11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 31da177e4SLinus Torvalds * (C) William Irwin, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/gfp.h> 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/module.h> 91da177e4SLinus Torvalds #include <linux/mm.h> 10e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 111da177e4SLinus Torvalds #include <linux/sysctl.h> 121da177e4SLinus Torvalds #include <linux/highmem.h> 13cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 141da177e4SLinus Torvalds #include <linux/nodemask.h> 1563551ae0SDavid Gibson #include <linux/pagemap.h> 165da7ca86SChristoph Lameter #include <linux/mempolicy.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 19aa888a74SAndi Kleen #include <linux/bootmem.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 21d6606683SLinus Torvalds 2263551ae0SDavid Gibson #include <asm/page.h> 2363551ae0SDavid Gibson #include <asm/pgtable.h> 2478a34ae2SAdrian Bunk #include <asm/io.h> 2563551ae0SDavid Gibson 2663551ae0SDavid Gibson #include <linux/hugetlb.h> 277835e98bSNick Piggin #include "internal.h" 281da177e4SLinus Torvalds 291da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 30396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 31396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 32a5516438SAndi Kleen 33e5ff2159SAndi Kleen static int max_hstate; 34e5ff2159SAndi Kleen unsigned int default_hstate_idx; 35e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 36e5ff2159SAndi Kleen 3753ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 3853ba51d2SJon Tollefson 39e5ff2159SAndi Kleen /* for command line parsing */ 40e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 41e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 42e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 43e5ff2159SAndi Kleen 44e5ff2159SAndi Kleen #define for_each_hstate(h) \ 45e5ff2159SAndi Kleen for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) 46396faf03SMel Gorman 473935baa9SDavid Gibson /* 483935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 493935baa9SDavid Gibson */ 503935baa9SDavid Gibson static DEFINE_SPINLOCK(hugetlb_lock); 510bd0f9fbSEric Paris 52e7c4b0bfSAndy Whitcroft /* 5396822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 5496822904SAndy Whitcroft * across the pages in a mapping. 5584afd99bSAndy Whitcroft * 5684afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 5784afd99bSAndy Whitcroft * and the hugetlb_instantion_mutex. To access or modify a region the caller 5884afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 5984afd99bSAndy Whitcroft * the hugetlb_instantiation mutex: 6084afd99bSAndy Whitcroft * 6184afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 6284afd99bSAndy Whitcroft * or 6384afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 6484afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 6596822904SAndy Whitcroft */ 6696822904SAndy Whitcroft struct file_region { 6796822904SAndy Whitcroft struct list_head link; 6896822904SAndy Whitcroft long from; 6996822904SAndy Whitcroft long to; 7096822904SAndy Whitcroft }; 7196822904SAndy Whitcroft 7296822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 7396822904SAndy Whitcroft { 7496822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 7596822904SAndy Whitcroft 7696822904SAndy Whitcroft /* Locate the region we are either in or before. */ 7796822904SAndy Whitcroft list_for_each_entry(rg, head, link) 7896822904SAndy Whitcroft if (f <= rg->to) 7996822904SAndy Whitcroft break; 8096822904SAndy Whitcroft 8196822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 8296822904SAndy Whitcroft if (f > rg->from) 8396822904SAndy Whitcroft f = rg->from; 8496822904SAndy Whitcroft 8596822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 8696822904SAndy Whitcroft nrg = rg; 8796822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 8896822904SAndy Whitcroft if (&rg->link == head) 8996822904SAndy Whitcroft break; 9096822904SAndy Whitcroft if (rg->from > t) 9196822904SAndy Whitcroft break; 9296822904SAndy Whitcroft 9396822904SAndy Whitcroft /* If this area reaches higher then extend our area to 9496822904SAndy Whitcroft * include it completely. If this is not the first area 9596822904SAndy Whitcroft * which we intend to reuse, free it. */ 9696822904SAndy Whitcroft if (rg->to > t) 9796822904SAndy Whitcroft t = rg->to; 9896822904SAndy Whitcroft if (rg != nrg) { 9996822904SAndy Whitcroft list_del(&rg->link); 10096822904SAndy Whitcroft kfree(rg); 10196822904SAndy Whitcroft } 10296822904SAndy Whitcroft } 10396822904SAndy Whitcroft nrg->from = f; 10496822904SAndy Whitcroft nrg->to = t; 10596822904SAndy Whitcroft return 0; 10696822904SAndy Whitcroft } 10796822904SAndy Whitcroft 10896822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 10996822904SAndy Whitcroft { 11096822904SAndy Whitcroft struct file_region *rg, *nrg; 11196822904SAndy Whitcroft long chg = 0; 11296822904SAndy Whitcroft 11396822904SAndy Whitcroft /* Locate the region we are before or in. */ 11496822904SAndy Whitcroft list_for_each_entry(rg, head, link) 11596822904SAndy Whitcroft if (f <= rg->to) 11696822904SAndy Whitcroft break; 11796822904SAndy Whitcroft 11896822904SAndy Whitcroft /* If we are below the current region then a new region is required. 11996822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 12096822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 12196822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 12296822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 12396822904SAndy Whitcroft if (!nrg) 12496822904SAndy Whitcroft return -ENOMEM; 12596822904SAndy Whitcroft nrg->from = f; 12696822904SAndy Whitcroft nrg->to = f; 12796822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 12896822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 12996822904SAndy Whitcroft 13096822904SAndy Whitcroft return t - f; 13196822904SAndy Whitcroft } 13296822904SAndy Whitcroft 13396822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 13496822904SAndy Whitcroft if (f > rg->from) 13596822904SAndy Whitcroft f = rg->from; 13696822904SAndy Whitcroft chg = t - f; 13796822904SAndy Whitcroft 13896822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 13996822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 14096822904SAndy Whitcroft if (&rg->link == head) 14196822904SAndy Whitcroft break; 14296822904SAndy Whitcroft if (rg->from > t) 14396822904SAndy Whitcroft return chg; 14496822904SAndy Whitcroft 14596822904SAndy Whitcroft /* We overlap with this area, if it extends futher than 14696822904SAndy Whitcroft * us then we must extend ourselves. Account for its 14796822904SAndy Whitcroft * existing reservation. */ 14896822904SAndy Whitcroft if (rg->to > t) { 14996822904SAndy Whitcroft chg += rg->to - t; 15096822904SAndy Whitcroft t = rg->to; 15196822904SAndy Whitcroft } 15296822904SAndy Whitcroft chg -= rg->to - rg->from; 15396822904SAndy Whitcroft } 15496822904SAndy Whitcroft return chg; 15596822904SAndy Whitcroft } 15696822904SAndy Whitcroft 15796822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 15896822904SAndy Whitcroft { 15996822904SAndy Whitcroft struct file_region *rg, *trg; 16096822904SAndy Whitcroft long chg = 0; 16196822904SAndy Whitcroft 16296822904SAndy Whitcroft /* Locate the region we are either in or before. */ 16396822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16496822904SAndy Whitcroft if (end <= rg->to) 16596822904SAndy Whitcroft break; 16696822904SAndy Whitcroft if (&rg->link == head) 16796822904SAndy Whitcroft return 0; 16896822904SAndy Whitcroft 16996822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 17096822904SAndy Whitcroft if (end > rg->from) { 17196822904SAndy Whitcroft chg = rg->to - end; 17296822904SAndy Whitcroft rg->to = end; 17396822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 17496822904SAndy Whitcroft } 17596822904SAndy Whitcroft 17696822904SAndy Whitcroft /* Drop any remaining regions. */ 17796822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 17896822904SAndy Whitcroft if (&rg->link == head) 17996822904SAndy Whitcroft break; 18096822904SAndy Whitcroft chg += rg->to - rg->from; 18196822904SAndy Whitcroft list_del(&rg->link); 18296822904SAndy Whitcroft kfree(rg); 18396822904SAndy Whitcroft } 18496822904SAndy Whitcroft return chg; 18596822904SAndy Whitcroft } 18696822904SAndy Whitcroft 18784afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 18884afd99bSAndy Whitcroft { 18984afd99bSAndy Whitcroft struct file_region *rg; 19084afd99bSAndy Whitcroft long chg = 0; 19184afd99bSAndy Whitcroft 19284afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 19384afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 19484afd99bSAndy Whitcroft int seg_from; 19584afd99bSAndy Whitcroft int seg_to; 19684afd99bSAndy Whitcroft 19784afd99bSAndy Whitcroft if (rg->to <= f) 19884afd99bSAndy Whitcroft continue; 19984afd99bSAndy Whitcroft if (rg->from >= t) 20084afd99bSAndy Whitcroft break; 20184afd99bSAndy Whitcroft 20284afd99bSAndy Whitcroft seg_from = max(rg->from, f); 20384afd99bSAndy Whitcroft seg_to = min(rg->to, t); 20484afd99bSAndy Whitcroft 20584afd99bSAndy Whitcroft chg += seg_to - seg_from; 20684afd99bSAndy Whitcroft } 20784afd99bSAndy Whitcroft 20884afd99bSAndy Whitcroft return chg; 20984afd99bSAndy Whitcroft } 21084afd99bSAndy Whitcroft 21196822904SAndy Whitcroft /* 212e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 213e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 214e7c4b0bfSAndy Whitcroft */ 215a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 216a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 217e7c4b0bfSAndy Whitcroft { 218a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 219a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 220e7c4b0bfSAndy Whitcroft } 221e7c4b0bfSAndy Whitcroft 22284afd99bSAndy Whitcroft /* 22308fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 22408fba699SMel Gorman * cases this will be same size as used by the page table entries. 22508fba699SMel Gorman */ 22608fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 22708fba699SMel Gorman { 22808fba699SMel Gorman struct hstate *hstate; 22908fba699SMel Gorman 23008fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 23108fba699SMel Gorman return PAGE_SIZE; 23208fba699SMel Gorman 23308fba699SMel Gorman hstate = hstate_vma(vma); 23408fba699SMel Gorman 23508fba699SMel Gorman return 1UL << (hstate->order + PAGE_SHIFT); 23608fba699SMel Gorman } 237f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 23808fba699SMel Gorman 23908fba699SMel Gorman /* 2403340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 2413340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 2423340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 2433340289dSMel Gorman * function is required. 2443340289dSMel Gorman */ 2453340289dSMel Gorman #ifndef vma_mmu_pagesize 2463340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 2473340289dSMel Gorman { 2483340289dSMel Gorman return vma_kernel_pagesize(vma); 2493340289dSMel Gorman } 2503340289dSMel Gorman #endif 2513340289dSMel Gorman 2523340289dSMel Gorman /* 25384afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 25484afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 25584afd99bSAndy Whitcroft * alignment. 25684afd99bSAndy Whitcroft */ 25784afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 25884afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 25904f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 26084afd99bSAndy Whitcroft 261a1e78772SMel Gorman /* 262a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 263a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 264a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 265a1e78772SMel Gorman * 266a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 267a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 268a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 269a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 27084afd99bSAndy Whitcroft * 27184afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 27284afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 27384afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 27484afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 27584afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 27684afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 27784afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 27884afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 279a1e78772SMel Gorman */ 280e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 281e7c4b0bfSAndy Whitcroft { 282e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 283e7c4b0bfSAndy Whitcroft } 284e7c4b0bfSAndy Whitcroft 285e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 286e7c4b0bfSAndy Whitcroft unsigned long value) 287e7c4b0bfSAndy Whitcroft { 288e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 289e7c4b0bfSAndy Whitcroft } 290e7c4b0bfSAndy Whitcroft 29184afd99bSAndy Whitcroft struct resv_map { 29284afd99bSAndy Whitcroft struct kref refs; 29384afd99bSAndy Whitcroft struct list_head regions; 29484afd99bSAndy Whitcroft }; 29584afd99bSAndy Whitcroft 2962a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 29784afd99bSAndy Whitcroft { 29884afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 29984afd99bSAndy Whitcroft if (!resv_map) 30084afd99bSAndy Whitcroft return NULL; 30184afd99bSAndy Whitcroft 30284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 30384afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 30484afd99bSAndy Whitcroft 30584afd99bSAndy Whitcroft return resv_map; 30684afd99bSAndy Whitcroft } 30784afd99bSAndy Whitcroft 3082a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 30984afd99bSAndy Whitcroft { 31084afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 31184afd99bSAndy Whitcroft 31284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 31384afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 31484afd99bSAndy Whitcroft kfree(resv_map); 31584afd99bSAndy Whitcroft } 31684afd99bSAndy Whitcroft 31784afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 318a1e78772SMel Gorman { 319a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 320f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 32184afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 32284afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 3232a4b3dedSHarvey Harrison return NULL; 324a1e78772SMel Gorman } 325a1e78772SMel Gorman 32684afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 327a1e78772SMel Gorman { 328a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 329f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 330a1e78772SMel Gorman 33184afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 33284afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 33304f2cbe3SMel Gorman } 33404f2cbe3SMel Gorman 33504f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 33604f2cbe3SMel Gorman { 33704f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 338f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 339e7c4b0bfSAndy Whitcroft 340e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 34104f2cbe3SMel Gorman } 34204f2cbe3SMel Gorman 34304f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 34404f2cbe3SMel Gorman { 34504f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 346e7c4b0bfSAndy Whitcroft 347e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 348a1e78772SMel Gorman } 349a1e78772SMel Gorman 350a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 351a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 352a5516438SAndi Kleen struct vm_area_struct *vma) 353a1e78772SMel Gorman { 354c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 355c37f9fb1SAndy Whitcroft return; 356c37f9fb1SAndy Whitcroft 357f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 358a1e78772SMel Gorman /* Shared mappings always use reserves */ 359a5516438SAndi Kleen h->resv_huge_pages--; 36084afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 361a1e78772SMel Gorman /* 362a1e78772SMel Gorman * Only the process that called mmap() has reserves for 363a1e78772SMel Gorman * private mappings. 364a1e78772SMel Gorman */ 365a5516438SAndi Kleen h->resv_huge_pages--; 366a1e78772SMel Gorman } 367a1e78772SMel Gorman } 368a1e78772SMel Gorman 36904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 370a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 371a1e78772SMel Gorman { 372a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 373f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 374a1e78772SMel Gorman vma->vm_private_data = (void *)0; 375a1e78772SMel Gorman } 376a1e78772SMel Gorman 377a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 3787f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 379a1e78772SMel Gorman { 380f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 381a1e78772SMel Gorman return 1; 3827f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3837f09ca51SMel Gorman return 1; 3847f09ca51SMel Gorman return 0; 385a1e78772SMel Gorman } 386a1e78772SMel Gorman 38769d177c2SAndy Whitcroft static void clear_gigantic_page(struct page *page, 38869d177c2SAndy Whitcroft unsigned long addr, unsigned long sz) 38969d177c2SAndy Whitcroft { 39069d177c2SAndy Whitcroft int i; 39169d177c2SAndy Whitcroft struct page *p = page; 39269d177c2SAndy Whitcroft 39369d177c2SAndy Whitcroft might_sleep(); 39469d177c2SAndy Whitcroft for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) { 39569d177c2SAndy Whitcroft cond_resched(); 39669d177c2SAndy Whitcroft clear_user_highpage(p, addr + i * PAGE_SIZE); 39769d177c2SAndy Whitcroft } 39869d177c2SAndy Whitcroft } 399a5516438SAndi Kleen static void clear_huge_page(struct page *page, 400a5516438SAndi Kleen unsigned long addr, unsigned long sz) 40179ac6ba4SDavid Gibson { 40279ac6ba4SDavid Gibson int i; 40379ac6ba4SDavid Gibson 404ebdd4aeaSHannes Eder if (unlikely(sz > MAX_ORDER_NR_PAGES)) { 405ebdd4aeaSHannes Eder clear_gigantic_page(page, addr, sz); 406ebdd4aeaSHannes Eder return; 407ebdd4aeaSHannes Eder } 40869d177c2SAndy Whitcroft 40979ac6ba4SDavid Gibson might_sleep(); 410a5516438SAndi Kleen for (i = 0; i < sz/PAGE_SIZE; i++) { 41179ac6ba4SDavid Gibson cond_resched(); 412281e0e3bSRalf Baechle clear_user_highpage(page + i, addr + i * PAGE_SIZE); 41379ac6ba4SDavid Gibson } 41479ac6ba4SDavid Gibson } 41579ac6ba4SDavid Gibson 41669d177c2SAndy Whitcroft static void copy_gigantic_page(struct page *dst, struct page *src, 41769d177c2SAndy Whitcroft unsigned long addr, struct vm_area_struct *vma) 41869d177c2SAndy Whitcroft { 41969d177c2SAndy Whitcroft int i; 42069d177c2SAndy Whitcroft struct hstate *h = hstate_vma(vma); 42169d177c2SAndy Whitcroft struct page *dst_base = dst; 42269d177c2SAndy Whitcroft struct page *src_base = src; 42369d177c2SAndy Whitcroft might_sleep(); 42469d177c2SAndy Whitcroft for (i = 0; i < pages_per_huge_page(h); ) { 42569d177c2SAndy Whitcroft cond_resched(); 42669d177c2SAndy Whitcroft copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 42769d177c2SAndy Whitcroft 42869d177c2SAndy Whitcroft i++; 42969d177c2SAndy Whitcroft dst = mem_map_next(dst, dst_base, i); 43069d177c2SAndy Whitcroft src = mem_map_next(src, src_base, i); 43169d177c2SAndy Whitcroft } 43269d177c2SAndy Whitcroft } 43379ac6ba4SDavid Gibson static void copy_huge_page(struct page *dst, struct page *src, 4349de455b2SAtsushi Nemoto unsigned long addr, struct vm_area_struct *vma) 43579ac6ba4SDavid Gibson { 43679ac6ba4SDavid Gibson int i; 437a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43879ac6ba4SDavid Gibson 439ebdd4aeaSHannes Eder if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 440ebdd4aeaSHannes Eder copy_gigantic_page(dst, src, addr, vma); 441ebdd4aeaSHannes Eder return; 442ebdd4aeaSHannes Eder } 44369d177c2SAndy Whitcroft 44479ac6ba4SDavid Gibson might_sleep(); 445a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 44679ac6ba4SDavid Gibson cond_resched(); 4479de455b2SAtsushi Nemoto copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 44879ac6ba4SDavid Gibson } 44979ac6ba4SDavid Gibson } 45079ac6ba4SDavid Gibson 451a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 4521da177e4SLinus Torvalds { 4531da177e4SLinus Torvalds int nid = page_to_nid(page); 454a5516438SAndi Kleen list_add(&page->lru, &h->hugepage_freelists[nid]); 455a5516438SAndi Kleen h->free_huge_pages++; 456a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 4571da177e4SLinus Torvalds } 4581da177e4SLinus Torvalds 459a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 460a5516438SAndi Kleen struct vm_area_struct *vma, 46104f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 4621da177e4SLinus Torvalds { 46331a5c6e4SNishanth Aravamudan int nid; 4641da177e4SLinus Torvalds struct page *page = NULL; 465480eccf9SLee Schermerhorn struct mempolicy *mpol; 46619770b32SMel Gorman nodemask_t *nodemask; 467396faf03SMel Gorman struct zonelist *zonelist = huge_zonelist(vma, address, 46819770b32SMel Gorman htlb_alloc_mask, &mpol, &nodemask); 469dd1a239fSMel Gorman struct zone *zone; 470dd1a239fSMel Gorman struct zoneref *z; 4711da177e4SLinus Torvalds 472a1e78772SMel Gorman /* 473a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 474a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 475a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 476a1e78772SMel Gorman */ 4777f09ca51SMel Gorman if (!vma_has_reserves(vma) && 478a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 479a1e78772SMel Gorman return NULL; 480a1e78772SMel Gorman 48104f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 482a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 48304f2cbe3SMel Gorman return NULL; 48404f2cbe3SMel Gorman 48519770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 48619770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 48754a6eb5cSMel Gorman nid = zone_to_nid(zone); 48854a6eb5cSMel Gorman if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 489a5516438SAndi Kleen !list_empty(&h->hugepage_freelists[nid])) { 490a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 4911da177e4SLinus Torvalds struct page, lru); 4921da177e4SLinus Torvalds list_del(&page->lru); 493a5516438SAndi Kleen h->free_huge_pages--; 494a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 49504f2cbe3SMel Gorman 49604f2cbe3SMel Gorman if (!avoid_reserve) 497a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 498a1e78772SMel Gorman 4995ab3ee7bSKen Chen break; 5001da177e4SLinus Torvalds } 5013abf7afdSAndrew Morton } 50252cd3b07SLee Schermerhorn mpol_cond_put(mpol); 5031da177e4SLinus Torvalds return page; 5041da177e4SLinus Torvalds } 5051da177e4SLinus Torvalds 506a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5076af2acb6SAdam Litke { 5086af2acb6SAdam Litke int i; 509a5516438SAndi Kleen 51018229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 51118229df5SAndy Whitcroft 512a5516438SAndi Kleen h->nr_huge_pages--; 513a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 514a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 5156af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 5166af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 5176af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 5186af2acb6SAdam Litke } 5196af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5206af2acb6SAdam Litke set_page_refcounted(page); 5217f2e9525SGerald Schaefer arch_release_hugepage(page); 522a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 5236af2acb6SAdam Litke } 5246af2acb6SAdam Litke 525e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 526e5ff2159SAndi Kleen { 527e5ff2159SAndi Kleen struct hstate *h; 528e5ff2159SAndi Kleen 529e5ff2159SAndi Kleen for_each_hstate(h) { 530e5ff2159SAndi Kleen if (huge_page_size(h) == size) 531e5ff2159SAndi Kleen return h; 532e5ff2159SAndi Kleen } 533e5ff2159SAndi Kleen return NULL; 534e5ff2159SAndi Kleen } 535e5ff2159SAndi Kleen 53627a85ef1SDavid Gibson static void free_huge_page(struct page *page) 53727a85ef1SDavid Gibson { 538a5516438SAndi Kleen /* 539a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 540a5516438SAndi Kleen * compound page destructor. 541a5516438SAndi Kleen */ 542e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 5437893d1d5SAdam Litke int nid = page_to_nid(page); 544c79fb75eSAdam Litke struct address_space *mapping; 54527a85ef1SDavid Gibson 546c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 547e5df70abSAndy Whitcroft set_page_private(page, 0); 5487893d1d5SAdam Litke BUG_ON(page_count(page)); 54927a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 55027a85ef1SDavid Gibson 55127a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 552aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 553a5516438SAndi Kleen update_and_free_page(h, page); 554a5516438SAndi Kleen h->surplus_huge_pages--; 555a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 5567893d1d5SAdam Litke } else { 557a5516438SAndi Kleen enqueue_huge_page(h, page); 5587893d1d5SAdam Litke } 55927a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 560c79fb75eSAdam Litke if (mapping) 5619a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 56227a85ef1SDavid Gibson } 56327a85ef1SDavid Gibson 564a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 565b7ba30c6SAndi Kleen { 566b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 567b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 568a5516438SAndi Kleen h->nr_huge_pages++; 569a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 570b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 571b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 572b7ba30c6SAndi Kleen } 573b7ba30c6SAndi Kleen 57420a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 57520a0307cSWu Fengguang { 57620a0307cSWu Fengguang int i; 57720a0307cSWu Fengguang int nr_pages = 1 << order; 57820a0307cSWu Fengguang struct page *p = page + 1; 57920a0307cSWu Fengguang 58020a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 58120a0307cSWu Fengguang set_compound_order(page, order); 58220a0307cSWu Fengguang __SetPageHead(page); 58320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 58420a0307cSWu Fengguang __SetPageTail(p); 58520a0307cSWu Fengguang p->first_page = page; 58620a0307cSWu Fengguang } 58720a0307cSWu Fengguang } 58820a0307cSWu Fengguang 58920a0307cSWu Fengguang int PageHuge(struct page *page) 59020a0307cSWu Fengguang { 59120a0307cSWu Fengguang compound_page_dtor *dtor; 59220a0307cSWu Fengguang 59320a0307cSWu Fengguang if (!PageCompound(page)) 59420a0307cSWu Fengguang return 0; 59520a0307cSWu Fengguang 59620a0307cSWu Fengguang page = compound_head(page); 59720a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 59820a0307cSWu Fengguang 59920a0307cSWu Fengguang return dtor == free_huge_page; 60020a0307cSWu Fengguang } 60120a0307cSWu Fengguang 602a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6031da177e4SLinus Torvalds { 6041da177e4SLinus Torvalds struct page *page; 605f96efd58SJoe Jin 606aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 607aa888a74SAndi Kleen return NULL; 608aa888a74SAndi Kleen 6096484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 610551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 611551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 612a5516438SAndi Kleen huge_page_order(h)); 6131da177e4SLinus Torvalds if (page) { 6147f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 615caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6167b8ee84dSHarvey Harrison return NULL; 6177f2e9525SGerald Schaefer } 618a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6191da177e4SLinus Torvalds } 62063b4613cSNishanth Aravamudan 62163b4613cSNishanth Aravamudan return page; 62263b4613cSNishanth Aravamudan } 62363b4613cSNishanth Aravamudan 6245ced66c9SAndi Kleen /* 6255ced66c9SAndi Kleen * Use a helper variable to find the next node and then 626e8c5c824SLee Schermerhorn * copy it back to next_nid_to_alloc afterwards: 6275ced66c9SAndi Kleen * otherwise there's a window in which a racer might 6286484eb3eSMel Gorman * pass invalid nid MAX_NUMNODES to alloc_pages_exact_node. 6295ced66c9SAndi Kleen * But we don't need to use a spin_lock here: it really 6305ced66c9SAndi Kleen * doesn't matter if occasionally a racer chooses the 6315ced66c9SAndi Kleen * same nid as we do. Move nid forward in the mask even 6325ced66c9SAndi Kleen * if we just successfully allocated a hugepage so that 6335ced66c9SAndi Kleen * the next caller gets hugepages on the next node. 6345ced66c9SAndi Kleen */ 635e8c5c824SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h) 6365ced66c9SAndi Kleen { 6375ced66c9SAndi Kleen int next_nid; 638e8c5c824SLee Schermerhorn next_nid = next_node(h->next_nid_to_alloc, node_online_map); 6395ced66c9SAndi Kleen if (next_nid == MAX_NUMNODES) 6405ced66c9SAndi Kleen next_nid = first_node(node_online_map); 641e8c5c824SLee Schermerhorn h->next_nid_to_alloc = next_nid; 6425ced66c9SAndi Kleen return next_nid; 6435ced66c9SAndi Kleen } 6445ced66c9SAndi Kleen 645a5516438SAndi Kleen static int alloc_fresh_huge_page(struct hstate *h) 64663b4613cSNishanth Aravamudan { 64763b4613cSNishanth Aravamudan struct page *page; 64863b4613cSNishanth Aravamudan int start_nid; 64963b4613cSNishanth Aravamudan int next_nid; 65063b4613cSNishanth Aravamudan int ret = 0; 65163b4613cSNishanth Aravamudan 652e8c5c824SLee Schermerhorn start_nid = h->next_nid_to_alloc; 653e8c5c824SLee Schermerhorn next_nid = start_nid; 65463b4613cSNishanth Aravamudan 65563b4613cSNishanth Aravamudan do { 656e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 65763b4613cSNishanth Aravamudan if (page) 65863b4613cSNishanth Aravamudan ret = 1; 659e8c5c824SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h); 660e8c5c824SLee Schermerhorn } while (!page && next_nid != start_nid); 66163b4613cSNishanth Aravamudan 6623b116300SAdam Litke if (ret) 6633b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 6643b116300SAdam Litke else 6653b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 6663b116300SAdam Litke 66763b4613cSNishanth Aravamudan return ret; 6681da177e4SLinus Torvalds } 6691da177e4SLinus Torvalds 670e8c5c824SLee Schermerhorn /* 671e8c5c824SLee Schermerhorn * helper for free_pool_huge_page() - find next node 672e8c5c824SLee Schermerhorn * from which to free a huge page 673e8c5c824SLee Schermerhorn */ 674e8c5c824SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h) 675e8c5c824SLee Schermerhorn { 676e8c5c824SLee Schermerhorn int next_nid; 677e8c5c824SLee Schermerhorn next_nid = next_node(h->next_nid_to_free, node_online_map); 678e8c5c824SLee Schermerhorn if (next_nid == MAX_NUMNODES) 679e8c5c824SLee Schermerhorn next_nid = first_node(node_online_map); 680e8c5c824SLee Schermerhorn h->next_nid_to_free = next_nid; 681e8c5c824SLee Schermerhorn return next_nid; 682e8c5c824SLee Schermerhorn } 683e8c5c824SLee Schermerhorn 684e8c5c824SLee Schermerhorn /* 685e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 686e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 687e8c5c824SLee Schermerhorn * balanced over allowed nodes. 688e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 689e8c5c824SLee Schermerhorn */ 690685f3457SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, bool acct_surplus) 691e8c5c824SLee Schermerhorn { 692e8c5c824SLee Schermerhorn int start_nid; 693e8c5c824SLee Schermerhorn int next_nid; 694e8c5c824SLee Schermerhorn int ret = 0; 695e8c5c824SLee Schermerhorn 696e8c5c824SLee Schermerhorn start_nid = h->next_nid_to_free; 697e8c5c824SLee Schermerhorn next_nid = start_nid; 698e8c5c824SLee Schermerhorn 699e8c5c824SLee Schermerhorn do { 700685f3457SLee Schermerhorn /* 701685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 702685f3457SLee Schermerhorn * nodes with surplus pages. 703685f3457SLee Schermerhorn */ 704685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 705685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 706e8c5c824SLee Schermerhorn struct page *page = 707e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 708e8c5c824SLee Schermerhorn struct page, lru); 709e8c5c824SLee Schermerhorn list_del(&page->lru); 710e8c5c824SLee Schermerhorn h->free_huge_pages--; 711e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 712685f3457SLee Schermerhorn if (acct_surplus) { 713685f3457SLee Schermerhorn h->surplus_huge_pages--; 714685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 715685f3457SLee Schermerhorn } 716e8c5c824SLee Schermerhorn update_and_free_page(h, page); 717e8c5c824SLee Schermerhorn ret = 1; 718e8c5c824SLee Schermerhorn } 719e8c5c824SLee Schermerhorn next_nid = hstate_next_node_to_free(h); 720e8c5c824SLee Schermerhorn } while (!ret && next_nid != start_nid); 721e8c5c824SLee Schermerhorn 722e8c5c824SLee Schermerhorn return ret; 723e8c5c824SLee Schermerhorn } 724e8c5c824SLee Schermerhorn 725a5516438SAndi Kleen static struct page *alloc_buddy_huge_page(struct hstate *h, 726a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 7277893d1d5SAdam Litke { 7287893d1d5SAdam Litke struct page *page; 729d1c3fb1fSNishanth Aravamudan unsigned int nid; 7307893d1d5SAdam Litke 731aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 732aa888a74SAndi Kleen return NULL; 733aa888a74SAndi Kleen 734d1c3fb1fSNishanth Aravamudan /* 735d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 736d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 737d1c3fb1fSNishanth Aravamudan * overcommit 738d1c3fb1fSNishanth Aravamudan * 739d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 740d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 741d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 742d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 743d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 744d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 745d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 746d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 747d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 748d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 749d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 750d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 751d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 752d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 753d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 754d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 755d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 756d1c3fb1fSNishanth Aravamudan */ 757d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 758a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 759d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 760d1c3fb1fSNishanth Aravamudan return NULL; 761d1c3fb1fSNishanth Aravamudan } else { 762a5516438SAndi Kleen h->nr_huge_pages++; 763a5516438SAndi Kleen h->surplus_huge_pages++; 764d1c3fb1fSNishanth Aravamudan } 765d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 766d1c3fb1fSNishanth Aravamudan 767551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 768551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 769a5516438SAndi Kleen huge_page_order(h)); 770d1c3fb1fSNishanth Aravamudan 771caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 772caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 773caff3a2cSGerald Schaefer return NULL; 774caff3a2cSGerald Schaefer } 775caff3a2cSGerald Schaefer 7767893d1d5SAdam Litke spin_lock(&hugetlb_lock); 777d1c3fb1fSNishanth Aravamudan if (page) { 7782668db91SAdam Litke /* 7792668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 7802668db91SAdam Litke * no users -- drop the buddy allocator's reference. 7812668db91SAdam Litke */ 7822668db91SAdam Litke put_page_testzero(page); 7832668db91SAdam Litke VM_BUG_ON(page_count(page)); 784d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 785d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 786d1c3fb1fSNishanth Aravamudan /* 787d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 788d1c3fb1fSNishanth Aravamudan */ 789a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 790a5516438SAndi Kleen h->surplus_huge_pages_node[nid]++; 7913b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 792d1c3fb1fSNishanth Aravamudan } else { 793a5516438SAndi Kleen h->nr_huge_pages--; 794a5516438SAndi Kleen h->surplus_huge_pages--; 7953b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7967893d1d5SAdam Litke } 797d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 7987893d1d5SAdam Litke 7997893d1d5SAdam Litke return page; 8007893d1d5SAdam Litke } 8017893d1d5SAdam Litke 802e4e574b7SAdam Litke /* 803e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 804e4e574b7SAdam Litke * of size 'delta'. 805e4e574b7SAdam Litke */ 806a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 807e4e574b7SAdam Litke { 808e4e574b7SAdam Litke struct list_head surplus_list; 809e4e574b7SAdam Litke struct page *page, *tmp; 810e4e574b7SAdam Litke int ret, i; 811e4e574b7SAdam Litke int needed, allocated; 812e4e574b7SAdam Litke 813a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 814ac09b3a1SAdam Litke if (needed <= 0) { 815a5516438SAndi Kleen h->resv_huge_pages += delta; 816e4e574b7SAdam Litke return 0; 817ac09b3a1SAdam Litke } 818e4e574b7SAdam Litke 819e4e574b7SAdam Litke allocated = 0; 820e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 821e4e574b7SAdam Litke 822e4e574b7SAdam Litke ret = -ENOMEM; 823e4e574b7SAdam Litke retry: 824e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 825e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 826a5516438SAndi Kleen page = alloc_buddy_huge_page(h, NULL, 0); 827e4e574b7SAdam Litke if (!page) { 828e4e574b7SAdam Litke /* 829e4e574b7SAdam Litke * We were not able to allocate enough pages to 830e4e574b7SAdam Litke * satisfy the entire reservation so we free what 831e4e574b7SAdam Litke * we've allocated so far. 832e4e574b7SAdam Litke */ 833e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 834e4e574b7SAdam Litke needed = 0; 835e4e574b7SAdam Litke goto free; 836e4e574b7SAdam Litke } 837e4e574b7SAdam Litke 838e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 839e4e574b7SAdam Litke } 840e4e574b7SAdam Litke allocated += needed; 841e4e574b7SAdam Litke 842e4e574b7SAdam Litke /* 843e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 844e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 845e4e574b7SAdam Litke */ 846e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 847a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 848a5516438SAndi Kleen (h->free_huge_pages + allocated); 849e4e574b7SAdam Litke if (needed > 0) 850e4e574b7SAdam Litke goto retry; 851e4e574b7SAdam Litke 852e4e574b7SAdam Litke /* 853e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 854e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 855e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 856ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 857ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 858ac09b3a1SAdam Litke * before they are reserved. 859e4e574b7SAdam Litke */ 860e4e574b7SAdam Litke needed += allocated; 861a5516438SAndi Kleen h->resv_huge_pages += delta; 862e4e574b7SAdam Litke ret = 0; 863e4e574b7SAdam Litke free: 86419fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 86519fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 86619fc3f0aSAdam Litke if ((--needed) < 0) 86719fc3f0aSAdam Litke break; 86819fc3f0aSAdam Litke list_del(&page->lru); 869a5516438SAndi Kleen enqueue_huge_page(h, page); 87019fc3f0aSAdam Litke } 87119fc3f0aSAdam Litke 87219fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 87319fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 87419fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 875e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 876e4e574b7SAdam Litke list_del(&page->lru); 877af767cbdSAdam Litke /* 8782668db91SAdam Litke * The page has a reference count of zero already, so 8792668db91SAdam Litke * call free_huge_page directly instead of using 8802668db91SAdam Litke * put_page. This must be done with hugetlb_lock 881af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 882af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 883af767cbdSAdam Litke */ 8842668db91SAdam Litke free_huge_page(page); 885af767cbdSAdam Litke } 88619fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 887e4e574b7SAdam Litke } 888e4e574b7SAdam Litke 889e4e574b7SAdam Litke return ret; 890e4e574b7SAdam Litke } 891e4e574b7SAdam Litke 892e4e574b7SAdam Litke /* 893e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 894e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 895e4e574b7SAdam Litke * never used. 896685f3457SLee Schermerhorn * Called with hugetlb_lock held. 897e4e574b7SAdam Litke */ 898a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 899a5516438SAndi Kleen unsigned long unused_resv_pages) 900e4e574b7SAdam Litke { 901e4e574b7SAdam Litke unsigned long nr_pages; 902e4e574b7SAdam Litke 903ac09b3a1SAdam Litke /* Uncommit the reservation */ 904a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 905ac09b3a1SAdam Litke 906aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 907aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 908aa888a74SAndi Kleen return; 909aa888a74SAndi Kleen 910a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 911e4e574b7SAdam Litke 912685f3457SLee Schermerhorn /* 913685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 914685f3457SLee Schermerhorn * evenly across all nodes. Iterate across all nodes until we 915685f3457SLee Schermerhorn * can no longer free unreserved surplus pages. This occurs when 916685f3457SLee Schermerhorn * the nodes with surplus pages have no free pages. 917685f3457SLee Schermerhorn * free_pool_huge_page() will balance the the frees across the 918685f3457SLee Schermerhorn * on-line nodes for us and will handle the hstate accounting. 919685f3457SLee Schermerhorn */ 920685f3457SLee Schermerhorn while (nr_pages--) { 921685f3457SLee Schermerhorn if (!free_pool_huge_page(h, 1)) 922685f3457SLee Schermerhorn break; 923e4e574b7SAdam Litke } 924e4e574b7SAdam Litke } 925e4e574b7SAdam Litke 926c37f9fb1SAndy Whitcroft /* 927c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 928c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 929c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 930c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 931c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 932c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 933c37f9fb1SAndy Whitcroft * No action is required on failure. 934c37f9fb1SAndy Whitcroft */ 935e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 936a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 937c37f9fb1SAndy Whitcroft { 938c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 939c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 940c37f9fb1SAndy Whitcroft 941f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 942a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 943c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 944c37f9fb1SAndy Whitcroft idx, idx + 1); 945c37f9fb1SAndy Whitcroft 94684afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 947c37f9fb1SAndy Whitcroft return 1; 948c37f9fb1SAndy Whitcroft 94984afd99bSAndy Whitcroft } else { 950e2f17d94SRoel Kluin long err; 951a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 95284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 95384afd99bSAndy Whitcroft 95484afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 95584afd99bSAndy Whitcroft if (err < 0) 95684afd99bSAndy Whitcroft return err; 957c37f9fb1SAndy Whitcroft return 0; 958c37f9fb1SAndy Whitcroft } 95984afd99bSAndy Whitcroft } 960a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 961a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 962c37f9fb1SAndy Whitcroft { 963c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 964c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 965c37f9fb1SAndy Whitcroft 966f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 967a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 968c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 96984afd99bSAndy Whitcroft 97084afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 971a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 97284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 97384afd99bSAndy Whitcroft 97484afd99bSAndy Whitcroft /* Mark this page used in the map. */ 97584afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 976c37f9fb1SAndy Whitcroft } 977c37f9fb1SAndy Whitcroft } 978c37f9fb1SAndy Whitcroft 979348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 98004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 981348ea204SAdam Litke { 982a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 983348ea204SAdam Litke struct page *page; 9842fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 985a1e78772SMel Gorman struct inode *inode = mapping->host; 986e2f17d94SRoel Kluin long chg; 9872fc39cecSAdam Litke 988a1e78772SMel Gorman /* 989a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 990a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 991a1e78772SMel Gorman * made before satisfying the allocation 992c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 993c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 994a1e78772SMel Gorman */ 995a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 996c37f9fb1SAndy Whitcroft if (chg < 0) 997c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 998c37f9fb1SAndy Whitcroft if (chg) 999a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 1000a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 100190d8b7e6SAdam Litke 1002a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1003a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1004a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1005a1e78772SMel Gorman 1006a1e78772SMel Gorman if (!page) { 1007a5516438SAndi Kleen page = alloc_buddy_huge_page(h, vma, addr); 1008a1e78772SMel Gorman if (!page) { 1009a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 1010a1e78772SMel Gorman return ERR_PTR(-VM_FAULT_OOM); 1011a1e78772SMel Gorman } 1012a1e78772SMel Gorman } 1013a1e78772SMel Gorman 1014348ea204SAdam Litke set_page_refcounted(page); 10152fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 1016a1e78772SMel Gorman 1017a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1018c37f9fb1SAndy Whitcroft 10197893d1d5SAdam Litke return page; 1020b45b5bd6SDavid Gibson } 1021b45b5bd6SDavid Gibson 102291f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1023aa888a74SAndi Kleen { 1024aa888a74SAndi Kleen struct huge_bootmem_page *m; 1025aa888a74SAndi Kleen int nr_nodes = nodes_weight(node_online_map); 1026aa888a74SAndi Kleen 1027aa888a74SAndi Kleen while (nr_nodes) { 1028aa888a74SAndi Kleen void *addr; 1029aa888a74SAndi Kleen 1030aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 1031e8c5c824SLee Schermerhorn NODE_DATA(h->next_nid_to_alloc), 1032aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1033aa888a74SAndi Kleen 1034aa888a74SAndi Kleen if (addr) { 1035aa888a74SAndi Kleen /* 1036aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1037aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1038aa888a74SAndi Kleen * puts them into the mem_map). 1039aa888a74SAndi Kleen */ 1040aa888a74SAndi Kleen m = addr; 1041aa888a74SAndi Kleen goto found; 1042aa888a74SAndi Kleen } 1043e8c5c824SLee Schermerhorn hstate_next_node_to_alloc(h); 1044aa888a74SAndi Kleen nr_nodes--; 1045aa888a74SAndi Kleen } 1046aa888a74SAndi Kleen return 0; 1047aa888a74SAndi Kleen 1048aa888a74SAndi Kleen found: 1049aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1050aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1051aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1052aa888a74SAndi Kleen m->hstate = h; 1053aa888a74SAndi Kleen return 1; 1054aa888a74SAndi Kleen } 1055aa888a74SAndi Kleen 105618229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 105718229df5SAndy Whitcroft { 105818229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 105918229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 106018229df5SAndy Whitcroft else 106118229df5SAndy Whitcroft prep_compound_page(page, order); 106218229df5SAndy Whitcroft } 106318229df5SAndy Whitcroft 1064aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1065aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1066aa888a74SAndi Kleen { 1067aa888a74SAndi Kleen struct huge_bootmem_page *m; 1068aa888a74SAndi Kleen 1069aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1070aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1071aa888a74SAndi Kleen struct hstate *h = m->hstate; 1072aa888a74SAndi Kleen __ClearPageReserved(page); 1073aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 107418229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1075aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1076aa888a74SAndi Kleen } 1077aa888a74SAndi Kleen } 1078aa888a74SAndi Kleen 10798faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 10801da177e4SLinus Torvalds { 10811da177e4SLinus Torvalds unsigned long i; 10821da177e4SLinus Torvalds 1083e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1084aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1085aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1086aa888a74SAndi Kleen break; 1087aa888a74SAndi Kleen } else if (!alloc_fresh_huge_page(h)) 10881da177e4SLinus Torvalds break; 10891da177e4SLinus Torvalds } 10908faa8b07SAndi Kleen h->max_huge_pages = i; 1091e5ff2159SAndi Kleen } 1092e5ff2159SAndi Kleen 1093e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1094e5ff2159SAndi Kleen { 1095e5ff2159SAndi Kleen struct hstate *h; 1096e5ff2159SAndi Kleen 1097e5ff2159SAndi Kleen for_each_hstate(h) { 10988faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 10998faa8b07SAndi Kleen if (h->order < MAX_ORDER) 11008faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1101e5ff2159SAndi Kleen } 1102e5ff2159SAndi Kleen } 1103e5ff2159SAndi Kleen 11044abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 11054abd32dbSAndi Kleen { 11064abd32dbSAndi Kleen if (n >= (1UL << 30)) 11074abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 11084abd32dbSAndi Kleen else if (n >= (1UL << 20)) 11094abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 11104abd32dbSAndi Kleen else 11114abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 11124abd32dbSAndi Kleen return buf; 11134abd32dbSAndi Kleen } 11144abd32dbSAndi Kleen 1115e5ff2159SAndi Kleen static void __init report_hugepages(void) 1116e5ff2159SAndi Kleen { 1117e5ff2159SAndi Kleen struct hstate *h; 1118e5ff2159SAndi Kleen 1119e5ff2159SAndi Kleen for_each_hstate(h) { 11204abd32dbSAndi Kleen char buf[32]; 11214abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11224abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11234abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11244abd32dbSAndi Kleen h->free_huge_pages); 1125e5ff2159SAndi Kleen } 1126e5ff2159SAndi Kleen } 1127e5ff2159SAndi Kleen 11281da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 1129a5516438SAndi Kleen static void try_to_free_low(struct hstate *h, unsigned long count) 11301da177e4SLinus Torvalds { 11314415cc8dSChristoph Lameter int i; 11324415cc8dSChristoph Lameter 1133aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1134aa888a74SAndi Kleen return; 1135aa888a74SAndi Kleen 11361da177e4SLinus Torvalds for (i = 0; i < MAX_NUMNODES; ++i) { 11371da177e4SLinus Torvalds struct page *page, *next; 1138a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1139a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1140a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11416b0c880dSAdam Litke return; 11421da177e4SLinus Torvalds if (PageHighMem(page)) 11431da177e4SLinus Torvalds continue; 11441da177e4SLinus Torvalds list_del(&page->lru); 1145e5ff2159SAndi Kleen update_and_free_page(h, page); 1146a5516438SAndi Kleen h->free_huge_pages--; 1147a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11481da177e4SLinus Torvalds } 11491da177e4SLinus Torvalds } 11501da177e4SLinus Torvalds } 11511da177e4SLinus Torvalds #else 1152a5516438SAndi Kleen static inline void try_to_free_low(struct hstate *h, unsigned long count) 11531da177e4SLinus Torvalds { 11541da177e4SLinus Torvalds } 11551da177e4SLinus Torvalds #endif 11561da177e4SLinus Torvalds 115720a0307cSWu Fengguang /* 115820a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 115920a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 116020a0307cSWu Fengguang * Returns 1 if an adjustment was made. 116120a0307cSWu Fengguang */ 116220a0307cSWu Fengguang static int adjust_pool_surplus(struct hstate *h, int delta) 116320a0307cSWu Fengguang { 1164e8c5c824SLee Schermerhorn int start_nid, next_nid; 116520a0307cSWu Fengguang int ret = 0; 116620a0307cSWu Fengguang 116720a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 116820a0307cSWu Fengguang 1169e8c5c824SLee Schermerhorn if (delta < 0) 1170e8c5c824SLee Schermerhorn start_nid = h->next_nid_to_alloc; 1171e8c5c824SLee Schermerhorn else 1172e8c5c824SLee Schermerhorn start_nid = h->next_nid_to_free; 1173e8c5c824SLee Schermerhorn next_nid = start_nid; 1174e8c5c824SLee Schermerhorn 1175e8c5c824SLee Schermerhorn do { 1176e8c5c824SLee Schermerhorn int nid = next_nid; 1177e8c5c824SLee Schermerhorn if (delta < 0) { 1178e8c5c824SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h); 1179e8c5c824SLee Schermerhorn /* 1180e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1181e8c5c824SLee Schermerhorn */ 1182e8c5c824SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) 118320a0307cSWu Fengguang continue; 1184e8c5c824SLee Schermerhorn } 1185e8c5c824SLee Schermerhorn if (delta > 0) { 1186e8c5c824SLee Schermerhorn next_nid = hstate_next_node_to_free(h); 1187e8c5c824SLee Schermerhorn /* 1188e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1189e8c5c824SLee Schermerhorn */ 1190e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 119120a0307cSWu Fengguang h->nr_huge_pages_node[nid]) 119220a0307cSWu Fengguang continue; 1193e8c5c824SLee Schermerhorn } 119420a0307cSWu Fengguang 119520a0307cSWu Fengguang h->surplus_huge_pages += delta; 119620a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 119720a0307cSWu Fengguang ret = 1; 119820a0307cSWu Fengguang break; 1199e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 120020a0307cSWu Fengguang 120120a0307cSWu Fengguang return ret; 120220a0307cSWu Fengguang } 120320a0307cSWu Fengguang 1204a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1205e5ff2159SAndi Kleen static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count) 12061da177e4SLinus Torvalds { 12077893d1d5SAdam Litke unsigned long min_count, ret; 12081da177e4SLinus Torvalds 1209aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1210aa888a74SAndi Kleen return h->max_huge_pages; 1211aa888a74SAndi Kleen 12127893d1d5SAdam Litke /* 12137893d1d5SAdam Litke * Increase the pool size 12147893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 12157893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1216d1c3fb1fSNishanth Aravamudan * 1217d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1218d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1219d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1220d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1221d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 12227893d1d5SAdam Litke */ 12231da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1224a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1225a5516438SAndi Kleen if (!adjust_pool_surplus(h, -1)) 12267893d1d5SAdam Litke break; 12277893d1d5SAdam Litke } 12287893d1d5SAdam Litke 1229a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 12307893d1d5SAdam Litke /* 12317893d1d5SAdam Litke * If this allocation races such that we no longer need the 12327893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12337893d1d5SAdam Litke * and reducing the surplus. 12347893d1d5SAdam Litke */ 12357893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 1236a5516438SAndi Kleen ret = alloc_fresh_huge_page(h); 12377893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12387893d1d5SAdam Litke if (!ret) 12397893d1d5SAdam Litke goto out; 12407893d1d5SAdam Litke 12417893d1d5SAdam Litke } 12427893d1d5SAdam Litke 12437893d1d5SAdam Litke /* 12447893d1d5SAdam Litke * Decrease the pool size 12457893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 12467893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 12477893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 12487893d1d5SAdam Litke * to the desired size as pages become free. 1249d1c3fb1fSNishanth Aravamudan * 1250d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1251d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1252d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1253d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1254d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1255d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1256d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 12577893d1d5SAdam Litke */ 1258a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 12596b0c880dSAdam Litke min_count = max(count, min_count); 1260a5516438SAndi Kleen try_to_free_low(h, min_count); 1261a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 1262685f3457SLee Schermerhorn if (!free_pool_huge_page(h, 0)) 12631da177e4SLinus Torvalds break; 12641da177e4SLinus Torvalds } 1265a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 1266a5516438SAndi Kleen if (!adjust_pool_surplus(h, 1)) 12677893d1d5SAdam Litke break; 12687893d1d5SAdam Litke } 12697893d1d5SAdam Litke out: 1270a5516438SAndi Kleen ret = persistent_huge_pages(h); 12711da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 12727893d1d5SAdam Litke return ret; 12731da177e4SLinus Torvalds } 12741da177e4SLinus Torvalds 1275a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1276a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1277a3437870SNishanth Aravamudan 1278a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1279a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1280a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1281a3437870SNishanth Aravamudan 1282a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1283a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1284a3437870SNishanth Aravamudan 1285a3437870SNishanth Aravamudan static struct hstate *kobj_to_hstate(struct kobject *kobj) 1286a3437870SNishanth Aravamudan { 1287a3437870SNishanth Aravamudan int i; 1288a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 1289a3437870SNishanth Aravamudan if (hstate_kobjs[i] == kobj) 1290a3437870SNishanth Aravamudan return &hstates[i]; 1291a3437870SNishanth Aravamudan BUG(); 1292a3437870SNishanth Aravamudan return NULL; 1293a3437870SNishanth Aravamudan } 1294a3437870SNishanth Aravamudan 1295a3437870SNishanth Aravamudan static ssize_t nr_hugepages_show(struct kobject *kobj, 1296a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1297a3437870SNishanth Aravamudan { 1298a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1299a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_huge_pages); 1300a3437870SNishanth Aravamudan } 1301a3437870SNishanth Aravamudan static ssize_t nr_hugepages_store(struct kobject *kobj, 1302a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1303a3437870SNishanth Aravamudan { 1304a3437870SNishanth Aravamudan int err; 1305a3437870SNishanth Aravamudan unsigned long input; 1306a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1307a3437870SNishanth Aravamudan 1308a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1309a3437870SNishanth Aravamudan if (err) 1310a3437870SNishanth Aravamudan return 0; 1311a3437870SNishanth Aravamudan 1312a3437870SNishanth Aravamudan h->max_huge_pages = set_max_huge_pages(h, input); 1313a3437870SNishanth Aravamudan 1314a3437870SNishanth Aravamudan return count; 1315a3437870SNishanth Aravamudan } 1316a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1317a3437870SNishanth Aravamudan 1318a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1319a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1320a3437870SNishanth Aravamudan { 1321a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1322a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1323a3437870SNishanth Aravamudan } 1324a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1325a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1326a3437870SNishanth Aravamudan { 1327a3437870SNishanth Aravamudan int err; 1328a3437870SNishanth Aravamudan unsigned long input; 1329a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1330a3437870SNishanth Aravamudan 1331a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1332a3437870SNishanth Aravamudan if (err) 1333a3437870SNishanth Aravamudan return 0; 1334a3437870SNishanth Aravamudan 1335a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1336a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1337a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1338a3437870SNishanth Aravamudan 1339a3437870SNishanth Aravamudan return count; 1340a3437870SNishanth Aravamudan } 1341a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1342a3437870SNishanth Aravamudan 1343a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1344a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1345a3437870SNishanth Aravamudan { 1346a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1347a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->free_huge_pages); 1348a3437870SNishanth Aravamudan } 1349a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1350a3437870SNishanth Aravamudan 1351a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1352a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1353a3437870SNishanth Aravamudan { 1354a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1355a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1356a3437870SNishanth Aravamudan } 1357a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1358a3437870SNishanth Aravamudan 1359a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1360a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1361a3437870SNishanth Aravamudan { 1362a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1363a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->surplus_huge_pages); 1364a3437870SNishanth Aravamudan } 1365a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1366a3437870SNishanth Aravamudan 1367a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1368a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1369a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1370a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1371a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1372a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 1373a3437870SNishanth Aravamudan NULL, 1374a3437870SNishanth Aravamudan }; 1375a3437870SNishanth Aravamudan 1376a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1377a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1378a3437870SNishanth Aravamudan }; 1379a3437870SNishanth Aravamudan 1380a3437870SNishanth Aravamudan static int __init hugetlb_sysfs_add_hstate(struct hstate *h) 1381a3437870SNishanth Aravamudan { 1382a3437870SNishanth Aravamudan int retval; 1383a3437870SNishanth Aravamudan 1384a3437870SNishanth Aravamudan hstate_kobjs[h - hstates] = kobject_create_and_add(h->name, 1385a3437870SNishanth Aravamudan hugepages_kobj); 1386a3437870SNishanth Aravamudan if (!hstate_kobjs[h - hstates]) 1387a3437870SNishanth Aravamudan return -ENOMEM; 1388a3437870SNishanth Aravamudan 1389a3437870SNishanth Aravamudan retval = sysfs_create_group(hstate_kobjs[h - hstates], 1390a3437870SNishanth Aravamudan &hstate_attr_group); 1391a3437870SNishanth Aravamudan if (retval) 1392a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1393a3437870SNishanth Aravamudan 1394a3437870SNishanth Aravamudan return retval; 1395a3437870SNishanth Aravamudan } 1396a3437870SNishanth Aravamudan 1397a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1398a3437870SNishanth Aravamudan { 1399a3437870SNishanth Aravamudan struct hstate *h; 1400a3437870SNishanth Aravamudan int err; 1401a3437870SNishanth Aravamudan 1402a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1403a3437870SNishanth Aravamudan if (!hugepages_kobj) 1404a3437870SNishanth Aravamudan return; 1405a3437870SNishanth Aravamudan 1406a3437870SNishanth Aravamudan for_each_hstate(h) { 1407a3437870SNishanth Aravamudan err = hugetlb_sysfs_add_hstate(h); 1408a3437870SNishanth Aravamudan if (err) 1409a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1410a3437870SNishanth Aravamudan h->name); 1411a3437870SNishanth Aravamudan } 1412a3437870SNishanth Aravamudan } 1413a3437870SNishanth Aravamudan 1414a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1415a3437870SNishanth Aravamudan { 1416a3437870SNishanth Aravamudan struct hstate *h; 1417a3437870SNishanth Aravamudan 1418a3437870SNishanth Aravamudan for_each_hstate(h) { 1419a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1420a3437870SNishanth Aravamudan } 1421a3437870SNishanth Aravamudan 1422a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1423a3437870SNishanth Aravamudan } 1424a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1425a3437870SNishanth Aravamudan 1426a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1427a3437870SNishanth Aravamudan { 14280ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 14290ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 14300ef89d25SBenjamin Herrenschmidt * there is no such support 14310ef89d25SBenjamin Herrenschmidt */ 14320ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 14330ef89d25SBenjamin Herrenschmidt return 0; 1434a3437870SNishanth Aravamudan 1435e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1436e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1437e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1438a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1439a3437870SNishanth Aravamudan } 1440e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1441e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1442e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1443a3437870SNishanth Aravamudan 1444a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1445a3437870SNishanth Aravamudan 1446aa888a74SAndi Kleen gather_bootmem_prealloc(); 1447aa888a74SAndi Kleen 1448a3437870SNishanth Aravamudan report_hugepages(); 1449a3437870SNishanth Aravamudan 1450a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1451a3437870SNishanth Aravamudan 1452a3437870SNishanth Aravamudan return 0; 1453a3437870SNishanth Aravamudan } 1454a3437870SNishanth Aravamudan module_init(hugetlb_init); 1455a3437870SNishanth Aravamudan 1456a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1457a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1458a3437870SNishanth Aravamudan { 1459a3437870SNishanth Aravamudan struct hstate *h; 14608faa8b07SAndi Kleen unsigned long i; 14618faa8b07SAndi Kleen 1462a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1463a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1464a3437870SNishanth Aravamudan return; 1465a3437870SNishanth Aravamudan } 1466a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1467a3437870SNishanth Aravamudan BUG_ON(order == 0); 1468a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1469a3437870SNishanth Aravamudan h->order = order; 1470a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 14718faa8b07SAndi Kleen h->nr_huge_pages = 0; 14728faa8b07SAndi Kleen h->free_huge_pages = 0; 14738faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 14748faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 1475e8c5c824SLee Schermerhorn h->next_nid_to_alloc = first_node(node_online_map); 1476e8c5c824SLee Schermerhorn h->next_nid_to_free = first_node(node_online_map); 1477a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1478a3437870SNishanth Aravamudan huge_page_size(h)/1024); 14798faa8b07SAndi Kleen 1480a3437870SNishanth Aravamudan parsed_hstate = h; 1481a3437870SNishanth Aravamudan } 1482a3437870SNishanth Aravamudan 1483e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1484a3437870SNishanth Aravamudan { 1485a3437870SNishanth Aravamudan unsigned long *mhp; 14868faa8b07SAndi Kleen static unsigned long *last_mhp; 1487a3437870SNishanth Aravamudan 1488a3437870SNishanth Aravamudan /* 1489a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1490a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1491a3437870SNishanth Aravamudan */ 1492a3437870SNishanth Aravamudan if (!max_hstate) 1493a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1494a3437870SNishanth Aravamudan else 1495a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1496a3437870SNishanth Aravamudan 14978faa8b07SAndi Kleen if (mhp == last_mhp) { 14988faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 14998faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 15008faa8b07SAndi Kleen return 1; 15018faa8b07SAndi Kleen } 15028faa8b07SAndi Kleen 1503a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1504a3437870SNishanth Aravamudan *mhp = 0; 1505a3437870SNishanth Aravamudan 15068faa8b07SAndi Kleen /* 15078faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 15088faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 15098faa8b07SAndi Kleen * use the bootmem allocator. 15108faa8b07SAndi Kleen */ 15118faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 15128faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 15138faa8b07SAndi Kleen 15148faa8b07SAndi Kleen last_mhp = mhp; 15158faa8b07SAndi Kleen 1516a3437870SNishanth Aravamudan return 1; 1517a3437870SNishanth Aravamudan } 1518e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1519e11bfbfcSNick Piggin 1520e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1521e11bfbfcSNick Piggin { 1522e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1523e11bfbfcSNick Piggin return 1; 1524e11bfbfcSNick Piggin } 1525e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1526a3437870SNishanth Aravamudan 15278a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 15288a213460SNishanth Aravamudan { 15298a213460SNishanth Aravamudan int node; 15308a213460SNishanth Aravamudan unsigned int nr = 0; 15318a213460SNishanth Aravamudan 15328a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 15338a213460SNishanth Aravamudan nr += array[node]; 15348a213460SNishanth Aravamudan 15358a213460SNishanth Aravamudan return nr; 15368a213460SNishanth Aravamudan } 15378a213460SNishanth Aravamudan 15388a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 15391da177e4SLinus Torvalds int hugetlb_sysctl_handler(struct ctl_table *table, int write, 15401da177e4SLinus Torvalds struct file *file, void __user *buffer, 15411da177e4SLinus Torvalds size_t *length, loff_t *ppos) 15421da177e4SLinus Torvalds { 1543e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1544e5ff2159SAndi Kleen unsigned long tmp; 1545e5ff2159SAndi Kleen 1546e5ff2159SAndi Kleen if (!write) 1547e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1548e5ff2159SAndi Kleen 1549e5ff2159SAndi Kleen table->data = &tmp; 1550e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 15511da177e4SLinus Torvalds proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 1552e5ff2159SAndi Kleen 1553e5ff2159SAndi Kleen if (write) 1554e5ff2159SAndi Kleen h->max_huge_pages = set_max_huge_pages(h, tmp); 1555e5ff2159SAndi Kleen 15561da177e4SLinus Torvalds return 0; 15571da177e4SLinus Torvalds } 1558396faf03SMel Gorman 1559396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 1560396faf03SMel Gorman struct file *file, void __user *buffer, 1561396faf03SMel Gorman size_t *length, loff_t *ppos) 1562396faf03SMel Gorman { 1563396faf03SMel Gorman proc_dointvec(table, write, file, buffer, length, ppos); 1564396faf03SMel Gorman if (hugepages_treat_as_movable) 1565396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1566396faf03SMel Gorman else 1567396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1568396faf03SMel Gorman return 0; 1569396faf03SMel Gorman } 1570396faf03SMel Gorman 1571a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 1572a3d0c6aaSNishanth Aravamudan struct file *file, void __user *buffer, 1573a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1574a3d0c6aaSNishanth Aravamudan { 1575a5516438SAndi Kleen struct hstate *h = &default_hstate; 1576e5ff2159SAndi Kleen unsigned long tmp; 1577e5ff2159SAndi Kleen 1578e5ff2159SAndi Kleen if (!write) 1579e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1580e5ff2159SAndi Kleen 1581e5ff2159SAndi Kleen table->data = &tmp; 1582e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 1583a3d0c6aaSNishanth Aravamudan proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 1584e5ff2159SAndi Kleen 1585e5ff2159SAndi Kleen if (write) { 1586064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1587e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1588a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1589e5ff2159SAndi Kleen } 1590e5ff2159SAndi Kleen 1591a3d0c6aaSNishanth Aravamudan return 0; 1592a3d0c6aaSNishanth Aravamudan } 1593a3d0c6aaSNishanth Aravamudan 15941da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 15951da177e4SLinus Torvalds 1596e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 15971da177e4SLinus Torvalds { 1598a5516438SAndi Kleen struct hstate *h = &default_hstate; 1599e1759c21SAlexey Dobriyan seq_printf(m, 16001da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 16011da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1602b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 16037893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 16044f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1605a5516438SAndi Kleen h->nr_huge_pages, 1606a5516438SAndi Kleen h->free_huge_pages, 1607a5516438SAndi Kleen h->resv_huge_pages, 1608a5516438SAndi Kleen h->surplus_huge_pages, 1609a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 16101da177e4SLinus Torvalds } 16111da177e4SLinus Torvalds 16121da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 16131da177e4SLinus Torvalds { 1614a5516438SAndi Kleen struct hstate *h = &default_hstate; 16151da177e4SLinus Torvalds return sprintf(buf, 16161da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1617a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1618a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1619a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1620a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1621a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 16221da177e4SLinus Torvalds } 16231da177e4SLinus Torvalds 16241da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 16251da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 16261da177e4SLinus Torvalds { 1627a5516438SAndi Kleen struct hstate *h = &default_hstate; 1628a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 16291da177e4SLinus Torvalds } 16301da177e4SLinus Torvalds 1631a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 1632fc1b8a73SMel Gorman { 1633fc1b8a73SMel Gorman int ret = -ENOMEM; 1634fc1b8a73SMel Gorman 1635fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 1636fc1b8a73SMel Gorman /* 1637fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 1638fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 1639fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 1640fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 1641fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 1642fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 1643fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 1644fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 1645fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 1646fc1b8a73SMel Gorman * 1647fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 1648fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 1649fc1b8a73SMel Gorman * we fall back to check against current free page availability as 1650fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 1651fc1b8a73SMel Gorman * semantics that cpuset has. 1652fc1b8a73SMel Gorman */ 1653fc1b8a73SMel Gorman if (delta > 0) { 1654a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 1655fc1b8a73SMel Gorman goto out; 1656fc1b8a73SMel Gorman 1657a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 1658a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 1659fc1b8a73SMel Gorman goto out; 1660fc1b8a73SMel Gorman } 1661fc1b8a73SMel Gorman } 1662fc1b8a73SMel Gorman 1663fc1b8a73SMel Gorman ret = 0; 1664fc1b8a73SMel Gorman if (delta < 0) 1665a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 1666fc1b8a73SMel Gorman 1667fc1b8a73SMel Gorman out: 1668fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 1669fc1b8a73SMel Gorman return ret; 1670fc1b8a73SMel Gorman } 1671fc1b8a73SMel Gorman 167284afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 167384afd99bSAndy Whitcroft { 167484afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 167584afd99bSAndy Whitcroft 167684afd99bSAndy Whitcroft /* 167784afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 167884afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 167984afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 168084afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 168184afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 168284afd99bSAndy Whitcroft * new reference here without additional locking. 168384afd99bSAndy Whitcroft */ 168484afd99bSAndy Whitcroft if (reservations) 168584afd99bSAndy Whitcroft kref_get(&reservations->refs); 168684afd99bSAndy Whitcroft } 168784afd99bSAndy Whitcroft 1688a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 1689a1e78772SMel Gorman { 1690a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 169184afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 169284afd99bSAndy Whitcroft unsigned long reserve; 169384afd99bSAndy Whitcroft unsigned long start; 169484afd99bSAndy Whitcroft unsigned long end; 169584afd99bSAndy Whitcroft 169684afd99bSAndy Whitcroft if (reservations) { 1697a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 1698a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 169984afd99bSAndy Whitcroft 170084afd99bSAndy Whitcroft reserve = (end - start) - 170184afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 170284afd99bSAndy Whitcroft 170384afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 170484afd99bSAndy Whitcroft 17057251ff78SAdam Litke if (reserve) { 1706a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 17077251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 17087251ff78SAdam Litke } 1709a1e78772SMel Gorman } 171084afd99bSAndy Whitcroft } 1711a1e78772SMel Gorman 17121da177e4SLinus Torvalds /* 17131da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 17141da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 17151da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 17161da177e4SLinus Torvalds * this far. 17171da177e4SLinus Torvalds */ 1718d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 17191da177e4SLinus Torvalds { 17201da177e4SLinus Torvalds BUG(); 1721d0217ac0SNick Piggin return 0; 17221da177e4SLinus Torvalds } 17231da177e4SLinus Torvalds 17241da177e4SLinus Torvalds struct vm_operations_struct hugetlb_vm_ops = { 1725d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 172684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 1727a1e78772SMel Gorman .close = hugetlb_vm_op_close, 17281da177e4SLinus Torvalds }; 17291da177e4SLinus Torvalds 17301e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 17311e8f889bSDavid Gibson int writable) 173263551ae0SDavid Gibson { 173363551ae0SDavid Gibson pte_t entry; 173463551ae0SDavid Gibson 17351e8f889bSDavid Gibson if (writable) { 173663551ae0SDavid Gibson entry = 173763551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 173863551ae0SDavid Gibson } else { 17397f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 174063551ae0SDavid Gibson } 174163551ae0SDavid Gibson entry = pte_mkyoung(entry); 174263551ae0SDavid Gibson entry = pte_mkhuge(entry); 174363551ae0SDavid Gibson 174463551ae0SDavid Gibson return entry; 174563551ae0SDavid Gibson } 174663551ae0SDavid Gibson 17471e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 17481e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 17491e8f889bSDavid Gibson { 17501e8f889bSDavid Gibson pte_t entry; 17511e8f889bSDavid Gibson 17527f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 17537f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 17541e8f889bSDavid Gibson update_mmu_cache(vma, address, entry); 17551e8f889bSDavid Gibson } 17568dab5241SBenjamin Herrenschmidt } 17571e8f889bSDavid Gibson 17581e8f889bSDavid Gibson 175963551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 176063551ae0SDavid Gibson struct vm_area_struct *vma) 176163551ae0SDavid Gibson { 176263551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 176363551ae0SDavid Gibson struct page *ptepage; 17641c59827dSHugh Dickins unsigned long addr; 17651e8f889bSDavid Gibson int cow; 1766a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1767a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 17681e8f889bSDavid Gibson 17691e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 177063551ae0SDavid Gibson 1771a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 1772c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 1773c74df32cSHugh Dickins if (!src_pte) 1774c74df32cSHugh Dickins continue; 1775a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 177663551ae0SDavid Gibson if (!dst_pte) 177763551ae0SDavid Gibson goto nomem; 1778c5c99429SLarry Woodman 1779c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 1780c5c99429SLarry Woodman if (dst_pte == src_pte) 1781c5c99429SLarry Woodman continue; 1782c5c99429SLarry Woodman 1783c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 178446478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 17857f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 17861e8f889bSDavid Gibson if (cow) 17877f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 17887f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 178963551ae0SDavid Gibson ptepage = pte_page(entry); 179063551ae0SDavid Gibson get_page(ptepage); 179163551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 17921c59827dSHugh Dickins } 17931c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 1794c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 179563551ae0SDavid Gibson } 179663551ae0SDavid Gibson return 0; 179763551ae0SDavid Gibson 179863551ae0SDavid Gibson nomem: 179963551ae0SDavid Gibson return -ENOMEM; 180063551ae0SDavid Gibson } 180163551ae0SDavid Gibson 1802502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 180304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 180463551ae0SDavid Gibson { 180563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 180663551ae0SDavid Gibson unsigned long address; 1807c7546f8fSDavid Gibson pte_t *ptep; 180863551ae0SDavid Gibson pte_t pte; 180963551ae0SDavid Gibson struct page *page; 1810fe1668aeSChen, Kenneth W struct page *tmp; 1811a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1812a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 1813a5516438SAndi Kleen 1814c0a499c2SChen, Kenneth W /* 1815c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 1816c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 1817c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 1818c0a499c2SChen, Kenneth W */ 1819fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 182063551ae0SDavid Gibson 182163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 1822a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 1823a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 182463551ae0SDavid Gibson 1825cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 1826508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 1827a5516438SAndi Kleen for (address = start; address < end; address += sz) { 1828c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 1829c7546f8fSDavid Gibson if (!ptep) 1830c7546f8fSDavid Gibson continue; 1831c7546f8fSDavid Gibson 183239dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 183339dde65cSChen, Kenneth W continue; 183439dde65cSChen, Kenneth W 183504f2cbe3SMel Gorman /* 183604f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 183704f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 183804f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 183904f2cbe3SMel Gorman */ 184004f2cbe3SMel Gorman if (ref_page) { 184104f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 184204f2cbe3SMel Gorman if (huge_pte_none(pte)) 184304f2cbe3SMel Gorman continue; 184404f2cbe3SMel Gorman page = pte_page(pte); 184504f2cbe3SMel Gorman if (page != ref_page) 184604f2cbe3SMel Gorman continue; 184704f2cbe3SMel Gorman 184804f2cbe3SMel Gorman /* 184904f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 185004f2cbe3SMel Gorman * future faults in this VMA will fail rather than 185104f2cbe3SMel Gorman * looking like data was lost 185204f2cbe3SMel Gorman */ 185304f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 185404f2cbe3SMel Gorman } 185504f2cbe3SMel Gorman 1856c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 18577f2e9525SGerald Schaefer if (huge_pte_none(pte)) 185863551ae0SDavid Gibson continue; 1859c7546f8fSDavid Gibson 186063551ae0SDavid Gibson page = pte_page(pte); 18616649a386SKen Chen if (pte_dirty(pte)) 18626649a386SKen Chen set_page_dirty(page); 1863fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 186463551ae0SDavid Gibson } 18651da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 1866508034a3SHugh Dickins flush_tlb_range(vma, start, end); 1867cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 1868fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 1869fe1668aeSChen, Kenneth W list_del(&page->lru); 1870fe1668aeSChen, Kenneth W put_page(page); 1871fe1668aeSChen, Kenneth W } 18721da177e4SLinus Torvalds } 187363551ae0SDavid Gibson 1874502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 187504f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 1876502717f4SChen, Kenneth W { 1877502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 187804f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 1879502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 1880502717f4SChen, Kenneth W } 1881502717f4SChen, Kenneth W 188204f2cbe3SMel Gorman /* 188304f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 188404f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 188504f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 188604f2cbe3SMel Gorman * same region. 188704f2cbe3SMel Gorman */ 18882a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 18892a4b3dedSHarvey Harrison struct page *page, unsigned long address) 189004f2cbe3SMel Gorman { 18917526674dSAdam Litke struct hstate *h = hstate_vma(vma); 189204f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 189304f2cbe3SMel Gorman struct address_space *mapping; 189404f2cbe3SMel Gorman struct prio_tree_iter iter; 189504f2cbe3SMel Gorman pgoff_t pgoff; 189604f2cbe3SMel Gorman 189704f2cbe3SMel Gorman /* 189804f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 189904f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 190004f2cbe3SMel Gorman */ 19017526674dSAdam Litke address = address & huge_page_mask(h); 190204f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 190304f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 190404f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 190504f2cbe3SMel Gorman 190604f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 190704f2cbe3SMel Gorman /* Do not unmap the current VMA */ 190804f2cbe3SMel Gorman if (iter_vma == vma) 190904f2cbe3SMel Gorman continue; 191004f2cbe3SMel Gorman 191104f2cbe3SMel Gorman /* 191204f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 191304f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 191404f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 191504f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 191604f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 191704f2cbe3SMel Gorman */ 191804f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 191904f2cbe3SMel Gorman unmap_hugepage_range(iter_vma, 19207526674dSAdam Litke address, address + huge_page_size(h), 192104f2cbe3SMel Gorman page); 192204f2cbe3SMel Gorman } 192304f2cbe3SMel Gorman 192404f2cbe3SMel Gorman return 1; 192504f2cbe3SMel Gorman } 192604f2cbe3SMel Gorman 19271e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 192804f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 192904f2cbe3SMel Gorman struct page *pagecache_page) 19301e8f889bSDavid Gibson { 1931a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 19321e8f889bSDavid Gibson struct page *old_page, *new_page; 193379ac6ba4SDavid Gibson int avoidcopy; 193404f2cbe3SMel Gorman int outside_reserve = 0; 19351e8f889bSDavid Gibson 19361e8f889bSDavid Gibson old_page = pte_page(pte); 19371e8f889bSDavid Gibson 193804f2cbe3SMel Gorman retry_avoidcopy: 19391e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 19401e8f889bSDavid Gibson * and just make the page writable */ 19411e8f889bSDavid Gibson avoidcopy = (page_count(old_page) == 1); 19421e8f889bSDavid Gibson if (avoidcopy) { 19431e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 194483c54070SNick Piggin return 0; 19451e8f889bSDavid Gibson } 19461e8f889bSDavid Gibson 194704f2cbe3SMel Gorman /* 194804f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 194904f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 195004f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 195104f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 195204f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 195304f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 195404f2cbe3SMel Gorman * of the full address range. 195504f2cbe3SMel Gorman */ 1956f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 195704f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 195804f2cbe3SMel Gorman old_page != pagecache_page) 195904f2cbe3SMel Gorman outside_reserve = 1; 196004f2cbe3SMel Gorman 19611e8f889bSDavid Gibson page_cache_get(old_page); 196204f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 19631e8f889bSDavid Gibson 19642fc39cecSAdam Litke if (IS_ERR(new_page)) { 19651e8f889bSDavid Gibson page_cache_release(old_page); 196604f2cbe3SMel Gorman 196704f2cbe3SMel Gorman /* 196804f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 196904f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 197004f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 197104f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 197204f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 197304f2cbe3SMel Gorman */ 197404f2cbe3SMel Gorman if (outside_reserve) { 197504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 197604f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 197704f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 197804f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 197904f2cbe3SMel Gorman goto retry_avoidcopy; 198004f2cbe3SMel Gorman } 198104f2cbe3SMel Gorman WARN_ON_ONCE(1); 198204f2cbe3SMel Gorman } 198304f2cbe3SMel Gorman 19842fc39cecSAdam Litke return -PTR_ERR(new_page); 19851e8f889bSDavid Gibson } 19861e8f889bSDavid Gibson 19871e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 19889de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 19890ed361deSNick Piggin __SetPageUptodate(new_page); 19901e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 19911e8f889bSDavid Gibson 1992a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 19937f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 19941e8f889bSDavid Gibson /* Break COW */ 19958fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 19961e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 19971e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 19981e8f889bSDavid Gibson /* Make the old page be freed below */ 19991e8f889bSDavid Gibson new_page = old_page; 20001e8f889bSDavid Gibson } 20011e8f889bSDavid Gibson page_cache_release(new_page); 20021e8f889bSDavid Gibson page_cache_release(old_page); 200383c54070SNick Piggin return 0; 20041e8f889bSDavid Gibson } 20051e8f889bSDavid Gibson 200604f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2007a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2008a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 200904f2cbe3SMel Gorman { 201004f2cbe3SMel Gorman struct address_space *mapping; 2011e7c4b0bfSAndy Whitcroft pgoff_t idx; 201204f2cbe3SMel Gorman 201304f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2014a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 201504f2cbe3SMel Gorman 201604f2cbe3SMel Gorman return find_lock_page(mapping, idx); 201704f2cbe3SMel Gorman } 201804f2cbe3SMel Gorman 2019a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2020788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2021ac9b9c66SHugh Dickins { 2022a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2023ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2024e7c4b0bfSAndy Whitcroft pgoff_t idx; 20254c887265SAdam Litke unsigned long size; 20264c887265SAdam Litke struct page *page; 20274c887265SAdam Litke struct address_space *mapping; 20281e8f889bSDavid Gibson pte_t new_pte; 20294c887265SAdam Litke 203004f2cbe3SMel Gorman /* 203104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 203204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 203304f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 203404f2cbe3SMel Gorman */ 203504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 203604f2cbe3SMel Gorman printk(KERN_WARNING 203704f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 203804f2cbe3SMel Gorman current->pid); 203904f2cbe3SMel Gorman return ret; 204004f2cbe3SMel Gorman } 204104f2cbe3SMel Gorman 20424c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2043a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 20444c887265SAdam Litke 20454c887265SAdam Litke /* 20464c887265SAdam Litke * Use page lock to guard against racing truncation 20474c887265SAdam Litke * before we get page_table_lock. 20484c887265SAdam Litke */ 20496bda666aSChristoph Lameter retry: 20506bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 20516bda666aSChristoph Lameter if (!page) { 2052a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2053ebed4bfcSHugh Dickins if (idx >= size) 2054ebed4bfcSHugh Dickins goto out; 205504f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 20562fc39cecSAdam Litke if (IS_ERR(page)) { 20572fc39cecSAdam Litke ret = -PTR_ERR(page); 20586bda666aSChristoph Lameter goto out; 20596bda666aSChristoph Lameter } 2060a5516438SAndi Kleen clear_huge_page(page, address, huge_page_size(h)); 20610ed361deSNick Piggin __SetPageUptodate(page); 2062ac9b9c66SHugh Dickins 2063f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 20646bda666aSChristoph Lameter int err; 206545c682a6SKen Chen struct inode *inode = mapping->host; 20666bda666aSChristoph Lameter 20676bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 20686bda666aSChristoph Lameter if (err) { 20696bda666aSChristoph Lameter put_page(page); 20706bda666aSChristoph Lameter if (err == -EEXIST) 20716bda666aSChristoph Lameter goto retry; 20726bda666aSChristoph Lameter goto out; 20736bda666aSChristoph Lameter } 207445c682a6SKen Chen 207545c682a6SKen Chen spin_lock(&inode->i_lock); 2076a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 207745c682a6SKen Chen spin_unlock(&inode->i_lock); 20786bda666aSChristoph Lameter } else 20796bda666aSChristoph Lameter lock_page(page); 20806bda666aSChristoph Lameter } 20811e8f889bSDavid Gibson 208257303d80SAndy Whitcroft /* 208357303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 208457303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 208557303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 208657303d80SAndy Whitcroft * the spinlock. 208757303d80SAndy Whitcroft */ 2088788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 20892b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 20902b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 20912b26736cSAndy Whitcroft goto backout_unlocked; 20922b26736cSAndy Whitcroft } 209357303d80SAndy Whitcroft 2094ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2095a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 20964c887265SAdam Litke if (idx >= size) 20974c887265SAdam Litke goto backout; 20984c887265SAdam Litke 209983c54070SNick Piggin ret = 0; 21007f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 21014c887265SAdam Litke goto backout; 21024c887265SAdam Litke 21031e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 21041e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 21051e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 21061e8f889bSDavid Gibson 2107788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 21081e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 210904f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 21101e8f889bSDavid Gibson } 21111e8f889bSDavid Gibson 2112ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 21134c887265SAdam Litke unlock_page(page); 21144c887265SAdam Litke out: 2115ac9b9c66SHugh Dickins return ret; 21164c887265SAdam Litke 21174c887265SAdam Litke backout: 21184c887265SAdam Litke spin_unlock(&mm->page_table_lock); 21192b26736cSAndy Whitcroft backout_unlocked: 21204c887265SAdam Litke unlock_page(page); 21214c887265SAdam Litke put_page(page); 21224c887265SAdam Litke goto out; 2123ac9b9c66SHugh Dickins } 2124ac9b9c66SHugh Dickins 212586e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2126788c7df4SHugh Dickins unsigned long address, unsigned int flags) 212786e5216fSAdam Litke { 212886e5216fSAdam Litke pte_t *ptep; 212986e5216fSAdam Litke pte_t entry; 21301e8f889bSDavid Gibson int ret; 213157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 21323935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2133a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 213486e5216fSAdam Litke 2135a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 213686e5216fSAdam Litke if (!ptep) 213786e5216fSAdam Litke return VM_FAULT_OOM; 213886e5216fSAdam Litke 21393935baa9SDavid Gibson /* 21403935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 21413935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 21423935baa9SDavid Gibson * the same page in the page cache. 21433935baa9SDavid Gibson */ 21443935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 21457f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 21467f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2147788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2148b4d1d99fSDavid Gibson goto out_mutex; 21493935baa9SDavid Gibson } 215086e5216fSAdam Litke 215183c54070SNick Piggin ret = 0; 21521e8f889bSDavid Gibson 215357303d80SAndy Whitcroft /* 215457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 215557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 215657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 215757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 215857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 215957303d80SAndy Whitcroft * consumed. 216057303d80SAndy Whitcroft */ 2161788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 21622b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 21632b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2164b4d1d99fSDavid Gibson goto out_mutex; 21652b26736cSAndy Whitcroft } 216657303d80SAndy Whitcroft 2167f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 216857303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 216957303d80SAndy Whitcroft vma, address); 217057303d80SAndy Whitcroft } 217157303d80SAndy Whitcroft 21721e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 21731e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2174b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2175b4d1d99fSDavid Gibson goto out_page_table_lock; 2176b4d1d99fSDavid Gibson 2177b4d1d99fSDavid Gibson 2178788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2179b4d1d99fSDavid Gibson if (!pte_write(entry)) { 218057303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 218157303d80SAndy Whitcroft pagecache_page); 2182b4d1d99fSDavid Gibson goto out_page_table_lock; 2183b4d1d99fSDavid Gibson } 2184b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2185b4d1d99fSDavid Gibson } 2186b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2187788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2188788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 2189b4d1d99fSDavid Gibson update_mmu_cache(vma, address, entry); 2190b4d1d99fSDavid Gibson 2191b4d1d99fSDavid Gibson out_page_table_lock: 21921e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 219357303d80SAndy Whitcroft 219457303d80SAndy Whitcroft if (pagecache_page) { 219557303d80SAndy Whitcroft unlock_page(pagecache_page); 219657303d80SAndy Whitcroft put_page(pagecache_page); 219757303d80SAndy Whitcroft } 219857303d80SAndy Whitcroft 2199b4d1d99fSDavid Gibson out_mutex: 22003935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 22011e8f889bSDavid Gibson 22021e8f889bSDavid Gibson return ret; 220386e5216fSAdam Litke } 220486e5216fSAdam Litke 2205ceb86879SAndi Kleen /* Can be overriden by architectures */ 2206ceb86879SAndi Kleen __attribute__((weak)) struct page * 2207ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2208ceb86879SAndi Kleen pud_t *pud, int write) 2209ceb86879SAndi Kleen { 2210ceb86879SAndi Kleen BUG(); 2211ceb86879SAndi Kleen return NULL; 2212ceb86879SAndi Kleen } 2213ceb86879SAndi Kleen 22144b2e38adSKOSAKI Motohiro static int huge_zeropage_ok(pte_t *ptep, int write, int shared) 22154b2e38adSKOSAKI Motohiro { 22164b2e38adSKOSAKI Motohiro if (!ptep || write || shared) 22174b2e38adSKOSAKI Motohiro return 0; 22184b2e38adSKOSAKI Motohiro else 22194b2e38adSKOSAKI Motohiro return huge_pte_none(huge_ptep_get(ptep)); 22204b2e38adSKOSAKI Motohiro } 22214b2e38adSKOSAKI Motohiro 222263551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 222363551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 22245b23dbe8SAdam Litke unsigned long *position, int *length, int i, 22255b23dbe8SAdam Litke int write) 222663551ae0SDavid Gibson { 2227d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2228d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 222963551ae0SDavid Gibson int remainder = *length; 2230a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 22314b2e38adSKOSAKI Motohiro int zeropage_ok = 0; 22324b2e38adSKOSAKI Motohiro int shared = vma->vm_flags & VM_SHARED; 223363551ae0SDavid Gibson 22341c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 223563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 223663551ae0SDavid Gibson pte_t *pte; 223763551ae0SDavid Gibson struct page *page; 223863551ae0SDavid Gibson 22394c887265SAdam Litke /* 22404c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 22414c887265SAdam Litke * each hugepage. We have to make * sure we get the 22424c887265SAdam Litke * first, for the page indexing below to work. 22434c887265SAdam Litke */ 2244a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 22454b2e38adSKOSAKI Motohiro if (huge_zeropage_ok(pte, write, shared)) 22464b2e38adSKOSAKI Motohiro zeropage_ok = 1; 224763551ae0SDavid Gibson 22484b2e38adSKOSAKI Motohiro if (!pte || 22494b2e38adSKOSAKI Motohiro (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) || 22507f2e9525SGerald Schaefer (write && !pte_write(huge_ptep_get(pte)))) { 22514c887265SAdam Litke int ret; 22524c887265SAdam Litke 22534c887265SAdam Litke spin_unlock(&mm->page_table_lock); 22545b23dbe8SAdam Litke ret = hugetlb_fault(mm, vma, vaddr, write); 22554c887265SAdam Litke spin_lock(&mm->page_table_lock); 2256a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 22574c887265SAdam Litke continue; 22584c887265SAdam Litke 22591c59827dSHugh Dickins remainder = 0; 22601c59827dSHugh Dickins if (!i) 22611c59827dSHugh Dickins i = -EFAULT; 22621c59827dSHugh Dickins break; 22631c59827dSHugh Dickins } 226463551ae0SDavid Gibson 2265a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 22667f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2267d5d4b0aaSChen, Kenneth W same_page: 2268d6692183SChen, Kenneth W if (pages) { 22694b2e38adSKOSAKI Motohiro if (zeropage_ok) 22704b2e38adSKOSAKI Motohiro pages[i] = ZERO_PAGE(0); 22714b2e38adSKOSAKI Motohiro else 227269d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 22734b2e38adSKOSAKI Motohiro get_page(pages[i]); 2274d6692183SChen, Kenneth W } 227563551ae0SDavid Gibson 227663551ae0SDavid Gibson if (vmas) 227763551ae0SDavid Gibson vmas[i] = vma; 227863551ae0SDavid Gibson 227963551ae0SDavid Gibson vaddr += PAGE_SIZE; 2280d5d4b0aaSChen, Kenneth W ++pfn_offset; 228163551ae0SDavid Gibson --remainder; 228263551ae0SDavid Gibson ++i; 2283d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2284a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2285d5d4b0aaSChen, Kenneth W /* 2286d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2287d5d4b0aaSChen, Kenneth W * of this compound page. 2288d5d4b0aaSChen, Kenneth W */ 2289d5d4b0aaSChen, Kenneth W goto same_page; 2290d5d4b0aaSChen, Kenneth W } 229163551ae0SDavid Gibson } 22921c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 229363551ae0SDavid Gibson *length = remainder; 229463551ae0SDavid Gibson *position = vaddr; 229563551ae0SDavid Gibson 229663551ae0SDavid Gibson return i; 229763551ae0SDavid Gibson } 22988f860591SZhang, Yanmin 22998f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 23008f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 23018f860591SZhang, Yanmin { 23028f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 23038f860591SZhang, Yanmin unsigned long start = address; 23048f860591SZhang, Yanmin pte_t *ptep; 23058f860591SZhang, Yanmin pte_t pte; 2306a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 23078f860591SZhang, Yanmin 23088f860591SZhang, Yanmin BUG_ON(address >= end); 23098f860591SZhang, Yanmin flush_cache_range(vma, address, end); 23108f860591SZhang, Yanmin 231139dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 23128f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2313a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 23148f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 23158f860591SZhang, Yanmin if (!ptep) 23168f860591SZhang, Yanmin continue; 231739dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 231839dde65cSChen, Kenneth W continue; 23197f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 23208f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 23218f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 23228f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 23238f860591SZhang, Yanmin } 23248f860591SZhang, Yanmin } 23258f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 232639dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 23278f860591SZhang, Yanmin 23288f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 23298f860591SZhang, Yanmin } 23308f860591SZhang, Yanmin 2331a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2332a1e78772SMel Gorman long from, long to, 23335a6fe125SMel Gorman struct vm_area_struct *vma, 23345a6fe125SMel Gorman int acctflag) 2335e4e574b7SAdam Litke { 233617c9d12eSMel Gorman long ret, chg; 2337a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2338e4e574b7SAdam Litke 2339a1e78772SMel Gorman /* 234017c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 234117c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 234217c9d12eSMel Gorman * and filesystem quota without using reserves 234317c9d12eSMel Gorman */ 234417c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 234517c9d12eSMel Gorman return 0; 234617c9d12eSMel Gorman 234717c9d12eSMel Gorman /* 2348a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2349a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2350a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2351a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2352a1e78772SMel Gorman */ 2353f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2354e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 23555a6fe125SMel Gorman else { 23565a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 23575a6fe125SMel Gorman if (!resv_map) 23585a6fe125SMel Gorman return -ENOMEM; 23595a6fe125SMel Gorman 236017c9d12eSMel Gorman chg = to - from; 236117c9d12eSMel Gorman 23625a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 23635a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 23645a6fe125SMel Gorman } 23655a6fe125SMel Gorman 236617c9d12eSMel Gorman if (chg < 0) 236717c9d12eSMel Gorman return chg; 236817c9d12eSMel Gorman 236917c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 237017c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 237117c9d12eSMel Gorman return -ENOSPC; 237217c9d12eSMel Gorman 237317c9d12eSMel Gorman /* 237417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 237517c9d12eSMel Gorman * Hand back the quota if there are not 237617c9d12eSMel Gorman */ 237717c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 237817c9d12eSMel Gorman if (ret < 0) { 237917c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 238017c9d12eSMel Gorman return ret; 238117c9d12eSMel Gorman } 238217c9d12eSMel Gorman 238317c9d12eSMel Gorman /* 238417c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 238517c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 238617c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 238717c9d12eSMel Gorman * the reservation was and the page cache tells how much of 238817c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 238917c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 239017c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 239117c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 239217c9d12eSMel Gorman * else has to be done for private mappings here 239317c9d12eSMel Gorman */ 2394f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 239517c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2396a43a8c39SChen, Kenneth W return 0; 2397a43a8c39SChen, Kenneth W } 2398a43a8c39SChen, Kenneth W 2399a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2400a43a8c39SChen, Kenneth W { 2401a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2402a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 240345c682a6SKen Chen 240445c682a6SKen Chen spin_lock(&inode->i_lock); 2405e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 240645c682a6SKen Chen spin_unlock(&inode->i_lock); 240745c682a6SKen Chen 240890d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2409a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2410a43a8c39SChen, Kenneth W } 2411