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 } 23708fba699SMel Gorman 23808fba699SMel Gorman /* 2393340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 2403340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 2413340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 2423340289dSMel Gorman * function is required. 2433340289dSMel Gorman */ 2443340289dSMel Gorman #ifndef vma_mmu_pagesize 2453340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 2463340289dSMel Gorman { 2473340289dSMel Gorman return vma_kernel_pagesize(vma); 2483340289dSMel Gorman } 2493340289dSMel Gorman #endif 2503340289dSMel Gorman 2513340289dSMel Gorman /* 25284afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 25384afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 25484afd99bSAndy Whitcroft * alignment. 25584afd99bSAndy Whitcroft */ 25684afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 25784afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 25804f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 25984afd99bSAndy Whitcroft 260a1e78772SMel Gorman /* 261a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 262a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 263a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 264a1e78772SMel Gorman * 265a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 266a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 267a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 268a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 26984afd99bSAndy Whitcroft * 27084afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 27184afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 27284afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 27384afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 27484afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 27584afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 27684afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 27784afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 278a1e78772SMel Gorman */ 279e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 280e7c4b0bfSAndy Whitcroft { 281e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 282e7c4b0bfSAndy Whitcroft } 283e7c4b0bfSAndy Whitcroft 284e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 285e7c4b0bfSAndy Whitcroft unsigned long value) 286e7c4b0bfSAndy Whitcroft { 287e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 288e7c4b0bfSAndy Whitcroft } 289e7c4b0bfSAndy Whitcroft 29084afd99bSAndy Whitcroft struct resv_map { 29184afd99bSAndy Whitcroft struct kref refs; 29284afd99bSAndy Whitcroft struct list_head regions; 29384afd99bSAndy Whitcroft }; 29484afd99bSAndy Whitcroft 2952a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 29684afd99bSAndy Whitcroft { 29784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 29884afd99bSAndy Whitcroft if (!resv_map) 29984afd99bSAndy Whitcroft return NULL; 30084afd99bSAndy Whitcroft 30184afd99bSAndy Whitcroft kref_init(&resv_map->refs); 30284afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 30384afd99bSAndy Whitcroft 30484afd99bSAndy Whitcroft return resv_map; 30584afd99bSAndy Whitcroft } 30684afd99bSAndy Whitcroft 3072a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 30884afd99bSAndy Whitcroft { 30984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 31084afd99bSAndy Whitcroft 31184afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 31284afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 31384afd99bSAndy Whitcroft kfree(resv_map); 31484afd99bSAndy Whitcroft } 31584afd99bSAndy Whitcroft 31684afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 317a1e78772SMel Gorman { 318a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 319f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 32084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 32184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 3222a4b3dedSHarvey Harrison return NULL; 323a1e78772SMel Gorman } 324a1e78772SMel Gorman 32584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 326a1e78772SMel Gorman { 327a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 328f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 329a1e78772SMel Gorman 33084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 33184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 33204f2cbe3SMel Gorman } 33304f2cbe3SMel Gorman 33404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 33504f2cbe3SMel Gorman { 33604f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 337f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 338e7c4b0bfSAndy Whitcroft 339e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 34004f2cbe3SMel Gorman } 34104f2cbe3SMel Gorman 34204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 34304f2cbe3SMel Gorman { 34404f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 345e7c4b0bfSAndy Whitcroft 346e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 347a1e78772SMel Gorman } 348a1e78772SMel Gorman 349a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 350a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 351a5516438SAndi Kleen struct vm_area_struct *vma) 352a1e78772SMel Gorman { 353c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 354c37f9fb1SAndy Whitcroft return; 355c37f9fb1SAndy Whitcroft 356f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 357a1e78772SMel Gorman /* Shared mappings always use reserves */ 358a5516438SAndi Kleen h->resv_huge_pages--; 35984afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 360a1e78772SMel Gorman /* 361a1e78772SMel Gorman * Only the process that called mmap() has reserves for 362a1e78772SMel Gorman * private mappings. 363a1e78772SMel Gorman */ 364a5516438SAndi Kleen h->resv_huge_pages--; 365a1e78772SMel Gorman } 366a1e78772SMel Gorman } 367a1e78772SMel Gorman 36804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 369a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 370a1e78772SMel Gorman { 371a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 372f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 373a1e78772SMel Gorman vma->vm_private_data = (void *)0; 374a1e78772SMel Gorman } 375a1e78772SMel Gorman 376a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 3777f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 378a1e78772SMel Gorman { 379f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 380a1e78772SMel Gorman return 1; 3817f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3827f09ca51SMel Gorman return 1; 3837f09ca51SMel Gorman return 0; 384a1e78772SMel Gorman } 385a1e78772SMel Gorman 38669d177c2SAndy Whitcroft static void clear_gigantic_page(struct page *page, 38769d177c2SAndy Whitcroft unsigned long addr, unsigned long sz) 38869d177c2SAndy Whitcroft { 38969d177c2SAndy Whitcroft int i; 39069d177c2SAndy Whitcroft struct page *p = page; 39169d177c2SAndy Whitcroft 39269d177c2SAndy Whitcroft might_sleep(); 39369d177c2SAndy Whitcroft for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) { 39469d177c2SAndy Whitcroft cond_resched(); 39569d177c2SAndy Whitcroft clear_user_highpage(p, addr + i * PAGE_SIZE); 39669d177c2SAndy Whitcroft } 39769d177c2SAndy Whitcroft } 398a5516438SAndi Kleen static void clear_huge_page(struct page *page, 399a5516438SAndi Kleen unsigned long addr, unsigned long sz) 40079ac6ba4SDavid Gibson { 40179ac6ba4SDavid Gibson int i; 40279ac6ba4SDavid Gibson 403ebdd4aeaSHannes Eder if (unlikely(sz > MAX_ORDER_NR_PAGES)) { 404ebdd4aeaSHannes Eder clear_gigantic_page(page, addr, sz); 405ebdd4aeaSHannes Eder return; 406ebdd4aeaSHannes Eder } 40769d177c2SAndy Whitcroft 40879ac6ba4SDavid Gibson might_sleep(); 409a5516438SAndi Kleen for (i = 0; i < sz/PAGE_SIZE; i++) { 41079ac6ba4SDavid Gibson cond_resched(); 411281e0e3bSRalf Baechle clear_user_highpage(page + i, addr + i * PAGE_SIZE); 41279ac6ba4SDavid Gibson } 41379ac6ba4SDavid Gibson } 41479ac6ba4SDavid Gibson 41569d177c2SAndy Whitcroft static void copy_gigantic_page(struct page *dst, struct page *src, 41669d177c2SAndy Whitcroft unsigned long addr, struct vm_area_struct *vma) 41769d177c2SAndy Whitcroft { 41869d177c2SAndy Whitcroft int i; 41969d177c2SAndy Whitcroft struct hstate *h = hstate_vma(vma); 42069d177c2SAndy Whitcroft struct page *dst_base = dst; 42169d177c2SAndy Whitcroft struct page *src_base = src; 42269d177c2SAndy Whitcroft might_sleep(); 42369d177c2SAndy Whitcroft for (i = 0; i < pages_per_huge_page(h); ) { 42469d177c2SAndy Whitcroft cond_resched(); 42569d177c2SAndy Whitcroft copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 42669d177c2SAndy Whitcroft 42769d177c2SAndy Whitcroft i++; 42869d177c2SAndy Whitcroft dst = mem_map_next(dst, dst_base, i); 42969d177c2SAndy Whitcroft src = mem_map_next(src, src_base, i); 43069d177c2SAndy Whitcroft } 43169d177c2SAndy Whitcroft } 43279ac6ba4SDavid Gibson static void copy_huge_page(struct page *dst, struct page *src, 4339de455b2SAtsushi Nemoto unsigned long addr, struct vm_area_struct *vma) 43479ac6ba4SDavid Gibson { 43579ac6ba4SDavid Gibson int i; 436a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43779ac6ba4SDavid Gibson 438ebdd4aeaSHannes Eder if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 439ebdd4aeaSHannes Eder copy_gigantic_page(dst, src, addr, vma); 440ebdd4aeaSHannes Eder return; 441ebdd4aeaSHannes Eder } 44269d177c2SAndy Whitcroft 44379ac6ba4SDavid Gibson might_sleep(); 444a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 44579ac6ba4SDavid Gibson cond_resched(); 4469de455b2SAtsushi Nemoto copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 44779ac6ba4SDavid Gibson } 44879ac6ba4SDavid Gibson } 44979ac6ba4SDavid Gibson 450a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 4511da177e4SLinus Torvalds { 4521da177e4SLinus Torvalds int nid = page_to_nid(page); 453a5516438SAndi Kleen list_add(&page->lru, &h->hugepage_freelists[nid]); 454a5516438SAndi Kleen h->free_huge_pages++; 455a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 4561da177e4SLinus Torvalds } 4571da177e4SLinus Torvalds 458a5516438SAndi Kleen static struct page *dequeue_huge_page(struct hstate *h) 459348e1e04SNishanth Aravamudan { 460348e1e04SNishanth Aravamudan int nid; 461348e1e04SNishanth Aravamudan struct page *page = NULL; 462348e1e04SNishanth Aravamudan 463348e1e04SNishanth Aravamudan for (nid = 0; nid < MAX_NUMNODES; ++nid) { 464a5516438SAndi Kleen if (!list_empty(&h->hugepage_freelists[nid])) { 465a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 466348e1e04SNishanth Aravamudan struct page, lru); 467348e1e04SNishanth Aravamudan list_del(&page->lru); 468a5516438SAndi Kleen h->free_huge_pages--; 469a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 470348e1e04SNishanth Aravamudan break; 471348e1e04SNishanth Aravamudan } 472348e1e04SNishanth Aravamudan } 473348e1e04SNishanth Aravamudan return page; 474348e1e04SNishanth Aravamudan } 475348e1e04SNishanth Aravamudan 476a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 477a5516438SAndi Kleen struct vm_area_struct *vma, 47804f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 4791da177e4SLinus Torvalds { 48031a5c6e4SNishanth Aravamudan int nid; 4811da177e4SLinus Torvalds struct page *page = NULL; 482480eccf9SLee Schermerhorn struct mempolicy *mpol; 48319770b32SMel Gorman nodemask_t *nodemask; 484396faf03SMel Gorman struct zonelist *zonelist = huge_zonelist(vma, address, 48519770b32SMel Gorman htlb_alloc_mask, &mpol, &nodemask); 486dd1a239fSMel Gorman struct zone *zone; 487dd1a239fSMel Gorman struct zoneref *z; 4881da177e4SLinus Torvalds 489a1e78772SMel Gorman /* 490a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 491a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 492a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 493a1e78772SMel Gorman */ 4947f09ca51SMel Gorman if (!vma_has_reserves(vma) && 495a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 496a1e78772SMel Gorman return NULL; 497a1e78772SMel Gorman 49804f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 499a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 50004f2cbe3SMel Gorman return NULL; 50104f2cbe3SMel Gorman 50219770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 50319770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 50454a6eb5cSMel Gorman nid = zone_to_nid(zone); 50554a6eb5cSMel Gorman if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 506a5516438SAndi Kleen !list_empty(&h->hugepage_freelists[nid])) { 507a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 5081da177e4SLinus Torvalds struct page, lru); 5091da177e4SLinus Torvalds list_del(&page->lru); 510a5516438SAndi Kleen h->free_huge_pages--; 511a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 51204f2cbe3SMel Gorman 51304f2cbe3SMel Gorman if (!avoid_reserve) 514a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 515a1e78772SMel Gorman 5165ab3ee7bSKen Chen break; 5171da177e4SLinus Torvalds } 5183abf7afdSAndrew Morton } 51952cd3b07SLee Schermerhorn mpol_cond_put(mpol); 5201da177e4SLinus Torvalds return page; 5211da177e4SLinus Torvalds } 5221da177e4SLinus Torvalds 523a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5246af2acb6SAdam Litke { 5256af2acb6SAdam Litke int i; 526a5516438SAndi Kleen 52718229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 52818229df5SAndy Whitcroft 529a5516438SAndi Kleen h->nr_huge_pages--; 530a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 531a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 5326af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 5336af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 5346af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 5356af2acb6SAdam Litke } 5366af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5376af2acb6SAdam Litke set_page_refcounted(page); 5387f2e9525SGerald Schaefer arch_release_hugepage(page); 539a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 5406af2acb6SAdam Litke } 5416af2acb6SAdam Litke 542e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 543e5ff2159SAndi Kleen { 544e5ff2159SAndi Kleen struct hstate *h; 545e5ff2159SAndi Kleen 546e5ff2159SAndi Kleen for_each_hstate(h) { 547e5ff2159SAndi Kleen if (huge_page_size(h) == size) 548e5ff2159SAndi Kleen return h; 549e5ff2159SAndi Kleen } 550e5ff2159SAndi Kleen return NULL; 551e5ff2159SAndi Kleen } 552e5ff2159SAndi Kleen 55327a85ef1SDavid Gibson static void free_huge_page(struct page *page) 55427a85ef1SDavid Gibson { 555a5516438SAndi Kleen /* 556a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 557a5516438SAndi Kleen * compound page destructor. 558a5516438SAndi Kleen */ 559e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 5607893d1d5SAdam Litke int nid = page_to_nid(page); 561c79fb75eSAdam Litke struct address_space *mapping; 56227a85ef1SDavid Gibson 563c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 564e5df70abSAndy Whitcroft set_page_private(page, 0); 5657893d1d5SAdam Litke BUG_ON(page_count(page)); 56627a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 56727a85ef1SDavid Gibson 56827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 569aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 570a5516438SAndi Kleen update_and_free_page(h, page); 571a5516438SAndi Kleen h->surplus_huge_pages--; 572a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 5737893d1d5SAdam Litke } else { 574a5516438SAndi Kleen enqueue_huge_page(h, page); 5757893d1d5SAdam Litke } 57627a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 577c79fb75eSAdam Litke if (mapping) 5789a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 57927a85ef1SDavid Gibson } 58027a85ef1SDavid Gibson 581a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 582b7ba30c6SAndi Kleen { 583b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 584b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 585a5516438SAndi Kleen h->nr_huge_pages++; 586a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 587b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 588b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 589b7ba30c6SAndi Kleen } 590b7ba30c6SAndi Kleen 59120a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 59220a0307cSWu Fengguang { 59320a0307cSWu Fengguang int i; 59420a0307cSWu Fengguang int nr_pages = 1 << order; 59520a0307cSWu Fengguang struct page *p = page + 1; 59620a0307cSWu Fengguang 59720a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 59820a0307cSWu Fengguang set_compound_order(page, order); 59920a0307cSWu Fengguang __SetPageHead(page); 60020a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 60120a0307cSWu Fengguang __SetPageTail(p); 60220a0307cSWu Fengguang p->first_page = page; 60320a0307cSWu Fengguang } 60420a0307cSWu Fengguang } 60520a0307cSWu Fengguang 60620a0307cSWu Fengguang int PageHuge(struct page *page) 60720a0307cSWu Fengguang { 60820a0307cSWu Fengguang compound_page_dtor *dtor; 60920a0307cSWu Fengguang 61020a0307cSWu Fengguang if (!PageCompound(page)) 61120a0307cSWu Fengguang return 0; 61220a0307cSWu Fengguang 61320a0307cSWu Fengguang page = compound_head(page); 61420a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 61520a0307cSWu Fengguang 61620a0307cSWu Fengguang return dtor == free_huge_page; 61720a0307cSWu Fengguang } 61820a0307cSWu Fengguang 619a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6201da177e4SLinus Torvalds { 6211da177e4SLinus Torvalds struct page *page; 622f96efd58SJoe Jin 623aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 624aa888a74SAndi Kleen return NULL; 625aa888a74SAndi Kleen 6266484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 627551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 628551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 629a5516438SAndi Kleen huge_page_order(h)); 6301da177e4SLinus Torvalds if (page) { 6317f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 632caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6337b8ee84dSHarvey Harrison return NULL; 6347f2e9525SGerald Schaefer } 635a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6361da177e4SLinus Torvalds } 63763b4613cSNishanth Aravamudan 63863b4613cSNishanth Aravamudan return page; 63963b4613cSNishanth Aravamudan } 64063b4613cSNishanth Aravamudan 6415ced66c9SAndi Kleen /* 6425ced66c9SAndi Kleen * Use a helper variable to find the next node and then 6435ced66c9SAndi Kleen * copy it back to hugetlb_next_nid afterwards: 6445ced66c9SAndi Kleen * otherwise there's a window in which a racer might 6456484eb3eSMel Gorman * pass invalid nid MAX_NUMNODES to alloc_pages_exact_node. 6465ced66c9SAndi Kleen * But we don't need to use a spin_lock here: it really 6475ced66c9SAndi Kleen * doesn't matter if occasionally a racer chooses the 6485ced66c9SAndi Kleen * same nid as we do. Move nid forward in the mask even 6495ced66c9SAndi Kleen * if we just successfully allocated a hugepage so that 6505ced66c9SAndi Kleen * the next caller gets hugepages on the next node. 6515ced66c9SAndi Kleen */ 6525ced66c9SAndi Kleen static int hstate_next_node(struct hstate *h) 6535ced66c9SAndi Kleen { 6545ced66c9SAndi Kleen int next_nid; 6555ced66c9SAndi Kleen next_nid = next_node(h->hugetlb_next_nid, node_online_map); 6565ced66c9SAndi Kleen if (next_nid == MAX_NUMNODES) 6575ced66c9SAndi Kleen next_nid = first_node(node_online_map); 6585ced66c9SAndi Kleen h->hugetlb_next_nid = next_nid; 6595ced66c9SAndi Kleen return next_nid; 6605ced66c9SAndi Kleen } 6615ced66c9SAndi Kleen 662a5516438SAndi Kleen static int alloc_fresh_huge_page(struct hstate *h) 66363b4613cSNishanth Aravamudan { 66463b4613cSNishanth Aravamudan struct page *page; 66563b4613cSNishanth Aravamudan int start_nid; 66663b4613cSNishanth Aravamudan int next_nid; 66763b4613cSNishanth Aravamudan int ret = 0; 66863b4613cSNishanth Aravamudan 669a5516438SAndi Kleen start_nid = h->hugetlb_next_nid; 67063b4613cSNishanth Aravamudan 67163b4613cSNishanth Aravamudan do { 672a5516438SAndi Kleen page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid); 67363b4613cSNishanth Aravamudan if (page) 67463b4613cSNishanth Aravamudan ret = 1; 6755ced66c9SAndi Kleen next_nid = hstate_next_node(h); 676a5516438SAndi Kleen } while (!page && h->hugetlb_next_nid != start_nid); 67763b4613cSNishanth Aravamudan 6783b116300SAdam Litke if (ret) 6793b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 6803b116300SAdam Litke else 6813b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 6823b116300SAdam Litke 68363b4613cSNishanth Aravamudan return ret; 6841da177e4SLinus Torvalds } 6851da177e4SLinus Torvalds 686a5516438SAndi Kleen static struct page *alloc_buddy_huge_page(struct hstate *h, 687a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 6887893d1d5SAdam Litke { 6897893d1d5SAdam Litke struct page *page; 690d1c3fb1fSNishanth Aravamudan unsigned int nid; 6917893d1d5SAdam Litke 692aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 693aa888a74SAndi Kleen return NULL; 694aa888a74SAndi Kleen 695d1c3fb1fSNishanth Aravamudan /* 696d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 697d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 698d1c3fb1fSNishanth Aravamudan * overcommit 699d1c3fb1fSNishanth Aravamudan * 700d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 701d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 702d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 703d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 704d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 705d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 706d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 707d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 708d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 709d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 710d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 711d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 712d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 713d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 714d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 715d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 716d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 717d1c3fb1fSNishanth Aravamudan */ 718d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 719a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 720d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 721d1c3fb1fSNishanth Aravamudan return NULL; 722d1c3fb1fSNishanth Aravamudan } else { 723a5516438SAndi Kleen h->nr_huge_pages++; 724a5516438SAndi Kleen h->surplus_huge_pages++; 725d1c3fb1fSNishanth Aravamudan } 726d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 727d1c3fb1fSNishanth Aravamudan 728551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 729551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 730a5516438SAndi Kleen huge_page_order(h)); 731d1c3fb1fSNishanth Aravamudan 732caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 733caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 734caff3a2cSGerald Schaefer return NULL; 735caff3a2cSGerald Schaefer } 736caff3a2cSGerald Schaefer 7377893d1d5SAdam Litke spin_lock(&hugetlb_lock); 738d1c3fb1fSNishanth Aravamudan if (page) { 7392668db91SAdam Litke /* 7402668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 7412668db91SAdam Litke * no users -- drop the buddy allocator's reference. 7422668db91SAdam Litke */ 7432668db91SAdam Litke put_page_testzero(page); 7442668db91SAdam Litke VM_BUG_ON(page_count(page)); 745d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 746d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 747d1c3fb1fSNishanth Aravamudan /* 748d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 749d1c3fb1fSNishanth Aravamudan */ 750a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 751a5516438SAndi Kleen h->surplus_huge_pages_node[nid]++; 7523b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 753d1c3fb1fSNishanth Aravamudan } else { 754a5516438SAndi Kleen h->nr_huge_pages--; 755a5516438SAndi Kleen h->surplus_huge_pages--; 7563b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7577893d1d5SAdam Litke } 758d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 7597893d1d5SAdam Litke 7607893d1d5SAdam Litke return page; 7617893d1d5SAdam Litke } 7627893d1d5SAdam Litke 763e4e574b7SAdam Litke /* 764e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 765e4e574b7SAdam Litke * of size 'delta'. 766e4e574b7SAdam Litke */ 767a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 768e4e574b7SAdam Litke { 769e4e574b7SAdam Litke struct list_head surplus_list; 770e4e574b7SAdam Litke struct page *page, *tmp; 771e4e574b7SAdam Litke int ret, i; 772e4e574b7SAdam Litke int needed, allocated; 773e4e574b7SAdam Litke 774a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 775ac09b3a1SAdam Litke if (needed <= 0) { 776a5516438SAndi Kleen h->resv_huge_pages += delta; 777e4e574b7SAdam Litke return 0; 778ac09b3a1SAdam Litke } 779e4e574b7SAdam Litke 780e4e574b7SAdam Litke allocated = 0; 781e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 782e4e574b7SAdam Litke 783e4e574b7SAdam Litke ret = -ENOMEM; 784e4e574b7SAdam Litke retry: 785e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 786e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 787a5516438SAndi Kleen page = alloc_buddy_huge_page(h, NULL, 0); 788e4e574b7SAdam Litke if (!page) { 789e4e574b7SAdam Litke /* 790e4e574b7SAdam Litke * We were not able to allocate enough pages to 791e4e574b7SAdam Litke * satisfy the entire reservation so we free what 792e4e574b7SAdam Litke * we've allocated so far. 793e4e574b7SAdam Litke */ 794e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 795e4e574b7SAdam Litke needed = 0; 796e4e574b7SAdam Litke goto free; 797e4e574b7SAdam Litke } 798e4e574b7SAdam Litke 799e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 800e4e574b7SAdam Litke } 801e4e574b7SAdam Litke allocated += needed; 802e4e574b7SAdam Litke 803e4e574b7SAdam Litke /* 804e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 805e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 806e4e574b7SAdam Litke */ 807e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 808a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 809a5516438SAndi Kleen (h->free_huge_pages + allocated); 810e4e574b7SAdam Litke if (needed > 0) 811e4e574b7SAdam Litke goto retry; 812e4e574b7SAdam Litke 813e4e574b7SAdam Litke /* 814e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 815e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 816e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 817ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 818ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 819ac09b3a1SAdam Litke * before they are reserved. 820e4e574b7SAdam Litke */ 821e4e574b7SAdam Litke needed += allocated; 822a5516438SAndi Kleen h->resv_huge_pages += delta; 823e4e574b7SAdam Litke ret = 0; 824e4e574b7SAdam Litke free: 82519fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 82619fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 82719fc3f0aSAdam Litke if ((--needed) < 0) 82819fc3f0aSAdam Litke break; 82919fc3f0aSAdam Litke list_del(&page->lru); 830a5516438SAndi Kleen enqueue_huge_page(h, page); 83119fc3f0aSAdam Litke } 83219fc3f0aSAdam Litke 83319fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 83419fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 83519fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 836e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 837e4e574b7SAdam Litke list_del(&page->lru); 838af767cbdSAdam Litke /* 8392668db91SAdam Litke * The page has a reference count of zero already, so 8402668db91SAdam Litke * call free_huge_page directly instead of using 8412668db91SAdam Litke * put_page. This must be done with hugetlb_lock 842af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 843af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 844af767cbdSAdam Litke */ 8452668db91SAdam Litke free_huge_page(page); 846af767cbdSAdam Litke } 84719fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 848e4e574b7SAdam Litke } 849e4e574b7SAdam Litke 850e4e574b7SAdam Litke return ret; 851e4e574b7SAdam Litke } 852e4e574b7SAdam Litke 853e4e574b7SAdam Litke /* 854e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 855e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 856e4e574b7SAdam Litke * never used. 857e4e574b7SAdam Litke */ 858a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 859a5516438SAndi Kleen unsigned long unused_resv_pages) 860e4e574b7SAdam Litke { 861e4e574b7SAdam Litke static int nid = -1; 862e4e574b7SAdam Litke struct page *page; 863e4e574b7SAdam Litke unsigned long nr_pages; 864e4e574b7SAdam Litke 86511320d17SNishanth Aravamudan /* 86611320d17SNishanth Aravamudan * We want to release as many surplus pages as possible, spread 86711320d17SNishanth Aravamudan * evenly across all nodes. Iterate across all nodes until we 86811320d17SNishanth Aravamudan * can no longer free unreserved surplus pages. This occurs when 86911320d17SNishanth Aravamudan * the nodes with surplus pages have no free pages. 87011320d17SNishanth Aravamudan */ 87162bc62a8SChristoph Lameter unsigned long remaining_iterations = nr_online_nodes; 87211320d17SNishanth Aravamudan 873ac09b3a1SAdam Litke /* Uncommit the reservation */ 874a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 875ac09b3a1SAdam Litke 876aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 877aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 878aa888a74SAndi Kleen return; 879aa888a74SAndi Kleen 880a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 881e4e574b7SAdam Litke 88211320d17SNishanth Aravamudan while (remaining_iterations-- && nr_pages) { 883e4e574b7SAdam Litke nid = next_node(nid, node_online_map); 884e4e574b7SAdam Litke if (nid == MAX_NUMNODES) 885e4e574b7SAdam Litke nid = first_node(node_online_map); 886e4e574b7SAdam Litke 887a5516438SAndi Kleen if (!h->surplus_huge_pages_node[nid]) 888e4e574b7SAdam Litke continue; 889e4e574b7SAdam Litke 890a5516438SAndi Kleen if (!list_empty(&h->hugepage_freelists[nid])) { 891a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 892e4e574b7SAdam Litke struct page, lru); 893e4e574b7SAdam Litke list_del(&page->lru); 894a5516438SAndi Kleen update_and_free_page(h, page); 895a5516438SAndi Kleen h->free_huge_pages--; 896a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 897a5516438SAndi Kleen h->surplus_huge_pages--; 898a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 899e4e574b7SAdam Litke nr_pages--; 90062bc62a8SChristoph Lameter remaining_iterations = nr_online_nodes; 901e4e574b7SAdam Litke } 902e4e574b7SAdam Litke } 903e4e574b7SAdam Litke } 904e4e574b7SAdam Litke 905c37f9fb1SAndy Whitcroft /* 906c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 907c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 908c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 909c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 910c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 911c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 912c37f9fb1SAndy Whitcroft * No action is required on failure. 913c37f9fb1SAndy Whitcroft */ 914e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 915a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 916c37f9fb1SAndy Whitcroft { 917c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 918c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 919c37f9fb1SAndy Whitcroft 920f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 921a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 922c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 923c37f9fb1SAndy Whitcroft idx, idx + 1); 924c37f9fb1SAndy Whitcroft 92584afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 926c37f9fb1SAndy Whitcroft return 1; 927c37f9fb1SAndy Whitcroft 92884afd99bSAndy Whitcroft } else { 929e2f17d94SRoel Kluin long err; 930a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 93184afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 93284afd99bSAndy Whitcroft 93384afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 93484afd99bSAndy Whitcroft if (err < 0) 93584afd99bSAndy Whitcroft return err; 936c37f9fb1SAndy Whitcroft return 0; 937c37f9fb1SAndy Whitcroft } 93884afd99bSAndy Whitcroft } 939a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 940a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 941c37f9fb1SAndy Whitcroft { 942c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 943c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 944c37f9fb1SAndy Whitcroft 945f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 946a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 947c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 94884afd99bSAndy Whitcroft 94984afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 950a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 95184afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 95284afd99bSAndy Whitcroft 95384afd99bSAndy Whitcroft /* Mark this page used in the map. */ 95484afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 955c37f9fb1SAndy Whitcroft } 956c37f9fb1SAndy Whitcroft } 957c37f9fb1SAndy Whitcroft 958348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 95904f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 960348ea204SAdam Litke { 961a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 962348ea204SAdam Litke struct page *page; 9632fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 964a1e78772SMel Gorman struct inode *inode = mapping->host; 965e2f17d94SRoel Kluin long chg; 9662fc39cecSAdam Litke 967a1e78772SMel Gorman /* 968a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 969a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 970a1e78772SMel Gorman * made before satisfying the allocation 971c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 972c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 973a1e78772SMel Gorman */ 974a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 975c37f9fb1SAndy Whitcroft if (chg < 0) 976c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 977c37f9fb1SAndy Whitcroft if (chg) 978a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 979a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 98090d8b7e6SAdam Litke 981a1e78772SMel Gorman spin_lock(&hugetlb_lock); 982a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 983a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 984a1e78772SMel Gorman 985a1e78772SMel Gorman if (!page) { 986a5516438SAndi Kleen page = alloc_buddy_huge_page(h, vma, addr); 987a1e78772SMel Gorman if (!page) { 988a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 989a1e78772SMel Gorman return ERR_PTR(-VM_FAULT_OOM); 990a1e78772SMel Gorman } 991a1e78772SMel Gorman } 992a1e78772SMel Gorman 993348ea204SAdam Litke set_page_refcounted(page); 9942fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 995a1e78772SMel Gorman 996a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 997c37f9fb1SAndy Whitcroft 9987893d1d5SAdam Litke return page; 999b45b5bd6SDavid Gibson } 1000b45b5bd6SDavid Gibson 100191f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1002aa888a74SAndi Kleen { 1003aa888a74SAndi Kleen struct huge_bootmem_page *m; 1004aa888a74SAndi Kleen int nr_nodes = nodes_weight(node_online_map); 1005aa888a74SAndi Kleen 1006aa888a74SAndi Kleen while (nr_nodes) { 1007aa888a74SAndi Kleen void *addr; 1008aa888a74SAndi Kleen 1009aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 1010aa888a74SAndi Kleen NODE_DATA(h->hugetlb_next_nid), 1011aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1012aa888a74SAndi Kleen 1013aa888a74SAndi Kleen if (addr) { 1014aa888a74SAndi Kleen /* 1015aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1016aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1017aa888a74SAndi Kleen * puts them into the mem_map). 1018aa888a74SAndi Kleen */ 1019aa888a74SAndi Kleen m = addr; 1020aa888a74SAndi Kleen goto found; 1021aa888a74SAndi Kleen } 1022aa888a74SAndi Kleen hstate_next_node(h); 1023aa888a74SAndi Kleen nr_nodes--; 1024aa888a74SAndi Kleen } 1025aa888a74SAndi Kleen return 0; 1026aa888a74SAndi Kleen 1027aa888a74SAndi Kleen found: 1028aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1029aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1030aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1031aa888a74SAndi Kleen m->hstate = h; 1032aa888a74SAndi Kleen return 1; 1033aa888a74SAndi Kleen } 1034aa888a74SAndi Kleen 103518229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 103618229df5SAndy Whitcroft { 103718229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 103818229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 103918229df5SAndy Whitcroft else 104018229df5SAndy Whitcroft prep_compound_page(page, order); 104118229df5SAndy Whitcroft } 104218229df5SAndy Whitcroft 1043aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1044aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1045aa888a74SAndi Kleen { 1046aa888a74SAndi Kleen struct huge_bootmem_page *m; 1047aa888a74SAndi Kleen 1048aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1049aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1050aa888a74SAndi Kleen struct hstate *h = m->hstate; 1051aa888a74SAndi Kleen __ClearPageReserved(page); 1052aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 105318229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1054aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1055aa888a74SAndi Kleen } 1056aa888a74SAndi Kleen } 1057aa888a74SAndi Kleen 10588faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 10591da177e4SLinus Torvalds { 10601da177e4SLinus Torvalds unsigned long i; 10611da177e4SLinus Torvalds 1062e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1063aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1064aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1065aa888a74SAndi Kleen break; 1066aa888a74SAndi Kleen } else if (!alloc_fresh_huge_page(h)) 10671da177e4SLinus Torvalds break; 10681da177e4SLinus Torvalds } 10698faa8b07SAndi Kleen h->max_huge_pages = i; 1070e5ff2159SAndi Kleen } 1071e5ff2159SAndi Kleen 1072e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1073e5ff2159SAndi Kleen { 1074e5ff2159SAndi Kleen struct hstate *h; 1075e5ff2159SAndi Kleen 1076e5ff2159SAndi Kleen for_each_hstate(h) { 10778faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 10788faa8b07SAndi Kleen if (h->order < MAX_ORDER) 10798faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1080e5ff2159SAndi Kleen } 1081e5ff2159SAndi Kleen } 1082e5ff2159SAndi Kleen 10834abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 10844abd32dbSAndi Kleen { 10854abd32dbSAndi Kleen if (n >= (1UL << 30)) 10864abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 10874abd32dbSAndi Kleen else if (n >= (1UL << 20)) 10884abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 10894abd32dbSAndi Kleen else 10904abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 10914abd32dbSAndi Kleen return buf; 10924abd32dbSAndi Kleen } 10934abd32dbSAndi Kleen 1094e5ff2159SAndi Kleen static void __init report_hugepages(void) 1095e5ff2159SAndi Kleen { 1096e5ff2159SAndi Kleen struct hstate *h; 1097e5ff2159SAndi Kleen 1098e5ff2159SAndi Kleen for_each_hstate(h) { 10994abd32dbSAndi Kleen char buf[32]; 11004abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11014abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11024abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11034abd32dbSAndi Kleen h->free_huge_pages); 1104e5ff2159SAndi Kleen } 1105e5ff2159SAndi Kleen } 1106e5ff2159SAndi Kleen 11071da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 1108a5516438SAndi Kleen static void try_to_free_low(struct hstate *h, unsigned long count) 11091da177e4SLinus Torvalds { 11104415cc8dSChristoph Lameter int i; 11114415cc8dSChristoph Lameter 1112aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1113aa888a74SAndi Kleen return; 1114aa888a74SAndi Kleen 11151da177e4SLinus Torvalds for (i = 0; i < MAX_NUMNODES; ++i) { 11161da177e4SLinus Torvalds struct page *page, *next; 1117a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1118a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1119a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11206b0c880dSAdam Litke return; 11211da177e4SLinus Torvalds if (PageHighMem(page)) 11221da177e4SLinus Torvalds continue; 11231da177e4SLinus Torvalds list_del(&page->lru); 1124e5ff2159SAndi Kleen update_and_free_page(h, page); 1125a5516438SAndi Kleen h->free_huge_pages--; 1126a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11271da177e4SLinus Torvalds } 11281da177e4SLinus Torvalds } 11291da177e4SLinus Torvalds } 11301da177e4SLinus Torvalds #else 1131a5516438SAndi Kleen static inline void try_to_free_low(struct hstate *h, unsigned long count) 11321da177e4SLinus Torvalds { 11331da177e4SLinus Torvalds } 11341da177e4SLinus Torvalds #endif 11351da177e4SLinus Torvalds 113620a0307cSWu Fengguang /* 113720a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 113820a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 113920a0307cSWu Fengguang * Returns 1 if an adjustment was made. 114020a0307cSWu Fengguang */ 114120a0307cSWu Fengguang static int adjust_pool_surplus(struct hstate *h, int delta) 114220a0307cSWu Fengguang { 114320a0307cSWu Fengguang static int prev_nid; 114420a0307cSWu Fengguang int nid = prev_nid; 114520a0307cSWu Fengguang int ret = 0; 114620a0307cSWu Fengguang 114720a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 114820a0307cSWu Fengguang do { 114920a0307cSWu Fengguang nid = next_node(nid, node_online_map); 115020a0307cSWu Fengguang if (nid == MAX_NUMNODES) 115120a0307cSWu Fengguang nid = first_node(node_online_map); 115220a0307cSWu Fengguang 115320a0307cSWu Fengguang /* To shrink on this node, there must be a surplus page */ 115420a0307cSWu Fengguang if (delta < 0 && !h->surplus_huge_pages_node[nid]) 115520a0307cSWu Fengguang continue; 115620a0307cSWu Fengguang /* Surplus cannot exceed the total number of pages */ 115720a0307cSWu Fengguang if (delta > 0 && h->surplus_huge_pages_node[nid] >= 115820a0307cSWu Fengguang h->nr_huge_pages_node[nid]) 115920a0307cSWu Fengguang continue; 116020a0307cSWu Fengguang 116120a0307cSWu Fengguang h->surplus_huge_pages += delta; 116220a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 116320a0307cSWu Fengguang ret = 1; 116420a0307cSWu Fengguang break; 116520a0307cSWu Fengguang } while (nid != prev_nid); 116620a0307cSWu Fengguang 116720a0307cSWu Fengguang prev_nid = nid; 116820a0307cSWu Fengguang return ret; 116920a0307cSWu Fengguang } 117020a0307cSWu Fengguang 1171a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1172e5ff2159SAndi Kleen static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count) 11731da177e4SLinus Torvalds { 11747893d1d5SAdam Litke unsigned long min_count, ret; 11751da177e4SLinus Torvalds 1176aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1177aa888a74SAndi Kleen return h->max_huge_pages; 1178aa888a74SAndi Kleen 11797893d1d5SAdam Litke /* 11807893d1d5SAdam Litke * Increase the pool size 11817893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 11827893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1183d1c3fb1fSNishanth Aravamudan * 1184d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1185d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1186d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1187d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1188d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 11897893d1d5SAdam Litke */ 11901da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1191a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1192a5516438SAndi Kleen if (!adjust_pool_surplus(h, -1)) 11937893d1d5SAdam Litke break; 11947893d1d5SAdam Litke } 11957893d1d5SAdam Litke 1196a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 11977893d1d5SAdam Litke /* 11987893d1d5SAdam Litke * If this allocation races such that we no longer need the 11997893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12007893d1d5SAdam Litke * and reducing the surplus. 12017893d1d5SAdam Litke */ 12027893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 1203a5516438SAndi Kleen ret = alloc_fresh_huge_page(h); 12047893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12057893d1d5SAdam Litke if (!ret) 12067893d1d5SAdam Litke goto out; 12077893d1d5SAdam Litke 12087893d1d5SAdam Litke } 12097893d1d5SAdam Litke 12107893d1d5SAdam Litke /* 12117893d1d5SAdam Litke * Decrease the pool size 12127893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 12137893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 12147893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 12157893d1d5SAdam Litke * to the desired size as pages become free. 1216d1c3fb1fSNishanth Aravamudan * 1217d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1218d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1219d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1220d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1221d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1222d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1223d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 12247893d1d5SAdam Litke */ 1225a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 12266b0c880dSAdam Litke min_count = max(count, min_count); 1227a5516438SAndi Kleen try_to_free_low(h, min_count); 1228a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 1229a5516438SAndi Kleen struct page *page = dequeue_huge_page(h); 12301da177e4SLinus Torvalds if (!page) 12311da177e4SLinus Torvalds break; 1232a5516438SAndi Kleen update_and_free_page(h, page); 12331da177e4SLinus Torvalds } 1234a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 1235a5516438SAndi Kleen if (!adjust_pool_surplus(h, 1)) 12367893d1d5SAdam Litke break; 12377893d1d5SAdam Litke } 12387893d1d5SAdam Litke out: 1239a5516438SAndi Kleen ret = persistent_huge_pages(h); 12401da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 12417893d1d5SAdam Litke return ret; 12421da177e4SLinus Torvalds } 12431da177e4SLinus Torvalds 1244a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1245a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1246a3437870SNishanth Aravamudan 1247a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1248a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1249a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1250a3437870SNishanth Aravamudan 1251a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1252a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1253a3437870SNishanth Aravamudan 1254a3437870SNishanth Aravamudan static struct hstate *kobj_to_hstate(struct kobject *kobj) 1255a3437870SNishanth Aravamudan { 1256a3437870SNishanth Aravamudan int i; 1257a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 1258a3437870SNishanth Aravamudan if (hstate_kobjs[i] == kobj) 1259a3437870SNishanth Aravamudan return &hstates[i]; 1260a3437870SNishanth Aravamudan BUG(); 1261a3437870SNishanth Aravamudan return NULL; 1262a3437870SNishanth Aravamudan } 1263a3437870SNishanth Aravamudan 1264a3437870SNishanth Aravamudan static ssize_t nr_hugepages_show(struct kobject *kobj, 1265a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1266a3437870SNishanth Aravamudan { 1267a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1268a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_huge_pages); 1269a3437870SNishanth Aravamudan } 1270a3437870SNishanth Aravamudan static ssize_t nr_hugepages_store(struct kobject *kobj, 1271a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1272a3437870SNishanth Aravamudan { 1273a3437870SNishanth Aravamudan int err; 1274a3437870SNishanth Aravamudan unsigned long input; 1275a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1276a3437870SNishanth Aravamudan 1277a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1278a3437870SNishanth Aravamudan if (err) 1279a3437870SNishanth Aravamudan return 0; 1280a3437870SNishanth Aravamudan 1281a3437870SNishanth Aravamudan h->max_huge_pages = set_max_huge_pages(h, input); 1282a3437870SNishanth Aravamudan 1283a3437870SNishanth Aravamudan return count; 1284a3437870SNishanth Aravamudan } 1285a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1286a3437870SNishanth Aravamudan 1287a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1288a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1289a3437870SNishanth Aravamudan { 1290a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1291a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1292a3437870SNishanth Aravamudan } 1293a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1294a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1295a3437870SNishanth Aravamudan { 1296a3437870SNishanth Aravamudan int err; 1297a3437870SNishanth Aravamudan unsigned long input; 1298a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1299a3437870SNishanth Aravamudan 1300a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1301a3437870SNishanth Aravamudan if (err) 1302a3437870SNishanth Aravamudan return 0; 1303a3437870SNishanth Aravamudan 1304a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1305a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1306a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1307a3437870SNishanth Aravamudan 1308a3437870SNishanth Aravamudan return count; 1309a3437870SNishanth Aravamudan } 1310a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1311a3437870SNishanth Aravamudan 1312a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1313a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1314a3437870SNishanth Aravamudan { 1315a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1316a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->free_huge_pages); 1317a3437870SNishanth Aravamudan } 1318a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1319a3437870SNishanth Aravamudan 1320a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1321a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1322a3437870SNishanth Aravamudan { 1323a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1324a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1325a3437870SNishanth Aravamudan } 1326a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1327a3437870SNishanth Aravamudan 1328a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1329a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1330a3437870SNishanth Aravamudan { 1331a3437870SNishanth Aravamudan struct hstate *h = kobj_to_hstate(kobj); 1332a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->surplus_huge_pages); 1333a3437870SNishanth Aravamudan } 1334a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1335a3437870SNishanth Aravamudan 1336a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1337a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1338a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1339a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1340a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1341a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 1342a3437870SNishanth Aravamudan NULL, 1343a3437870SNishanth Aravamudan }; 1344a3437870SNishanth Aravamudan 1345a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1346a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1347a3437870SNishanth Aravamudan }; 1348a3437870SNishanth Aravamudan 1349a3437870SNishanth Aravamudan static int __init hugetlb_sysfs_add_hstate(struct hstate *h) 1350a3437870SNishanth Aravamudan { 1351a3437870SNishanth Aravamudan int retval; 1352a3437870SNishanth Aravamudan 1353a3437870SNishanth Aravamudan hstate_kobjs[h - hstates] = kobject_create_and_add(h->name, 1354a3437870SNishanth Aravamudan hugepages_kobj); 1355a3437870SNishanth Aravamudan if (!hstate_kobjs[h - hstates]) 1356a3437870SNishanth Aravamudan return -ENOMEM; 1357a3437870SNishanth Aravamudan 1358a3437870SNishanth Aravamudan retval = sysfs_create_group(hstate_kobjs[h - hstates], 1359a3437870SNishanth Aravamudan &hstate_attr_group); 1360a3437870SNishanth Aravamudan if (retval) 1361a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1362a3437870SNishanth Aravamudan 1363a3437870SNishanth Aravamudan return retval; 1364a3437870SNishanth Aravamudan } 1365a3437870SNishanth Aravamudan 1366a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1367a3437870SNishanth Aravamudan { 1368a3437870SNishanth Aravamudan struct hstate *h; 1369a3437870SNishanth Aravamudan int err; 1370a3437870SNishanth Aravamudan 1371a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1372a3437870SNishanth Aravamudan if (!hugepages_kobj) 1373a3437870SNishanth Aravamudan return; 1374a3437870SNishanth Aravamudan 1375a3437870SNishanth Aravamudan for_each_hstate(h) { 1376a3437870SNishanth Aravamudan err = hugetlb_sysfs_add_hstate(h); 1377a3437870SNishanth Aravamudan if (err) 1378a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1379a3437870SNishanth Aravamudan h->name); 1380a3437870SNishanth Aravamudan } 1381a3437870SNishanth Aravamudan } 1382a3437870SNishanth Aravamudan 1383a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1384a3437870SNishanth Aravamudan { 1385a3437870SNishanth Aravamudan struct hstate *h; 1386a3437870SNishanth Aravamudan 1387a3437870SNishanth Aravamudan for_each_hstate(h) { 1388a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1389a3437870SNishanth Aravamudan } 1390a3437870SNishanth Aravamudan 1391a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1392a3437870SNishanth Aravamudan } 1393a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1394a3437870SNishanth Aravamudan 1395a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1396a3437870SNishanth Aravamudan { 13970ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 13980ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 13990ef89d25SBenjamin Herrenschmidt * there is no such support 14000ef89d25SBenjamin Herrenschmidt */ 14010ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 14020ef89d25SBenjamin Herrenschmidt return 0; 1403a3437870SNishanth Aravamudan 1404e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1405e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1406e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1407a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1408a3437870SNishanth Aravamudan } 1409e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1410e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1411e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1412a3437870SNishanth Aravamudan 1413a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1414a3437870SNishanth Aravamudan 1415aa888a74SAndi Kleen gather_bootmem_prealloc(); 1416aa888a74SAndi Kleen 1417a3437870SNishanth Aravamudan report_hugepages(); 1418a3437870SNishanth Aravamudan 1419a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1420a3437870SNishanth Aravamudan 1421a3437870SNishanth Aravamudan return 0; 1422a3437870SNishanth Aravamudan } 1423a3437870SNishanth Aravamudan module_init(hugetlb_init); 1424a3437870SNishanth Aravamudan 1425a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1426a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1427a3437870SNishanth Aravamudan { 1428a3437870SNishanth Aravamudan struct hstate *h; 14298faa8b07SAndi Kleen unsigned long i; 14308faa8b07SAndi Kleen 1431a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1432a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1433a3437870SNishanth Aravamudan return; 1434a3437870SNishanth Aravamudan } 1435a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1436a3437870SNishanth Aravamudan BUG_ON(order == 0); 1437a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1438a3437870SNishanth Aravamudan h->order = order; 1439a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 14408faa8b07SAndi Kleen h->nr_huge_pages = 0; 14418faa8b07SAndi Kleen h->free_huge_pages = 0; 14428faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 14438faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 14448faa8b07SAndi Kleen h->hugetlb_next_nid = first_node(node_online_map); 1445a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1446a3437870SNishanth Aravamudan huge_page_size(h)/1024); 14478faa8b07SAndi Kleen 1448a3437870SNishanth Aravamudan parsed_hstate = h; 1449a3437870SNishanth Aravamudan } 1450a3437870SNishanth Aravamudan 1451e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1452a3437870SNishanth Aravamudan { 1453a3437870SNishanth Aravamudan unsigned long *mhp; 14548faa8b07SAndi Kleen static unsigned long *last_mhp; 1455a3437870SNishanth Aravamudan 1456a3437870SNishanth Aravamudan /* 1457a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1458a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1459a3437870SNishanth Aravamudan */ 1460a3437870SNishanth Aravamudan if (!max_hstate) 1461a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1462a3437870SNishanth Aravamudan else 1463a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1464a3437870SNishanth Aravamudan 14658faa8b07SAndi Kleen if (mhp == last_mhp) { 14668faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 14678faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 14688faa8b07SAndi Kleen return 1; 14698faa8b07SAndi Kleen } 14708faa8b07SAndi Kleen 1471a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1472a3437870SNishanth Aravamudan *mhp = 0; 1473a3437870SNishanth Aravamudan 14748faa8b07SAndi Kleen /* 14758faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 14768faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 14778faa8b07SAndi Kleen * use the bootmem allocator. 14788faa8b07SAndi Kleen */ 14798faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 14808faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 14818faa8b07SAndi Kleen 14828faa8b07SAndi Kleen last_mhp = mhp; 14838faa8b07SAndi Kleen 1484a3437870SNishanth Aravamudan return 1; 1485a3437870SNishanth Aravamudan } 1486e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1487e11bfbfcSNick Piggin 1488e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1489e11bfbfcSNick Piggin { 1490e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1491e11bfbfcSNick Piggin return 1; 1492e11bfbfcSNick Piggin } 1493e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1494a3437870SNishanth Aravamudan 14958a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 14968a213460SNishanth Aravamudan { 14978a213460SNishanth Aravamudan int node; 14988a213460SNishanth Aravamudan unsigned int nr = 0; 14998a213460SNishanth Aravamudan 15008a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 15018a213460SNishanth Aravamudan nr += array[node]; 15028a213460SNishanth Aravamudan 15038a213460SNishanth Aravamudan return nr; 15048a213460SNishanth Aravamudan } 15058a213460SNishanth Aravamudan 15068a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 15071da177e4SLinus Torvalds int hugetlb_sysctl_handler(struct ctl_table *table, int write, 15081da177e4SLinus Torvalds struct file *file, void __user *buffer, 15091da177e4SLinus Torvalds size_t *length, loff_t *ppos) 15101da177e4SLinus Torvalds { 1511e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1512e5ff2159SAndi Kleen unsigned long tmp; 1513e5ff2159SAndi Kleen 1514e5ff2159SAndi Kleen if (!write) 1515e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1516e5ff2159SAndi Kleen 1517e5ff2159SAndi Kleen table->data = &tmp; 1518e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 15191da177e4SLinus Torvalds proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 1520e5ff2159SAndi Kleen 1521e5ff2159SAndi Kleen if (write) 1522e5ff2159SAndi Kleen h->max_huge_pages = set_max_huge_pages(h, tmp); 1523e5ff2159SAndi Kleen 15241da177e4SLinus Torvalds return 0; 15251da177e4SLinus Torvalds } 1526396faf03SMel Gorman 1527396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 1528396faf03SMel Gorman struct file *file, void __user *buffer, 1529396faf03SMel Gorman size_t *length, loff_t *ppos) 1530396faf03SMel Gorman { 1531396faf03SMel Gorman proc_dointvec(table, write, file, buffer, length, ppos); 1532396faf03SMel Gorman if (hugepages_treat_as_movable) 1533396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1534396faf03SMel Gorman else 1535396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1536396faf03SMel Gorman return 0; 1537396faf03SMel Gorman } 1538396faf03SMel Gorman 1539a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 1540a3d0c6aaSNishanth Aravamudan struct file *file, void __user *buffer, 1541a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1542a3d0c6aaSNishanth Aravamudan { 1543a5516438SAndi Kleen struct hstate *h = &default_hstate; 1544e5ff2159SAndi Kleen unsigned long tmp; 1545e5ff2159SAndi Kleen 1546e5ff2159SAndi Kleen if (!write) 1547e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1548e5ff2159SAndi Kleen 1549e5ff2159SAndi Kleen table->data = &tmp; 1550e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 1551a3d0c6aaSNishanth Aravamudan proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 1552e5ff2159SAndi Kleen 1553e5ff2159SAndi Kleen if (write) { 1554064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1555e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1556a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1557e5ff2159SAndi Kleen } 1558e5ff2159SAndi Kleen 1559a3d0c6aaSNishanth Aravamudan return 0; 1560a3d0c6aaSNishanth Aravamudan } 1561a3d0c6aaSNishanth Aravamudan 15621da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 15631da177e4SLinus Torvalds 1564e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 15651da177e4SLinus Torvalds { 1566a5516438SAndi Kleen struct hstate *h = &default_hstate; 1567e1759c21SAlexey Dobriyan seq_printf(m, 15681da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 15691da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1570b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 15717893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 15724f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1573a5516438SAndi Kleen h->nr_huge_pages, 1574a5516438SAndi Kleen h->free_huge_pages, 1575a5516438SAndi Kleen h->resv_huge_pages, 1576a5516438SAndi Kleen h->surplus_huge_pages, 1577a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 15781da177e4SLinus Torvalds } 15791da177e4SLinus Torvalds 15801da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 15811da177e4SLinus Torvalds { 1582a5516438SAndi Kleen struct hstate *h = &default_hstate; 15831da177e4SLinus Torvalds return sprintf(buf, 15841da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1585a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1586a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1587a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1588a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1589a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 15901da177e4SLinus Torvalds } 15911da177e4SLinus Torvalds 15921da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 15931da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 15941da177e4SLinus Torvalds { 1595a5516438SAndi Kleen struct hstate *h = &default_hstate; 1596a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 15971da177e4SLinus Torvalds } 15981da177e4SLinus Torvalds 1599a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 1600fc1b8a73SMel Gorman { 1601fc1b8a73SMel Gorman int ret = -ENOMEM; 1602fc1b8a73SMel Gorman 1603fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 1604fc1b8a73SMel Gorman /* 1605fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 1606fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 1607fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 1608fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 1609fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 1610fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 1611fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 1612fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 1613fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 1614fc1b8a73SMel Gorman * 1615fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 1616fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 1617fc1b8a73SMel Gorman * we fall back to check against current free page availability as 1618fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 1619fc1b8a73SMel Gorman * semantics that cpuset has. 1620fc1b8a73SMel Gorman */ 1621fc1b8a73SMel Gorman if (delta > 0) { 1622a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 1623fc1b8a73SMel Gorman goto out; 1624fc1b8a73SMel Gorman 1625a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 1626a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 1627fc1b8a73SMel Gorman goto out; 1628fc1b8a73SMel Gorman } 1629fc1b8a73SMel Gorman } 1630fc1b8a73SMel Gorman 1631fc1b8a73SMel Gorman ret = 0; 1632fc1b8a73SMel Gorman if (delta < 0) 1633a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 1634fc1b8a73SMel Gorman 1635fc1b8a73SMel Gorman out: 1636fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 1637fc1b8a73SMel Gorman return ret; 1638fc1b8a73SMel Gorman } 1639fc1b8a73SMel Gorman 164084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 164184afd99bSAndy Whitcroft { 164284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 164384afd99bSAndy Whitcroft 164484afd99bSAndy Whitcroft /* 164584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 164684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 164784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 164884afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 164984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 165084afd99bSAndy Whitcroft * new reference here without additional locking. 165184afd99bSAndy Whitcroft */ 165284afd99bSAndy Whitcroft if (reservations) 165384afd99bSAndy Whitcroft kref_get(&reservations->refs); 165484afd99bSAndy Whitcroft } 165584afd99bSAndy Whitcroft 1656a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 1657a1e78772SMel Gorman { 1658a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 165984afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 166084afd99bSAndy Whitcroft unsigned long reserve; 166184afd99bSAndy Whitcroft unsigned long start; 166284afd99bSAndy Whitcroft unsigned long end; 166384afd99bSAndy Whitcroft 166484afd99bSAndy Whitcroft if (reservations) { 1665a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 1666a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 166784afd99bSAndy Whitcroft 166884afd99bSAndy Whitcroft reserve = (end - start) - 166984afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 167084afd99bSAndy Whitcroft 167184afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 167284afd99bSAndy Whitcroft 16737251ff78SAdam Litke if (reserve) { 1674a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 16757251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 16767251ff78SAdam Litke } 1677a1e78772SMel Gorman } 167884afd99bSAndy Whitcroft } 1679a1e78772SMel Gorman 16801da177e4SLinus Torvalds /* 16811da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 16821da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 16831da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 16841da177e4SLinus Torvalds * this far. 16851da177e4SLinus Torvalds */ 1686d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 16871da177e4SLinus Torvalds { 16881da177e4SLinus Torvalds BUG(); 1689d0217ac0SNick Piggin return 0; 16901da177e4SLinus Torvalds } 16911da177e4SLinus Torvalds 16921da177e4SLinus Torvalds struct vm_operations_struct hugetlb_vm_ops = { 1693d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 169484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 1695a1e78772SMel Gorman .close = hugetlb_vm_op_close, 16961da177e4SLinus Torvalds }; 16971da177e4SLinus Torvalds 16981e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 16991e8f889bSDavid Gibson int writable) 170063551ae0SDavid Gibson { 170163551ae0SDavid Gibson pte_t entry; 170263551ae0SDavid Gibson 17031e8f889bSDavid Gibson if (writable) { 170463551ae0SDavid Gibson entry = 170563551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 170663551ae0SDavid Gibson } else { 17077f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 170863551ae0SDavid Gibson } 170963551ae0SDavid Gibson entry = pte_mkyoung(entry); 171063551ae0SDavid Gibson entry = pte_mkhuge(entry); 171163551ae0SDavid Gibson 171263551ae0SDavid Gibson return entry; 171363551ae0SDavid Gibson } 171463551ae0SDavid Gibson 17151e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 17161e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 17171e8f889bSDavid Gibson { 17181e8f889bSDavid Gibson pte_t entry; 17191e8f889bSDavid Gibson 17207f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 17217f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 17221e8f889bSDavid Gibson update_mmu_cache(vma, address, entry); 17231e8f889bSDavid Gibson } 17248dab5241SBenjamin Herrenschmidt } 17251e8f889bSDavid Gibson 17261e8f889bSDavid Gibson 172763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 172863551ae0SDavid Gibson struct vm_area_struct *vma) 172963551ae0SDavid Gibson { 173063551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 173163551ae0SDavid Gibson struct page *ptepage; 17321c59827dSHugh Dickins unsigned long addr; 17331e8f889bSDavid Gibson int cow; 1734a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1735a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 17361e8f889bSDavid Gibson 17371e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 173863551ae0SDavid Gibson 1739a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 1740c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 1741c74df32cSHugh Dickins if (!src_pte) 1742c74df32cSHugh Dickins continue; 1743a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 174463551ae0SDavid Gibson if (!dst_pte) 174563551ae0SDavid Gibson goto nomem; 1746c5c99429SLarry Woodman 1747c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 1748c5c99429SLarry Woodman if (dst_pte == src_pte) 1749c5c99429SLarry Woodman continue; 1750c5c99429SLarry Woodman 1751c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 175246478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 17537f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 17541e8f889bSDavid Gibson if (cow) 17557f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 17567f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 175763551ae0SDavid Gibson ptepage = pte_page(entry); 175863551ae0SDavid Gibson get_page(ptepage); 175963551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 17601c59827dSHugh Dickins } 17611c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 1762c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 176363551ae0SDavid Gibson } 176463551ae0SDavid Gibson return 0; 176563551ae0SDavid Gibson 176663551ae0SDavid Gibson nomem: 176763551ae0SDavid Gibson return -ENOMEM; 176863551ae0SDavid Gibson } 176963551ae0SDavid Gibson 1770502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 177104f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 177263551ae0SDavid Gibson { 177363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 177463551ae0SDavid Gibson unsigned long address; 1775c7546f8fSDavid Gibson pte_t *ptep; 177663551ae0SDavid Gibson pte_t pte; 177763551ae0SDavid Gibson struct page *page; 1778fe1668aeSChen, Kenneth W struct page *tmp; 1779a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1780a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 1781a5516438SAndi Kleen 1782c0a499c2SChen, Kenneth W /* 1783c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 1784c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 1785c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 1786c0a499c2SChen, Kenneth W */ 1787fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 178863551ae0SDavid Gibson 178963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 1790a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 1791a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 179263551ae0SDavid Gibson 1793cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 1794508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 1795a5516438SAndi Kleen for (address = start; address < end; address += sz) { 1796c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 1797c7546f8fSDavid Gibson if (!ptep) 1798c7546f8fSDavid Gibson continue; 1799c7546f8fSDavid Gibson 180039dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 180139dde65cSChen, Kenneth W continue; 180239dde65cSChen, Kenneth W 180304f2cbe3SMel Gorman /* 180404f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 180504f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 180604f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 180704f2cbe3SMel Gorman */ 180804f2cbe3SMel Gorman if (ref_page) { 180904f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 181004f2cbe3SMel Gorman if (huge_pte_none(pte)) 181104f2cbe3SMel Gorman continue; 181204f2cbe3SMel Gorman page = pte_page(pte); 181304f2cbe3SMel Gorman if (page != ref_page) 181404f2cbe3SMel Gorman continue; 181504f2cbe3SMel Gorman 181604f2cbe3SMel Gorman /* 181704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 181804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 181904f2cbe3SMel Gorman * looking like data was lost 182004f2cbe3SMel Gorman */ 182104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 182204f2cbe3SMel Gorman } 182304f2cbe3SMel Gorman 1824c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 18257f2e9525SGerald Schaefer if (huge_pte_none(pte)) 182663551ae0SDavid Gibson continue; 1827c7546f8fSDavid Gibson 182863551ae0SDavid Gibson page = pte_page(pte); 18296649a386SKen Chen if (pte_dirty(pte)) 18306649a386SKen Chen set_page_dirty(page); 1831fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 183263551ae0SDavid Gibson } 18331da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 1834508034a3SHugh Dickins flush_tlb_range(vma, start, end); 1835cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 1836fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 1837fe1668aeSChen, Kenneth W list_del(&page->lru); 1838fe1668aeSChen, Kenneth W put_page(page); 1839fe1668aeSChen, Kenneth W } 18401da177e4SLinus Torvalds } 184163551ae0SDavid Gibson 1842502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 184304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 1844502717f4SChen, Kenneth W { 1845502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 184604f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 1847502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 1848502717f4SChen, Kenneth W } 1849502717f4SChen, Kenneth W 185004f2cbe3SMel Gorman /* 185104f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 185204f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 185304f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 185404f2cbe3SMel Gorman * same region. 185504f2cbe3SMel Gorman */ 18562a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 18572a4b3dedSHarvey Harrison struct page *page, unsigned long address) 185804f2cbe3SMel Gorman { 18597526674dSAdam Litke struct hstate *h = hstate_vma(vma); 186004f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 186104f2cbe3SMel Gorman struct address_space *mapping; 186204f2cbe3SMel Gorman struct prio_tree_iter iter; 186304f2cbe3SMel Gorman pgoff_t pgoff; 186404f2cbe3SMel Gorman 186504f2cbe3SMel Gorman /* 186604f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 186704f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 186804f2cbe3SMel Gorman */ 18697526674dSAdam Litke address = address & huge_page_mask(h); 187004f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 187104f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 187204f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 187304f2cbe3SMel Gorman 187404f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 187504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 187604f2cbe3SMel Gorman if (iter_vma == vma) 187704f2cbe3SMel Gorman continue; 187804f2cbe3SMel Gorman 187904f2cbe3SMel Gorman /* 188004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 188104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 188204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 188304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 188404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 188504f2cbe3SMel Gorman */ 188604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 188704f2cbe3SMel Gorman unmap_hugepage_range(iter_vma, 18887526674dSAdam Litke address, address + huge_page_size(h), 188904f2cbe3SMel Gorman page); 189004f2cbe3SMel Gorman } 189104f2cbe3SMel Gorman 189204f2cbe3SMel Gorman return 1; 189304f2cbe3SMel Gorman } 189404f2cbe3SMel Gorman 18951e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 189604f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 189704f2cbe3SMel Gorman struct page *pagecache_page) 18981e8f889bSDavid Gibson { 1899a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 19001e8f889bSDavid Gibson struct page *old_page, *new_page; 190179ac6ba4SDavid Gibson int avoidcopy; 190204f2cbe3SMel Gorman int outside_reserve = 0; 19031e8f889bSDavid Gibson 19041e8f889bSDavid Gibson old_page = pte_page(pte); 19051e8f889bSDavid Gibson 190604f2cbe3SMel Gorman retry_avoidcopy: 19071e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 19081e8f889bSDavid Gibson * and just make the page writable */ 19091e8f889bSDavid Gibson avoidcopy = (page_count(old_page) == 1); 19101e8f889bSDavid Gibson if (avoidcopy) { 19111e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 191283c54070SNick Piggin return 0; 19131e8f889bSDavid Gibson } 19141e8f889bSDavid Gibson 191504f2cbe3SMel Gorman /* 191604f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 191704f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 191804f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 191904f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 192004f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 192104f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 192204f2cbe3SMel Gorman * of the full address range. 192304f2cbe3SMel Gorman */ 1924f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 192504f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 192604f2cbe3SMel Gorman old_page != pagecache_page) 192704f2cbe3SMel Gorman outside_reserve = 1; 192804f2cbe3SMel Gorman 19291e8f889bSDavid Gibson page_cache_get(old_page); 193004f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 19311e8f889bSDavid Gibson 19322fc39cecSAdam Litke if (IS_ERR(new_page)) { 19331e8f889bSDavid Gibson page_cache_release(old_page); 193404f2cbe3SMel Gorman 193504f2cbe3SMel Gorman /* 193604f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 193704f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 193804f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 193904f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 194004f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 194104f2cbe3SMel Gorman */ 194204f2cbe3SMel Gorman if (outside_reserve) { 194304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 194404f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 194504f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 194604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 194704f2cbe3SMel Gorman goto retry_avoidcopy; 194804f2cbe3SMel Gorman } 194904f2cbe3SMel Gorman WARN_ON_ONCE(1); 195004f2cbe3SMel Gorman } 195104f2cbe3SMel Gorman 19522fc39cecSAdam Litke return -PTR_ERR(new_page); 19531e8f889bSDavid Gibson } 19541e8f889bSDavid Gibson 19551e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 19569de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 19570ed361deSNick Piggin __SetPageUptodate(new_page); 19581e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 19591e8f889bSDavid Gibson 1960a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 19617f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 19621e8f889bSDavid Gibson /* Break COW */ 19638fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 19641e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 19651e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 19661e8f889bSDavid Gibson /* Make the old page be freed below */ 19671e8f889bSDavid Gibson new_page = old_page; 19681e8f889bSDavid Gibson } 19691e8f889bSDavid Gibson page_cache_release(new_page); 19701e8f889bSDavid Gibson page_cache_release(old_page); 197183c54070SNick Piggin return 0; 19721e8f889bSDavid Gibson } 19731e8f889bSDavid Gibson 197404f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 1975a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 1976a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 197704f2cbe3SMel Gorman { 197804f2cbe3SMel Gorman struct address_space *mapping; 1979e7c4b0bfSAndy Whitcroft pgoff_t idx; 198004f2cbe3SMel Gorman 198104f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 1982a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 198304f2cbe3SMel Gorman 198404f2cbe3SMel Gorman return find_lock_page(mapping, idx); 198504f2cbe3SMel Gorman } 198604f2cbe3SMel Gorman 1987a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 1988788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 1989ac9b9c66SHugh Dickins { 1990a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1991ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 1992e7c4b0bfSAndy Whitcroft pgoff_t idx; 19934c887265SAdam Litke unsigned long size; 19944c887265SAdam Litke struct page *page; 19954c887265SAdam Litke struct address_space *mapping; 19961e8f889bSDavid Gibson pte_t new_pte; 19974c887265SAdam Litke 199804f2cbe3SMel Gorman /* 199904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 200004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 200104f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 200204f2cbe3SMel Gorman */ 200304f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 200404f2cbe3SMel Gorman printk(KERN_WARNING 200504f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 200604f2cbe3SMel Gorman current->pid); 200704f2cbe3SMel Gorman return ret; 200804f2cbe3SMel Gorman } 200904f2cbe3SMel Gorman 20104c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2011a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 20124c887265SAdam Litke 20134c887265SAdam Litke /* 20144c887265SAdam Litke * Use page lock to guard against racing truncation 20154c887265SAdam Litke * before we get page_table_lock. 20164c887265SAdam Litke */ 20176bda666aSChristoph Lameter retry: 20186bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 20196bda666aSChristoph Lameter if (!page) { 2020a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2021ebed4bfcSHugh Dickins if (idx >= size) 2022ebed4bfcSHugh Dickins goto out; 202304f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 20242fc39cecSAdam Litke if (IS_ERR(page)) { 20252fc39cecSAdam Litke ret = -PTR_ERR(page); 20266bda666aSChristoph Lameter goto out; 20276bda666aSChristoph Lameter } 2028a5516438SAndi Kleen clear_huge_page(page, address, huge_page_size(h)); 20290ed361deSNick Piggin __SetPageUptodate(page); 2030ac9b9c66SHugh Dickins 2031f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 20326bda666aSChristoph Lameter int err; 203345c682a6SKen Chen struct inode *inode = mapping->host; 20346bda666aSChristoph Lameter 20356bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 20366bda666aSChristoph Lameter if (err) { 20376bda666aSChristoph Lameter put_page(page); 20386bda666aSChristoph Lameter if (err == -EEXIST) 20396bda666aSChristoph Lameter goto retry; 20406bda666aSChristoph Lameter goto out; 20416bda666aSChristoph Lameter } 204245c682a6SKen Chen 204345c682a6SKen Chen spin_lock(&inode->i_lock); 2044a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 204545c682a6SKen Chen spin_unlock(&inode->i_lock); 20466bda666aSChristoph Lameter } else 20476bda666aSChristoph Lameter lock_page(page); 20486bda666aSChristoph Lameter } 20491e8f889bSDavid Gibson 205057303d80SAndy Whitcroft /* 205157303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 205257303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 205357303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 205457303d80SAndy Whitcroft * the spinlock. 205557303d80SAndy Whitcroft */ 2056788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 20572b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 20582b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 20592b26736cSAndy Whitcroft goto backout_unlocked; 20602b26736cSAndy Whitcroft } 206157303d80SAndy Whitcroft 2062ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2063a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 20644c887265SAdam Litke if (idx >= size) 20654c887265SAdam Litke goto backout; 20664c887265SAdam Litke 206783c54070SNick Piggin ret = 0; 20687f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 20694c887265SAdam Litke goto backout; 20704c887265SAdam Litke 20711e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 20721e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 20731e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 20741e8f889bSDavid Gibson 2075788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 20761e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 207704f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 20781e8f889bSDavid Gibson } 20791e8f889bSDavid Gibson 2080ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 20814c887265SAdam Litke unlock_page(page); 20824c887265SAdam Litke out: 2083ac9b9c66SHugh Dickins return ret; 20844c887265SAdam Litke 20854c887265SAdam Litke backout: 20864c887265SAdam Litke spin_unlock(&mm->page_table_lock); 20872b26736cSAndy Whitcroft backout_unlocked: 20884c887265SAdam Litke unlock_page(page); 20894c887265SAdam Litke put_page(page); 20904c887265SAdam Litke goto out; 2091ac9b9c66SHugh Dickins } 2092ac9b9c66SHugh Dickins 209386e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2094788c7df4SHugh Dickins unsigned long address, unsigned int flags) 209586e5216fSAdam Litke { 209686e5216fSAdam Litke pte_t *ptep; 209786e5216fSAdam Litke pte_t entry; 20981e8f889bSDavid Gibson int ret; 209957303d80SAndy Whitcroft struct page *pagecache_page = NULL; 21003935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2101a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 210286e5216fSAdam Litke 2103a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 210486e5216fSAdam Litke if (!ptep) 210586e5216fSAdam Litke return VM_FAULT_OOM; 210686e5216fSAdam Litke 21073935baa9SDavid Gibson /* 21083935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 21093935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 21103935baa9SDavid Gibson * the same page in the page cache. 21113935baa9SDavid Gibson */ 21123935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 21137f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 21147f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2115788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2116b4d1d99fSDavid Gibson goto out_mutex; 21173935baa9SDavid Gibson } 211886e5216fSAdam Litke 211983c54070SNick Piggin ret = 0; 21201e8f889bSDavid Gibson 212157303d80SAndy Whitcroft /* 212257303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 212357303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 212457303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 212557303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 212657303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 212757303d80SAndy Whitcroft * consumed. 212857303d80SAndy Whitcroft */ 2129788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 21302b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 21312b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2132b4d1d99fSDavid Gibson goto out_mutex; 21332b26736cSAndy Whitcroft } 213457303d80SAndy Whitcroft 2135f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 213657303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 213757303d80SAndy Whitcroft vma, address); 213857303d80SAndy Whitcroft } 213957303d80SAndy Whitcroft 21401e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 21411e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2142b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2143b4d1d99fSDavid Gibson goto out_page_table_lock; 2144b4d1d99fSDavid Gibson 2145b4d1d99fSDavid Gibson 2146788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2147b4d1d99fSDavid Gibson if (!pte_write(entry)) { 214857303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 214957303d80SAndy Whitcroft pagecache_page); 2150b4d1d99fSDavid Gibson goto out_page_table_lock; 2151b4d1d99fSDavid Gibson } 2152b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2153b4d1d99fSDavid Gibson } 2154b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2155788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2156788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 2157b4d1d99fSDavid Gibson update_mmu_cache(vma, address, entry); 2158b4d1d99fSDavid Gibson 2159b4d1d99fSDavid Gibson out_page_table_lock: 21601e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 216157303d80SAndy Whitcroft 216257303d80SAndy Whitcroft if (pagecache_page) { 216357303d80SAndy Whitcroft unlock_page(pagecache_page); 216457303d80SAndy Whitcroft put_page(pagecache_page); 216557303d80SAndy Whitcroft } 216657303d80SAndy Whitcroft 2167b4d1d99fSDavid Gibson out_mutex: 21683935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 21691e8f889bSDavid Gibson 21701e8f889bSDavid Gibson return ret; 217186e5216fSAdam Litke } 217286e5216fSAdam Litke 2173ceb86879SAndi Kleen /* Can be overriden by architectures */ 2174ceb86879SAndi Kleen __attribute__((weak)) struct page * 2175ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2176ceb86879SAndi Kleen pud_t *pud, int write) 2177ceb86879SAndi Kleen { 2178ceb86879SAndi Kleen BUG(); 2179ceb86879SAndi Kleen return NULL; 2180ceb86879SAndi Kleen } 2181ceb86879SAndi Kleen 21824b2e38adSKOSAKI Motohiro static int huge_zeropage_ok(pte_t *ptep, int write, int shared) 21834b2e38adSKOSAKI Motohiro { 21844b2e38adSKOSAKI Motohiro if (!ptep || write || shared) 21854b2e38adSKOSAKI Motohiro return 0; 21864b2e38adSKOSAKI Motohiro else 21874b2e38adSKOSAKI Motohiro return huge_pte_none(huge_ptep_get(ptep)); 21884b2e38adSKOSAKI Motohiro } 21894b2e38adSKOSAKI Motohiro 219063551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 219163551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 21925b23dbe8SAdam Litke unsigned long *position, int *length, int i, 21935b23dbe8SAdam Litke int write) 219463551ae0SDavid Gibson { 2195d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2196d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 219763551ae0SDavid Gibson int remainder = *length; 2198a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 21994b2e38adSKOSAKI Motohiro int zeropage_ok = 0; 22004b2e38adSKOSAKI Motohiro int shared = vma->vm_flags & VM_SHARED; 220163551ae0SDavid Gibson 22021c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 220363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 220463551ae0SDavid Gibson pte_t *pte; 220563551ae0SDavid Gibson struct page *page; 220663551ae0SDavid Gibson 22074c887265SAdam Litke /* 22084c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 22094c887265SAdam Litke * each hugepage. We have to make * sure we get the 22104c887265SAdam Litke * first, for the page indexing below to work. 22114c887265SAdam Litke */ 2212a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 22134b2e38adSKOSAKI Motohiro if (huge_zeropage_ok(pte, write, shared)) 22144b2e38adSKOSAKI Motohiro zeropage_ok = 1; 221563551ae0SDavid Gibson 22164b2e38adSKOSAKI Motohiro if (!pte || 22174b2e38adSKOSAKI Motohiro (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) || 22187f2e9525SGerald Schaefer (write && !pte_write(huge_ptep_get(pte)))) { 22194c887265SAdam Litke int ret; 22204c887265SAdam Litke 22214c887265SAdam Litke spin_unlock(&mm->page_table_lock); 22225b23dbe8SAdam Litke ret = hugetlb_fault(mm, vma, vaddr, write); 22234c887265SAdam Litke spin_lock(&mm->page_table_lock); 2224a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 22254c887265SAdam Litke continue; 22264c887265SAdam Litke 22271c59827dSHugh Dickins remainder = 0; 22281c59827dSHugh Dickins if (!i) 22291c59827dSHugh Dickins i = -EFAULT; 22301c59827dSHugh Dickins break; 22311c59827dSHugh Dickins } 223263551ae0SDavid Gibson 2233a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 22347f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2235d5d4b0aaSChen, Kenneth W same_page: 2236d6692183SChen, Kenneth W if (pages) { 22374b2e38adSKOSAKI Motohiro if (zeropage_ok) 22384b2e38adSKOSAKI Motohiro pages[i] = ZERO_PAGE(0); 22394b2e38adSKOSAKI Motohiro else 224069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 22414b2e38adSKOSAKI Motohiro get_page(pages[i]); 2242d6692183SChen, Kenneth W } 224363551ae0SDavid Gibson 224463551ae0SDavid Gibson if (vmas) 224563551ae0SDavid Gibson vmas[i] = vma; 224663551ae0SDavid Gibson 224763551ae0SDavid Gibson vaddr += PAGE_SIZE; 2248d5d4b0aaSChen, Kenneth W ++pfn_offset; 224963551ae0SDavid Gibson --remainder; 225063551ae0SDavid Gibson ++i; 2251d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2252a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2253d5d4b0aaSChen, Kenneth W /* 2254d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2255d5d4b0aaSChen, Kenneth W * of this compound page. 2256d5d4b0aaSChen, Kenneth W */ 2257d5d4b0aaSChen, Kenneth W goto same_page; 2258d5d4b0aaSChen, Kenneth W } 225963551ae0SDavid Gibson } 22601c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 226163551ae0SDavid Gibson *length = remainder; 226263551ae0SDavid Gibson *position = vaddr; 226363551ae0SDavid Gibson 226463551ae0SDavid Gibson return i; 226563551ae0SDavid Gibson } 22668f860591SZhang, Yanmin 22678f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 22688f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 22698f860591SZhang, Yanmin { 22708f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 22718f860591SZhang, Yanmin unsigned long start = address; 22728f860591SZhang, Yanmin pte_t *ptep; 22738f860591SZhang, Yanmin pte_t pte; 2274a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 22758f860591SZhang, Yanmin 22768f860591SZhang, Yanmin BUG_ON(address >= end); 22778f860591SZhang, Yanmin flush_cache_range(vma, address, end); 22788f860591SZhang, Yanmin 227939dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 22808f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2281a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 22828f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 22838f860591SZhang, Yanmin if (!ptep) 22848f860591SZhang, Yanmin continue; 228539dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 228639dde65cSChen, Kenneth W continue; 22877f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 22888f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 22898f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 22908f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 22918f860591SZhang, Yanmin } 22928f860591SZhang, Yanmin } 22938f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 229439dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 22958f860591SZhang, Yanmin 22968f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 22978f860591SZhang, Yanmin } 22988f860591SZhang, Yanmin 2299a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2300a1e78772SMel Gorman long from, long to, 23015a6fe125SMel Gorman struct vm_area_struct *vma, 23025a6fe125SMel Gorman int acctflag) 2303e4e574b7SAdam Litke { 230417c9d12eSMel Gorman long ret, chg; 2305a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2306e4e574b7SAdam Litke 2307a1e78772SMel Gorman /* 230817c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 230917c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 231017c9d12eSMel Gorman * and filesystem quota without using reserves 231117c9d12eSMel Gorman */ 231217c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 231317c9d12eSMel Gorman return 0; 231417c9d12eSMel Gorman 231517c9d12eSMel Gorman /* 2316a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2317a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2318a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2319a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2320a1e78772SMel Gorman */ 2321f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2322e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 23235a6fe125SMel Gorman else { 23245a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 23255a6fe125SMel Gorman if (!resv_map) 23265a6fe125SMel Gorman return -ENOMEM; 23275a6fe125SMel Gorman 232817c9d12eSMel Gorman chg = to - from; 232917c9d12eSMel Gorman 23305a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 23315a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 23325a6fe125SMel Gorman } 23335a6fe125SMel Gorman 233417c9d12eSMel Gorman if (chg < 0) 233517c9d12eSMel Gorman return chg; 233617c9d12eSMel Gorman 233717c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 233817c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 233917c9d12eSMel Gorman return -ENOSPC; 234017c9d12eSMel Gorman 234117c9d12eSMel Gorman /* 234217c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 234317c9d12eSMel Gorman * Hand back the quota if there are not 234417c9d12eSMel Gorman */ 234517c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 234617c9d12eSMel Gorman if (ret < 0) { 234717c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 234817c9d12eSMel Gorman return ret; 234917c9d12eSMel Gorman } 235017c9d12eSMel Gorman 235117c9d12eSMel Gorman /* 235217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 235317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 235417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 235517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 235617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 235717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 235817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 235917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 236017c9d12eSMel Gorman * else has to be done for private mappings here 236117c9d12eSMel Gorman */ 2362f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 236317c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2364a43a8c39SChen, Kenneth W return 0; 2365a43a8c39SChen, Kenneth W } 2366a43a8c39SChen, Kenneth W 2367a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2368a43a8c39SChen, Kenneth W { 2369a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2370a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 237145c682a6SKen Chen 237245c682a6SKen Chen spin_lock(&inode->i_lock); 2373a5516438SAndi Kleen inode->i_blocks -= blocks_per_huge_page(h); 237445c682a6SKen Chen spin_unlock(&inode->i_lock); 237545c682a6SKen Chen 237690d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2377a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2378a43a8c39SChen, Kenneth W } 2379