Aravamudan { 2078a5516438SAndi Kleen struct hstate *h = &default_hstate; 2079e5ff2159SAndi Kleen unsigned long tmp; 208008d4a246SMichal Hocko int ret; 2081e5ff2159SAndi Kleen 2082e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2083e5ff2159SAndi Kleen 2084adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2085adbe8726SEric B Munson return -EINVAL; 2086adbe8726SEric B Munson 2087e5ff2159SAndi Kleen table->data = &tmp; 2088e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 208908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 209008d4a246SMichal Hocko if (ret) 209108d4a246SMichal Hocko goto out; 2092e5ff2159SAndi Kleen 2093e5ff2159SAndi Kleen if (write) { 2094064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2095e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2096a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2097e5ff2159SAndi Kleen } 209808d4a246SMichal Hocko out: 209908d4a246SMichal Hocko return ret; 2100a3d0c6aaSNishanth Aravamudan } 2101a3d0c6aaSNishanth Aravamudan 21021da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 21031da177e4SLinus Torvalds 2104e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 21051da177e4SLinus Torvalds { 2106a5516438SAndi Kleen struct hstate *h = &default_hstate; 2107e1759c21SAlexey Dobriyan seq_printf(m, 21081da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 21091da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2110b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 21117893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 21124f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2113a5516438SAndi Kleen h->nr_huge_pages, 2114a5516438SAndi Kleen h->free_huge_pages, 2115a5516438SAndi Kleen h->resv_huge_pages, 2116a5516438SAndi Kleen h->surplus_huge_pages, 2117a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 21181da177e4SLinus Torvalds } 21191da177e4SLinus Torvalds 21201da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 21211da177e4SLinus Torvalds { 2122a5516438SAndi Kleen struct hstate *h = &default_hstate; 21231da177e4SLinus Torvalds return sprintf(buf, 21241da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2125a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2126a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2127a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2128a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2129a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 21301da177e4SLinus Torvalds } 21311da177e4SLinus Torvalds 21321da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 21331da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 21341da177e4SLinus Torvalds { 2135a5516438SAndi Kleen struct hstate *h = &default_hstate; 2136a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 21371da177e4SLinus Torvalds } 21381da177e4SLinus Torvalds 2139a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2140fc1b8a73SMel Gorman { 2141fc1b8a73SMel Gorman int ret = -ENOMEM; 2142fc1b8a73SMel Gorman 2143fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2144fc1b8a73SMel Gorman /* 2145fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2146fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2147fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2148fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2149fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2150fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2151fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2152fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2153fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2154fc1b8a73SMel Gorman * 2155fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2156fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2157fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2158fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2159fc1b8a73SMel Gorman * semantics that cpuset has. 2160fc1b8a73SMel Gorman */ 2161fc1b8a73SMel Gorman if (delta > 0) { 2162a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2163fc1b8a73SMel Gorman goto out; 2164fc1b8a73SMel Gorman 2165a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2166a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2167fc1b8a73SMel Gorman goto out; 2168fc1b8a73SMel Gorman } 2169fc1b8a73SMel Gorman } 2170fc1b8a73SMel Gorman 2171fc1b8a73SMel Gorman ret = 0; 2172fc1b8a73SMel Gorman if (delta < 0) 2173a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2174fc1b8a73SMel Gorman 2175fc1b8a73SMel Gorman out: 2176fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2177fc1b8a73SMel Gorman return ret; 2178fc1b8a73SMel Gorman } 2179fc1b8a73SMel Gorman 218084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 218184afd99bSAndy Whitcroft { 218284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 218384afd99bSAndy Whitcroft 218484afd99bSAndy Whitcroft /* 218584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 218684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 218784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 218825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 218984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 219084afd99bSAndy Whitcroft * new reference here without additional locking. 219184afd99bSAndy Whitcroft */ 219284afd99bSAndy Whitcroft if (reservations) 219384afd99bSAndy Whitcroft kref_get(&reservations->refs); 219484afd99bSAndy Whitcroft } 219584afd99bSAndy Whitcroft 2196c50ac050SDave Hansen static void resv_map_put(struct vm_area_struct *vma) 2197c50ac050SDave Hansen { 2198c50ac050SDave Hansen struct resv_map *reservations = vma_resv_map(vma); 2199c50ac050SDave Hansen 2200c50ac050SDave Hansen if (!reservations) 2201c50ac050SDave Hansen return; 2202c50ac050SDave Hansen kref_put(&reservations->refs, resv_map_release); 2203c50ac050SDave Hansen } 2204c50ac050SDave Hansen 2205a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2206a1e78772SMel Gorman { 2207a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 220884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 220990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 221084afd99bSAndy Whitcroft unsigned long reserve; 221184afd99bSAndy Whitcroft unsigned long start; 221284afd99bSAndy Whitcroft unsigned long end; 221384afd99bSAndy Whitcroft 221484afd99bSAndy Whitcroft if (reservations) { 2215a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2216a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 221784afd99bSAndy Whitcroft 221884afd99bSAndy Whitcroft reserve = (end - start) - 221984afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 222084afd99bSAndy Whitcroft 2221c50ac050SDave Hansen resv_map_put(vma); 222284afd99bSAndy Whitcroft 22237251ff78SAdam Litke if (reserve) { 2224a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 222590481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 22267251ff78SAdam Litke } 2227a1e78772SMel Gorman } 222884afd99bSAndy Whitcroft } 2229a1e78772SMel Gorman 22301da177e4SLinus Torvalds /* 22311da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 22321da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 22331da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 22341da177e4SLinus Torvalds * this far. 22351da177e4SLinus Torvalds */ 2236d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 22371da177e4SLinus Torvalds { 22381da177e4SLinus Torvalds BUG(); 2239d0217ac0SNick Piggin return 0; 22401da177e4SLinus Torvalds } 22411da177e4SLinus Torvalds 2242f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2243d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 224484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2245a1e78772SMel Gorman .close = hugetlb_vm_op_close, 22461da177e4SLinus Torvalds }; 22471da177e4SLinus Torvalds 22481e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 22491e8f889bSDavid Gibson int writable) 225063551ae0SDavid Gibson { 225163551ae0SDavid Gibson pte_t entry; 225263551ae0SDavid Gibson 22531e8f889bSDavid Gibson if (writable) { 225463551ae0SDavid Gibson entry = 225563551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 225663551ae0SDavid Gibson } else { 22577f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 225863551ae0SDavid Gibson } 225963551ae0SDavid Gibson entry = pte_mkyoung(entry); 226063551ae0SDavid Gibson entry = pte_mkhuge(entry); 2261d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 226263551ae0SDavid Gibson 226363551ae0SDavid Gibson return entry; 226463551ae0SDavid Gibson } 226563551ae0SDavid Gibson 22661e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 22671e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 22681e8f889bSDavid Gibson { 22691e8f889bSDavid Gibson pte_t entry; 22701e8f889bSDavid Gibson 22717f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 227232f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 22734b3073e1SRussell King update_mmu_cache(vma, address, ptep); 22741e8f889bSDavid Gibson } 22751e8f889bSDavid Gibson 22761e8f889bSDavid Gibson 227763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 227863551ae0SDavid Gibson struct vm_area_struct *vma) 227963551ae0SDavid Gibson { 228063551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 228163551ae0SDavid Gibson struct page *ptepage; 22821c59827dSHugh Dickins unsigned long addr; 22831e8f889bSDavid Gibson int cow; 2284a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2285a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 22861e8f889bSDavid Gibson 22871e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 228863551ae0SDavid Gibson 2289a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2290c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2291c74df32cSHugh Dickins if (!src_pte) 2292c74df32cSHugh Dickins continue; 2293a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 229463551ae0SDavid Gibson if (!dst_pte) 229563551ae0SDavid Gibson goto nomem; 2296c5c99429SLarry Woodman 2297c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2298c5c99429SLarry Woodman if (dst_pte == src_pte) 2299c5c99429SLarry Woodman continue; 2300c5c99429SLarry Woodman 2301c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 230246478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 23037f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 23041e8f889bSDavid Gibson if (cow) 23057f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 23067f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 230763551ae0SDavid Gibson ptepage = pte_page(entry); 230863551ae0SDavid Gibson get_page(ptepage); 23090fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 231063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 23111c59827dSHugh Dickins } 23121c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2313c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 231463551ae0SDavid Gibson } 231563551ae0SDavid Gibson return 0; 231663551ae0SDavid Gibson 231763551ae0SDavid Gibson nomem: 231863551ae0SDavid Gibson return -ENOMEM; 231963551ae0SDavid Gibson } 232063551ae0SDavid Gibson 2321290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2322290408d4SNaoya Horiguchi { 2323290408d4SNaoya Horiguchi swp_entry_t swp; 2324290408d4SNaoya Horiguchi 2325290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2326290408d4SNaoya Horiguchi return 0; 2327290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 232832f84528SChris Forbes if (non_swap_entry(swp) && is_migration_entry(swp)) 2329290408d4SNaoya Horiguchi return 1; 233032f84528SChris Forbes else 2331290408d4SNaoya Horiguchi return 0; 2332290408d4SNaoya Horiguchi } 2333290408d4SNaoya Horiguchi 2334fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2335fd6a03edSNaoya Horiguchi { 2336fd6a03edSNaoya Horiguchi swp_entry_t swp; 2337fd6a03edSNaoya Horiguchi 2338fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2339fd6a03edSNaoya Horiguchi return 0; 2340fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 234132f84528SChris Forbes if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2342fd6a03edSNaoya Horiguchi return 1; 234332f84528SChris Forbes else 2344fd6a03edSNaoya Horiguchi return 0; 2345fd6a03edSNaoya Horiguchi } 2346fd6a03edSNaoya Horiguchi 234724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 234824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 234924669e58SAneesh Kumar K.V struct page *ref_page) 235063551ae0SDavid Gibson { 235124669e58SAneesh Kumar K.V int force_flush = 0; 235263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 235363551ae0SDavid Gibson unsigned long address; 2354c7546f8fSDavid Gibson pte_t *ptep; 235563551ae0SDavid Gibson pte_t pte; 235663551ae0SDavid Gibson struct page *page; 2357a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2358a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2359a5516438SAndi Kleen 236063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2361a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2362a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 236363551ae0SDavid Gibson 236424669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 2365cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 236624669e58SAneesh Kumar K.V again: 2367508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2368a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2369c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2370c7546f8fSDavid Gibson if (!ptep) 2371c7546f8fSDavid Gibson continue; 2372c7546f8fSDavid Gibson 237339dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 237439dde65cSChen, Kenneth W continue; 237539dde65cSChen, Kenneth W 23766629326bSHillf Danton pte = huge_ptep_get(ptep); 23776629326bSHillf Danton if (huge_pte_none(pte)) 23786629326bSHillf Danton continue; 23796629326bSHillf Danton 23806629326bSHillf Danton /* 23816629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 23826629326bSHillf Danton */ 23836629326bSHillf Danton if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 23846629326bSHillf Danton continue; 23856629326bSHillf Danton 23866629326bSHillf Danton page = pte_page(pte); 238704f2cbe3SMel Gorman /* 238804f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 238904f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 239004f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 239104f2cbe3SMel Gorman */ 239204f2cbe3SMel Gorman if (ref_page) { 239304f2cbe3SMel Gorman if (page != ref_page) 239404f2cbe3SMel Gorman continue; 239504f2cbe3SMel Gorman 239604f2cbe3SMel Gorman /* 239704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 239804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 239904f2cbe3SMel Gorman * looking like data was lost 240004f2cbe3SMel Gorman */ 240104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 240204f2cbe3SMel Gorman } 240304f2cbe3SMel Gorman 2404c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 240524669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 24066649a386SKen Chen if (pte_dirty(pte)) 24076649a386SKen Chen set_page_dirty(page); 24089e81130bSHillf Danton 240924669e58SAneesh Kumar K.V page_remove_rmap(page); 241024669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 241124669e58SAneesh Kumar K.V if (force_flush) 241224669e58SAneesh Kumar K.V break; 24139e81130bSHillf Danton /* Bail out after unmapping reference page if supplied */ 24149e81130bSHillf Danton if (ref_page) 24159e81130bSHillf Danton break; 241663551ae0SDavid Gibson } 2417cd2934a3SAl Viro spin_unlock(&mm->page_table_lock); 241824669e58SAneesh Kumar K.V /* 241924669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 242024669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 242124669e58SAneesh Kumar K.V * and page-free while holding it. 242224669e58SAneesh Kumar K.V */ 242324669e58SAneesh Kumar K.V if (force_flush) { 242424669e58SAneesh Kumar K.V force_flush = 0; 242524669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 242624669e58SAneesh Kumar K.V if (address < end && !ref_page) 242724669e58SAneesh Kumar K.V goto again; 2428fe1668aeSChen, Kenneth W } 242924669e58SAneesh Kumar K.V mmu_notifier_invalidate_range_end(mm, start, end); 243024669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 24311da177e4SLinus Torvalds } 243263551ae0SDavid Gibson 2433d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2434d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2435d833352aSMel Gorman unsigned long end, struct page *ref_page) 2436d833352aSMel Gorman { 2437d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2438d833352aSMel Gorman 2439d833352aSMel Gorman /* 2440d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2441d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2442d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2443d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2444d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2445d833352aSMel Gorman * is to clear it before releasing the i_mmap_mutex. This works 2446d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2447d833352aSMel Gorman * destroyed and the i_mmap_mutex is held. 2448d833352aSMel Gorman */ 2449d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2450d833352aSMel Gorman } 2451d833352aSMel Gorman 2452502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 245304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2454502717f4SChen, Kenneth W { 245524669e58SAneesh Kumar K.V struct mm_struct *mm; 245624669e58SAneesh Kumar K.V struct mmu_gather tlb; 245724669e58SAneesh Kumar K.V 245824669e58SAneesh Kumar K.V mm = vma->vm_mm; 245924669e58SAneesh Kumar K.V 246024669e58SAneesh Kumar K.V tlb_gather_mmu(&tlb, mm, 0); 246124669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 246224669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2463502717f4SChen, Kenneth W } 2464502717f4SChen, Kenneth W 246504f2cbe3SMel Gorman /* 246604f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 246704f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 246804f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 246904f2cbe3SMel Gorman * same region. 247004f2cbe3SMel Gorman */ 24712a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 24722a4b3dedSHarvey Harrison struct page *page, unsigned long address) 247304f2cbe3SMel Gorman { 24747526674dSAdam Litke struct hstate *h = hstate_vma(vma); 247504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 247604f2cbe3SMel Gorman struct address_space *mapping; 247704f2cbe3SMel Gorman pgoff_t pgoff; 247804f2cbe3SMel Gorman 247904f2cbe3SMel Gorman /* 248004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 248104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 248204f2cbe3SMel Gorman */ 24837526674dSAdam Litke address = address & huge_page_mask(h); 24840c176d52SHillf Danton pgoff = vma_hugecache_offset(h, vma, address); 248590481622SDavid Gibson mapping = vma->vm_file->f_dentry->d_inode->i_mapping; 248604f2cbe3SMel Gorman 24874eb2b1dcSMel Gorman /* 24884eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 24894eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 24904eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 24914eb2b1dcSMel Gorman */ 24923d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 24936b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 249404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 249504f2cbe3SMel Gorman if (iter_vma == vma) 249604f2cbe3SMel Gorman continue; 249704f2cbe3SMel Gorman 249804f2cbe3SMel Gorman /* 249904f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 250004f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 250104f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 250204f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 250304f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 250404f2cbe3SMel Gorman */ 250504f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 250624669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 250724669e58SAneesh Kumar K.V address + huge_page_size(h), page); 250804f2cbe3SMel Gorman } 25093d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 251004f2cbe3SMel Gorman 251104f2cbe3SMel Gorman return 1; 251204f2cbe3SMel Gorman } 251304f2cbe3SMel Gorman 25140fe6e20bSNaoya Horiguchi /* 25150fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2516ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2517ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2518ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 25190fe6e20bSNaoya Horiguchi */ 25201e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 252104f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 252204f2cbe3SMel Gorman struct page *pagecache_page) 25231e8f889bSDavid Gibson { 2524a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 25251e8f889bSDavid Gibson struct page *old_page, *new_page; 252679ac6ba4SDavid Gibson int avoidcopy; 252704f2cbe3SMel Gorman int outside_reserve = 0; 25281e8f889bSDavid Gibson 25291e8f889bSDavid Gibson old_page = pte_page(pte); 25301e8f889bSDavid Gibson 253104f2cbe3SMel Gorman retry_avoidcopy: 25321e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 25331e8f889bSDavid Gibson * and just make the page writable */ 25340fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 25351e8f889bSDavid Gibson if (avoidcopy) { 25360fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 25370fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 25381e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 253983c54070SNick Piggin return 0; 25401e8f889bSDavid Gibson } 25411e8f889bSDavid Gibson 254204f2cbe3SMel Gorman /* 254304f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 254404f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 254504f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 254604f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 254704f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 254804f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 254904f2cbe3SMel Gorman * of the full address range. 255004f2cbe3SMel Gorman */ 2551f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 255204f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 255304f2cbe3SMel Gorman old_page != pagecache_page) 255404f2cbe3SMel Gorman outside_reserve = 1; 255504f2cbe3SMel Gorman 25561e8f889bSDavid Gibson page_cache_get(old_page); 2557b76c8cfbSLarry Woodman 2558b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2559b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 256004f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 25611e8f889bSDavid Gibson 25622fc39cecSAdam Litke if (IS_ERR(new_page)) { 256376dcee75SAneesh Kumar K.V long err = PTR_ERR(new_page); 25641e8f889bSDavid Gibson page_cache_release(old_page); 256504f2cbe3SMel Gorman 256604f2cbe3SMel Gorman /* 256704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 256804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 256904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 257004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 257104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 257204f2cbe3SMel Gorman */ 257304f2cbe3SMel Gorman if (outside_reserve) { 257404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 257504f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 257604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2577b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2578a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2579a734bcc8SHillf Danton if (likely(pte_same(huge_ptep_get(ptep), pte))) 258004f2cbe3SMel Gorman goto retry_avoidcopy; 2581a734bcc8SHillf Danton /* 2582a734bcc8SHillf Danton * race occurs while re-acquiring page_table_lock, and 2583a734bcc8SHillf Danton * our job is done. 2584a734bcc8SHillf Danton */ 2585a734bcc8SHillf Danton return 0; 258604f2cbe3SMel Gorman } 258704f2cbe3SMel Gorman WARN_ON_ONCE(1); 258804f2cbe3SMel Gorman } 258904f2cbe3SMel Gorman 2590b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2591b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 259276dcee75SAneesh Kumar K.V if (err == -ENOMEM) 259376dcee75SAneesh Kumar K.V return VM_FAULT_OOM; 259476dcee75SAneesh Kumar K.V else 259576dcee75SAneesh Kumar K.V return VM_FAULT_SIGBUS; 25961e8f889bSDavid Gibson } 25971e8f889bSDavid Gibson 25980fe6e20bSNaoya Horiguchi /* 25990fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 26000fe6e20bSNaoya Horiguchi * anon_vma prepared. 26010fe6e20bSNaoya Horiguchi */ 260244e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2603ea4039a3SHillf Danton page_cache_release(new_page); 2604ea4039a3SHillf Danton page_cache_release(old_page); 260544e2aa93SDean Nelson /* Caller expects lock to be held */ 260644e2aa93SDean Nelson spin_lock(&mm->page_table_lock); 26070fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 260844e2aa93SDean Nelson } 26090fe6e20bSNaoya Horiguchi 261047ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 261147ad8475SAndrea Arcangeli pages_per_huge_page(h)); 26120ed361deSNick Piggin __SetPageUptodate(new_page); 26131e8f889bSDavid Gibson 2614b76c8cfbSLarry Woodman /* 2615b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2616b76c8cfbSLarry Woodman * before the page tables are altered 2617b76c8cfbSLarry Woodman */ 2618b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2619a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 26207f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 26211e8f889bSDavid Gibson /* Break COW */ 26223edd4fc9SDoug Doan mmu_notifier_invalidate_range_start(mm, 26233edd4fc9SDoug Doan address & huge_page_mask(h), 26243edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 26258fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 26261e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 26271e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 26280fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2629cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 26301e8f889bSDavid Gibson /* Make the old page be freed below */ 26311e8f889bSDavid Gibson new_page = old_page; 26323edd4fc9SDoug Doan mmu_notifier_invalidate_range_end(mm, 26333edd4fc9SDoug Doan address & huge_page_mask(h), 26343edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 26351e8f889bSDavid Gibson } 26361e8f889bSDavid Gibson page_cache_release(new_page); 26371e8f889bSDavid Gibson page_cache_release(old_page); 263883c54070SNick Piggin return 0; 26391e8f889bSDavid Gibson } 26401e8f889bSDavid Gibson 264104f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2642a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2643a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 264404f2cbe3SMel Gorman { 264504f2cbe3SMel Gorman struct address_space *mapping; 2646e7c4b0bfSAndy Whitcroft pgoff_t idx; 264704f2cbe3SMel Gorman 264804f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2649a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 265004f2cbe3SMel Gorman 265104f2cbe3SMel Gorman return find_lock_page(mapping, idx); 265204f2cbe3SMel Gorman } 265304f2cbe3SMel Gorman 26543ae77f43SHugh Dickins /* 26553ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 26563ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 26573ae77f43SHugh Dickins */ 26583ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 26592a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 26602a15efc9SHugh Dickins { 26612a15efc9SHugh Dickins struct address_space *mapping; 26622a15efc9SHugh Dickins pgoff_t idx; 26632a15efc9SHugh Dickins struct page *page; 26642a15efc9SHugh Dickins 26652a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 26662a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 26672a15efc9SHugh Dickins 26682a15efc9SHugh Dickins page = find_get_page(mapping, idx); 26692a15efc9SHugh Dickins if (page) 26702a15efc9SHugh Dickins put_page(page); 26712a15efc9SHugh Dickins return page != NULL; 26722a15efc9SHugh Dickins } 26732a15efc9SHugh Dickins 2674a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2675788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2676ac9b9c66SHugh Dickins { 2677a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2678ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2679409eb8c2SHillf Danton int anon_rmap = 0; 2680e7c4b0bfSAndy Whitcroft pgoff_t idx; 26814c887265SAdam Litke unsigned long size; 26824c887265SAdam Litke struct page *page; 26834c887265SAdam Litke struct address_space *mapping; 26841e8f889bSDavid Gibson pte_t new_pte; 26854c887265SAdam Litke 268604f2cbe3SMel Gorman /* 268704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 268804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 268925985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 269004f2cbe3SMel Gorman */ 269104f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 269204f2cbe3SMel Gorman printk(KERN_WARNING 269304f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 269404f2cbe3SMel Gorman current->pid); 269504f2cbe3SMel Gorman return ret; 269604f2cbe3SMel Gorman } 269704f2cbe3SMel Gorman 26984c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2699a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 27004c887265SAdam Litke 27014c887265SAdam Litke /* 27024c887265SAdam Litke * Use page lock to guard against racing truncation 27034c887265SAdam Litke * before we get page_table_lock. 27044c887265SAdam Litke */ 27056bda666aSChristoph Lameter retry: 27066bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 27076bda666aSChristoph Lameter if (!page) { 2708a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2709ebed4bfcSHugh Dickins if (idx >= size) 2710ebed4bfcSHugh Dickins goto out; 271104f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 27122fc39cecSAdam Litke if (IS_ERR(page)) { 271376dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 271476dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 271576dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 271676dcee75SAneesh Kumar K.V else 271776dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 27186bda666aSChristoph Lameter goto out; 27196bda666aSChristoph Lameter } 272047ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 27210ed361deSNick Piggin __SetPageUptodate(page); 2722ac9b9c66SHugh Dickins 2723f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 27246bda666aSChristoph Lameter int err; 272545c682a6SKen Chen struct inode *inode = mapping->host; 27266bda666aSChristoph Lameter 27276bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 27286bda666aSChristoph Lameter if (err) { 27296bda666aSChristoph Lameter put_page(page); 27306bda666aSChristoph Lameter if (err == -EEXIST) 27316bda666aSChristoph Lameter goto retry; 27326bda666aSChristoph Lameter goto out; 27336bda666aSChristoph Lameter } 273445c682a6SKen Chen 273545c682a6SKen Chen spin_lock(&inode->i_lock); 2736a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 273745c682a6SKen Chen spin_unlock(&inode->i_lock); 273823be7468SMel Gorman } else { 27396bda666aSChristoph Lameter lock_page(page); 27400fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 27410fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 27420fe6e20bSNaoya Horiguchi goto backout_unlocked; 274323be7468SMel Gorman } 2744409eb8c2SHillf Danton anon_rmap = 1; 27450fe6e20bSNaoya Horiguchi } 27460fe6e20bSNaoya Horiguchi } else { 274757303d80SAndy Whitcroft /* 2748998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2749998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2750998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2751fd6a03edSNaoya Horiguchi */ 2752fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2753aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2754972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2755fd6a03edSNaoya Horiguchi goto backout_unlocked; 27566bda666aSChristoph Lameter } 2757998b4382SNaoya Horiguchi } 27581e8f889bSDavid Gibson 275957303d80SAndy Whitcroft /* 276057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 276157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 276257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 276357303d80SAndy Whitcroft * the spinlock. 276457303d80SAndy Whitcroft */ 2765788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 27662b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 27672b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 27682b26736cSAndy Whitcroft goto backout_unlocked; 27692b26736cSAndy Whitcroft } 277057303d80SAndy Whitcroft 2771ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2772a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 27734c887265SAdam Litke if (idx >= size) 27744c887265SAdam Litke goto backout; 27754c887265SAdam Litke 277683c54070SNick Piggin ret = 0; 27777f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 27784c887265SAdam Litke goto backout; 27794c887265SAdam Litke 2780409eb8c2SHillf Danton if (anon_rmap) 2781409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 2782409eb8c2SHillf Danton else 2783409eb8c2SHillf Danton page_dup_rmap(page); 27841e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 27851e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 27861e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 27871e8f889bSDavid Gibson 2788788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 27891e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 279004f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 27911e8f889bSDavid Gibson } 27921e8f889bSDavid Gibson 2793ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 27944c887265SAdam Litke unlock_page(page); 27954c887265SAdam Litke out: 2796ac9b9c66SHugh Dickins return ret; 27974c887265SAdam Litke 27984c887265SAdam Litke backout: 27994c887265SAdam Litke spin_unlock(&mm->page_table_lock); 28002b26736cSAndy Whitcroft backout_unlocked: 28014c887265SAdam Litke unlock_page(page); 28024c887265SAdam Litke put_page(page); 28034c887265SAdam Litke goto out; 2804ac9b9c66SHugh Dickins } 2805ac9b9c66SHugh Dickins 280686e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2807788c7df4SHugh Dickins unsigned long address, unsigned int flags) 280886e5216fSAdam Litke { 280986e5216fSAdam Litke pte_t *ptep; 281086e5216fSAdam Litke pte_t entry; 28111e8f889bSDavid Gibson int ret; 28120fe6e20bSNaoya Horiguchi struct page *page = NULL; 281357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 28143935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2815a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 281686e5216fSAdam Litke 28171e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 28181e16a539SKAMEZAWA Hiroyuki 2819fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2820fd6a03edSNaoya Horiguchi if (ptep) { 2821fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2822290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 2823290408d4SNaoya Horiguchi migration_entry_wait(mm, (pmd_t *)ptep, address); 2824290408d4SNaoya Horiguchi return 0; 2825290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2826aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2827972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2828fd6a03edSNaoya Horiguchi } 2829fd6a03edSNaoya Horiguchi 2830a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 283186e5216fSAdam Litke if (!ptep) 283286e5216fSAdam Litke return VM_FAULT_OOM; 283386e5216fSAdam Litke 28343935baa9SDavid Gibson /* 28353935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 28363935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 28373935baa9SDavid Gibson * the same page in the page cache. 28383935baa9SDavid Gibson */ 28393935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 28407f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 28417f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2842788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2843b4d1d99fSDavid Gibson goto out_mutex; 28443935baa9SDavid Gibson } 284586e5216fSAdam Litke 284683c54070SNick Piggin ret = 0; 28471e8f889bSDavid Gibson 284857303d80SAndy Whitcroft /* 284957303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 285057303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 285157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 285257303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 285357303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 285457303d80SAndy Whitcroft * consumed. 285557303d80SAndy Whitcroft */ 2856788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 28572b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28582b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2859b4d1d99fSDavid Gibson goto out_mutex; 28602b26736cSAndy Whitcroft } 286157303d80SAndy Whitcroft 2862f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 286357303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 286457303d80SAndy Whitcroft vma, address); 286557303d80SAndy Whitcroft } 286657303d80SAndy Whitcroft 286756c9cfb1SNaoya Horiguchi /* 286856c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 286956c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 287056c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 287156c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 287256c9cfb1SNaoya Horiguchi * so no worry about deadlock. 287356c9cfb1SNaoya Horiguchi */ 28740fe6e20bSNaoya Horiguchi page = pte_page(entry); 287566aebce7SChris Metcalf get_page(page); 287656c9cfb1SNaoya Horiguchi if (page != pagecache_page) 28770fe6e20bSNaoya Horiguchi lock_page(page); 28780fe6e20bSNaoya Horiguchi 28791e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 28801e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2881b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2882b4d1d99fSDavid Gibson goto out_page_table_lock; 2883b4d1d99fSDavid Gibson 2884b4d1d99fSDavid Gibson 2885788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2886b4d1d99fSDavid Gibson if (!pte_write(entry)) { 288757303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 288857303d80SAndy Whitcroft pagecache_page); 2889b4d1d99fSDavid Gibson goto out_page_table_lock; 2890b4d1d99fSDavid Gibson } 2891b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2892b4d1d99fSDavid Gibson } 2893b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2894788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2895788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 28964b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2897b4d1d99fSDavid Gibson 2898b4d1d99fSDavid Gibson out_page_table_lock: 28991e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 290057303d80SAndy Whitcroft 290157303d80SAndy Whitcroft if (pagecache_page) { 290257303d80SAndy Whitcroft unlock_page(pagecache_page); 290357303d80SAndy Whitcroft put_page(pagecache_page); 290457303d80SAndy Whitcroft } 29051f64d69cSDean Nelson if (page != pagecache_page) 290656c9cfb1SNaoya Horiguchi unlock_page(page); 290766aebce7SChris Metcalf put_page(page); 290857303d80SAndy Whitcroft 2909b4d1d99fSDavid Gibson out_mutex: 29103935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 29111e8f889bSDavid Gibson 29121e8f889bSDavid Gibson return ret; 291386e5216fSAdam Litke } 291486e5216fSAdam Litke 2915ceb86879SAndi Kleen /* Can be overriden by architectures */ 2916ceb86879SAndi Kleen __attribute__((weak)) struct page * 2917ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2918ceb86879SAndi Kleen pud_t *pud, int write) 2919ceb86879SAndi Kleen { 2920ceb86879SAndi Kleen BUG(); 2921ceb86879SAndi Kleen return NULL; 2922ceb86879SAndi Kleen } 2923ceb86879SAndi Kleen 292463551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 292563551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 29265b23dbe8SAdam Litke unsigned long *position, int *length, int i, 29272a15efc9SHugh Dickins unsigned int flags) 292863551ae0SDavid Gibson { 2929d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2930d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 293163551ae0SDavid Gibson int remainder = *length; 2932a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 293363551ae0SDavid Gibson 29341c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 293563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 293663551ae0SDavid Gibson pte_t *pte; 29372a15efc9SHugh Dickins int absent; 293863551ae0SDavid Gibson struct page *page; 293963551ae0SDavid Gibson 29404c887265SAdam Litke /* 29414c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 29422a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 29434c887265SAdam Litke * first, for the page indexing below to work. 29444c887265SAdam Litke */ 2945a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 29462a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 294763551ae0SDavid Gibson 29482a15efc9SHugh Dickins /* 29492a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 29503ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 29513ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 29523ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 29533ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 29542a15efc9SHugh Dickins */ 29553ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 29563ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 29572a15efc9SHugh Dickins remainder = 0; 29582a15efc9SHugh Dickins break; 29592a15efc9SHugh Dickins } 29602a15efc9SHugh Dickins 29612a15efc9SHugh Dickins if (absent || 29622a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 29634c887265SAdam Litke int ret; 29644c887265SAdam Litke 29654c887265SAdam Litke spin_unlock(&mm->page_table_lock); 29662a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 29672a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 29684c887265SAdam Litke spin_lock(&mm->page_table_lock); 2969a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 29704c887265SAdam Litke continue; 29714c887265SAdam Litke 29721c59827dSHugh Dickins remainder = 0; 29731c59827dSHugh Dickins break; 29741c59827dSHugh Dickins } 297563551ae0SDavid Gibson 2976a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 29777f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2978d5d4b0aaSChen, Kenneth W same_page: 2979d6692183SChen, Kenneth W if (pages) { 298069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 29814b2e38adSKOSAKI Motohiro get_page(pages[i]); 2982d6692183SChen, Kenneth W } 298363551ae0SDavid Gibson 298463551ae0SDavid Gibson if (vmas) 298563551ae0SDavid Gibson vmas[i] = vma; 298663551ae0SDavid Gibson 298763551ae0SDavid Gibson vaddr += PAGE_SIZE; 2988d5d4b0aaSChen, Kenneth W ++pfn_offset; 298963551ae0SDavid Gibson --remainder; 299063551ae0SDavid Gibson ++i; 2991d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2992a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2993d5d4b0aaSChen, Kenneth W /* 2994d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2995d5d4b0aaSChen, Kenneth W * of this compound page. 2996d5d4b0aaSChen, Kenneth W */ 2997d5d4b0aaSChen, Kenneth W goto same_page; 2998d5d4b0aaSChen, Kenneth W } 299963551ae0SDavid Gibson } 30001c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 300163551ae0SDavid Gibson *length = remainder; 300263551ae0SDavid Gibson *position = vaddr; 300363551ae0SDavid Gibson 30042a15efc9SHugh Dickins return i ? i : -EFAULT; 300563551ae0SDavid Gibson } 30068f860591SZhang, Yanmin 30078f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 30088f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 30098f860591SZhang, Yanmin { 30108f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 30118f860591SZhang, Yanmin unsigned long start = address; 30128f860591SZhang, Yanmin pte_t *ptep; 30138f860591SZhang, Yanmin pte_t pte; 3014a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 30158f860591SZhang, Yanmin 30168f860591SZhang, Yanmin BUG_ON(address >= end); 30178f860591SZhang, Yanmin flush_cache_range(vma, address, end); 30188f860591SZhang, Yanmin 30193d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 30208f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 3021a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 30228f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 30238f860591SZhang, Yanmin if (!ptep) 30248f860591SZhang, Yanmin continue; 302539dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 302639dde65cSChen, Kenneth W continue; 30277f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 30288f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 30298f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 30308f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 30318f860591SZhang, Yanmin } 30328f860591SZhang, Yanmin } 30338f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 3034d833352aSMel Gorman /* 3035d833352aSMel Gorman * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3036d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3037d833352aSMel Gorman * once we release i_mmap_mutex, another task can do the final put_page 3038d833352aSMel Gorman * and that page table be reused and filled with junk. 3039d833352aSMel Gorman */ 30408f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 3041d833352aSMel Gorman mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 30428f860591SZhang, Yanmin } 30438f860591SZhang, Yanmin 3044a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3045a1e78772SMel Gorman long from, long to, 30465a6fe125SMel Gorman struct vm_area_struct *vma, 3047ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3048e4e574b7SAdam Litke { 304917c9d12eSMel Gorman long ret, chg; 3050a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 305190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 3052e4e574b7SAdam Litke 3053a1e78772SMel Gorman /* 305417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 305517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 305690481622SDavid Gibson * without using reserves 305717c9d12eSMel Gorman */ 3058ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 305917c9d12eSMel Gorman return 0; 306017c9d12eSMel Gorman 306117c9d12eSMel Gorman /* 3062a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3063a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3064a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3065a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3066a1e78772SMel Gorman */ 3067f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 3068e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 30695a6fe125SMel Gorman else { 30705a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 30715a6fe125SMel Gorman if (!resv_map) 30725a6fe125SMel Gorman return -ENOMEM; 30735a6fe125SMel Gorman 307417c9d12eSMel Gorman chg = to - from; 307517c9d12eSMel Gorman 30765a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 30775a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 30785a6fe125SMel Gorman } 30795a6fe125SMel Gorman 3080c50ac050SDave Hansen if (chg < 0) { 3081c50ac050SDave Hansen ret = chg; 3082c50ac050SDave Hansen goto out_err; 3083c50ac050SDave Hansen } 308417c9d12eSMel Gorman 308590481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3086c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3087c50ac050SDave Hansen ret = -ENOSPC; 3088c50ac050SDave Hansen goto out_err; 3089c50ac050SDave Hansen } 309017c9d12eSMel Gorman 309117c9d12eSMel Gorman /* 309217c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 309390481622SDavid Gibson * Hand the pages back to the subpool if there are not 309417c9d12eSMel Gorman */ 309517c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 309617c9d12eSMel Gorman if (ret < 0) { 309790481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3098c50ac050SDave Hansen goto out_err; 309917c9d12eSMel Gorman } 310017c9d12eSMel Gorman 310117c9d12eSMel Gorman /* 310217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 310317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 310417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 310517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 310617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 310717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 310817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 310917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 311017c9d12eSMel Gorman * else has to be done for private mappings here 311117c9d12eSMel Gorman */ 3112f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 311317c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 3114a43a8c39SChen, Kenneth W return 0; 3115c50ac050SDave Hansen out_err: 31164523e145SDave Hansen if (vma) 3117c50ac050SDave Hansen resv_map_put(vma); 3118c50ac050SDave Hansen return ret; 3119a43a8c39SChen, Kenneth W } 3120a43a8c39SChen, Kenneth W 3121a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3122a43a8c39SChen, Kenneth W { 3123a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 3124a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 312590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 312645c682a6SKen Chen 312745c682a6SKen Chen spin_lock(&inode->i_lock); 3128e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 312945c682a6SKen Chen spin_unlock(&inode->i_lock); 313045c682a6SKen Chen 313190481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3132a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3133a43a8c39SChen, Kenneth W } 313493f70f90SNaoya Horiguchi 3135d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3136d5bd9106SAndi Kleen 31376de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 31386de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 31396de2b1aaSNaoya Horiguchi { 31406de2b1aaSNaoya Horiguchi struct page *page; 31416de2b1aaSNaoya Horiguchi struct page *tmp; 31426de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 31436de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 31446de2b1aaSNaoya Horiguchi 31456de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 31466de2b1aaSNaoya Horiguchi if (page == hpage) 31476de2b1aaSNaoya Horiguchi return 1; 31486de2b1aaSNaoya Horiguchi return 0; 31496de2b1aaSNaoya Horiguchi } 31506de2b1aaSNaoya Horiguchi 315193f70f90SNaoya Horiguchi /* 315293f70f90SNaoya Horiguchi * This function is called from memory failure code. 315393f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 315493f70f90SNaoya Horiguchi */ 31556de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 315693f70f90SNaoya Horiguchi { 315793f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 315893f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 31596de2b1aaSNaoya Horiguchi int ret = -EBUSY; 316093f70f90SNaoya Horiguchi 316193f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 31626de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 316393f70f90SNaoya Horiguchi list_del(&hpage->lru); 31648c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 316593f70f90SNaoya Horiguchi h->free_huge_pages--; 316693f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 31676de2b1aaSNaoya Horiguchi ret = 0; 316893f70f90SNaoya Horiguchi } 31696de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 31706de2b1aaSNaoya Horiguchi return ret; 31716de2b1aaSNaoya Horiguchi } 31726de2b1aaSNaoya Horiguchi #endif 3173