11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 31da177e4SLinus Torvalds * (C) William Irwin, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 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> 279a305230SLee Schermerhorn #include <linux/node.h> 287835e98bSNick Piggin #include "internal.h" 291da177e4SLinus Torvalds 301da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 31396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 32396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 33a5516438SAndi Kleen 34e5ff2159SAndi Kleen static int max_hstate; 35e5ff2159SAndi Kleen unsigned int default_hstate_idx; 36e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 37e5ff2159SAndi Kleen 3853ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 3953ba51d2SJon Tollefson 40e5ff2159SAndi Kleen /* for command line parsing */ 41e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 42e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 43e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 44e5ff2159SAndi Kleen 45e5ff2159SAndi Kleen #define for_each_hstate(h) \ 46e5ff2159SAndi Kleen for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) 47396faf03SMel Gorman 483935baa9SDavid Gibson /* 493935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 503935baa9SDavid Gibson */ 513935baa9SDavid Gibson static DEFINE_SPINLOCK(hugetlb_lock); 520bd0f9fbSEric Paris 53e7c4b0bfSAndy Whitcroft /* 5496822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 5596822904SAndy Whitcroft * across the pages in a mapping. 5684afd99bSAndy Whitcroft * 5784afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 5884afd99bSAndy Whitcroft * and the hugetlb_instantion_mutex. To access or modify a region the caller 5984afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 6084afd99bSAndy Whitcroft * the hugetlb_instantiation mutex: 6184afd99bSAndy Whitcroft * 6284afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 6384afd99bSAndy Whitcroft * or 6484afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 6584afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 6696822904SAndy Whitcroft */ 6796822904SAndy Whitcroft struct file_region { 6896822904SAndy Whitcroft struct list_head link; 6996822904SAndy Whitcroft long from; 7096822904SAndy Whitcroft long to; 7196822904SAndy Whitcroft }; 7296822904SAndy Whitcroft 7396822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 7496822904SAndy Whitcroft { 7596822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 7696822904SAndy Whitcroft 7796822904SAndy Whitcroft /* Locate the region we are either in or before. */ 7896822904SAndy Whitcroft list_for_each_entry(rg, head, link) 7996822904SAndy Whitcroft if (f <= rg->to) 8096822904SAndy Whitcroft break; 8196822904SAndy Whitcroft 8296822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 8396822904SAndy Whitcroft if (f > rg->from) 8496822904SAndy Whitcroft f = rg->from; 8596822904SAndy Whitcroft 8696822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 8796822904SAndy Whitcroft nrg = rg; 8896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 8996822904SAndy Whitcroft if (&rg->link == head) 9096822904SAndy Whitcroft break; 9196822904SAndy Whitcroft if (rg->from > t) 9296822904SAndy Whitcroft break; 9396822904SAndy Whitcroft 9496822904SAndy Whitcroft /* If this area reaches higher then extend our area to 9596822904SAndy Whitcroft * include it completely. If this is not the first area 9696822904SAndy Whitcroft * which we intend to reuse, free it. */ 9796822904SAndy Whitcroft if (rg->to > t) 9896822904SAndy Whitcroft t = rg->to; 9996822904SAndy Whitcroft if (rg != nrg) { 10096822904SAndy Whitcroft list_del(&rg->link); 10196822904SAndy Whitcroft kfree(rg); 10296822904SAndy Whitcroft } 10396822904SAndy Whitcroft } 10496822904SAndy Whitcroft nrg->from = f; 10596822904SAndy Whitcroft nrg->to = t; 10696822904SAndy Whitcroft return 0; 10796822904SAndy Whitcroft } 10896822904SAndy Whitcroft 10996822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 11096822904SAndy Whitcroft { 11196822904SAndy Whitcroft struct file_region *rg, *nrg; 11296822904SAndy Whitcroft long chg = 0; 11396822904SAndy Whitcroft 11496822904SAndy Whitcroft /* Locate the region we are before or in. */ 11596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 11696822904SAndy Whitcroft if (f <= rg->to) 11796822904SAndy Whitcroft break; 11896822904SAndy Whitcroft 11996822904SAndy Whitcroft /* If we are below the current region then a new region is required. 12096822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 12196822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 12296822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 12396822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 12496822904SAndy Whitcroft if (!nrg) 12596822904SAndy Whitcroft return -ENOMEM; 12696822904SAndy Whitcroft nrg->from = f; 12796822904SAndy Whitcroft nrg->to = f; 12896822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 12996822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 13096822904SAndy Whitcroft 13196822904SAndy Whitcroft return t - f; 13296822904SAndy Whitcroft } 13396822904SAndy Whitcroft 13496822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 13596822904SAndy Whitcroft if (f > rg->from) 13696822904SAndy Whitcroft f = rg->from; 13796822904SAndy Whitcroft chg = t - f; 13896822904SAndy Whitcroft 13996822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 14096822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 14196822904SAndy Whitcroft if (&rg->link == head) 14296822904SAndy Whitcroft break; 14396822904SAndy Whitcroft if (rg->from > t) 14496822904SAndy Whitcroft return chg; 14596822904SAndy Whitcroft 14696822904SAndy Whitcroft /* We overlap with this area, if it extends futher than 14796822904SAndy Whitcroft * us then we must extend ourselves. Account for its 14896822904SAndy Whitcroft * existing reservation. */ 14996822904SAndy Whitcroft if (rg->to > t) { 15096822904SAndy Whitcroft chg += rg->to - t; 15196822904SAndy Whitcroft t = rg->to; 15296822904SAndy Whitcroft } 15396822904SAndy Whitcroft chg -= rg->to - rg->from; 15496822904SAndy Whitcroft } 15596822904SAndy Whitcroft return chg; 15696822904SAndy Whitcroft } 15796822904SAndy Whitcroft 15896822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 15996822904SAndy Whitcroft { 16096822904SAndy Whitcroft struct file_region *rg, *trg; 16196822904SAndy Whitcroft long chg = 0; 16296822904SAndy Whitcroft 16396822904SAndy Whitcroft /* Locate the region we are either in or before. */ 16496822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16596822904SAndy Whitcroft if (end <= rg->to) 16696822904SAndy Whitcroft break; 16796822904SAndy Whitcroft if (&rg->link == head) 16896822904SAndy Whitcroft return 0; 16996822904SAndy Whitcroft 17096822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 17196822904SAndy Whitcroft if (end > rg->from) { 17296822904SAndy Whitcroft chg = rg->to - end; 17396822904SAndy Whitcroft rg->to = end; 17496822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 17596822904SAndy Whitcroft } 17696822904SAndy Whitcroft 17796822904SAndy Whitcroft /* Drop any remaining regions. */ 17896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 17996822904SAndy Whitcroft if (&rg->link == head) 18096822904SAndy Whitcroft break; 18196822904SAndy Whitcroft chg += rg->to - rg->from; 18296822904SAndy Whitcroft list_del(&rg->link); 18396822904SAndy Whitcroft kfree(rg); 18496822904SAndy Whitcroft } 18596822904SAndy Whitcroft return chg; 18696822904SAndy Whitcroft } 18796822904SAndy Whitcroft 18884afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 18984afd99bSAndy Whitcroft { 19084afd99bSAndy Whitcroft struct file_region *rg; 19184afd99bSAndy Whitcroft long chg = 0; 19284afd99bSAndy Whitcroft 19384afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 19484afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 19584afd99bSAndy Whitcroft int seg_from; 19684afd99bSAndy Whitcroft int seg_to; 19784afd99bSAndy Whitcroft 19884afd99bSAndy Whitcroft if (rg->to <= f) 19984afd99bSAndy Whitcroft continue; 20084afd99bSAndy Whitcroft if (rg->from >= t) 20184afd99bSAndy Whitcroft break; 20284afd99bSAndy Whitcroft 20384afd99bSAndy Whitcroft seg_from = max(rg->from, f); 20484afd99bSAndy Whitcroft seg_to = min(rg->to, t); 20584afd99bSAndy Whitcroft 20684afd99bSAndy Whitcroft chg += seg_to - seg_from; 20784afd99bSAndy Whitcroft } 20884afd99bSAndy Whitcroft 20984afd99bSAndy Whitcroft return chg; 21084afd99bSAndy Whitcroft } 21184afd99bSAndy Whitcroft 21296822904SAndy Whitcroft /* 213e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 214e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 215e7c4b0bfSAndy Whitcroft */ 216a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 217a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 218e7c4b0bfSAndy Whitcroft { 219a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 220a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 221e7c4b0bfSAndy Whitcroft } 222e7c4b0bfSAndy Whitcroft 22384afd99bSAndy Whitcroft /* 22408fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 22508fba699SMel Gorman * cases this will be same size as used by the page table entries. 22608fba699SMel Gorman */ 22708fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 22808fba699SMel Gorman { 22908fba699SMel Gorman struct hstate *hstate; 23008fba699SMel Gorman 23108fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 23208fba699SMel Gorman return PAGE_SIZE; 23308fba699SMel Gorman 23408fba699SMel Gorman hstate = hstate_vma(vma); 23508fba699SMel Gorman 23608fba699SMel Gorman return 1UL << (hstate->order + PAGE_SHIFT); 23708fba699SMel Gorman } 238f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 23908fba699SMel Gorman 24008fba699SMel Gorman /* 2413340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 2423340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 2433340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 2443340289dSMel Gorman * function is required. 2453340289dSMel Gorman */ 2463340289dSMel Gorman #ifndef vma_mmu_pagesize 2473340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 2483340289dSMel Gorman { 2493340289dSMel Gorman return vma_kernel_pagesize(vma); 2503340289dSMel Gorman } 2513340289dSMel Gorman #endif 2523340289dSMel Gorman 2533340289dSMel Gorman /* 25484afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 25584afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 25684afd99bSAndy Whitcroft * alignment. 25784afd99bSAndy Whitcroft */ 25884afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 25984afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 26004f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 26184afd99bSAndy Whitcroft 262a1e78772SMel Gorman /* 263a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 264a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 265a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 266a1e78772SMel Gorman * 267a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 268a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 269a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 270a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 27184afd99bSAndy Whitcroft * 27284afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 27384afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 27484afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 27584afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 27684afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 27784afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 27884afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 27984afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 280a1e78772SMel Gorman */ 281e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 282e7c4b0bfSAndy Whitcroft { 283e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 284e7c4b0bfSAndy Whitcroft } 285e7c4b0bfSAndy Whitcroft 286e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 287e7c4b0bfSAndy Whitcroft unsigned long value) 288e7c4b0bfSAndy Whitcroft { 289e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 290e7c4b0bfSAndy Whitcroft } 291e7c4b0bfSAndy Whitcroft 29284afd99bSAndy Whitcroft struct resv_map { 29384afd99bSAndy Whitcroft struct kref refs; 29484afd99bSAndy Whitcroft struct list_head regions; 29584afd99bSAndy Whitcroft }; 29684afd99bSAndy Whitcroft 2972a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 29884afd99bSAndy Whitcroft { 29984afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 30084afd99bSAndy Whitcroft if (!resv_map) 30184afd99bSAndy Whitcroft return NULL; 30284afd99bSAndy Whitcroft 30384afd99bSAndy Whitcroft kref_init(&resv_map->refs); 30484afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 30584afd99bSAndy Whitcroft 30684afd99bSAndy Whitcroft return resv_map; 30784afd99bSAndy Whitcroft } 30884afd99bSAndy Whitcroft 3092a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 31084afd99bSAndy Whitcroft { 31184afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 31284afd99bSAndy Whitcroft 31384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 31484afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 31584afd99bSAndy Whitcroft kfree(resv_map); 31684afd99bSAndy Whitcroft } 31784afd99bSAndy Whitcroft 31884afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 319a1e78772SMel Gorman { 320a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 321f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 32284afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 32384afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 3242a4b3dedSHarvey Harrison return NULL; 325a1e78772SMel Gorman } 326a1e78772SMel Gorman 32784afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 328a1e78772SMel Gorman { 329a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 330f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 331a1e78772SMel Gorman 33284afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 33384afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 33404f2cbe3SMel Gorman } 33504f2cbe3SMel Gorman 33604f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 33704f2cbe3SMel Gorman { 33804f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 339f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 340e7c4b0bfSAndy Whitcroft 341e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 34204f2cbe3SMel Gorman } 34304f2cbe3SMel Gorman 34404f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 34504f2cbe3SMel Gorman { 34604f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 347e7c4b0bfSAndy Whitcroft 348e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 349a1e78772SMel Gorman } 350a1e78772SMel Gorman 351a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 352a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 353a5516438SAndi Kleen struct vm_area_struct *vma) 354a1e78772SMel Gorman { 355c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 356c37f9fb1SAndy Whitcroft return; 357c37f9fb1SAndy Whitcroft 358f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 359a1e78772SMel Gorman /* Shared mappings always use reserves */ 360a5516438SAndi Kleen h->resv_huge_pages--; 36184afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 362a1e78772SMel Gorman /* 363a1e78772SMel Gorman * Only the process that called mmap() has reserves for 364a1e78772SMel Gorman * private mappings. 365a1e78772SMel Gorman */ 366a5516438SAndi Kleen h->resv_huge_pages--; 367a1e78772SMel Gorman } 368a1e78772SMel Gorman } 369a1e78772SMel Gorman 37004f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 371a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 372a1e78772SMel Gorman { 373a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 374f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 375a1e78772SMel Gorman vma->vm_private_data = (void *)0; 376a1e78772SMel Gorman } 377a1e78772SMel Gorman 378a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 3797f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 380a1e78772SMel Gorman { 381f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 382a1e78772SMel Gorman return 1; 3837f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3847f09ca51SMel Gorman return 1; 3857f09ca51SMel Gorman return 0; 386a1e78772SMel Gorman } 387a1e78772SMel Gorman 38869d177c2SAndy Whitcroft static void clear_gigantic_page(struct page *page, 38969d177c2SAndy Whitcroft unsigned long addr, unsigned long sz) 39069d177c2SAndy Whitcroft { 39169d177c2SAndy Whitcroft int i; 39269d177c2SAndy Whitcroft struct page *p = page; 39369d177c2SAndy Whitcroft 39469d177c2SAndy Whitcroft might_sleep(); 39569d177c2SAndy Whitcroft for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) { 39669d177c2SAndy Whitcroft cond_resched(); 39769d177c2SAndy Whitcroft clear_user_highpage(p, addr + i * PAGE_SIZE); 39869d177c2SAndy Whitcroft } 39969d177c2SAndy Whitcroft } 400a5516438SAndi Kleen static void clear_huge_page(struct page *page, 401a5516438SAndi Kleen unsigned long addr, unsigned long sz) 40279ac6ba4SDavid Gibson { 40379ac6ba4SDavid Gibson int i; 40479ac6ba4SDavid Gibson 40574dbdd23SAndrea Arcangeli if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) { 406ebdd4aeaSHannes Eder clear_gigantic_page(page, addr, sz); 407ebdd4aeaSHannes Eder return; 408ebdd4aeaSHannes Eder } 40969d177c2SAndy Whitcroft 41079ac6ba4SDavid Gibson might_sleep(); 411a5516438SAndi Kleen for (i = 0; i < sz/PAGE_SIZE; i++) { 41279ac6ba4SDavid Gibson cond_resched(); 413281e0e3bSRalf Baechle clear_user_highpage(page + i, addr + i * PAGE_SIZE); 41479ac6ba4SDavid Gibson } 41579ac6ba4SDavid Gibson } 41679ac6ba4SDavid Gibson 41769d177c2SAndy Whitcroft static void copy_gigantic_page(struct page *dst, struct page *src, 41869d177c2SAndy Whitcroft unsigned long addr, struct vm_area_struct *vma) 41969d177c2SAndy Whitcroft { 42069d177c2SAndy Whitcroft int i; 42169d177c2SAndy Whitcroft struct hstate *h = hstate_vma(vma); 42269d177c2SAndy Whitcroft struct page *dst_base = dst; 42369d177c2SAndy Whitcroft struct page *src_base = src; 42469d177c2SAndy Whitcroft might_sleep(); 42569d177c2SAndy Whitcroft for (i = 0; i < pages_per_huge_page(h); ) { 42669d177c2SAndy Whitcroft cond_resched(); 42769d177c2SAndy Whitcroft copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 42869d177c2SAndy Whitcroft 42969d177c2SAndy Whitcroft i++; 43069d177c2SAndy Whitcroft dst = mem_map_next(dst, dst_base, i); 43169d177c2SAndy Whitcroft src = mem_map_next(src, src_base, i); 43269d177c2SAndy Whitcroft } 43369d177c2SAndy Whitcroft } 43479ac6ba4SDavid Gibson static void copy_huge_page(struct page *dst, struct page *src, 4359de455b2SAtsushi Nemoto unsigned long addr, struct vm_area_struct *vma) 43679ac6ba4SDavid Gibson { 43779ac6ba4SDavid Gibson int i; 438a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43979ac6ba4SDavid Gibson 440ebdd4aeaSHannes Eder if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 441ebdd4aeaSHannes Eder copy_gigantic_page(dst, src, addr, vma); 442ebdd4aeaSHannes Eder return; 443ebdd4aeaSHannes Eder } 44469d177c2SAndy Whitcroft 44579ac6ba4SDavid Gibson might_sleep(); 446a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 44779ac6ba4SDavid Gibson cond_resched(); 4489de455b2SAtsushi Nemoto copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 44979ac6ba4SDavid Gibson } 45079ac6ba4SDavid Gibson } 45179ac6ba4SDavid Gibson 452a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 4531da177e4SLinus Torvalds { 4541da177e4SLinus Torvalds int nid = page_to_nid(page); 455a5516438SAndi Kleen list_add(&page->lru, &h->hugepage_freelists[nid]); 456a5516438SAndi Kleen h->free_huge_pages++; 457a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 4581da177e4SLinus Torvalds } 4591da177e4SLinus Torvalds 460a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 461a5516438SAndi Kleen struct vm_area_struct *vma, 46204f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 4631da177e4SLinus Torvalds { 46431a5c6e4SNishanth Aravamudan int nid; 4651da177e4SLinus Torvalds struct page *page = NULL; 466480eccf9SLee Schermerhorn struct mempolicy *mpol; 46719770b32SMel Gorman nodemask_t *nodemask; 468396faf03SMel Gorman struct zonelist *zonelist = huge_zonelist(vma, address, 46919770b32SMel Gorman htlb_alloc_mask, &mpol, &nodemask); 470dd1a239fSMel Gorman struct zone *zone; 471dd1a239fSMel Gorman struct zoneref *z; 4721da177e4SLinus Torvalds 473a1e78772SMel Gorman /* 474a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 475a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 476a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 477a1e78772SMel Gorman */ 4787f09ca51SMel Gorman if (!vma_has_reserves(vma) && 479a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 480a1e78772SMel Gorman return NULL; 481a1e78772SMel Gorman 48204f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 483a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 48404f2cbe3SMel Gorman return NULL; 48504f2cbe3SMel Gorman 48619770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 48719770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 48854a6eb5cSMel Gorman nid = zone_to_nid(zone); 48954a6eb5cSMel Gorman if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 490a5516438SAndi Kleen !list_empty(&h->hugepage_freelists[nid])) { 491a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 4921da177e4SLinus Torvalds struct page, lru); 4931da177e4SLinus Torvalds list_del(&page->lru); 494a5516438SAndi Kleen h->free_huge_pages--; 495a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 49604f2cbe3SMel Gorman 49704f2cbe3SMel Gorman if (!avoid_reserve) 498a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 499a1e78772SMel Gorman 5005ab3ee7bSKen Chen break; 5011da177e4SLinus Torvalds } 5023abf7afdSAndrew Morton } 50352cd3b07SLee Schermerhorn mpol_cond_put(mpol); 5041da177e4SLinus Torvalds return page; 5051da177e4SLinus Torvalds } 5061da177e4SLinus Torvalds 507a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5086af2acb6SAdam Litke { 5096af2acb6SAdam Litke int i; 510a5516438SAndi Kleen 51118229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 51218229df5SAndy Whitcroft 513a5516438SAndi Kleen h->nr_huge_pages--; 514a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 515a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 5166af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 5176af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 5186af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 5196af2acb6SAdam Litke } 5206af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5216af2acb6SAdam Litke set_page_refcounted(page); 5227f2e9525SGerald Schaefer arch_release_hugepage(page); 523a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 5246af2acb6SAdam Litke } 5256af2acb6SAdam Litke 526e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 527e5ff2159SAndi Kleen { 528e5ff2159SAndi Kleen struct hstate *h; 529e5ff2159SAndi Kleen 530e5ff2159SAndi Kleen for_each_hstate(h) { 531e5ff2159SAndi Kleen if (huge_page_size(h) == size) 532e5ff2159SAndi Kleen return h; 533e5ff2159SAndi Kleen } 534e5ff2159SAndi Kleen return NULL; 535e5ff2159SAndi Kleen } 536e5ff2159SAndi Kleen 53727a85ef1SDavid Gibson static void free_huge_page(struct page *page) 53827a85ef1SDavid Gibson { 539a5516438SAndi Kleen /* 540a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 541a5516438SAndi Kleen * compound page destructor. 542a5516438SAndi Kleen */ 543e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 5447893d1d5SAdam Litke int nid = page_to_nid(page); 545c79fb75eSAdam Litke struct address_space *mapping; 54627a85ef1SDavid Gibson 547c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 548e5df70abSAndy Whitcroft set_page_private(page, 0); 54923be7468SMel Gorman page->mapping = NULL; 5507893d1d5SAdam Litke BUG_ON(page_count(page)); 55127a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 55227a85ef1SDavid Gibson 55327a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 554aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 555a5516438SAndi Kleen update_and_free_page(h, page); 556a5516438SAndi Kleen h->surplus_huge_pages--; 557a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 5587893d1d5SAdam Litke } else { 559a5516438SAndi Kleen enqueue_huge_page(h, page); 5607893d1d5SAdam Litke } 56127a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 562c79fb75eSAdam Litke if (mapping) 5639a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 56427a85ef1SDavid Gibson } 56527a85ef1SDavid Gibson 566a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 567b7ba30c6SAndi Kleen { 568b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 569b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 570a5516438SAndi Kleen h->nr_huge_pages++; 571a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 572b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 573b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 574b7ba30c6SAndi Kleen } 575b7ba30c6SAndi Kleen 57620a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 57720a0307cSWu Fengguang { 57820a0307cSWu Fengguang int i; 57920a0307cSWu Fengguang int nr_pages = 1 << order; 58020a0307cSWu Fengguang struct page *p = page + 1; 58120a0307cSWu Fengguang 58220a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 58320a0307cSWu Fengguang set_compound_order(page, order); 58420a0307cSWu Fengguang __SetPageHead(page); 58520a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 58620a0307cSWu Fengguang __SetPageTail(p); 58720a0307cSWu Fengguang p->first_page = page; 58820a0307cSWu Fengguang } 58920a0307cSWu Fengguang } 59020a0307cSWu Fengguang 59120a0307cSWu Fengguang int PageHuge(struct page *page) 59220a0307cSWu Fengguang { 59320a0307cSWu Fengguang compound_page_dtor *dtor; 59420a0307cSWu Fengguang 59520a0307cSWu Fengguang if (!PageCompound(page)) 59620a0307cSWu Fengguang return 0; 59720a0307cSWu Fengguang 59820a0307cSWu Fengguang page = compound_head(page); 59920a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 60020a0307cSWu Fengguang 60120a0307cSWu Fengguang return dtor == free_huge_page; 60220a0307cSWu Fengguang } 60320a0307cSWu Fengguang 604a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6051da177e4SLinus Torvalds { 6061da177e4SLinus Torvalds struct page *page; 607f96efd58SJoe Jin 608aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 609aa888a74SAndi Kleen return NULL; 610aa888a74SAndi Kleen 6116484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 612551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 613551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 614a5516438SAndi Kleen huge_page_order(h)); 6151da177e4SLinus Torvalds if (page) { 6167f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 617caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6187b8ee84dSHarvey Harrison return NULL; 6197f2e9525SGerald Schaefer } 620a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6211da177e4SLinus Torvalds } 62263b4613cSNishanth Aravamudan 62363b4613cSNishanth Aravamudan return page; 62463b4613cSNishanth Aravamudan } 62563b4613cSNishanth Aravamudan 6265ced66c9SAndi Kleen /* 6276ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 6286ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 6296ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 6306ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 6316ae11b27SLee Schermerhorn * node for alloc or free. 6329a76db09SLee Schermerhorn */ 6336ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 6349a76db09SLee Schermerhorn { 6356ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 6369a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 6376ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 6389a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 6399a76db09SLee Schermerhorn 6409a76db09SLee Schermerhorn return nid; 6419a76db09SLee Schermerhorn } 6429a76db09SLee Schermerhorn 6436ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 6445ced66c9SAndi Kleen { 6456ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 6466ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 6479a76db09SLee Schermerhorn return nid; 6485ced66c9SAndi Kleen } 6495ced66c9SAndi Kleen 6506ae11b27SLee Schermerhorn /* 6516ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 6526ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 6536ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 6546ae11b27SLee Schermerhorn * mask. 6556ae11b27SLee Schermerhorn */ 6566ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 6576ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 6586ae11b27SLee Schermerhorn { 6596ae11b27SLee Schermerhorn int nid; 6606ae11b27SLee Schermerhorn 6616ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 6626ae11b27SLee Schermerhorn 6636ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 6646ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 6656ae11b27SLee Schermerhorn 6666ae11b27SLee Schermerhorn return nid; 6676ae11b27SLee Schermerhorn } 6686ae11b27SLee Schermerhorn 6696ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 67063b4613cSNishanth Aravamudan { 67163b4613cSNishanth Aravamudan struct page *page; 67263b4613cSNishanth Aravamudan int start_nid; 67363b4613cSNishanth Aravamudan int next_nid; 67463b4613cSNishanth Aravamudan int ret = 0; 67563b4613cSNishanth Aravamudan 6766ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 677e8c5c824SLee Schermerhorn next_nid = start_nid; 67863b4613cSNishanth Aravamudan 67963b4613cSNishanth Aravamudan do { 680e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 6819a76db09SLee Schermerhorn if (page) { 68263b4613cSNishanth Aravamudan ret = 1; 6839a76db09SLee Schermerhorn break; 6849a76db09SLee Schermerhorn } 6856ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 6869a76db09SLee Schermerhorn } while (next_nid != start_nid); 68763b4613cSNishanth Aravamudan 6883b116300SAdam Litke if (ret) 6893b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 6903b116300SAdam Litke else 6913b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 6923b116300SAdam Litke 69363b4613cSNishanth Aravamudan return ret; 6941da177e4SLinus Torvalds } 6951da177e4SLinus Torvalds 696e8c5c824SLee Schermerhorn /* 6976ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 6986ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 6996ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7006ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 701e8c5c824SLee Schermerhorn */ 7026ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 703e8c5c824SLee Schermerhorn { 7046ae11b27SLee Schermerhorn int nid; 7059a76db09SLee Schermerhorn 7066ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7076ae11b27SLee Schermerhorn 7086ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7096ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7106ae11b27SLee Schermerhorn 7119a76db09SLee Schermerhorn return nid; 712e8c5c824SLee Schermerhorn } 713e8c5c824SLee Schermerhorn 714e8c5c824SLee Schermerhorn /* 715e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 716e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 717e8c5c824SLee Schermerhorn * balanced over allowed nodes. 718e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 719e8c5c824SLee Schermerhorn */ 7206ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 7216ae11b27SLee Schermerhorn bool acct_surplus) 722e8c5c824SLee Schermerhorn { 723e8c5c824SLee Schermerhorn int start_nid; 724e8c5c824SLee Schermerhorn int next_nid; 725e8c5c824SLee Schermerhorn int ret = 0; 726e8c5c824SLee Schermerhorn 7276ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 728e8c5c824SLee Schermerhorn next_nid = start_nid; 729e8c5c824SLee Schermerhorn 730e8c5c824SLee Schermerhorn do { 731685f3457SLee Schermerhorn /* 732685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 733685f3457SLee Schermerhorn * nodes with surplus pages. 734685f3457SLee Schermerhorn */ 735685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 736685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 737e8c5c824SLee Schermerhorn struct page *page = 738e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 739e8c5c824SLee Schermerhorn struct page, lru); 740e8c5c824SLee Schermerhorn list_del(&page->lru); 741e8c5c824SLee Schermerhorn h->free_huge_pages--; 742e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 743685f3457SLee Schermerhorn if (acct_surplus) { 744685f3457SLee Schermerhorn h->surplus_huge_pages--; 745685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 746685f3457SLee Schermerhorn } 747e8c5c824SLee Schermerhorn update_and_free_page(h, page); 748e8c5c824SLee Schermerhorn ret = 1; 7499a76db09SLee Schermerhorn break; 750e8c5c824SLee Schermerhorn } 7516ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 7529a76db09SLee Schermerhorn } while (next_nid != start_nid); 753e8c5c824SLee Schermerhorn 754e8c5c824SLee Schermerhorn return ret; 755e8c5c824SLee Schermerhorn } 756e8c5c824SLee Schermerhorn 757a5516438SAndi Kleen static struct page *alloc_buddy_huge_page(struct hstate *h, 758a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 7597893d1d5SAdam Litke { 7607893d1d5SAdam Litke struct page *page; 761d1c3fb1fSNishanth Aravamudan unsigned int nid; 7627893d1d5SAdam Litke 763aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 764aa888a74SAndi Kleen return NULL; 765aa888a74SAndi Kleen 766d1c3fb1fSNishanth Aravamudan /* 767d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 768d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 769d1c3fb1fSNishanth Aravamudan * overcommit 770d1c3fb1fSNishanth Aravamudan * 771d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 772d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 773d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 774d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 775d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 776d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 777d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 778d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 779d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 780d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 781d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 782d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 783d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 784d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 785d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 786d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 787d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 788d1c3fb1fSNishanth Aravamudan */ 789d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 790a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 791d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 792d1c3fb1fSNishanth Aravamudan return NULL; 793d1c3fb1fSNishanth Aravamudan } else { 794a5516438SAndi Kleen h->nr_huge_pages++; 795a5516438SAndi Kleen h->surplus_huge_pages++; 796d1c3fb1fSNishanth Aravamudan } 797d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 798d1c3fb1fSNishanth Aravamudan 799551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 800551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 801a5516438SAndi Kleen huge_page_order(h)); 802d1c3fb1fSNishanth Aravamudan 803caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 804caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 805caff3a2cSGerald Schaefer return NULL; 806caff3a2cSGerald Schaefer } 807caff3a2cSGerald Schaefer 8087893d1d5SAdam Litke spin_lock(&hugetlb_lock); 809d1c3fb1fSNishanth Aravamudan if (page) { 8102668db91SAdam Litke /* 8112668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 8122668db91SAdam Litke * no users -- drop the buddy allocator's reference. 8132668db91SAdam Litke */ 8142668db91SAdam Litke put_page_testzero(page); 8152668db91SAdam Litke VM_BUG_ON(page_count(page)); 816d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 817d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 818d1c3fb1fSNishanth Aravamudan /* 819d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 820d1c3fb1fSNishanth Aravamudan */ 821a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 822a5516438SAndi Kleen h->surplus_huge_pages_node[nid]++; 8233b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 824d1c3fb1fSNishanth Aravamudan } else { 825a5516438SAndi Kleen h->nr_huge_pages--; 826a5516438SAndi Kleen h->surplus_huge_pages--; 8273b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 8287893d1d5SAdam Litke } 829d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 8307893d1d5SAdam Litke 8317893d1d5SAdam Litke return page; 8327893d1d5SAdam Litke } 8337893d1d5SAdam Litke 834e4e574b7SAdam Litke /* 835e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 836e4e574b7SAdam Litke * of size 'delta'. 837e4e574b7SAdam Litke */ 838a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 839e4e574b7SAdam Litke { 840e4e574b7SAdam Litke struct list_head surplus_list; 841e4e574b7SAdam Litke struct page *page, *tmp; 842e4e574b7SAdam Litke int ret, i; 843e4e574b7SAdam Litke int needed, allocated; 844e4e574b7SAdam Litke 845a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 846ac09b3a1SAdam Litke if (needed <= 0) { 847a5516438SAndi Kleen h->resv_huge_pages += delta; 848e4e574b7SAdam Litke return 0; 849ac09b3a1SAdam Litke } 850e4e574b7SAdam Litke 851e4e574b7SAdam Litke allocated = 0; 852e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 853e4e574b7SAdam Litke 854e4e574b7SAdam Litke ret = -ENOMEM; 855e4e574b7SAdam Litke retry: 856e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 857e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 858a5516438SAndi Kleen page = alloc_buddy_huge_page(h, NULL, 0); 859e4e574b7SAdam Litke if (!page) { 860e4e574b7SAdam Litke /* 861e4e574b7SAdam Litke * We were not able to allocate enough pages to 862e4e574b7SAdam Litke * satisfy the entire reservation so we free what 863e4e574b7SAdam Litke * we've allocated so far. 864e4e574b7SAdam Litke */ 865e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 866e4e574b7SAdam Litke needed = 0; 867e4e574b7SAdam Litke goto free; 868e4e574b7SAdam Litke } 869e4e574b7SAdam Litke 870e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 871e4e574b7SAdam Litke } 872e4e574b7SAdam Litke allocated += needed; 873e4e574b7SAdam Litke 874e4e574b7SAdam Litke /* 875e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 876e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 877e4e574b7SAdam Litke */ 878e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 879a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 880a5516438SAndi Kleen (h->free_huge_pages + allocated); 881e4e574b7SAdam Litke if (needed > 0) 882e4e574b7SAdam Litke goto retry; 883e4e574b7SAdam Litke 884e4e574b7SAdam Litke /* 885e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 886e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 887e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 888ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 889ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 890ac09b3a1SAdam Litke * before they are reserved. 891e4e574b7SAdam Litke */ 892e4e574b7SAdam Litke needed += allocated; 893a5516438SAndi Kleen h->resv_huge_pages += delta; 894e4e574b7SAdam Litke ret = 0; 895e4e574b7SAdam Litke free: 89619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 89719fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 89819fc3f0aSAdam Litke if ((--needed) < 0) 89919fc3f0aSAdam Litke break; 90019fc3f0aSAdam Litke list_del(&page->lru); 901a5516438SAndi Kleen enqueue_huge_page(h, page); 90219fc3f0aSAdam Litke } 90319fc3f0aSAdam Litke 90419fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 90519fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 90619fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 907e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 908e4e574b7SAdam Litke list_del(&page->lru); 909af767cbdSAdam Litke /* 9102668db91SAdam Litke * The page has a reference count of zero already, so 9112668db91SAdam Litke * call free_huge_page directly instead of using 9122668db91SAdam Litke * put_page. This must be done with hugetlb_lock 913af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 914af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 915af767cbdSAdam Litke */ 9162668db91SAdam Litke free_huge_page(page); 917af767cbdSAdam Litke } 91819fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 919e4e574b7SAdam Litke } 920e4e574b7SAdam Litke 921e4e574b7SAdam Litke return ret; 922e4e574b7SAdam Litke } 923e4e574b7SAdam Litke 924e4e574b7SAdam Litke /* 925e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 926e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 927e4e574b7SAdam Litke * never used. 928685f3457SLee Schermerhorn * Called with hugetlb_lock held. 929e4e574b7SAdam Litke */ 930a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 931a5516438SAndi Kleen unsigned long unused_resv_pages) 932e4e574b7SAdam Litke { 933e4e574b7SAdam Litke unsigned long nr_pages; 934e4e574b7SAdam Litke 935ac09b3a1SAdam Litke /* Uncommit the reservation */ 936a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 937ac09b3a1SAdam Litke 938aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 939aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 940aa888a74SAndi Kleen return; 941aa888a74SAndi Kleen 942a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 943e4e574b7SAdam Litke 944685f3457SLee Schermerhorn /* 945685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 9469b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 9479b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 9489b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 9499b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 9509b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 951685f3457SLee Schermerhorn */ 952685f3457SLee Schermerhorn while (nr_pages--) { 9539b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 954685f3457SLee Schermerhorn break; 955e4e574b7SAdam Litke } 956e4e574b7SAdam Litke } 957e4e574b7SAdam Litke 958c37f9fb1SAndy Whitcroft /* 959c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 960c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 961c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 962c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 963c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 964c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 965c37f9fb1SAndy Whitcroft * No action is required on failure. 966c37f9fb1SAndy Whitcroft */ 967e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 968a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 969c37f9fb1SAndy Whitcroft { 970c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 971c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 972c37f9fb1SAndy Whitcroft 973f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 974a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 975c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 976c37f9fb1SAndy Whitcroft idx, idx + 1); 977c37f9fb1SAndy Whitcroft 97884afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 979c37f9fb1SAndy Whitcroft return 1; 980c37f9fb1SAndy Whitcroft 98184afd99bSAndy Whitcroft } else { 982e2f17d94SRoel Kluin long err; 983a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 98484afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 98584afd99bSAndy Whitcroft 98684afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 98784afd99bSAndy Whitcroft if (err < 0) 98884afd99bSAndy Whitcroft return err; 989c37f9fb1SAndy Whitcroft return 0; 990c37f9fb1SAndy Whitcroft } 99184afd99bSAndy Whitcroft } 992a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 993a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 994c37f9fb1SAndy Whitcroft { 995c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 996c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 997c37f9fb1SAndy Whitcroft 998f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 999a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1000c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 100184afd99bSAndy Whitcroft 100284afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1003a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 100484afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 100584afd99bSAndy Whitcroft 100684afd99bSAndy Whitcroft /* Mark this page used in the map. */ 100784afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1008c37f9fb1SAndy Whitcroft } 1009c37f9fb1SAndy Whitcroft } 1010c37f9fb1SAndy Whitcroft 1011348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 101204f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1013348ea204SAdam Litke { 1014a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1015348ea204SAdam Litke struct page *page; 10162fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 1017a1e78772SMel Gorman struct inode *inode = mapping->host; 1018e2f17d94SRoel Kluin long chg; 10192fc39cecSAdam Litke 1020a1e78772SMel Gorman /* 1021a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 1022a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 1023a1e78772SMel Gorman * made before satisfying the allocation 1024c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 1025c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 1026a1e78772SMel Gorman */ 1027a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1028c37f9fb1SAndy Whitcroft if (chg < 0) 1029c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 1030c37f9fb1SAndy Whitcroft if (chg) 1031a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 1032a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 103390d8b7e6SAdam Litke 1034a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1035a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1036a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1037a1e78772SMel Gorman 1038a1e78772SMel Gorman if (!page) { 1039a5516438SAndi Kleen page = alloc_buddy_huge_page(h, vma, addr); 1040a1e78772SMel Gorman if (!page) { 1041a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 1042a1e78772SMel Gorman return ERR_PTR(-VM_FAULT_OOM); 1043a1e78772SMel Gorman } 1044a1e78772SMel Gorman } 1045a1e78772SMel Gorman 1046348ea204SAdam Litke set_page_refcounted(page); 10472fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 1048a1e78772SMel Gorman 1049a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1050c37f9fb1SAndy Whitcroft 10517893d1d5SAdam Litke return page; 1052b45b5bd6SDavid Gibson } 1053b45b5bd6SDavid Gibson 105491f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1055aa888a74SAndi Kleen { 1056aa888a74SAndi Kleen struct huge_bootmem_page *m; 10579b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1058aa888a74SAndi Kleen 1059aa888a74SAndi Kleen while (nr_nodes) { 1060aa888a74SAndi Kleen void *addr; 1061aa888a74SAndi Kleen 1062aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 10636ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 10649b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1065aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1066aa888a74SAndi Kleen 1067aa888a74SAndi Kleen if (addr) { 1068aa888a74SAndi Kleen /* 1069aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1070aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1071aa888a74SAndi Kleen * puts them into the mem_map). 1072aa888a74SAndi Kleen */ 1073aa888a74SAndi Kleen m = addr; 1074aa888a74SAndi Kleen goto found; 1075aa888a74SAndi Kleen } 1076aa888a74SAndi Kleen nr_nodes--; 1077aa888a74SAndi Kleen } 1078aa888a74SAndi Kleen return 0; 1079aa888a74SAndi Kleen 1080aa888a74SAndi Kleen found: 1081aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1082aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1083aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1084aa888a74SAndi Kleen m->hstate = h; 1085aa888a74SAndi Kleen return 1; 1086aa888a74SAndi Kleen } 1087aa888a74SAndi Kleen 108818229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 108918229df5SAndy Whitcroft { 109018229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 109118229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 109218229df5SAndy Whitcroft else 109318229df5SAndy Whitcroft prep_compound_page(page, order); 109418229df5SAndy Whitcroft } 109518229df5SAndy Whitcroft 1096aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1097aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1098aa888a74SAndi Kleen { 1099aa888a74SAndi Kleen struct huge_bootmem_page *m; 1100aa888a74SAndi Kleen 1101aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1102aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1103aa888a74SAndi Kleen struct hstate *h = m->hstate; 1104aa888a74SAndi Kleen __ClearPageReserved(page); 1105aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 110618229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1107aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1108aa888a74SAndi Kleen } 1109aa888a74SAndi Kleen } 1110aa888a74SAndi Kleen 11118faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 11121da177e4SLinus Torvalds { 11131da177e4SLinus Torvalds unsigned long i; 11141da177e4SLinus Torvalds 1115e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1116aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1117aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1118aa888a74SAndi Kleen break; 11199b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 11209b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 11211da177e4SLinus Torvalds break; 11221da177e4SLinus Torvalds } 11238faa8b07SAndi Kleen h->max_huge_pages = i; 1124e5ff2159SAndi Kleen } 1125e5ff2159SAndi Kleen 1126e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1127e5ff2159SAndi Kleen { 1128e5ff2159SAndi Kleen struct hstate *h; 1129e5ff2159SAndi Kleen 1130e5ff2159SAndi Kleen for_each_hstate(h) { 11318faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 11328faa8b07SAndi Kleen if (h->order < MAX_ORDER) 11338faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1134e5ff2159SAndi Kleen } 1135e5ff2159SAndi Kleen } 1136e5ff2159SAndi Kleen 11374abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 11384abd32dbSAndi Kleen { 11394abd32dbSAndi Kleen if (n >= (1UL << 30)) 11404abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 11414abd32dbSAndi Kleen else if (n >= (1UL << 20)) 11424abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 11434abd32dbSAndi Kleen else 11444abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 11454abd32dbSAndi Kleen return buf; 11464abd32dbSAndi Kleen } 11474abd32dbSAndi Kleen 1148e5ff2159SAndi Kleen static void __init report_hugepages(void) 1149e5ff2159SAndi Kleen { 1150e5ff2159SAndi Kleen struct hstate *h; 1151e5ff2159SAndi Kleen 1152e5ff2159SAndi Kleen for_each_hstate(h) { 11534abd32dbSAndi Kleen char buf[32]; 11544abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11554abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11564abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11574abd32dbSAndi Kleen h->free_huge_pages); 1158e5ff2159SAndi Kleen } 1159e5ff2159SAndi Kleen } 1160e5ff2159SAndi Kleen 11611da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 11626ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 11636ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11641da177e4SLinus Torvalds { 11654415cc8dSChristoph Lameter int i; 11664415cc8dSChristoph Lameter 1167aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1168aa888a74SAndi Kleen return; 1169aa888a74SAndi Kleen 11706ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 11711da177e4SLinus Torvalds struct page *page, *next; 1172a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1173a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1174a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11756b0c880dSAdam Litke return; 11761da177e4SLinus Torvalds if (PageHighMem(page)) 11771da177e4SLinus Torvalds continue; 11781da177e4SLinus Torvalds list_del(&page->lru); 1179e5ff2159SAndi Kleen update_and_free_page(h, page); 1180a5516438SAndi Kleen h->free_huge_pages--; 1181a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11821da177e4SLinus Torvalds } 11831da177e4SLinus Torvalds } 11841da177e4SLinus Torvalds } 11851da177e4SLinus Torvalds #else 11866ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 11876ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11881da177e4SLinus Torvalds { 11891da177e4SLinus Torvalds } 11901da177e4SLinus Torvalds #endif 11911da177e4SLinus Torvalds 119220a0307cSWu Fengguang /* 119320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 119420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 119520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 119620a0307cSWu Fengguang */ 11976ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 11986ae11b27SLee Schermerhorn int delta) 119920a0307cSWu Fengguang { 1200e8c5c824SLee Schermerhorn int start_nid, next_nid; 120120a0307cSWu Fengguang int ret = 0; 120220a0307cSWu Fengguang 120320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 120420a0307cSWu Fengguang 1205e8c5c824SLee Schermerhorn if (delta < 0) 12066ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1207e8c5c824SLee Schermerhorn else 12086ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1209e8c5c824SLee Schermerhorn next_nid = start_nid; 1210e8c5c824SLee Schermerhorn 1211e8c5c824SLee Schermerhorn do { 1212e8c5c824SLee Schermerhorn int nid = next_nid; 1213e8c5c824SLee Schermerhorn if (delta < 0) { 1214e8c5c824SLee Schermerhorn /* 1215e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1216e8c5c824SLee Schermerhorn */ 12179a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 12186ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 12196ae11b27SLee Schermerhorn nodes_allowed); 122020a0307cSWu Fengguang continue; 1221e8c5c824SLee Schermerhorn } 12229a76db09SLee Schermerhorn } 1223e8c5c824SLee Schermerhorn if (delta > 0) { 1224e8c5c824SLee Schermerhorn /* 1225e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1226e8c5c824SLee Schermerhorn */ 1227e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 12289a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 12296ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 12306ae11b27SLee Schermerhorn nodes_allowed); 123120a0307cSWu Fengguang continue; 1232e8c5c824SLee Schermerhorn } 12339a76db09SLee Schermerhorn } 123420a0307cSWu Fengguang 123520a0307cSWu Fengguang h->surplus_huge_pages += delta; 123620a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 123720a0307cSWu Fengguang ret = 1; 123820a0307cSWu Fengguang break; 1239e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 124020a0307cSWu Fengguang 124120a0307cSWu Fengguang return ret; 124220a0307cSWu Fengguang } 124320a0307cSWu Fengguang 1244a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 12456ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 12466ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12471da177e4SLinus Torvalds { 12487893d1d5SAdam Litke unsigned long min_count, ret; 12491da177e4SLinus Torvalds 1250aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1251aa888a74SAndi Kleen return h->max_huge_pages; 1252aa888a74SAndi Kleen 12537893d1d5SAdam Litke /* 12547893d1d5SAdam Litke * Increase the pool size 12557893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 12567893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1257d1c3fb1fSNishanth Aravamudan * 1258d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1259d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1260d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1261d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1262d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 12637893d1d5SAdam Litke */ 12641da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1265a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 12666ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 12677893d1d5SAdam Litke break; 12687893d1d5SAdam Litke } 12697893d1d5SAdam Litke 1270a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 12717893d1d5SAdam Litke /* 12727893d1d5SAdam Litke * If this allocation races such that we no longer need the 12737893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12747893d1d5SAdam Litke * and reducing the surplus. 12757893d1d5SAdam Litke */ 12767893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 12776ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 12787893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12797893d1d5SAdam Litke if (!ret) 12807893d1d5SAdam Litke goto out; 12817893d1d5SAdam Litke 1282536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1283536240f2SMel Gorman if (signal_pending(current)) 1284536240f2SMel Gorman goto out; 12857893d1d5SAdam Litke } 12867893d1d5SAdam Litke 12877893d1d5SAdam Litke /* 12887893d1d5SAdam Litke * Decrease the pool size 12897893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 12907893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 12917893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 12927893d1d5SAdam Litke * to the desired size as pages become free. 1293d1c3fb1fSNishanth Aravamudan * 1294d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1295d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1296d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1297d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1298d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1299d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1300d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 13017893d1d5SAdam Litke */ 1302a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 13036b0c880dSAdam Litke min_count = max(count, min_count); 13046ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1305a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 13066ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 13071da177e4SLinus Torvalds break; 13081da177e4SLinus Torvalds } 1309a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 13106ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 13117893d1d5SAdam Litke break; 13127893d1d5SAdam Litke } 13137893d1d5SAdam Litke out: 1314a5516438SAndi Kleen ret = persistent_huge_pages(h); 13151da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 13167893d1d5SAdam Litke return ret; 13171da177e4SLinus Torvalds } 13181da177e4SLinus Torvalds 1319a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1320a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1321a3437870SNishanth Aravamudan 1322a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1323a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1324a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1325a3437870SNishanth Aravamudan 1326a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1327a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1328a3437870SNishanth Aravamudan 13299a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 13309a305230SLee Schermerhorn 13319a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1332a3437870SNishanth Aravamudan { 1333a3437870SNishanth Aravamudan int i; 13349a305230SLee Schermerhorn 1335a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 13369a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 13379a305230SLee Schermerhorn if (nidp) 13389a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1339a3437870SNishanth Aravamudan return &hstates[i]; 13409a305230SLee Schermerhorn } 13419a305230SLee Schermerhorn 13429a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1343a3437870SNishanth Aravamudan } 1344a3437870SNishanth Aravamudan 134506808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1346a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1347a3437870SNishanth Aravamudan { 13489a305230SLee Schermerhorn struct hstate *h; 13499a305230SLee Schermerhorn unsigned long nr_huge_pages; 13509a305230SLee Schermerhorn int nid; 13519a305230SLee Schermerhorn 13529a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13539a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 13549a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 13559a305230SLee Schermerhorn else 13569a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 13579a305230SLee Schermerhorn 13589a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1359a3437870SNishanth Aravamudan } 136006808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 136106808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 136206808b08SLee Schermerhorn const char *buf, size_t len) 1363a3437870SNishanth Aravamudan { 1364a3437870SNishanth Aravamudan int err; 13659a305230SLee Schermerhorn int nid; 136606808b08SLee Schermerhorn unsigned long count; 13679a305230SLee Schermerhorn struct hstate *h; 1368bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1369a3437870SNishanth Aravamudan 137006808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 1371a3437870SNishanth Aravamudan if (err) 1372a3437870SNishanth Aravamudan return 0; 1373a3437870SNishanth Aravamudan 13749a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13759a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 13769a305230SLee Schermerhorn /* 13779a305230SLee Schermerhorn * global hstate attribute 13789a305230SLee Schermerhorn */ 13799a305230SLee Schermerhorn if (!(obey_mempolicy && 13809a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 138106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 13829a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 138306808b08SLee Schermerhorn } 13849a305230SLee Schermerhorn } else if (nodes_allowed) { 13859a305230SLee Schermerhorn /* 13869a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 13879a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 13889a305230SLee Schermerhorn */ 13899a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 13909a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 13919a305230SLee Schermerhorn } else 13929a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 13939a305230SLee Schermerhorn 139406808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1395a3437870SNishanth Aravamudan 13969b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 139706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 139806808b08SLee Schermerhorn 139906808b08SLee Schermerhorn return len; 140006808b08SLee Schermerhorn } 140106808b08SLee Schermerhorn 140206808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 140306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 140406808b08SLee Schermerhorn { 140506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 140606808b08SLee Schermerhorn } 140706808b08SLee Schermerhorn 140806808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 140906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 141006808b08SLee Schermerhorn { 141106808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1412a3437870SNishanth Aravamudan } 1413a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1414a3437870SNishanth Aravamudan 141506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 141606808b08SLee Schermerhorn 141706808b08SLee Schermerhorn /* 141806808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 141906808b08SLee Schermerhorn * huge page alloc/free. 142006808b08SLee Schermerhorn */ 142106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 142206808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 142306808b08SLee Schermerhorn { 142406808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 142506808b08SLee Schermerhorn } 142606808b08SLee Schermerhorn 142706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 142806808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 142906808b08SLee Schermerhorn { 143006808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 143106808b08SLee Schermerhorn } 143206808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 143306808b08SLee Schermerhorn #endif 143406808b08SLee Schermerhorn 143506808b08SLee Schermerhorn 1436a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1437a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1438a3437870SNishanth Aravamudan { 14399a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1440a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1441a3437870SNishanth Aravamudan } 1442a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1443a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1444a3437870SNishanth Aravamudan { 1445a3437870SNishanth Aravamudan int err; 1446a3437870SNishanth Aravamudan unsigned long input; 14479a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1448a3437870SNishanth Aravamudan 1449a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1450a3437870SNishanth Aravamudan if (err) 1451a3437870SNishanth Aravamudan return 0; 1452a3437870SNishanth Aravamudan 1453a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1454a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1455a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1456a3437870SNishanth Aravamudan 1457a3437870SNishanth Aravamudan return count; 1458a3437870SNishanth Aravamudan } 1459a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1460a3437870SNishanth Aravamudan 1461a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1462a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1463a3437870SNishanth Aravamudan { 14649a305230SLee Schermerhorn struct hstate *h; 14659a305230SLee Schermerhorn unsigned long free_huge_pages; 14669a305230SLee Schermerhorn int nid; 14679a305230SLee Schermerhorn 14689a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14699a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14709a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 14719a305230SLee Schermerhorn else 14729a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 14739a305230SLee Schermerhorn 14749a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1475a3437870SNishanth Aravamudan } 1476a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1477a3437870SNishanth Aravamudan 1478a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1479a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1480a3437870SNishanth Aravamudan { 14819a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1482a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1483a3437870SNishanth Aravamudan } 1484a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1485a3437870SNishanth Aravamudan 1486a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1487a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1488a3437870SNishanth Aravamudan { 14899a305230SLee Schermerhorn struct hstate *h; 14909a305230SLee Schermerhorn unsigned long surplus_huge_pages; 14919a305230SLee Schermerhorn int nid; 14929a305230SLee Schermerhorn 14939a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14949a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14959a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 14969a305230SLee Schermerhorn else 14979a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 14989a305230SLee Schermerhorn 14999a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1500a3437870SNishanth Aravamudan } 1501a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1502a3437870SNishanth Aravamudan 1503a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1504a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1505a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1506a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1507a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1508a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 150906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 151006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 151106808b08SLee Schermerhorn #endif 1512a3437870SNishanth Aravamudan NULL, 1513a3437870SNishanth Aravamudan }; 1514a3437870SNishanth Aravamudan 1515a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1516a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1517a3437870SNishanth Aravamudan }; 1518a3437870SNishanth Aravamudan 1519094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 15209a305230SLee Schermerhorn struct kobject **hstate_kobjs, 15219a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1522a3437870SNishanth Aravamudan { 1523a3437870SNishanth Aravamudan int retval; 15249a305230SLee Schermerhorn int hi = h - hstates; 1525a3437870SNishanth Aravamudan 15269a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 15279a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1528a3437870SNishanth Aravamudan return -ENOMEM; 1529a3437870SNishanth Aravamudan 15309a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1531a3437870SNishanth Aravamudan if (retval) 15329a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1533a3437870SNishanth Aravamudan 1534a3437870SNishanth Aravamudan return retval; 1535a3437870SNishanth Aravamudan } 1536a3437870SNishanth Aravamudan 1537a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1538a3437870SNishanth Aravamudan { 1539a3437870SNishanth Aravamudan struct hstate *h; 1540a3437870SNishanth Aravamudan int err; 1541a3437870SNishanth Aravamudan 1542a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1543a3437870SNishanth Aravamudan if (!hugepages_kobj) 1544a3437870SNishanth Aravamudan return; 1545a3437870SNishanth Aravamudan 1546a3437870SNishanth Aravamudan for_each_hstate(h) { 15479a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 15489a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1549a3437870SNishanth Aravamudan if (err) 1550a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1551a3437870SNishanth Aravamudan h->name); 1552a3437870SNishanth Aravamudan } 1553a3437870SNishanth Aravamudan } 1554a3437870SNishanth Aravamudan 15559a305230SLee Schermerhorn #ifdef CONFIG_NUMA 15569a305230SLee Schermerhorn 15579a305230SLee Schermerhorn /* 15589a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 15599a305230SLee Schermerhorn * with node sysdevs in node_devices[] using a parallel array. The array 15609a305230SLee Schermerhorn * index of a node sysdev or _hstate == node id. 15619a305230SLee Schermerhorn * This is here to avoid any static dependency of the node sysdev driver, in 15629a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 15639a305230SLee Schermerhorn */ 15649a305230SLee Schermerhorn struct node_hstate { 15659a305230SLee Schermerhorn struct kobject *hugepages_kobj; 15669a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 15679a305230SLee Schermerhorn }; 15689a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 15699a305230SLee Schermerhorn 15709a305230SLee Schermerhorn /* 15719a305230SLee Schermerhorn * A subset of global hstate attributes for node sysdevs 15729a305230SLee Schermerhorn */ 15739a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 15749a305230SLee Schermerhorn &nr_hugepages_attr.attr, 15759a305230SLee Schermerhorn &free_hugepages_attr.attr, 15769a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 15779a305230SLee Schermerhorn NULL, 15789a305230SLee Schermerhorn }; 15799a305230SLee Schermerhorn 15809a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 15819a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 15829a305230SLee Schermerhorn }; 15839a305230SLee Schermerhorn 15849a305230SLee Schermerhorn /* 15859a305230SLee Schermerhorn * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. 15869a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 15879a305230SLee Schermerhorn */ 15889a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 15899a305230SLee Schermerhorn { 15909a305230SLee Schermerhorn int nid; 15919a305230SLee Schermerhorn 15929a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 15939a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 15949a305230SLee Schermerhorn int i; 15959a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 15969a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 15979a305230SLee Schermerhorn if (nidp) 15989a305230SLee Schermerhorn *nidp = nid; 15999a305230SLee Schermerhorn return &hstates[i]; 16009a305230SLee Schermerhorn } 16019a305230SLee Schermerhorn } 16029a305230SLee Schermerhorn 16039a305230SLee Schermerhorn BUG(); 16049a305230SLee Schermerhorn return NULL; 16059a305230SLee Schermerhorn } 16069a305230SLee Schermerhorn 16079a305230SLee Schermerhorn /* 16089a305230SLee Schermerhorn * Unregister hstate attributes from a single node sysdev. 16099a305230SLee Schermerhorn * No-op if no hstate attributes attached. 16109a305230SLee Schermerhorn */ 16119a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 16129a305230SLee Schermerhorn { 16139a305230SLee Schermerhorn struct hstate *h; 16149a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16159a305230SLee Schermerhorn 16169a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16179b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 16189a305230SLee Schermerhorn 16199a305230SLee Schermerhorn for_each_hstate(h) 16209a305230SLee Schermerhorn if (nhs->hstate_kobjs[h - hstates]) { 16219a305230SLee Schermerhorn kobject_put(nhs->hstate_kobjs[h - hstates]); 16229a305230SLee Schermerhorn nhs->hstate_kobjs[h - hstates] = NULL; 16239a305230SLee Schermerhorn } 16249a305230SLee Schermerhorn 16259a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 16269a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 16279a305230SLee Schermerhorn } 16289a305230SLee Schermerhorn 16299a305230SLee Schermerhorn /* 16309a305230SLee Schermerhorn * hugetlb module exit: unregister hstate attributes from node sysdevs 16319a305230SLee Schermerhorn * that have them. 16329a305230SLee Schermerhorn */ 16339a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 16349a305230SLee Schermerhorn { 16359a305230SLee Schermerhorn int nid; 16369a305230SLee Schermerhorn 16379a305230SLee Schermerhorn /* 16389a305230SLee Schermerhorn * disable node sysdev registrations. 16399a305230SLee Schermerhorn */ 16409a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 16419a305230SLee Schermerhorn 16429a305230SLee Schermerhorn /* 16439a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 16449a305230SLee Schermerhorn */ 16459a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 16469a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 16479a305230SLee Schermerhorn } 16489a305230SLee Schermerhorn 16499a305230SLee Schermerhorn /* 16509a305230SLee Schermerhorn * Register hstate attributes for a single node sysdev. 16519a305230SLee Schermerhorn * No-op if attributes already registered. 16529a305230SLee Schermerhorn */ 16539a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 16549a305230SLee Schermerhorn { 16559a305230SLee Schermerhorn struct hstate *h; 16569a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16579a305230SLee Schermerhorn int err; 16589a305230SLee Schermerhorn 16599a305230SLee Schermerhorn if (nhs->hugepages_kobj) 16609a305230SLee Schermerhorn return; /* already allocated */ 16619a305230SLee Schermerhorn 16629a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 16639a305230SLee Schermerhorn &node->sysdev.kobj); 16649a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16659a305230SLee Schermerhorn return; 16669a305230SLee Schermerhorn 16679a305230SLee Schermerhorn for_each_hstate(h) { 16689a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 16699a305230SLee Schermerhorn nhs->hstate_kobjs, 16709a305230SLee Schermerhorn &per_node_hstate_attr_group); 16719a305230SLee Schermerhorn if (err) { 16729a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 16739a305230SLee Schermerhorn " for node %d\n", 16749a305230SLee Schermerhorn h->name, node->sysdev.id); 16759a305230SLee Schermerhorn hugetlb_unregister_node(node); 16769a305230SLee Schermerhorn break; 16779a305230SLee Schermerhorn } 16789a305230SLee Schermerhorn } 16799a305230SLee Schermerhorn } 16809a305230SLee Schermerhorn 16819a305230SLee Schermerhorn /* 16829b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 16839b5e5d0fSLee Schermerhorn * sysdevs of nodes that have memory. All on-line nodes should have 16849b5e5d0fSLee Schermerhorn * registered their associated sysdev by this time. 16859a305230SLee Schermerhorn */ 16869a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 16879a305230SLee Schermerhorn { 16889a305230SLee Schermerhorn int nid; 16899a305230SLee Schermerhorn 16909b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 16919a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 16929a305230SLee Schermerhorn if (node->sysdev.id == nid) 16939a305230SLee Schermerhorn hugetlb_register_node(node); 16949a305230SLee Schermerhorn } 16959a305230SLee Schermerhorn 16969a305230SLee Schermerhorn /* 16979a305230SLee Schermerhorn * Let the node sysdev driver know we're here so it can 16989a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 16999a305230SLee Schermerhorn */ 17009a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 17019a305230SLee Schermerhorn hugetlb_unregister_node); 17029a305230SLee Schermerhorn } 17039a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 17049a305230SLee Schermerhorn 17059a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17069a305230SLee Schermerhorn { 17079a305230SLee Schermerhorn BUG(); 17089a305230SLee Schermerhorn if (nidp) 17099a305230SLee Schermerhorn *nidp = -1; 17109a305230SLee Schermerhorn return NULL; 17119a305230SLee Schermerhorn } 17129a305230SLee Schermerhorn 17139a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 17149a305230SLee Schermerhorn 17159a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 17169a305230SLee Schermerhorn 17179a305230SLee Schermerhorn #endif 17189a305230SLee Schermerhorn 1719a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1720a3437870SNishanth Aravamudan { 1721a3437870SNishanth Aravamudan struct hstate *h; 1722a3437870SNishanth Aravamudan 17239a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 17249a305230SLee Schermerhorn 1725a3437870SNishanth Aravamudan for_each_hstate(h) { 1726a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1727a3437870SNishanth Aravamudan } 1728a3437870SNishanth Aravamudan 1729a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1730a3437870SNishanth Aravamudan } 1731a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1732a3437870SNishanth Aravamudan 1733a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1734a3437870SNishanth Aravamudan { 17350ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 17360ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 17370ef89d25SBenjamin Herrenschmidt * there is no such support 17380ef89d25SBenjamin Herrenschmidt */ 17390ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 17400ef89d25SBenjamin Herrenschmidt return 0; 1741a3437870SNishanth Aravamudan 1742e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1743e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1744e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1745a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1746a3437870SNishanth Aravamudan } 1747e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1748e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1749e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1750a3437870SNishanth Aravamudan 1751a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1752a3437870SNishanth Aravamudan 1753aa888a74SAndi Kleen gather_bootmem_prealloc(); 1754aa888a74SAndi Kleen 1755a3437870SNishanth Aravamudan report_hugepages(); 1756a3437870SNishanth Aravamudan 1757a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1758a3437870SNishanth Aravamudan 17599a305230SLee Schermerhorn hugetlb_register_all_nodes(); 17609a305230SLee Schermerhorn 1761a3437870SNishanth Aravamudan return 0; 1762a3437870SNishanth Aravamudan } 1763a3437870SNishanth Aravamudan module_init(hugetlb_init); 1764a3437870SNishanth Aravamudan 1765a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1766a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1767a3437870SNishanth Aravamudan { 1768a3437870SNishanth Aravamudan struct hstate *h; 17698faa8b07SAndi Kleen unsigned long i; 17708faa8b07SAndi Kleen 1771a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1772a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1773a3437870SNishanth Aravamudan return; 1774a3437870SNishanth Aravamudan } 1775a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1776a3437870SNishanth Aravamudan BUG_ON(order == 0); 1777a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1778a3437870SNishanth Aravamudan h->order = order; 1779a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 17808faa8b07SAndi Kleen h->nr_huge_pages = 0; 17818faa8b07SAndi Kleen h->free_huge_pages = 0; 17828faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 17838faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 17849b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 17859b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1786a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1787a3437870SNishanth Aravamudan huge_page_size(h)/1024); 17888faa8b07SAndi Kleen 1789a3437870SNishanth Aravamudan parsed_hstate = h; 1790a3437870SNishanth Aravamudan } 1791a3437870SNishanth Aravamudan 1792e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1793a3437870SNishanth Aravamudan { 1794a3437870SNishanth Aravamudan unsigned long *mhp; 17958faa8b07SAndi Kleen static unsigned long *last_mhp; 1796a3437870SNishanth Aravamudan 1797a3437870SNishanth Aravamudan /* 1798a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1799a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1800a3437870SNishanth Aravamudan */ 1801a3437870SNishanth Aravamudan if (!max_hstate) 1802a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1803a3437870SNishanth Aravamudan else 1804a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1805a3437870SNishanth Aravamudan 18068faa8b07SAndi Kleen if (mhp == last_mhp) { 18078faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 18088faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 18098faa8b07SAndi Kleen return 1; 18108faa8b07SAndi Kleen } 18118faa8b07SAndi Kleen 1812a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1813a3437870SNishanth Aravamudan *mhp = 0; 1814a3437870SNishanth Aravamudan 18158faa8b07SAndi Kleen /* 18168faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 18178faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 18188faa8b07SAndi Kleen * use the bootmem allocator. 18198faa8b07SAndi Kleen */ 18208faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 18218faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 18228faa8b07SAndi Kleen 18238faa8b07SAndi Kleen last_mhp = mhp; 18248faa8b07SAndi Kleen 1825a3437870SNishanth Aravamudan return 1; 1826a3437870SNishanth Aravamudan } 1827e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1828e11bfbfcSNick Piggin 1829e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1830e11bfbfcSNick Piggin { 1831e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1832e11bfbfcSNick Piggin return 1; 1833e11bfbfcSNick Piggin } 1834e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1835a3437870SNishanth Aravamudan 18368a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 18378a213460SNishanth Aravamudan { 18388a213460SNishanth Aravamudan int node; 18398a213460SNishanth Aravamudan unsigned int nr = 0; 18408a213460SNishanth Aravamudan 18418a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 18428a213460SNishanth Aravamudan nr += array[node]; 18438a213460SNishanth Aravamudan 18448a213460SNishanth Aravamudan return nr; 18458a213460SNishanth Aravamudan } 18468a213460SNishanth Aravamudan 18478a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 184806808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 184906808b08SLee Schermerhorn struct ctl_table *table, int write, 185006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 18511da177e4SLinus Torvalds { 1852e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1853e5ff2159SAndi Kleen unsigned long tmp; 1854e5ff2159SAndi Kleen 1855e5ff2159SAndi Kleen if (!write) 1856e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1857e5ff2159SAndi Kleen 1858e5ff2159SAndi Kleen table->data = &tmp; 1859e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 18608d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1861e5ff2159SAndi Kleen 186206808b08SLee Schermerhorn if (write) { 1863bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 1864bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 186506808b08SLee Schermerhorn if (!(obey_mempolicy && 186606808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 186706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 186806808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 186906808b08SLee Schermerhorn } 187006808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 187106808b08SLee Schermerhorn 187206808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 187306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 187406808b08SLee Schermerhorn } 1875e5ff2159SAndi Kleen 18761da177e4SLinus Torvalds return 0; 18771da177e4SLinus Torvalds } 1878396faf03SMel Gorman 187906808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 188006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 188106808b08SLee Schermerhorn { 188206808b08SLee Schermerhorn 188306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 188406808b08SLee Schermerhorn buffer, length, ppos); 188506808b08SLee Schermerhorn } 188606808b08SLee Schermerhorn 188706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 188806808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 188906808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 189006808b08SLee Schermerhorn { 189106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 189206808b08SLee Schermerhorn buffer, length, ppos); 189306808b08SLee Schermerhorn } 189406808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 189506808b08SLee Schermerhorn 1896396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 18978d65af78SAlexey Dobriyan void __user *buffer, 1898396faf03SMel Gorman size_t *length, loff_t *ppos) 1899396faf03SMel Gorman { 19008d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 1901396faf03SMel Gorman if (hugepages_treat_as_movable) 1902396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1903396faf03SMel Gorman else 1904396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1905396faf03SMel Gorman return 0; 1906396faf03SMel Gorman } 1907396faf03SMel Gorman 1908a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 19098d65af78SAlexey Dobriyan void __user *buffer, 1910a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1911a3d0c6aaSNishanth Aravamudan { 1912a5516438SAndi Kleen struct hstate *h = &default_hstate; 1913e5ff2159SAndi Kleen unsigned long tmp; 1914e5ff2159SAndi Kleen 1915e5ff2159SAndi Kleen if (!write) 1916e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1917e5ff2159SAndi Kleen 1918e5ff2159SAndi Kleen table->data = &tmp; 1919e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 19208d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1921e5ff2159SAndi Kleen 1922e5ff2159SAndi Kleen if (write) { 1923064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1924e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1925a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1926e5ff2159SAndi Kleen } 1927e5ff2159SAndi Kleen 1928a3d0c6aaSNishanth Aravamudan return 0; 1929a3d0c6aaSNishanth Aravamudan } 1930a3d0c6aaSNishanth Aravamudan 19311da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 19321da177e4SLinus Torvalds 1933e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 19341da177e4SLinus Torvalds { 1935a5516438SAndi Kleen struct hstate *h = &default_hstate; 1936e1759c21SAlexey Dobriyan seq_printf(m, 19371da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 19381da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1939b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 19407893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 19414f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1942a5516438SAndi Kleen h->nr_huge_pages, 1943a5516438SAndi Kleen h->free_huge_pages, 1944a5516438SAndi Kleen h->resv_huge_pages, 1945a5516438SAndi Kleen h->surplus_huge_pages, 1946a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 19471da177e4SLinus Torvalds } 19481da177e4SLinus Torvalds 19491da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 19501da177e4SLinus Torvalds { 1951a5516438SAndi Kleen struct hstate *h = &default_hstate; 19521da177e4SLinus Torvalds return sprintf(buf, 19531da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1954a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1955a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1956a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1957a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1958a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 19591da177e4SLinus Torvalds } 19601da177e4SLinus Torvalds 19611da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 19621da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 19631da177e4SLinus Torvalds { 1964a5516438SAndi Kleen struct hstate *h = &default_hstate; 1965a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 19661da177e4SLinus Torvalds } 19671da177e4SLinus Torvalds 1968a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 1969fc1b8a73SMel Gorman { 1970fc1b8a73SMel Gorman int ret = -ENOMEM; 1971fc1b8a73SMel Gorman 1972fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 1973fc1b8a73SMel Gorman /* 1974fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 1975fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 1976fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 1977fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 1978fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 1979fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 1980fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 1981fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 1982fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 1983fc1b8a73SMel Gorman * 1984fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 1985fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 1986fc1b8a73SMel Gorman * we fall back to check against current free page availability as 1987fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 1988fc1b8a73SMel Gorman * semantics that cpuset has. 1989fc1b8a73SMel Gorman */ 1990fc1b8a73SMel Gorman if (delta > 0) { 1991a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 1992fc1b8a73SMel Gorman goto out; 1993fc1b8a73SMel Gorman 1994a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 1995a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 1996fc1b8a73SMel Gorman goto out; 1997fc1b8a73SMel Gorman } 1998fc1b8a73SMel Gorman } 1999fc1b8a73SMel Gorman 2000fc1b8a73SMel Gorman ret = 0; 2001fc1b8a73SMel Gorman if (delta < 0) 2002a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2003fc1b8a73SMel Gorman 2004fc1b8a73SMel Gorman out: 2005fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2006fc1b8a73SMel Gorman return ret; 2007fc1b8a73SMel Gorman } 2008fc1b8a73SMel Gorman 200984afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 201084afd99bSAndy Whitcroft { 201184afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 201284afd99bSAndy Whitcroft 201384afd99bSAndy Whitcroft /* 201484afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 201584afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 201684afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 201784afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 201884afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 201984afd99bSAndy Whitcroft * new reference here without additional locking. 202084afd99bSAndy Whitcroft */ 202184afd99bSAndy Whitcroft if (reservations) 202284afd99bSAndy Whitcroft kref_get(&reservations->refs); 202384afd99bSAndy Whitcroft } 202484afd99bSAndy Whitcroft 2025a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2026a1e78772SMel Gorman { 2027a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 202884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 202984afd99bSAndy Whitcroft unsigned long reserve; 203084afd99bSAndy Whitcroft unsigned long start; 203184afd99bSAndy Whitcroft unsigned long end; 203284afd99bSAndy Whitcroft 203384afd99bSAndy Whitcroft if (reservations) { 2034a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2035a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 203684afd99bSAndy Whitcroft 203784afd99bSAndy Whitcroft reserve = (end - start) - 203884afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 203984afd99bSAndy Whitcroft 204084afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 204184afd99bSAndy Whitcroft 20427251ff78SAdam Litke if (reserve) { 2043a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 20447251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 20457251ff78SAdam Litke } 2046a1e78772SMel Gorman } 204784afd99bSAndy Whitcroft } 2048a1e78772SMel Gorman 20491da177e4SLinus Torvalds /* 20501da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 20511da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 20521da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 20531da177e4SLinus Torvalds * this far. 20541da177e4SLinus Torvalds */ 2055d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 20561da177e4SLinus Torvalds { 20571da177e4SLinus Torvalds BUG(); 2058d0217ac0SNick Piggin return 0; 20591da177e4SLinus Torvalds } 20601da177e4SLinus Torvalds 2061f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2062d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 206384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2064a1e78772SMel Gorman .close = hugetlb_vm_op_close, 20651da177e4SLinus Torvalds }; 20661da177e4SLinus Torvalds 20671e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 20681e8f889bSDavid Gibson int writable) 206963551ae0SDavid Gibson { 207063551ae0SDavid Gibson pte_t entry; 207163551ae0SDavid Gibson 20721e8f889bSDavid Gibson if (writable) { 207363551ae0SDavid Gibson entry = 207463551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 207563551ae0SDavid Gibson } else { 20767f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 207763551ae0SDavid Gibson } 207863551ae0SDavid Gibson entry = pte_mkyoung(entry); 207963551ae0SDavid Gibson entry = pte_mkhuge(entry); 208063551ae0SDavid Gibson 208163551ae0SDavid Gibson return entry; 208263551ae0SDavid Gibson } 208363551ae0SDavid Gibson 20841e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 20851e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 20861e8f889bSDavid Gibson { 20871e8f889bSDavid Gibson pte_t entry; 20881e8f889bSDavid Gibson 20897f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 20907f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 20914b3073e1SRussell King update_mmu_cache(vma, address, ptep); 20921e8f889bSDavid Gibson } 20938dab5241SBenjamin Herrenschmidt } 20941e8f889bSDavid Gibson 20951e8f889bSDavid Gibson 209663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 209763551ae0SDavid Gibson struct vm_area_struct *vma) 209863551ae0SDavid Gibson { 209963551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 210063551ae0SDavid Gibson struct page *ptepage; 21011c59827dSHugh Dickins unsigned long addr; 21021e8f889bSDavid Gibson int cow; 2103a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2104a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 21051e8f889bSDavid Gibson 21061e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 210763551ae0SDavid Gibson 2108a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2109c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2110c74df32cSHugh Dickins if (!src_pte) 2111c74df32cSHugh Dickins continue; 2112a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 211363551ae0SDavid Gibson if (!dst_pte) 211463551ae0SDavid Gibson goto nomem; 2115c5c99429SLarry Woodman 2116c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2117c5c99429SLarry Woodman if (dst_pte == src_pte) 2118c5c99429SLarry Woodman continue; 2119c5c99429SLarry Woodman 2120c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 212146478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 21227f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 21231e8f889bSDavid Gibson if (cow) 21247f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 21257f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 212663551ae0SDavid Gibson ptepage = pte_page(entry); 212763551ae0SDavid Gibson get_page(ptepage); 212863551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 21291c59827dSHugh Dickins } 21301c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2131c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 213263551ae0SDavid Gibson } 213363551ae0SDavid Gibson return 0; 213463551ae0SDavid Gibson 213563551ae0SDavid Gibson nomem: 213663551ae0SDavid Gibson return -ENOMEM; 213763551ae0SDavid Gibson } 213863551ae0SDavid Gibson 2139502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 214004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 214163551ae0SDavid Gibson { 214263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 214363551ae0SDavid Gibson unsigned long address; 2144c7546f8fSDavid Gibson pte_t *ptep; 214563551ae0SDavid Gibson pte_t pte; 214663551ae0SDavid Gibson struct page *page; 2147fe1668aeSChen, Kenneth W struct page *tmp; 2148a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2149a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2150a5516438SAndi Kleen 2151c0a499c2SChen, Kenneth W /* 2152c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 2153c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 2154c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 2155c0a499c2SChen, Kenneth W */ 2156fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 215763551ae0SDavid Gibson 215863551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2159a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2160a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 216163551ae0SDavid Gibson 2162cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 2163508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2164a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2165c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2166c7546f8fSDavid Gibson if (!ptep) 2167c7546f8fSDavid Gibson continue; 2168c7546f8fSDavid Gibson 216939dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 217039dde65cSChen, Kenneth W continue; 217139dde65cSChen, Kenneth W 217204f2cbe3SMel Gorman /* 217304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 217404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 217504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 217604f2cbe3SMel Gorman */ 217704f2cbe3SMel Gorman if (ref_page) { 217804f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 217904f2cbe3SMel Gorman if (huge_pte_none(pte)) 218004f2cbe3SMel Gorman continue; 218104f2cbe3SMel Gorman page = pte_page(pte); 218204f2cbe3SMel Gorman if (page != ref_page) 218304f2cbe3SMel Gorman continue; 218404f2cbe3SMel Gorman 218504f2cbe3SMel Gorman /* 218604f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 218704f2cbe3SMel Gorman * future faults in this VMA will fail rather than 218804f2cbe3SMel Gorman * looking like data was lost 218904f2cbe3SMel Gorman */ 219004f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 219104f2cbe3SMel Gorman } 219204f2cbe3SMel Gorman 2193c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 21947f2e9525SGerald Schaefer if (huge_pte_none(pte)) 219563551ae0SDavid Gibson continue; 2196c7546f8fSDavid Gibson 219763551ae0SDavid Gibson page = pte_page(pte); 21986649a386SKen Chen if (pte_dirty(pte)) 21996649a386SKen Chen set_page_dirty(page); 2200fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 220163551ae0SDavid Gibson } 22021da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 2203508034a3SHugh Dickins flush_tlb_range(vma, start, end); 2204cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 2205fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 2206fe1668aeSChen, Kenneth W list_del(&page->lru); 2207fe1668aeSChen, Kenneth W put_page(page); 2208fe1668aeSChen, Kenneth W } 22091da177e4SLinus Torvalds } 221063551ae0SDavid Gibson 2211502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 221204f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2213502717f4SChen, Kenneth W { 2214502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 221504f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 2216502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2217502717f4SChen, Kenneth W } 2218502717f4SChen, Kenneth W 221904f2cbe3SMel Gorman /* 222004f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 222104f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 222204f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 222304f2cbe3SMel Gorman * same region. 222404f2cbe3SMel Gorman */ 22252a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 22262a4b3dedSHarvey Harrison struct page *page, unsigned long address) 222704f2cbe3SMel Gorman { 22287526674dSAdam Litke struct hstate *h = hstate_vma(vma); 222904f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 223004f2cbe3SMel Gorman struct address_space *mapping; 223104f2cbe3SMel Gorman struct prio_tree_iter iter; 223204f2cbe3SMel Gorman pgoff_t pgoff; 223304f2cbe3SMel Gorman 223404f2cbe3SMel Gorman /* 223504f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 223604f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 223704f2cbe3SMel Gorman */ 22387526674dSAdam Litke address = address & huge_page_mask(h); 223904f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 224004f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 224104f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 224204f2cbe3SMel Gorman 22434eb2b1dcSMel Gorman /* 22444eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 22454eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 22464eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 22474eb2b1dcSMel Gorman */ 22484eb2b1dcSMel Gorman spin_lock(&mapping->i_mmap_lock); 224904f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 225004f2cbe3SMel Gorman /* Do not unmap the current VMA */ 225104f2cbe3SMel Gorman if (iter_vma == vma) 225204f2cbe3SMel Gorman continue; 225304f2cbe3SMel Gorman 225404f2cbe3SMel Gorman /* 225504f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 225604f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 225704f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 225804f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 225904f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 226004f2cbe3SMel Gorman */ 226104f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 22624eb2b1dcSMel Gorman __unmap_hugepage_range(iter_vma, 22637526674dSAdam Litke address, address + huge_page_size(h), 226404f2cbe3SMel Gorman page); 226504f2cbe3SMel Gorman } 22664eb2b1dcSMel Gorman spin_unlock(&mapping->i_mmap_lock); 226704f2cbe3SMel Gorman 226804f2cbe3SMel Gorman return 1; 226904f2cbe3SMel Gorman } 227004f2cbe3SMel Gorman 22711e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 227204f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 227304f2cbe3SMel Gorman struct page *pagecache_page) 22741e8f889bSDavid Gibson { 2275a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 22761e8f889bSDavid Gibson struct page *old_page, *new_page; 227779ac6ba4SDavid Gibson int avoidcopy; 227804f2cbe3SMel Gorman int outside_reserve = 0; 22791e8f889bSDavid Gibson 22801e8f889bSDavid Gibson old_page = pte_page(pte); 22811e8f889bSDavid Gibson 228204f2cbe3SMel Gorman retry_avoidcopy: 22831e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 22841e8f889bSDavid Gibson * and just make the page writable */ 22851e8f889bSDavid Gibson avoidcopy = (page_count(old_page) == 1); 22861e8f889bSDavid Gibson if (avoidcopy) { 22871e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 228883c54070SNick Piggin return 0; 22891e8f889bSDavid Gibson } 22901e8f889bSDavid Gibson 229104f2cbe3SMel Gorman /* 229204f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 229304f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 229404f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 229504f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 229604f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 229704f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 229804f2cbe3SMel Gorman * of the full address range. 229904f2cbe3SMel Gorman */ 2300f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 230104f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 230204f2cbe3SMel Gorman old_page != pagecache_page) 230304f2cbe3SMel Gorman outside_reserve = 1; 230404f2cbe3SMel Gorman 23051e8f889bSDavid Gibson page_cache_get(old_page); 2306b76c8cfbSLarry Woodman 2307b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2308b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 230904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 23101e8f889bSDavid Gibson 23112fc39cecSAdam Litke if (IS_ERR(new_page)) { 23121e8f889bSDavid Gibson page_cache_release(old_page); 231304f2cbe3SMel Gorman 231404f2cbe3SMel Gorman /* 231504f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 231604f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 231704f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 231804f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 231904f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 232004f2cbe3SMel Gorman */ 232104f2cbe3SMel Gorman if (outside_reserve) { 232204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 232304f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 232404f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 232504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2326b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 232704f2cbe3SMel Gorman goto retry_avoidcopy; 232804f2cbe3SMel Gorman } 232904f2cbe3SMel Gorman WARN_ON_ONCE(1); 233004f2cbe3SMel Gorman } 233104f2cbe3SMel Gorman 2332b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2333b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 23342fc39cecSAdam Litke return -PTR_ERR(new_page); 23351e8f889bSDavid Gibson } 23361e8f889bSDavid Gibson 23379de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 23380ed361deSNick Piggin __SetPageUptodate(new_page); 23391e8f889bSDavid Gibson 2340b76c8cfbSLarry Woodman /* 2341b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2342b76c8cfbSLarry Woodman * before the page tables are altered 2343b76c8cfbSLarry Woodman */ 2344b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2345a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 23467f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 23471e8f889bSDavid Gibson /* Break COW */ 23488fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 23491e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 23501e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 23511e8f889bSDavid Gibson /* Make the old page be freed below */ 23521e8f889bSDavid Gibson new_page = old_page; 23531e8f889bSDavid Gibson } 23541e8f889bSDavid Gibson page_cache_release(new_page); 23551e8f889bSDavid Gibson page_cache_release(old_page); 235683c54070SNick Piggin return 0; 23571e8f889bSDavid Gibson } 23581e8f889bSDavid Gibson 235904f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2360a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2361a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 236204f2cbe3SMel Gorman { 236304f2cbe3SMel Gorman struct address_space *mapping; 2364e7c4b0bfSAndy Whitcroft pgoff_t idx; 236504f2cbe3SMel Gorman 236604f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2367a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 236804f2cbe3SMel Gorman 236904f2cbe3SMel Gorman return find_lock_page(mapping, idx); 237004f2cbe3SMel Gorman } 237104f2cbe3SMel Gorman 23723ae77f43SHugh Dickins /* 23733ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 23743ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 23753ae77f43SHugh Dickins */ 23763ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 23772a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 23782a15efc9SHugh Dickins { 23792a15efc9SHugh Dickins struct address_space *mapping; 23802a15efc9SHugh Dickins pgoff_t idx; 23812a15efc9SHugh Dickins struct page *page; 23822a15efc9SHugh Dickins 23832a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 23842a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 23852a15efc9SHugh Dickins 23862a15efc9SHugh Dickins page = find_get_page(mapping, idx); 23872a15efc9SHugh Dickins if (page) 23882a15efc9SHugh Dickins put_page(page); 23892a15efc9SHugh Dickins return page != NULL; 23902a15efc9SHugh Dickins } 23912a15efc9SHugh Dickins 2392a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2393788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2394ac9b9c66SHugh Dickins { 2395a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2396ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2397e7c4b0bfSAndy Whitcroft pgoff_t idx; 23984c887265SAdam Litke unsigned long size; 23994c887265SAdam Litke struct page *page; 24004c887265SAdam Litke struct address_space *mapping; 24011e8f889bSDavid Gibson pte_t new_pte; 24024c887265SAdam Litke 240304f2cbe3SMel Gorman /* 240404f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 240504f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 240604f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 240704f2cbe3SMel Gorman */ 240804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 240904f2cbe3SMel Gorman printk(KERN_WARNING 241004f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 241104f2cbe3SMel Gorman current->pid); 241204f2cbe3SMel Gorman return ret; 241304f2cbe3SMel Gorman } 241404f2cbe3SMel Gorman 24154c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2416a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 24174c887265SAdam Litke 24184c887265SAdam Litke /* 24194c887265SAdam Litke * Use page lock to guard against racing truncation 24204c887265SAdam Litke * before we get page_table_lock. 24214c887265SAdam Litke */ 24226bda666aSChristoph Lameter retry: 24236bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 24246bda666aSChristoph Lameter if (!page) { 2425a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2426ebed4bfcSHugh Dickins if (idx >= size) 2427ebed4bfcSHugh Dickins goto out; 242804f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 24292fc39cecSAdam Litke if (IS_ERR(page)) { 24302fc39cecSAdam Litke ret = -PTR_ERR(page); 24316bda666aSChristoph Lameter goto out; 24326bda666aSChristoph Lameter } 2433a5516438SAndi Kleen clear_huge_page(page, address, huge_page_size(h)); 24340ed361deSNick Piggin __SetPageUptodate(page); 2435ac9b9c66SHugh Dickins 2436f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 24376bda666aSChristoph Lameter int err; 243845c682a6SKen Chen struct inode *inode = mapping->host; 24396bda666aSChristoph Lameter 24406bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 24416bda666aSChristoph Lameter if (err) { 24426bda666aSChristoph Lameter put_page(page); 24436bda666aSChristoph Lameter if (err == -EEXIST) 24446bda666aSChristoph Lameter goto retry; 24456bda666aSChristoph Lameter goto out; 24466bda666aSChristoph Lameter } 244745c682a6SKen Chen 244845c682a6SKen Chen spin_lock(&inode->i_lock); 2449a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 245045c682a6SKen Chen spin_unlock(&inode->i_lock); 245123be7468SMel Gorman } else { 24526bda666aSChristoph Lameter lock_page(page); 245323be7468SMel Gorman page->mapping = HUGETLB_POISON; 245423be7468SMel Gorman } 24556bda666aSChristoph Lameter } 24561e8f889bSDavid Gibson 245757303d80SAndy Whitcroft /* 245857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 245957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 246057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 246157303d80SAndy Whitcroft * the spinlock. 246257303d80SAndy Whitcroft */ 2463788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 24642b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 24652b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 24662b26736cSAndy Whitcroft goto backout_unlocked; 24672b26736cSAndy Whitcroft } 246857303d80SAndy Whitcroft 2469ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2470a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 24714c887265SAdam Litke if (idx >= size) 24724c887265SAdam Litke goto backout; 24734c887265SAdam Litke 247483c54070SNick Piggin ret = 0; 24757f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 24764c887265SAdam Litke goto backout; 24774c887265SAdam Litke 24781e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 24791e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 24801e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 24811e8f889bSDavid Gibson 2482788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 24831e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 248404f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 24851e8f889bSDavid Gibson } 24861e8f889bSDavid Gibson 2487ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 24884c887265SAdam Litke unlock_page(page); 24894c887265SAdam Litke out: 2490ac9b9c66SHugh Dickins return ret; 24914c887265SAdam Litke 24924c887265SAdam Litke backout: 24934c887265SAdam Litke spin_unlock(&mm->page_table_lock); 24942b26736cSAndy Whitcroft backout_unlocked: 24954c887265SAdam Litke unlock_page(page); 24964c887265SAdam Litke put_page(page); 24974c887265SAdam Litke goto out; 2498ac9b9c66SHugh Dickins } 2499ac9b9c66SHugh Dickins 250086e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2501788c7df4SHugh Dickins unsigned long address, unsigned int flags) 250286e5216fSAdam Litke { 250386e5216fSAdam Litke pte_t *ptep; 250486e5216fSAdam Litke pte_t entry; 25051e8f889bSDavid Gibson int ret; 250657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 25073935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2508a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 250986e5216fSAdam Litke 2510a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 251186e5216fSAdam Litke if (!ptep) 251286e5216fSAdam Litke return VM_FAULT_OOM; 251386e5216fSAdam Litke 25143935baa9SDavid Gibson /* 25153935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 25163935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 25173935baa9SDavid Gibson * the same page in the page cache. 25183935baa9SDavid Gibson */ 25193935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 25207f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 25217f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2522788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2523b4d1d99fSDavid Gibson goto out_mutex; 25243935baa9SDavid Gibson } 252586e5216fSAdam Litke 252683c54070SNick Piggin ret = 0; 25271e8f889bSDavid Gibson 252857303d80SAndy Whitcroft /* 252957303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 253057303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 253157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 253257303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 253357303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 253457303d80SAndy Whitcroft * consumed. 253557303d80SAndy Whitcroft */ 2536788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 25372b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 25382b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2539b4d1d99fSDavid Gibson goto out_mutex; 25402b26736cSAndy Whitcroft } 254157303d80SAndy Whitcroft 2542f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 254357303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 254457303d80SAndy Whitcroft vma, address); 254557303d80SAndy Whitcroft } 254657303d80SAndy Whitcroft 25471e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 25481e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2549b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2550b4d1d99fSDavid Gibson goto out_page_table_lock; 2551b4d1d99fSDavid Gibson 2552b4d1d99fSDavid Gibson 2553788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2554b4d1d99fSDavid Gibson if (!pte_write(entry)) { 255557303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 255657303d80SAndy Whitcroft pagecache_page); 2557b4d1d99fSDavid Gibson goto out_page_table_lock; 2558b4d1d99fSDavid Gibson } 2559b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2560b4d1d99fSDavid Gibson } 2561b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2562788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2563788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 25644b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2565b4d1d99fSDavid Gibson 2566b4d1d99fSDavid Gibson out_page_table_lock: 25671e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 256857303d80SAndy Whitcroft 256957303d80SAndy Whitcroft if (pagecache_page) { 257057303d80SAndy Whitcroft unlock_page(pagecache_page); 257157303d80SAndy Whitcroft put_page(pagecache_page); 257257303d80SAndy Whitcroft } 257357303d80SAndy Whitcroft 2574b4d1d99fSDavid Gibson out_mutex: 25753935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 25761e8f889bSDavid Gibson 25771e8f889bSDavid Gibson return ret; 257886e5216fSAdam Litke } 257986e5216fSAdam Litke 2580ceb86879SAndi Kleen /* Can be overriden by architectures */ 2581ceb86879SAndi Kleen __attribute__((weak)) struct page * 2582ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2583ceb86879SAndi Kleen pud_t *pud, int write) 2584ceb86879SAndi Kleen { 2585ceb86879SAndi Kleen BUG(); 2586ceb86879SAndi Kleen return NULL; 2587ceb86879SAndi Kleen } 2588ceb86879SAndi Kleen 258963551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 259063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 25915b23dbe8SAdam Litke unsigned long *position, int *length, int i, 25922a15efc9SHugh Dickins unsigned int flags) 259363551ae0SDavid Gibson { 2594d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2595d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 259663551ae0SDavid Gibson int remainder = *length; 2597a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 259863551ae0SDavid Gibson 25991c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 260063551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 260163551ae0SDavid Gibson pte_t *pte; 26022a15efc9SHugh Dickins int absent; 260363551ae0SDavid Gibson struct page *page; 260463551ae0SDavid Gibson 26054c887265SAdam Litke /* 26064c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 26072a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 26084c887265SAdam Litke * first, for the page indexing below to work. 26094c887265SAdam Litke */ 2610a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 26112a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 261263551ae0SDavid Gibson 26132a15efc9SHugh Dickins /* 26142a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 26153ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 26163ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 26173ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 26183ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 26192a15efc9SHugh Dickins */ 26203ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 26213ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 26222a15efc9SHugh Dickins remainder = 0; 26232a15efc9SHugh Dickins break; 26242a15efc9SHugh Dickins } 26252a15efc9SHugh Dickins 26262a15efc9SHugh Dickins if (absent || 26272a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 26284c887265SAdam Litke int ret; 26294c887265SAdam Litke 26304c887265SAdam Litke spin_unlock(&mm->page_table_lock); 26312a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 26322a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 26334c887265SAdam Litke spin_lock(&mm->page_table_lock); 2634a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 26354c887265SAdam Litke continue; 26364c887265SAdam Litke 26371c59827dSHugh Dickins remainder = 0; 26381c59827dSHugh Dickins break; 26391c59827dSHugh Dickins } 264063551ae0SDavid Gibson 2641a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 26427f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2643d5d4b0aaSChen, Kenneth W same_page: 2644d6692183SChen, Kenneth W if (pages) { 264569d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 26464b2e38adSKOSAKI Motohiro get_page(pages[i]); 2647d6692183SChen, Kenneth W } 264863551ae0SDavid Gibson 264963551ae0SDavid Gibson if (vmas) 265063551ae0SDavid Gibson vmas[i] = vma; 265163551ae0SDavid Gibson 265263551ae0SDavid Gibson vaddr += PAGE_SIZE; 2653d5d4b0aaSChen, Kenneth W ++pfn_offset; 265463551ae0SDavid Gibson --remainder; 265563551ae0SDavid Gibson ++i; 2656d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2657a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2658d5d4b0aaSChen, Kenneth W /* 2659d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2660d5d4b0aaSChen, Kenneth W * of this compound page. 2661d5d4b0aaSChen, Kenneth W */ 2662d5d4b0aaSChen, Kenneth W goto same_page; 2663d5d4b0aaSChen, Kenneth W } 266463551ae0SDavid Gibson } 26651c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 266663551ae0SDavid Gibson *length = remainder; 266763551ae0SDavid Gibson *position = vaddr; 266863551ae0SDavid Gibson 26692a15efc9SHugh Dickins return i ? i : -EFAULT; 267063551ae0SDavid Gibson } 26718f860591SZhang, Yanmin 26728f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 26738f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 26748f860591SZhang, Yanmin { 26758f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 26768f860591SZhang, Yanmin unsigned long start = address; 26778f860591SZhang, Yanmin pte_t *ptep; 26788f860591SZhang, Yanmin pte_t pte; 2679a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 26808f860591SZhang, Yanmin 26818f860591SZhang, Yanmin BUG_ON(address >= end); 26828f860591SZhang, Yanmin flush_cache_range(vma, address, end); 26838f860591SZhang, Yanmin 268439dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 26858f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2686a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 26878f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 26888f860591SZhang, Yanmin if (!ptep) 26898f860591SZhang, Yanmin continue; 269039dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 269139dde65cSChen, Kenneth W continue; 26927f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 26938f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 26948f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 26958f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 26968f860591SZhang, Yanmin } 26978f860591SZhang, Yanmin } 26988f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 269939dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 27008f860591SZhang, Yanmin 27018f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 27028f860591SZhang, Yanmin } 27038f860591SZhang, Yanmin 2704a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2705a1e78772SMel Gorman long from, long to, 27065a6fe125SMel Gorman struct vm_area_struct *vma, 27075a6fe125SMel Gorman int acctflag) 2708e4e574b7SAdam Litke { 270917c9d12eSMel Gorman long ret, chg; 2710a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2711e4e574b7SAdam Litke 2712a1e78772SMel Gorman /* 271317c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 271417c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 271517c9d12eSMel Gorman * and filesystem quota without using reserves 271617c9d12eSMel Gorman */ 271717c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 271817c9d12eSMel Gorman return 0; 271917c9d12eSMel Gorman 272017c9d12eSMel Gorman /* 2721a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2722a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2723a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2724a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2725a1e78772SMel Gorman */ 2726f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2727e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 27285a6fe125SMel Gorman else { 27295a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 27305a6fe125SMel Gorman if (!resv_map) 27315a6fe125SMel Gorman return -ENOMEM; 27325a6fe125SMel Gorman 273317c9d12eSMel Gorman chg = to - from; 273417c9d12eSMel Gorman 27355a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 27365a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 27375a6fe125SMel Gorman } 27385a6fe125SMel Gorman 273917c9d12eSMel Gorman if (chg < 0) 274017c9d12eSMel Gorman return chg; 274117c9d12eSMel Gorman 274217c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 274317c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 274417c9d12eSMel Gorman return -ENOSPC; 274517c9d12eSMel Gorman 274617c9d12eSMel Gorman /* 274717c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 274817c9d12eSMel Gorman * Hand back the quota if there are not 274917c9d12eSMel Gorman */ 275017c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 275117c9d12eSMel Gorman if (ret < 0) { 275217c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 275317c9d12eSMel Gorman return ret; 275417c9d12eSMel Gorman } 275517c9d12eSMel Gorman 275617c9d12eSMel Gorman /* 275717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 275817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 275917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 276017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 276117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 276217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 276317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 276417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 276517c9d12eSMel Gorman * else has to be done for private mappings here 276617c9d12eSMel Gorman */ 2767f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 276817c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2769a43a8c39SChen, Kenneth W return 0; 2770a43a8c39SChen, Kenneth W } 2771a43a8c39SChen, Kenneth W 2772a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2773a43a8c39SChen, Kenneth W { 2774a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2775a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 277645c682a6SKen Chen 277745c682a6SKen Chen spin_lock(&inode->i_lock); 2778e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 277945c682a6SKen Chen spin_unlock(&inode->i_lock); 278045c682a6SKen Chen 278190d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2782a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2783a43a8c39SChen, Kenneth W } 2784