11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 31da177e4SLinus Torvalds * (C) William Irwin, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24d6606683SLinus Torvalds 2563551ae0SDavid Gibson #include <asm/page.h> 2663551ae0SDavid Gibson #include <asm/pgtable.h> 2778a34ae2SAdrian Bunk #include <asm/io.h> 2863551ae0SDavid Gibson 2963551ae0SDavid Gibson #include <linux/hugetlb.h> 309a305230SLee Schermerhorn #include <linux/node.h> 317835e98bSNick Piggin #include "internal.h" 321da177e4SLinus Torvalds 331da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 34396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 35396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 36a5516438SAndi Kleen 37e5ff2159SAndi Kleen static int max_hstate; 38e5ff2159SAndi Kleen unsigned int default_hstate_idx; 39e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 40e5ff2159SAndi Kleen 4153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4253ba51d2SJon Tollefson 43e5ff2159SAndi Kleen /* for command line parsing */ 44e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 45e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 46e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 47e5ff2159SAndi Kleen 48e5ff2159SAndi Kleen #define for_each_hstate(h) \ 49e5ff2159SAndi Kleen for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) 50396faf03SMel Gorman 513935baa9SDavid Gibson /* 523935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 533935baa9SDavid Gibson */ 543935baa9SDavid Gibson static DEFINE_SPINLOCK(hugetlb_lock); 550bd0f9fbSEric Paris 56e7c4b0bfSAndy Whitcroft /* 5796822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 5896822904SAndy Whitcroft * across the pages in a mapping. 5984afd99bSAndy Whitcroft * 6084afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 6184afd99bSAndy Whitcroft * and the hugetlb_instantion_mutex. To access or modify a region the caller 6284afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 6384afd99bSAndy Whitcroft * the hugetlb_instantiation mutex: 6484afd99bSAndy Whitcroft * 6584afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 6684afd99bSAndy Whitcroft * or 6784afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 6884afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 6996822904SAndy Whitcroft */ 7096822904SAndy Whitcroft struct file_region { 7196822904SAndy Whitcroft struct list_head link; 7296822904SAndy Whitcroft long from; 7396822904SAndy Whitcroft long to; 7496822904SAndy Whitcroft }; 7596822904SAndy Whitcroft 7696822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 7796822904SAndy Whitcroft { 7896822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 7996822904SAndy Whitcroft 8096822904SAndy Whitcroft /* Locate the region we are either in or before. */ 8196822904SAndy Whitcroft list_for_each_entry(rg, head, link) 8296822904SAndy Whitcroft if (f <= rg->to) 8396822904SAndy Whitcroft break; 8496822904SAndy Whitcroft 8596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 8696822904SAndy Whitcroft if (f > rg->from) 8796822904SAndy Whitcroft f = rg->from; 8896822904SAndy Whitcroft 8996822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 9096822904SAndy Whitcroft nrg = rg; 9196822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 9296822904SAndy Whitcroft if (&rg->link == head) 9396822904SAndy Whitcroft break; 9496822904SAndy Whitcroft if (rg->from > t) 9596822904SAndy Whitcroft break; 9696822904SAndy Whitcroft 9796822904SAndy Whitcroft /* If this area reaches higher then extend our area to 9896822904SAndy Whitcroft * include it completely. If this is not the first area 9996822904SAndy Whitcroft * which we intend to reuse, free it. */ 10096822904SAndy Whitcroft if (rg->to > t) 10196822904SAndy Whitcroft t = rg->to; 10296822904SAndy Whitcroft if (rg != nrg) { 10396822904SAndy Whitcroft list_del(&rg->link); 10496822904SAndy Whitcroft kfree(rg); 10596822904SAndy Whitcroft } 10696822904SAndy Whitcroft } 10796822904SAndy Whitcroft nrg->from = f; 10896822904SAndy Whitcroft nrg->to = t; 10996822904SAndy Whitcroft return 0; 11096822904SAndy Whitcroft } 11196822904SAndy Whitcroft 11296822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 11396822904SAndy Whitcroft { 11496822904SAndy Whitcroft struct file_region *rg, *nrg; 11596822904SAndy Whitcroft long chg = 0; 11696822904SAndy Whitcroft 11796822904SAndy Whitcroft /* Locate the region we are before or in. */ 11896822904SAndy Whitcroft list_for_each_entry(rg, head, link) 11996822904SAndy Whitcroft if (f <= rg->to) 12096822904SAndy Whitcroft break; 12196822904SAndy Whitcroft 12296822904SAndy Whitcroft /* If we are below the current region then a new region is required. 12396822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 12496822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 12596822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 12696822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 12796822904SAndy Whitcroft if (!nrg) 12896822904SAndy Whitcroft return -ENOMEM; 12996822904SAndy Whitcroft nrg->from = f; 13096822904SAndy Whitcroft nrg->to = f; 13196822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 13296822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 13396822904SAndy Whitcroft 13496822904SAndy Whitcroft return t - f; 13596822904SAndy Whitcroft } 13696822904SAndy Whitcroft 13796822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 13896822904SAndy Whitcroft if (f > rg->from) 13996822904SAndy Whitcroft f = rg->from; 14096822904SAndy Whitcroft chg = t - f; 14196822904SAndy Whitcroft 14296822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 14396822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 14496822904SAndy Whitcroft if (&rg->link == head) 14596822904SAndy Whitcroft break; 14696822904SAndy Whitcroft if (rg->from > t) 14796822904SAndy Whitcroft return chg; 14896822904SAndy Whitcroft 14996822904SAndy Whitcroft /* We overlap with this area, if it extends futher than 15096822904SAndy Whitcroft * us then we must extend ourselves. Account for its 15196822904SAndy Whitcroft * existing reservation. */ 15296822904SAndy Whitcroft if (rg->to > t) { 15396822904SAndy Whitcroft chg += rg->to - t; 15496822904SAndy Whitcroft t = rg->to; 15596822904SAndy Whitcroft } 15696822904SAndy Whitcroft chg -= rg->to - rg->from; 15796822904SAndy Whitcroft } 15896822904SAndy Whitcroft return chg; 15996822904SAndy Whitcroft } 16096822904SAndy Whitcroft 16196822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 16296822904SAndy Whitcroft { 16396822904SAndy Whitcroft struct file_region *rg, *trg; 16496822904SAndy Whitcroft long chg = 0; 16596822904SAndy Whitcroft 16696822904SAndy Whitcroft /* Locate the region we are either in or before. */ 16796822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16896822904SAndy Whitcroft if (end <= rg->to) 16996822904SAndy Whitcroft break; 17096822904SAndy Whitcroft if (&rg->link == head) 17196822904SAndy Whitcroft return 0; 17296822904SAndy Whitcroft 17396822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 17496822904SAndy Whitcroft if (end > rg->from) { 17596822904SAndy Whitcroft chg = rg->to - end; 17696822904SAndy Whitcroft rg->to = end; 17796822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 17896822904SAndy Whitcroft } 17996822904SAndy Whitcroft 18096822904SAndy Whitcroft /* Drop any remaining regions. */ 18196822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 18296822904SAndy Whitcroft if (&rg->link == head) 18396822904SAndy Whitcroft break; 18496822904SAndy Whitcroft chg += rg->to - rg->from; 18596822904SAndy Whitcroft list_del(&rg->link); 18696822904SAndy Whitcroft kfree(rg); 18796822904SAndy Whitcroft } 18896822904SAndy Whitcroft return chg; 18996822904SAndy Whitcroft } 19096822904SAndy Whitcroft 19184afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 19284afd99bSAndy Whitcroft { 19384afd99bSAndy Whitcroft struct file_region *rg; 19484afd99bSAndy Whitcroft long chg = 0; 19584afd99bSAndy Whitcroft 19684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 19784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 19884afd99bSAndy Whitcroft int seg_from; 19984afd99bSAndy Whitcroft int seg_to; 20084afd99bSAndy Whitcroft 20184afd99bSAndy Whitcroft if (rg->to <= f) 20284afd99bSAndy Whitcroft continue; 20384afd99bSAndy Whitcroft if (rg->from >= t) 20484afd99bSAndy Whitcroft break; 20584afd99bSAndy Whitcroft 20684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 20784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 20884afd99bSAndy Whitcroft 20984afd99bSAndy Whitcroft chg += seg_to - seg_from; 21084afd99bSAndy Whitcroft } 21184afd99bSAndy Whitcroft 21284afd99bSAndy Whitcroft return chg; 21384afd99bSAndy Whitcroft } 21484afd99bSAndy Whitcroft 21596822904SAndy Whitcroft /* 216e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 217e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 218e7c4b0bfSAndy Whitcroft */ 219a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 220a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 221e7c4b0bfSAndy Whitcroft { 222a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 223a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 224e7c4b0bfSAndy Whitcroft } 225e7c4b0bfSAndy Whitcroft 2260fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 2270fe6e20bSNaoya Horiguchi unsigned long address) 2280fe6e20bSNaoya Horiguchi { 2290fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 2300fe6e20bSNaoya Horiguchi } 2310fe6e20bSNaoya Horiguchi 23284afd99bSAndy Whitcroft /* 23308fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 23408fba699SMel Gorman * cases this will be same size as used by the page table entries. 23508fba699SMel Gorman */ 23608fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 23708fba699SMel Gorman { 23808fba699SMel Gorman struct hstate *hstate; 23908fba699SMel Gorman 24008fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 24108fba699SMel Gorman return PAGE_SIZE; 24208fba699SMel Gorman 24308fba699SMel Gorman hstate = hstate_vma(vma); 24408fba699SMel Gorman 24508fba699SMel Gorman return 1UL << (hstate->order + PAGE_SHIFT); 24608fba699SMel Gorman } 247f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 24808fba699SMel Gorman 24908fba699SMel Gorman /* 2503340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 2513340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 2523340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 2533340289dSMel Gorman * function is required. 2543340289dSMel Gorman */ 2553340289dSMel Gorman #ifndef vma_mmu_pagesize 2563340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 2573340289dSMel Gorman { 2583340289dSMel Gorman return vma_kernel_pagesize(vma); 2593340289dSMel Gorman } 2603340289dSMel Gorman #endif 2613340289dSMel Gorman 2623340289dSMel Gorman /* 26384afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 26484afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 26584afd99bSAndy Whitcroft * alignment. 26684afd99bSAndy Whitcroft */ 26784afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 26884afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 26904f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 27084afd99bSAndy Whitcroft 271a1e78772SMel Gorman /* 272a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 273a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 274a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 275a1e78772SMel Gorman * 276a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 277a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 278a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 279a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 28084afd99bSAndy Whitcroft * 28184afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 28284afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 28384afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 28484afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 28584afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 28684afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 28784afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 28884afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 289a1e78772SMel Gorman */ 290e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 291e7c4b0bfSAndy Whitcroft { 292e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 293e7c4b0bfSAndy Whitcroft } 294e7c4b0bfSAndy Whitcroft 295e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 296e7c4b0bfSAndy Whitcroft unsigned long value) 297e7c4b0bfSAndy Whitcroft { 298e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 299e7c4b0bfSAndy Whitcroft } 300e7c4b0bfSAndy Whitcroft 30184afd99bSAndy Whitcroft struct resv_map { 30284afd99bSAndy Whitcroft struct kref refs; 30384afd99bSAndy Whitcroft struct list_head regions; 30484afd99bSAndy Whitcroft }; 30584afd99bSAndy Whitcroft 3062a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 30784afd99bSAndy Whitcroft { 30884afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 30984afd99bSAndy Whitcroft if (!resv_map) 31084afd99bSAndy Whitcroft return NULL; 31184afd99bSAndy Whitcroft 31284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 31384afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 31484afd99bSAndy Whitcroft 31584afd99bSAndy Whitcroft return resv_map; 31684afd99bSAndy Whitcroft } 31784afd99bSAndy Whitcroft 3182a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 31984afd99bSAndy Whitcroft { 32084afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 32184afd99bSAndy Whitcroft 32284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 32384afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 32484afd99bSAndy Whitcroft kfree(resv_map); 32584afd99bSAndy Whitcroft } 32684afd99bSAndy Whitcroft 32784afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 328a1e78772SMel Gorman { 329a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 330f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 33184afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 33284afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 3332a4b3dedSHarvey Harrison return NULL; 334a1e78772SMel Gorman } 335a1e78772SMel Gorman 33684afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 337a1e78772SMel Gorman { 338a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 339f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 340a1e78772SMel Gorman 34184afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 34284afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 34304f2cbe3SMel Gorman } 34404f2cbe3SMel Gorman 34504f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 34604f2cbe3SMel Gorman { 34704f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 348f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 349e7c4b0bfSAndy Whitcroft 350e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 35104f2cbe3SMel Gorman } 35204f2cbe3SMel Gorman 35304f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 35404f2cbe3SMel Gorman { 35504f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 356e7c4b0bfSAndy Whitcroft 357e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 358a1e78772SMel Gorman } 359a1e78772SMel Gorman 360a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 361a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 362a5516438SAndi Kleen struct vm_area_struct *vma) 363a1e78772SMel Gorman { 364c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 365c37f9fb1SAndy Whitcroft return; 366c37f9fb1SAndy Whitcroft 367f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 368a1e78772SMel Gorman /* Shared mappings always use reserves */ 369a5516438SAndi Kleen h->resv_huge_pages--; 37084afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 371a1e78772SMel Gorman /* 372a1e78772SMel Gorman * Only the process that called mmap() has reserves for 373a1e78772SMel Gorman * private mappings. 374a1e78772SMel Gorman */ 375a5516438SAndi Kleen h->resv_huge_pages--; 376a1e78772SMel Gorman } 377a1e78772SMel Gorman } 378a1e78772SMel Gorman 37904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 380a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 381a1e78772SMel Gorman { 382a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 383f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 384a1e78772SMel Gorman vma->vm_private_data = (void *)0; 385a1e78772SMel Gorman } 386a1e78772SMel Gorman 387a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 3887f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 389a1e78772SMel Gorman { 390f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 391a1e78772SMel Gorman return 1; 3927f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3937f09ca51SMel Gorman return 1; 3947f09ca51SMel Gorman return 0; 395a1e78772SMel Gorman } 396a1e78772SMel Gorman 39769d177c2SAndy Whitcroft static void clear_gigantic_page(struct page *page, 39869d177c2SAndy Whitcroft unsigned long addr, unsigned long sz) 39969d177c2SAndy Whitcroft { 40069d177c2SAndy Whitcroft int i; 40169d177c2SAndy Whitcroft struct page *p = page; 40269d177c2SAndy Whitcroft 40369d177c2SAndy Whitcroft might_sleep(); 40469d177c2SAndy Whitcroft for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) { 40569d177c2SAndy Whitcroft cond_resched(); 40669d177c2SAndy Whitcroft clear_user_highpage(p, addr + i * PAGE_SIZE); 40769d177c2SAndy Whitcroft } 40869d177c2SAndy Whitcroft } 409a5516438SAndi Kleen static void clear_huge_page(struct page *page, 410a5516438SAndi Kleen unsigned long addr, unsigned long sz) 41179ac6ba4SDavid Gibson { 41279ac6ba4SDavid Gibson int i; 41379ac6ba4SDavid Gibson 41474dbdd23SAndrea Arcangeli if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) { 415ebdd4aeaSHannes Eder clear_gigantic_page(page, addr, sz); 416ebdd4aeaSHannes Eder return; 417ebdd4aeaSHannes Eder } 41869d177c2SAndy Whitcroft 41979ac6ba4SDavid Gibson might_sleep(); 420a5516438SAndi Kleen for (i = 0; i < sz/PAGE_SIZE; i++) { 42179ac6ba4SDavid Gibson cond_resched(); 422281e0e3bSRalf Baechle clear_user_highpage(page + i, addr + i * PAGE_SIZE); 42379ac6ba4SDavid Gibson } 42479ac6ba4SDavid Gibson } 42579ac6ba4SDavid Gibson 42669d177c2SAndy Whitcroft static void copy_gigantic_page(struct page *dst, struct page *src, 42769d177c2SAndy Whitcroft unsigned long addr, struct vm_area_struct *vma) 42869d177c2SAndy Whitcroft { 42969d177c2SAndy Whitcroft int i; 43069d177c2SAndy Whitcroft struct hstate *h = hstate_vma(vma); 43169d177c2SAndy Whitcroft struct page *dst_base = dst; 43269d177c2SAndy Whitcroft struct page *src_base = src; 43369d177c2SAndy Whitcroft might_sleep(); 43469d177c2SAndy Whitcroft for (i = 0; i < pages_per_huge_page(h); ) { 43569d177c2SAndy Whitcroft cond_resched(); 43669d177c2SAndy Whitcroft copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 43769d177c2SAndy Whitcroft 43869d177c2SAndy Whitcroft i++; 43969d177c2SAndy Whitcroft dst = mem_map_next(dst, dst_base, i); 44069d177c2SAndy Whitcroft src = mem_map_next(src, src_base, i); 44169d177c2SAndy Whitcroft } 44269d177c2SAndy Whitcroft } 44379ac6ba4SDavid Gibson static void copy_huge_page(struct page *dst, struct page *src, 4449de455b2SAtsushi Nemoto unsigned long addr, struct vm_area_struct *vma) 44579ac6ba4SDavid Gibson { 44679ac6ba4SDavid Gibson int i; 447a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44879ac6ba4SDavid Gibson 449ebdd4aeaSHannes Eder if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 450ebdd4aeaSHannes Eder copy_gigantic_page(dst, src, addr, vma); 451ebdd4aeaSHannes Eder return; 452ebdd4aeaSHannes Eder } 45369d177c2SAndy Whitcroft 45479ac6ba4SDavid Gibson might_sleep(); 455a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 45679ac6ba4SDavid Gibson cond_resched(); 4579de455b2SAtsushi Nemoto copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 45879ac6ba4SDavid Gibson } 45979ac6ba4SDavid Gibson } 46079ac6ba4SDavid Gibson 461a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 4621da177e4SLinus Torvalds { 4631da177e4SLinus Torvalds int nid = page_to_nid(page); 464a5516438SAndi Kleen list_add(&page->lru, &h->hugepage_freelists[nid]); 465a5516438SAndi Kleen h->free_huge_pages++; 466a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 4671da177e4SLinus Torvalds } 4681da177e4SLinus Torvalds 469a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 470a5516438SAndi Kleen struct vm_area_struct *vma, 47104f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 4721da177e4SLinus Torvalds { 47331a5c6e4SNishanth Aravamudan int nid; 4741da177e4SLinus Torvalds struct page *page = NULL; 475480eccf9SLee Schermerhorn struct mempolicy *mpol; 47619770b32SMel Gorman nodemask_t *nodemask; 477c0ff7453SMiao Xie struct zonelist *zonelist; 478dd1a239fSMel Gorman struct zone *zone; 479dd1a239fSMel Gorman struct zoneref *z; 4801da177e4SLinus Torvalds 481c0ff7453SMiao Xie get_mems_allowed(); 482c0ff7453SMiao Xie zonelist = huge_zonelist(vma, address, 483c0ff7453SMiao Xie htlb_alloc_mask, &mpol, &nodemask); 484a1e78772SMel Gorman /* 485a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 486a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 487a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 488a1e78772SMel Gorman */ 4897f09ca51SMel Gorman if (!vma_has_reserves(vma) && 490a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 491c0ff7453SMiao Xie goto err; 492a1e78772SMel Gorman 49304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 494a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 495c0ff7453SMiao Xie goto err;; 49604f2cbe3SMel Gorman 49719770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 49819770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 49954a6eb5cSMel Gorman nid = zone_to_nid(zone); 50054a6eb5cSMel Gorman if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 501a5516438SAndi Kleen !list_empty(&h->hugepage_freelists[nid])) { 502a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 5031da177e4SLinus Torvalds struct page, lru); 5041da177e4SLinus Torvalds list_del(&page->lru); 505a5516438SAndi Kleen h->free_huge_pages--; 506a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 50704f2cbe3SMel Gorman 50804f2cbe3SMel Gorman if (!avoid_reserve) 509a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 510a1e78772SMel Gorman 5115ab3ee7bSKen Chen break; 5121da177e4SLinus Torvalds } 5133abf7afdSAndrew Morton } 514c0ff7453SMiao Xie err: 51552cd3b07SLee Schermerhorn mpol_cond_put(mpol); 516c0ff7453SMiao Xie put_mems_allowed(); 5171da177e4SLinus Torvalds return page; 5181da177e4SLinus Torvalds } 5191da177e4SLinus Torvalds 520a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5216af2acb6SAdam Litke { 5226af2acb6SAdam Litke int i; 523a5516438SAndi Kleen 52418229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 52518229df5SAndy Whitcroft 526a5516438SAndi Kleen h->nr_huge_pages--; 527a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 528a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 5296af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 5306af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 5316af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 5326af2acb6SAdam Litke } 5336af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5346af2acb6SAdam Litke set_page_refcounted(page); 5357f2e9525SGerald Schaefer arch_release_hugepage(page); 536a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 5376af2acb6SAdam Litke } 5386af2acb6SAdam Litke 539e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 540e5ff2159SAndi Kleen { 541e5ff2159SAndi Kleen struct hstate *h; 542e5ff2159SAndi Kleen 543e5ff2159SAndi Kleen for_each_hstate(h) { 544e5ff2159SAndi Kleen if (huge_page_size(h) == size) 545e5ff2159SAndi Kleen return h; 546e5ff2159SAndi Kleen } 547e5ff2159SAndi Kleen return NULL; 548e5ff2159SAndi Kleen } 549e5ff2159SAndi Kleen 55027a85ef1SDavid Gibson static void free_huge_page(struct page *page) 55127a85ef1SDavid Gibson { 552a5516438SAndi Kleen /* 553a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 554a5516438SAndi Kleen * compound page destructor. 555a5516438SAndi Kleen */ 556e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 5577893d1d5SAdam Litke int nid = page_to_nid(page); 558c79fb75eSAdam Litke struct address_space *mapping; 55927a85ef1SDavid Gibson 560c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 561e5df70abSAndy Whitcroft set_page_private(page, 0); 56223be7468SMel Gorman page->mapping = NULL; 5637893d1d5SAdam Litke BUG_ON(page_count(page)); 5640fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 56527a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 56627a85ef1SDavid Gibson 56727a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 568aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 569a5516438SAndi Kleen update_and_free_page(h, page); 570a5516438SAndi Kleen h->surplus_huge_pages--; 571a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 5727893d1d5SAdam Litke } else { 573a5516438SAndi Kleen enqueue_huge_page(h, page); 5747893d1d5SAdam Litke } 57527a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 576c79fb75eSAdam Litke if (mapping) 5779a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 57827a85ef1SDavid Gibson } 57927a85ef1SDavid Gibson 580a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 581b7ba30c6SAndi Kleen { 582b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 583b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 584a5516438SAndi Kleen h->nr_huge_pages++; 585a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 586b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 587b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 588b7ba30c6SAndi Kleen } 589b7ba30c6SAndi Kleen 59020a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 59120a0307cSWu Fengguang { 59220a0307cSWu Fengguang int i; 59320a0307cSWu Fengguang int nr_pages = 1 << order; 59420a0307cSWu Fengguang struct page *p = page + 1; 59520a0307cSWu Fengguang 59620a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 59720a0307cSWu Fengguang set_compound_order(page, order); 59820a0307cSWu Fengguang __SetPageHead(page); 59920a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 60020a0307cSWu Fengguang __SetPageTail(p); 60120a0307cSWu Fengguang p->first_page = page; 60220a0307cSWu Fengguang } 60320a0307cSWu Fengguang } 60420a0307cSWu Fengguang 60520a0307cSWu Fengguang int PageHuge(struct page *page) 60620a0307cSWu Fengguang { 60720a0307cSWu Fengguang compound_page_dtor *dtor; 60820a0307cSWu Fengguang 60920a0307cSWu Fengguang if (!PageCompound(page)) 61020a0307cSWu Fengguang return 0; 61120a0307cSWu Fengguang 61220a0307cSWu Fengguang page = compound_head(page); 61320a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 61420a0307cSWu Fengguang 61520a0307cSWu Fengguang return dtor == free_huge_page; 61620a0307cSWu Fengguang } 61720a0307cSWu Fengguang 618a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6191da177e4SLinus Torvalds { 6201da177e4SLinus Torvalds struct page *page; 621f96efd58SJoe Jin 622aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 623aa888a74SAndi Kleen return NULL; 624aa888a74SAndi Kleen 6256484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 626551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 627551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 628a5516438SAndi Kleen huge_page_order(h)); 6291da177e4SLinus Torvalds if (page) { 6307f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 631caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6327b8ee84dSHarvey Harrison return NULL; 6337f2e9525SGerald Schaefer } 634a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6351da177e4SLinus Torvalds } 63663b4613cSNishanth Aravamudan 63763b4613cSNishanth Aravamudan return page; 63863b4613cSNishanth Aravamudan } 63963b4613cSNishanth Aravamudan 6405ced66c9SAndi Kleen /* 6416ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 6426ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 6436ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 6446ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 6456ae11b27SLee Schermerhorn * node for alloc or free. 6469a76db09SLee Schermerhorn */ 6476ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 6489a76db09SLee Schermerhorn { 6496ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 6509a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 6516ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 6529a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 6539a76db09SLee Schermerhorn 6549a76db09SLee Schermerhorn return nid; 6559a76db09SLee Schermerhorn } 6569a76db09SLee Schermerhorn 6576ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 6585ced66c9SAndi Kleen { 6596ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 6606ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 6619a76db09SLee Schermerhorn return nid; 6625ced66c9SAndi Kleen } 6635ced66c9SAndi Kleen 6646ae11b27SLee Schermerhorn /* 6656ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 6666ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 6676ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 6686ae11b27SLee Schermerhorn * mask. 6696ae11b27SLee Schermerhorn */ 6706ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 6716ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 6726ae11b27SLee Schermerhorn { 6736ae11b27SLee Schermerhorn int nid; 6746ae11b27SLee Schermerhorn 6756ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 6766ae11b27SLee Schermerhorn 6776ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 6786ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 6796ae11b27SLee Schermerhorn 6806ae11b27SLee Schermerhorn return nid; 6816ae11b27SLee Schermerhorn } 6826ae11b27SLee Schermerhorn 6836ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 68463b4613cSNishanth Aravamudan { 68563b4613cSNishanth Aravamudan struct page *page; 68663b4613cSNishanth Aravamudan int start_nid; 68763b4613cSNishanth Aravamudan int next_nid; 68863b4613cSNishanth Aravamudan int ret = 0; 68963b4613cSNishanth Aravamudan 6906ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 691e8c5c824SLee Schermerhorn next_nid = start_nid; 69263b4613cSNishanth Aravamudan 69363b4613cSNishanth Aravamudan do { 694e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 6959a76db09SLee Schermerhorn if (page) { 69663b4613cSNishanth Aravamudan ret = 1; 6979a76db09SLee Schermerhorn break; 6989a76db09SLee Schermerhorn } 6996ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 7009a76db09SLee Schermerhorn } while (next_nid != start_nid); 70163b4613cSNishanth Aravamudan 7023b116300SAdam Litke if (ret) 7033b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7043b116300SAdam Litke else 7053b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7063b116300SAdam Litke 70763b4613cSNishanth Aravamudan return ret; 7081da177e4SLinus Torvalds } 7091da177e4SLinus Torvalds 710e8c5c824SLee Schermerhorn /* 7116ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7126ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7136ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7146ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 715e8c5c824SLee Schermerhorn */ 7166ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 717e8c5c824SLee Schermerhorn { 7186ae11b27SLee Schermerhorn int nid; 7199a76db09SLee Schermerhorn 7206ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7216ae11b27SLee Schermerhorn 7226ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7236ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7246ae11b27SLee Schermerhorn 7259a76db09SLee Schermerhorn return nid; 726e8c5c824SLee Schermerhorn } 727e8c5c824SLee Schermerhorn 728e8c5c824SLee Schermerhorn /* 729e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 730e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 731e8c5c824SLee Schermerhorn * balanced over allowed nodes. 732e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 733e8c5c824SLee Schermerhorn */ 7346ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 7356ae11b27SLee Schermerhorn bool acct_surplus) 736e8c5c824SLee Schermerhorn { 737e8c5c824SLee Schermerhorn int start_nid; 738e8c5c824SLee Schermerhorn int next_nid; 739e8c5c824SLee Schermerhorn int ret = 0; 740e8c5c824SLee Schermerhorn 7416ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 742e8c5c824SLee Schermerhorn next_nid = start_nid; 743e8c5c824SLee Schermerhorn 744e8c5c824SLee Schermerhorn do { 745685f3457SLee Schermerhorn /* 746685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 747685f3457SLee Schermerhorn * nodes with surplus pages. 748685f3457SLee Schermerhorn */ 749685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 750685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 751e8c5c824SLee Schermerhorn struct page *page = 752e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 753e8c5c824SLee Schermerhorn struct page, lru); 754e8c5c824SLee Schermerhorn list_del(&page->lru); 755e8c5c824SLee Schermerhorn h->free_huge_pages--; 756e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 757685f3457SLee Schermerhorn if (acct_surplus) { 758685f3457SLee Schermerhorn h->surplus_huge_pages--; 759685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 760685f3457SLee Schermerhorn } 761e8c5c824SLee Schermerhorn update_and_free_page(h, page); 762e8c5c824SLee Schermerhorn ret = 1; 7639a76db09SLee Schermerhorn break; 764e8c5c824SLee Schermerhorn } 7656ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 7669a76db09SLee Schermerhorn } while (next_nid != start_nid); 767e8c5c824SLee Schermerhorn 768e8c5c824SLee Schermerhorn return ret; 769e8c5c824SLee Schermerhorn } 770e8c5c824SLee Schermerhorn 771a5516438SAndi Kleen static struct page *alloc_buddy_huge_page(struct hstate *h, 772a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 7737893d1d5SAdam Litke { 7747893d1d5SAdam Litke struct page *page; 775d1c3fb1fSNishanth Aravamudan unsigned int nid; 7767893d1d5SAdam Litke 777aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 778aa888a74SAndi Kleen return NULL; 779aa888a74SAndi Kleen 780d1c3fb1fSNishanth Aravamudan /* 781d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 782d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 783d1c3fb1fSNishanth Aravamudan * overcommit 784d1c3fb1fSNishanth Aravamudan * 785d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 786d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 787d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 788d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 789d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 790d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 791d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 792d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 793d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 794d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 795d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 796d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 797d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 798d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 799d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 800d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 801d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 802d1c3fb1fSNishanth Aravamudan */ 803d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 804a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 805d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 806d1c3fb1fSNishanth Aravamudan return NULL; 807d1c3fb1fSNishanth Aravamudan } else { 808a5516438SAndi Kleen h->nr_huge_pages++; 809a5516438SAndi Kleen h->surplus_huge_pages++; 810d1c3fb1fSNishanth Aravamudan } 811d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 812d1c3fb1fSNishanth Aravamudan 813551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 814551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 815a5516438SAndi Kleen huge_page_order(h)); 816d1c3fb1fSNishanth Aravamudan 817caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 818caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 819caff3a2cSGerald Schaefer return NULL; 820caff3a2cSGerald Schaefer } 821caff3a2cSGerald Schaefer 8227893d1d5SAdam Litke spin_lock(&hugetlb_lock); 823d1c3fb1fSNishanth Aravamudan if (page) { 8242668db91SAdam Litke /* 8252668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 8262668db91SAdam Litke * no users -- drop the buddy allocator's reference. 8272668db91SAdam Litke */ 8282668db91SAdam Litke put_page_testzero(page); 8292668db91SAdam Litke VM_BUG_ON(page_count(page)); 830d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 831d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 832d1c3fb1fSNishanth Aravamudan /* 833d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 834d1c3fb1fSNishanth Aravamudan */ 835a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 836a5516438SAndi Kleen h->surplus_huge_pages_node[nid]++; 8373b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 838d1c3fb1fSNishanth Aravamudan } else { 839a5516438SAndi Kleen h->nr_huge_pages--; 840a5516438SAndi Kleen h->surplus_huge_pages--; 8413b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 8427893d1d5SAdam Litke } 843d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 8447893d1d5SAdam Litke 8457893d1d5SAdam Litke return page; 8467893d1d5SAdam Litke } 8477893d1d5SAdam Litke 848e4e574b7SAdam Litke /* 849e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 850e4e574b7SAdam Litke * of size 'delta'. 851e4e574b7SAdam Litke */ 852a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 853e4e574b7SAdam Litke { 854e4e574b7SAdam Litke struct list_head surplus_list; 855e4e574b7SAdam Litke struct page *page, *tmp; 856e4e574b7SAdam Litke int ret, i; 857e4e574b7SAdam Litke int needed, allocated; 858e4e574b7SAdam Litke 859a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 860ac09b3a1SAdam Litke if (needed <= 0) { 861a5516438SAndi Kleen h->resv_huge_pages += delta; 862e4e574b7SAdam Litke return 0; 863ac09b3a1SAdam Litke } 864e4e574b7SAdam Litke 865e4e574b7SAdam Litke allocated = 0; 866e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 867e4e574b7SAdam Litke 868e4e574b7SAdam Litke ret = -ENOMEM; 869e4e574b7SAdam Litke retry: 870e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 871e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 872a5516438SAndi Kleen page = alloc_buddy_huge_page(h, NULL, 0); 873e4e574b7SAdam Litke if (!page) { 874e4e574b7SAdam Litke /* 875e4e574b7SAdam Litke * We were not able to allocate enough pages to 876e4e574b7SAdam Litke * satisfy the entire reservation so we free what 877e4e574b7SAdam Litke * we've allocated so far. 878e4e574b7SAdam Litke */ 879e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 880e4e574b7SAdam Litke needed = 0; 881e4e574b7SAdam Litke goto free; 882e4e574b7SAdam Litke } 883e4e574b7SAdam Litke 884e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 885e4e574b7SAdam Litke } 886e4e574b7SAdam Litke allocated += needed; 887e4e574b7SAdam Litke 888e4e574b7SAdam Litke /* 889e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 890e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 891e4e574b7SAdam Litke */ 892e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 893a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 894a5516438SAndi Kleen (h->free_huge_pages + allocated); 895e4e574b7SAdam Litke if (needed > 0) 896e4e574b7SAdam Litke goto retry; 897e4e574b7SAdam Litke 898e4e574b7SAdam Litke /* 899e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 900e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 901e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 902ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 903ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 904ac09b3a1SAdam Litke * before they are reserved. 905e4e574b7SAdam Litke */ 906e4e574b7SAdam Litke needed += allocated; 907a5516438SAndi Kleen h->resv_huge_pages += delta; 908e4e574b7SAdam Litke ret = 0; 909e4e574b7SAdam Litke free: 91019fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 91119fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 91219fc3f0aSAdam Litke if ((--needed) < 0) 91319fc3f0aSAdam Litke break; 91419fc3f0aSAdam Litke list_del(&page->lru); 915a5516438SAndi Kleen enqueue_huge_page(h, page); 91619fc3f0aSAdam Litke } 91719fc3f0aSAdam Litke 91819fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 91919fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 92019fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 921e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 922e4e574b7SAdam Litke list_del(&page->lru); 923af767cbdSAdam Litke /* 9242668db91SAdam Litke * The page has a reference count of zero already, so 9252668db91SAdam Litke * call free_huge_page directly instead of using 9262668db91SAdam Litke * put_page. This must be done with hugetlb_lock 927af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 928af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 929af767cbdSAdam Litke */ 9302668db91SAdam Litke free_huge_page(page); 931af767cbdSAdam Litke } 93219fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 933e4e574b7SAdam Litke } 934e4e574b7SAdam Litke 935e4e574b7SAdam Litke return ret; 936e4e574b7SAdam Litke } 937e4e574b7SAdam Litke 938e4e574b7SAdam Litke /* 939e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 940e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 941e4e574b7SAdam Litke * never used. 942685f3457SLee Schermerhorn * Called with hugetlb_lock held. 943e4e574b7SAdam Litke */ 944a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 945a5516438SAndi Kleen unsigned long unused_resv_pages) 946e4e574b7SAdam Litke { 947e4e574b7SAdam Litke unsigned long nr_pages; 948e4e574b7SAdam Litke 949ac09b3a1SAdam Litke /* Uncommit the reservation */ 950a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 951ac09b3a1SAdam Litke 952aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 953aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 954aa888a74SAndi Kleen return; 955aa888a74SAndi Kleen 956a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 957e4e574b7SAdam Litke 958685f3457SLee Schermerhorn /* 959685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 9609b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 9619b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 9629b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 9639b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 9649b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 965685f3457SLee Schermerhorn */ 966685f3457SLee Schermerhorn while (nr_pages--) { 9679b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 968685f3457SLee Schermerhorn break; 969e4e574b7SAdam Litke } 970e4e574b7SAdam Litke } 971e4e574b7SAdam Litke 972c37f9fb1SAndy Whitcroft /* 973c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 974c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 975c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 976c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 977c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 978c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 979c37f9fb1SAndy Whitcroft * No action is required on failure. 980c37f9fb1SAndy Whitcroft */ 981e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 982a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 983c37f9fb1SAndy Whitcroft { 984c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 985c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 986c37f9fb1SAndy Whitcroft 987f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 988a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 989c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 990c37f9fb1SAndy Whitcroft idx, idx + 1); 991c37f9fb1SAndy Whitcroft 99284afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 993c37f9fb1SAndy Whitcroft return 1; 994c37f9fb1SAndy Whitcroft 99584afd99bSAndy Whitcroft } else { 996e2f17d94SRoel Kluin long err; 997a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 99884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 99984afd99bSAndy Whitcroft 100084afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 100184afd99bSAndy Whitcroft if (err < 0) 100284afd99bSAndy Whitcroft return err; 1003c37f9fb1SAndy Whitcroft return 0; 1004c37f9fb1SAndy Whitcroft } 100584afd99bSAndy Whitcroft } 1006a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1007a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1008c37f9fb1SAndy Whitcroft { 1009c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1010c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1011c37f9fb1SAndy Whitcroft 1012f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1013a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1014c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 101584afd99bSAndy Whitcroft 101684afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1017a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 101884afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 101984afd99bSAndy Whitcroft 102084afd99bSAndy Whitcroft /* Mark this page used in the map. */ 102184afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1022c37f9fb1SAndy Whitcroft } 1023c37f9fb1SAndy Whitcroft } 1024c37f9fb1SAndy Whitcroft 1025348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 102604f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1027348ea204SAdam Litke { 1028a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1029348ea204SAdam Litke struct page *page; 10302fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 1031a1e78772SMel Gorman struct inode *inode = mapping->host; 1032e2f17d94SRoel Kluin long chg; 10332fc39cecSAdam Litke 1034a1e78772SMel Gorman /* 1035a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 1036a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 1037a1e78772SMel Gorman * made before satisfying the allocation 1038c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 1039c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 1040a1e78772SMel Gorman */ 1041a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1042c37f9fb1SAndy Whitcroft if (chg < 0) 1043c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 1044c37f9fb1SAndy Whitcroft if (chg) 1045a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 1046a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 104790d8b7e6SAdam Litke 1048a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1049a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1050a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1051a1e78772SMel Gorman 1052a1e78772SMel Gorman if (!page) { 1053a5516438SAndi Kleen page = alloc_buddy_huge_page(h, vma, addr); 1054a1e78772SMel Gorman if (!page) { 1055a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 10564a6018f7SMel Gorman return ERR_PTR(-VM_FAULT_SIGBUS); 1057a1e78772SMel Gorman } 1058a1e78772SMel Gorman } 1059a1e78772SMel Gorman 1060348ea204SAdam Litke set_page_refcounted(page); 10612fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 1062a1e78772SMel Gorman 1063a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1064c37f9fb1SAndy Whitcroft 10657893d1d5SAdam Litke return page; 1066b45b5bd6SDavid Gibson } 1067b45b5bd6SDavid Gibson 106891f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1069aa888a74SAndi Kleen { 1070aa888a74SAndi Kleen struct huge_bootmem_page *m; 10719b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1072aa888a74SAndi Kleen 1073aa888a74SAndi Kleen while (nr_nodes) { 1074aa888a74SAndi Kleen void *addr; 1075aa888a74SAndi Kleen 1076aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 10776ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 10789b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1079aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1080aa888a74SAndi Kleen 1081aa888a74SAndi Kleen if (addr) { 1082aa888a74SAndi Kleen /* 1083aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1084aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1085aa888a74SAndi Kleen * puts them into the mem_map). 1086aa888a74SAndi Kleen */ 1087aa888a74SAndi Kleen m = addr; 1088aa888a74SAndi Kleen goto found; 1089aa888a74SAndi Kleen } 1090aa888a74SAndi Kleen nr_nodes--; 1091aa888a74SAndi Kleen } 1092aa888a74SAndi Kleen return 0; 1093aa888a74SAndi Kleen 1094aa888a74SAndi Kleen found: 1095aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1096aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1097aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1098aa888a74SAndi Kleen m->hstate = h; 1099aa888a74SAndi Kleen return 1; 1100aa888a74SAndi Kleen } 1101aa888a74SAndi Kleen 110218229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 110318229df5SAndy Whitcroft { 110418229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 110518229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 110618229df5SAndy Whitcroft else 110718229df5SAndy Whitcroft prep_compound_page(page, order); 110818229df5SAndy Whitcroft } 110918229df5SAndy Whitcroft 1110aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1111aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1112aa888a74SAndi Kleen { 1113aa888a74SAndi Kleen struct huge_bootmem_page *m; 1114aa888a74SAndi Kleen 1115aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1116aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1117aa888a74SAndi Kleen struct hstate *h = m->hstate; 1118aa888a74SAndi Kleen __ClearPageReserved(page); 1119aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 112018229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1121aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1122aa888a74SAndi Kleen } 1123aa888a74SAndi Kleen } 1124aa888a74SAndi Kleen 11258faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 11261da177e4SLinus Torvalds { 11271da177e4SLinus Torvalds unsigned long i; 11281da177e4SLinus Torvalds 1129e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1130aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1131aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1132aa888a74SAndi Kleen break; 11339b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 11349b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 11351da177e4SLinus Torvalds break; 11361da177e4SLinus Torvalds } 11378faa8b07SAndi Kleen h->max_huge_pages = i; 1138e5ff2159SAndi Kleen } 1139e5ff2159SAndi Kleen 1140e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1141e5ff2159SAndi Kleen { 1142e5ff2159SAndi Kleen struct hstate *h; 1143e5ff2159SAndi Kleen 1144e5ff2159SAndi Kleen for_each_hstate(h) { 11458faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 11468faa8b07SAndi Kleen if (h->order < MAX_ORDER) 11478faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1148e5ff2159SAndi Kleen } 1149e5ff2159SAndi Kleen } 1150e5ff2159SAndi Kleen 11514abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 11524abd32dbSAndi Kleen { 11534abd32dbSAndi Kleen if (n >= (1UL << 30)) 11544abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 11554abd32dbSAndi Kleen else if (n >= (1UL << 20)) 11564abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 11574abd32dbSAndi Kleen else 11584abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 11594abd32dbSAndi Kleen return buf; 11604abd32dbSAndi Kleen } 11614abd32dbSAndi Kleen 1162e5ff2159SAndi Kleen static void __init report_hugepages(void) 1163e5ff2159SAndi Kleen { 1164e5ff2159SAndi Kleen struct hstate *h; 1165e5ff2159SAndi Kleen 1166e5ff2159SAndi Kleen for_each_hstate(h) { 11674abd32dbSAndi Kleen char buf[32]; 11684abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11694abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11704abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11714abd32dbSAndi Kleen h->free_huge_pages); 1172e5ff2159SAndi Kleen } 1173e5ff2159SAndi Kleen } 1174e5ff2159SAndi Kleen 11751da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 11766ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 11776ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11781da177e4SLinus Torvalds { 11794415cc8dSChristoph Lameter int i; 11804415cc8dSChristoph Lameter 1181aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1182aa888a74SAndi Kleen return; 1183aa888a74SAndi Kleen 11846ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 11851da177e4SLinus Torvalds struct page *page, *next; 1186a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1187a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1188a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11896b0c880dSAdam Litke return; 11901da177e4SLinus Torvalds if (PageHighMem(page)) 11911da177e4SLinus Torvalds continue; 11921da177e4SLinus Torvalds list_del(&page->lru); 1193e5ff2159SAndi Kleen update_and_free_page(h, page); 1194a5516438SAndi Kleen h->free_huge_pages--; 1195a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11961da177e4SLinus Torvalds } 11971da177e4SLinus Torvalds } 11981da177e4SLinus Torvalds } 11991da177e4SLinus Torvalds #else 12006ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 12016ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12021da177e4SLinus Torvalds { 12031da177e4SLinus Torvalds } 12041da177e4SLinus Torvalds #endif 12051da177e4SLinus Torvalds 120620a0307cSWu Fengguang /* 120720a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 120820a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 120920a0307cSWu Fengguang * Returns 1 if an adjustment was made. 121020a0307cSWu Fengguang */ 12116ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 12126ae11b27SLee Schermerhorn int delta) 121320a0307cSWu Fengguang { 1214e8c5c824SLee Schermerhorn int start_nid, next_nid; 121520a0307cSWu Fengguang int ret = 0; 121620a0307cSWu Fengguang 121720a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 121820a0307cSWu Fengguang 1219e8c5c824SLee Schermerhorn if (delta < 0) 12206ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1221e8c5c824SLee Schermerhorn else 12226ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1223e8c5c824SLee Schermerhorn next_nid = start_nid; 1224e8c5c824SLee Schermerhorn 1225e8c5c824SLee Schermerhorn do { 1226e8c5c824SLee Schermerhorn int nid = next_nid; 1227e8c5c824SLee Schermerhorn if (delta < 0) { 1228e8c5c824SLee Schermerhorn /* 1229e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1230e8c5c824SLee Schermerhorn */ 12319a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 12326ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 12336ae11b27SLee Schermerhorn nodes_allowed); 123420a0307cSWu Fengguang continue; 1235e8c5c824SLee Schermerhorn } 12369a76db09SLee Schermerhorn } 1237e8c5c824SLee Schermerhorn if (delta > 0) { 1238e8c5c824SLee Schermerhorn /* 1239e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1240e8c5c824SLee Schermerhorn */ 1241e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 12429a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 12436ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 12446ae11b27SLee Schermerhorn nodes_allowed); 124520a0307cSWu Fengguang continue; 1246e8c5c824SLee Schermerhorn } 12479a76db09SLee Schermerhorn } 124820a0307cSWu Fengguang 124920a0307cSWu Fengguang h->surplus_huge_pages += delta; 125020a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 125120a0307cSWu Fengguang ret = 1; 125220a0307cSWu Fengguang break; 1253e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 125420a0307cSWu Fengguang 125520a0307cSWu Fengguang return ret; 125620a0307cSWu Fengguang } 125720a0307cSWu Fengguang 1258a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 12596ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 12606ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12611da177e4SLinus Torvalds { 12627893d1d5SAdam Litke unsigned long min_count, ret; 12631da177e4SLinus Torvalds 1264aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1265aa888a74SAndi Kleen return h->max_huge_pages; 1266aa888a74SAndi Kleen 12677893d1d5SAdam Litke /* 12687893d1d5SAdam Litke * Increase the pool size 12697893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 12707893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1271d1c3fb1fSNishanth Aravamudan * 1272d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1273d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1274d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1275d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1276d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 12777893d1d5SAdam Litke */ 12781da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1279a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 12806ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 12817893d1d5SAdam Litke break; 12827893d1d5SAdam Litke } 12837893d1d5SAdam Litke 1284a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 12857893d1d5SAdam Litke /* 12867893d1d5SAdam Litke * If this allocation races such that we no longer need the 12877893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12887893d1d5SAdam Litke * and reducing the surplus. 12897893d1d5SAdam Litke */ 12907893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 12916ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 12927893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12937893d1d5SAdam Litke if (!ret) 12947893d1d5SAdam Litke goto out; 12957893d1d5SAdam Litke 1296536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1297536240f2SMel Gorman if (signal_pending(current)) 1298536240f2SMel Gorman goto out; 12997893d1d5SAdam Litke } 13007893d1d5SAdam Litke 13017893d1d5SAdam Litke /* 13027893d1d5SAdam Litke * Decrease the pool size 13037893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 13047893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 13057893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 13067893d1d5SAdam Litke * to the desired size as pages become free. 1307d1c3fb1fSNishanth Aravamudan * 1308d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1309d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1310d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1311d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1312d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1313d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1314d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 13157893d1d5SAdam Litke */ 1316a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 13176b0c880dSAdam Litke min_count = max(count, min_count); 13186ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1319a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 13206ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 13211da177e4SLinus Torvalds break; 13221da177e4SLinus Torvalds } 1323a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 13246ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 13257893d1d5SAdam Litke break; 13267893d1d5SAdam Litke } 13277893d1d5SAdam Litke out: 1328a5516438SAndi Kleen ret = persistent_huge_pages(h); 13291da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 13307893d1d5SAdam Litke return ret; 13311da177e4SLinus Torvalds } 13321da177e4SLinus Torvalds 1333a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1334a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1335a3437870SNishanth Aravamudan 1336a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1337a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1338a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1339a3437870SNishanth Aravamudan 1340a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1341a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1342a3437870SNishanth Aravamudan 13439a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 13449a305230SLee Schermerhorn 13459a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1346a3437870SNishanth Aravamudan { 1347a3437870SNishanth Aravamudan int i; 13489a305230SLee Schermerhorn 1349a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 13509a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 13519a305230SLee Schermerhorn if (nidp) 13529a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1353a3437870SNishanth Aravamudan return &hstates[i]; 13549a305230SLee Schermerhorn } 13559a305230SLee Schermerhorn 13569a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1357a3437870SNishanth Aravamudan } 1358a3437870SNishanth Aravamudan 135906808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1360a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1361a3437870SNishanth Aravamudan { 13629a305230SLee Schermerhorn struct hstate *h; 13639a305230SLee Schermerhorn unsigned long nr_huge_pages; 13649a305230SLee Schermerhorn int nid; 13659a305230SLee Schermerhorn 13669a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13679a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 13689a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 13699a305230SLee Schermerhorn else 13709a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 13719a305230SLee Schermerhorn 13729a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1373a3437870SNishanth Aravamudan } 137406808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 137506808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 137606808b08SLee Schermerhorn const char *buf, size_t len) 1377a3437870SNishanth Aravamudan { 1378a3437870SNishanth Aravamudan int err; 13799a305230SLee Schermerhorn int nid; 138006808b08SLee Schermerhorn unsigned long count; 13819a305230SLee Schermerhorn struct hstate *h; 1382bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1383a3437870SNishanth Aravamudan 138406808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 1385a3437870SNishanth Aravamudan if (err) 1386a3437870SNishanth Aravamudan return 0; 1387a3437870SNishanth Aravamudan 13889a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13899a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 13909a305230SLee Schermerhorn /* 13919a305230SLee Schermerhorn * global hstate attribute 13929a305230SLee Schermerhorn */ 13939a305230SLee Schermerhorn if (!(obey_mempolicy && 13949a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 139506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 13969a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 139706808b08SLee Schermerhorn } 13989a305230SLee Schermerhorn } else if (nodes_allowed) { 13999a305230SLee Schermerhorn /* 14009a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 14019a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 14029a305230SLee Schermerhorn */ 14039a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 14049a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 14059a305230SLee Schermerhorn } else 14069a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 14079a305230SLee Schermerhorn 140806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1409a3437870SNishanth Aravamudan 14109b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 141106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 141206808b08SLee Schermerhorn 141306808b08SLee Schermerhorn return len; 141406808b08SLee Schermerhorn } 141506808b08SLee Schermerhorn 141606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 141706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 141806808b08SLee Schermerhorn { 141906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 142006808b08SLee Schermerhorn } 142106808b08SLee Schermerhorn 142206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 142306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 142406808b08SLee Schermerhorn { 142506808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1426a3437870SNishanth Aravamudan } 1427a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1428a3437870SNishanth Aravamudan 142906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 143006808b08SLee Schermerhorn 143106808b08SLee Schermerhorn /* 143206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 143306808b08SLee Schermerhorn * huge page alloc/free. 143406808b08SLee Schermerhorn */ 143506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 143606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 143706808b08SLee Schermerhorn { 143806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 143906808b08SLee Schermerhorn } 144006808b08SLee Schermerhorn 144106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 144206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 144306808b08SLee Schermerhorn { 144406808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 144506808b08SLee Schermerhorn } 144606808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 144706808b08SLee Schermerhorn #endif 144806808b08SLee Schermerhorn 144906808b08SLee Schermerhorn 1450a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1451a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1452a3437870SNishanth Aravamudan { 14539a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1454a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1455a3437870SNishanth Aravamudan } 1456a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1457a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1458a3437870SNishanth Aravamudan { 1459a3437870SNishanth Aravamudan int err; 1460a3437870SNishanth Aravamudan unsigned long input; 14619a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1462a3437870SNishanth Aravamudan 1463a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1464a3437870SNishanth Aravamudan if (err) 1465a3437870SNishanth Aravamudan return 0; 1466a3437870SNishanth Aravamudan 1467a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1468a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1469a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1470a3437870SNishanth Aravamudan 1471a3437870SNishanth Aravamudan return count; 1472a3437870SNishanth Aravamudan } 1473a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1474a3437870SNishanth Aravamudan 1475a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1476a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1477a3437870SNishanth Aravamudan { 14789a305230SLee Schermerhorn struct hstate *h; 14799a305230SLee Schermerhorn unsigned long free_huge_pages; 14809a305230SLee Schermerhorn int nid; 14819a305230SLee Schermerhorn 14829a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14839a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14849a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 14859a305230SLee Schermerhorn else 14869a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 14879a305230SLee Schermerhorn 14889a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1489a3437870SNishanth Aravamudan } 1490a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1491a3437870SNishanth Aravamudan 1492a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1493a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1494a3437870SNishanth Aravamudan { 14959a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1496a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1497a3437870SNishanth Aravamudan } 1498a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1499a3437870SNishanth Aravamudan 1500a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1501a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1502a3437870SNishanth Aravamudan { 15039a305230SLee Schermerhorn struct hstate *h; 15049a305230SLee Schermerhorn unsigned long surplus_huge_pages; 15059a305230SLee Schermerhorn int nid; 15069a305230SLee Schermerhorn 15079a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15089a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15099a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 15109a305230SLee Schermerhorn else 15119a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 15129a305230SLee Schermerhorn 15139a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1514a3437870SNishanth Aravamudan } 1515a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1516a3437870SNishanth Aravamudan 1517a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1518a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1519a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1520a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1521a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1522a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 152306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 152406808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 152506808b08SLee Schermerhorn #endif 1526a3437870SNishanth Aravamudan NULL, 1527a3437870SNishanth Aravamudan }; 1528a3437870SNishanth Aravamudan 1529a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1530a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1531a3437870SNishanth Aravamudan }; 1532a3437870SNishanth Aravamudan 1533094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 15349a305230SLee Schermerhorn struct kobject **hstate_kobjs, 15359a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1536a3437870SNishanth Aravamudan { 1537a3437870SNishanth Aravamudan int retval; 15389a305230SLee Schermerhorn int hi = h - hstates; 1539a3437870SNishanth Aravamudan 15409a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 15419a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1542a3437870SNishanth Aravamudan return -ENOMEM; 1543a3437870SNishanth Aravamudan 15449a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1545a3437870SNishanth Aravamudan if (retval) 15469a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1547a3437870SNishanth Aravamudan 1548a3437870SNishanth Aravamudan return retval; 1549a3437870SNishanth Aravamudan } 1550a3437870SNishanth Aravamudan 1551a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1552a3437870SNishanth Aravamudan { 1553a3437870SNishanth Aravamudan struct hstate *h; 1554a3437870SNishanth Aravamudan int err; 1555a3437870SNishanth Aravamudan 1556a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1557a3437870SNishanth Aravamudan if (!hugepages_kobj) 1558a3437870SNishanth Aravamudan return; 1559a3437870SNishanth Aravamudan 1560a3437870SNishanth Aravamudan for_each_hstate(h) { 15619a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 15629a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1563a3437870SNishanth Aravamudan if (err) 1564a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1565a3437870SNishanth Aravamudan h->name); 1566a3437870SNishanth Aravamudan } 1567a3437870SNishanth Aravamudan } 1568a3437870SNishanth Aravamudan 15699a305230SLee Schermerhorn #ifdef CONFIG_NUMA 15709a305230SLee Schermerhorn 15719a305230SLee Schermerhorn /* 15729a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 15739a305230SLee Schermerhorn * with node sysdevs in node_devices[] using a parallel array. The array 15749a305230SLee Schermerhorn * index of a node sysdev or _hstate == node id. 15759a305230SLee Schermerhorn * This is here to avoid any static dependency of the node sysdev driver, in 15769a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 15779a305230SLee Schermerhorn */ 15789a305230SLee Schermerhorn struct node_hstate { 15799a305230SLee Schermerhorn struct kobject *hugepages_kobj; 15809a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 15819a305230SLee Schermerhorn }; 15829a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 15839a305230SLee Schermerhorn 15849a305230SLee Schermerhorn /* 15859a305230SLee Schermerhorn * A subset of global hstate attributes for node sysdevs 15869a305230SLee Schermerhorn */ 15879a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 15889a305230SLee Schermerhorn &nr_hugepages_attr.attr, 15899a305230SLee Schermerhorn &free_hugepages_attr.attr, 15909a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 15919a305230SLee Schermerhorn NULL, 15929a305230SLee Schermerhorn }; 15939a305230SLee Schermerhorn 15949a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 15959a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 15969a305230SLee Schermerhorn }; 15979a305230SLee Schermerhorn 15989a305230SLee Schermerhorn /* 15999a305230SLee Schermerhorn * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. 16009a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 16019a305230SLee Schermerhorn */ 16029a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 16039a305230SLee Schermerhorn { 16049a305230SLee Schermerhorn int nid; 16059a305230SLee Schermerhorn 16069a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 16079a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 16089a305230SLee Schermerhorn int i; 16099a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 16109a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 16119a305230SLee Schermerhorn if (nidp) 16129a305230SLee Schermerhorn *nidp = nid; 16139a305230SLee Schermerhorn return &hstates[i]; 16149a305230SLee Schermerhorn } 16159a305230SLee Schermerhorn } 16169a305230SLee Schermerhorn 16179a305230SLee Schermerhorn BUG(); 16189a305230SLee Schermerhorn return NULL; 16199a305230SLee Schermerhorn } 16209a305230SLee Schermerhorn 16219a305230SLee Schermerhorn /* 16229a305230SLee Schermerhorn * Unregister hstate attributes from a single node sysdev. 16239a305230SLee Schermerhorn * No-op if no hstate attributes attached. 16249a305230SLee Schermerhorn */ 16259a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 16269a305230SLee Schermerhorn { 16279a305230SLee Schermerhorn struct hstate *h; 16289a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16299a305230SLee Schermerhorn 16309a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16319b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 16329a305230SLee Schermerhorn 16339a305230SLee Schermerhorn for_each_hstate(h) 16349a305230SLee Schermerhorn if (nhs->hstate_kobjs[h - hstates]) { 16359a305230SLee Schermerhorn kobject_put(nhs->hstate_kobjs[h - hstates]); 16369a305230SLee Schermerhorn nhs->hstate_kobjs[h - hstates] = NULL; 16379a305230SLee Schermerhorn } 16389a305230SLee Schermerhorn 16399a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 16409a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 16419a305230SLee Schermerhorn } 16429a305230SLee Schermerhorn 16439a305230SLee Schermerhorn /* 16449a305230SLee Schermerhorn * hugetlb module exit: unregister hstate attributes from node sysdevs 16459a305230SLee Schermerhorn * that have them. 16469a305230SLee Schermerhorn */ 16479a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 16489a305230SLee Schermerhorn { 16499a305230SLee Schermerhorn int nid; 16509a305230SLee Schermerhorn 16519a305230SLee Schermerhorn /* 16529a305230SLee Schermerhorn * disable node sysdev registrations. 16539a305230SLee Schermerhorn */ 16549a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 16559a305230SLee Schermerhorn 16569a305230SLee Schermerhorn /* 16579a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 16589a305230SLee Schermerhorn */ 16599a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 16609a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 16619a305230SLee Schermerhorn } 16629a305230SLee Schermerhorn 16639a305230SLee Schermerhorn /* 16649a305230SLee Schermerhorn * Register hstate attributes for a single node sysdev. 16659a305230SLee Schermerhorn * No-op if attributes already registered. 16669a305230SLee Schermerhorn */ 16679a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 16689a305230SLee Schermerhorn { 16699a305230SLee Schermerhorn struct hstate *h; 16709a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16719a305230SLee Schermerhorn int err; 16729a305230SLee Schermerhorn 16739a305230SLee Schermerhorn if (nhs->hugepages_kobj) 16749a305230SLee Schermerhorn return; /* already allocated */ 16759a305230SLee Schermerhorn 16769a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 16779a305230SLee Schermerhorn &node->sysdev.kobj); 16789a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16799a305230SLee Schermerhorn return; 16809a305230SLee Schermerhorn 16819a305230SLee Schermerhorn for_each_hstate(h) { 16829a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 16839a305230SLee Schermerhorn nhs->hstate_kobjs, 16849a305230SLee Schermerhorn &per_node_hstate_attr_group); 16859a305230SLee Schermerhorn if (err) { 16869a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 16879a305230SLee Schermerhorn " for node %d\n", 16889a305230SLee Schermerhorn h->name, node->sysdev.id); 16899a305230SLee Schermerhorn hugetlb_unregister_node(node); 16909a305230SLee Schermerhorn break; 16919a305230SLee Schermerhorn } 16929a305230SLee Schermerhorn } 16939a305230SLee Schermerhorn } 16949a305230SLee Schermerhorn 16959a305230SLee Schermerhorn /* 16969b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 16979b5e5d0fSLee Schermerhorn * sysdevs of nodes that have memory. All on-line nodes should have 16989b5e5d0fSLee Schermerhorn * registered their associated sysdev by this time. 16999a305230SLee Schermerhorn */ 17009a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 17019a305230SLee Schermerhorn { 17029a305230SLee Schermerhorn int nid; 17039a305230SLee Schermerhorn 17049b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 17059a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 17069a305230SLee Schermerhorn if (node->sysdev.id == nid) 17079a305230SLee Schermerhorn hugetlb_register_node(node); 17089a305230SLee Schermerhorn } 17099a305230SLee Schermerhorn 17109a305230SLee Schermerhorn /* 17119a305230SLee Schermerhorn * Let the node sysdev driver know we're here so it can 17129a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 17139a305230SLee Schermerhorn */ 17149a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 17159a305230SLee Schermerhorn hugetlb_unregister_node); 17169a305230SLee Schermerhorn } 17179a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 17189a305230SLee Schermerhorn 17199a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17209a305230SLee Schermerhorn { 17219a305230SLee Schermerhorn BUG(); 17229a305230SLee Schermerhorn if (nidp) 17239a305230SLee Schermerhorn *nidp = -1; 17249a305230SLee Schermerhorn return NULL; 17259a305230SLee Schermerhorn } 17269a305230SLee Schermerhorn 17279a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 17289a305230SLee Schermerhorn 17299a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 17309a305230SLee Schermerhorn 17319a305230SLee Schermerhorn #endif 17329a305230SLee Schermerhorn 1733a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1734a3437870SNishanth Aravamudan { 1735a3437870SNishanth Aravamudan struct hstate *h; 1736a3437870SNishanth Aravamudan 17379a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 17389a305230SLee Schermerhorn 1739a3437870SNishanth Aravamudan for_each_hstate(h) { 1740a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1741a3437870SNishanth Aravamudan } 1742a3437870SNishanth Aravamudan 1743a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1744a3437870SNishanth Aravamudan } 1745a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1746a3437870SNishanth Aravamudan 1747a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1748a3437870SNishanth Aravamudan { 17490ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 17500ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 17510ef89d25SBenjamin Herrenschmidt * there is no such support 17520ef89d25SBenjamin Herrenschmidt */ 17530ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 17540ef89d25SBenjamin Herrenschmidt return 0; 1755a3437870SNishanth Aravamudan 1756e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1757e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1758e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1759a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1760a3437870SNishanth Aravamudan } 1761e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1762e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1763e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1764a3437870SNishanth Aravamudan 1765a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1766a3437870SNishanth Aravamudan 1767aa888a74SAndi Kleen gather_bootmem_prealloc(); 1768aa888a74SAndi Kleen 1769a3437870SNishanth Aravamudan report_hugepages(); 1770a3437870SNishanth Aravamudan 1771a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1772a3437870SNishanth Aravamudan 17739a305230SLee Schermerhorn hugetlb_register_all_nodes(); 17749a305230SLee Schermerhorn 1775a3437870SNishanth Aravamudan return 0; 1776a3437870SNishanth Aravamudan } 1777a3437870SNishanth Aravamudan module_init(hugetlb_init); 1778a3437870SNishanth Aravamudan 1779a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1780a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1781a3437870SNishanth Aravamudan { 1782a3437870SNishanth Aravamudan struct hstate *h; 17838faa8b07SAndi Kleen unsigned long i; 17848faa8b07SAndi Kleen 1785a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1786a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1787a3437870SNishanth Aravamudan return; 1788a3437870SNishanth Aravamudan } 1789a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1790a3437870SNishanth Aravamudan BUG_ON(order == 0); 1791a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1792a3437870SNishanth Aravamudan h->order = order; 1793a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 17948faa8b07SAndi Kleen h->nr_huge_pages = 0; 17958faa8b07SAndi Kleen h->free_huge_pages = 0; 17968faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 17978faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 17989b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 17999b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1800a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1801a3437870SNishanth Aravamudan huge_page_size(h)/1024); 18028faa8b07SAndi Kleen 1803a3437870SNishanth Aravamudan parsed_hstate = h; 1804a3437870SNishanth Aravamudan } 1805a3437870SNishanth Aravamudan 1806e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1807a3437870SNishanth Aravamudan { 1808a3437870SNishanth Aravamudan unsigned long *mhp; 18098faa8b07SAndi Kleen static unsigned long *last_mhp; 1810a3437870SNishanth Aravamudan 1811a3437870SNishanth Aravamudan /* 1812a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1813a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1814a3437870SNishanth Aravamudan */ 1815a3437870SNishanth Aravamudan if (!max_hstate) 1816a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1817a3437870SNishanth Aravamudan else 1818a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1819a3437870SNishanth Aravamudan 18208faa8b07SAndi Kleen if (mhp == last_mhp) { 18218faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 18228faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 18238faa8b07SAndi Kleen return 1; 18248faa8b07SAndi Kleen } 18258faa8b07SAndi Kleen 1826a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1827a3437870SNishanth Aravamudan *mhp = 0; 1828a3437870SNishanth Aravamudan 18298faa8b07SAndi Kleen /* 18308faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 18318faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 18328faa8b07SAndi Kleen * use the bootmem allocator. 18338faa8b07SAndi Kleen */ 18348faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 18358faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 18368faa8b07SAndi Kleen 18378faa8b07SAndi Kleen last_mhp = mhp; 18388faa8b07SAndi Kleen 1839a3437870SNishanth Aravamudan return 1; 1840a3437870SNishanth Aravamudan } 1841e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1842e11bfbfcSNick Piggin 1843e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1844e11bfbfcSNick Piggin { 1845e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1846e11bfbfcSNick Piggin return 1; 1847e11bfbfcSNick Piggin } 1848e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1849a3437870SNishanth Aravamudan 18508a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 18518a213460SNishanth Aravamudan { 18528a213460SNishanth Aravamudan int node; 18538a213460SNishanth Aravamudan unsigned int nr = 0; 18548a213460SNishanth Aravamudan 18558a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 18568a213460SNishanth Aravamudan nr += array[node]; 18578a213460SNishanth Aravamudan 18588a213460SNishanth Aravamudan return nr; 18598a213460SNishanth Aravamudan } 18608a213460SNishanth Aravamudan 18618a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 186206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 186306808b08SLee Schermerhorn struct ctl_table *table, int write, 186406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 18651da177e4SLinus Torvalds { 1866e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1867e5ff2159SAndi Kleen unsigned long tmp; 1868e5ff2159SAndi Kleen 1869e5ff2159SAndi Kleen if (!write) 1870e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1871e5ff2159SAndi Kleen 1872e5ff2159SAndi Kleen table->data = &tmp; 1873e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 18748d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1875e5ff2159SAndi Kleen 187606808b08SLee Schermerhorn if (write) { 1877bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 1878bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 187906808b08SLee Schermerhorn if (!(obey_mempolicy && 188006808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 188106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 188206808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 188306808b08SLee Schermerhorn } 188406808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 188506808b08SLee Schermerhorn 188606808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 188706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 188806808b08SLee Schermerhorn } 1889e5ff2159SAndi Kleen 18901da177e4SLinus Torvalds return 0; 18911da177e4SLinus Torvalds } 1892396faf03SMel Gorman 189306808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 189406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 189506808b08SLee Schermerhorn { 189606808b08SLee Schermerhorn 189706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 189806808b08SLee Schermerhorn buffer, length, ppos); 189906808b08SLee Schermerhorn } 190006808b08SLee Schermerhorn 190106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 190206808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 190306808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 190406808b08SLee Schermerhorn { 190506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 190606808b08SLee Schermerhorn buffer, length, ppos); 190706808b08SLee Schermerhorn } 190806808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 190906808b08SLee Schermerhorn 1910396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 19118d65af78SAlexey Dobriyan void __user *buffer, 1912396faf03SMel Gorman size_t *length, loff_t *ppos) 1913396faf03SMel Gorman { 19148d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 1915396faf03SMel Gorman if (hugepages_treat_as_movable) 1916396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1917396faf03SMel Gorman else 1918396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1919396faf03SMel Gorman return 0; 1920396faf03SMel Gorman } 1921396faf03SMel Gorman 1922a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 19238d65af78SAlexey Dobriyan void __user *buffer, 1924a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1925a3d0c6aaSNishanth Aravamudan { 1926a5516438SAndi Kleen struct hstate *h = &default_hstate; 1927e5ff2159SAndi Kleen unsigned long tmp; 1928e5ff2159SAndi Kleen 1929e5ff2159SAndi Kleen if (!write) 1930e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1931e5ff2159SAndi Kleen 1932e5ff2159SAndi Kleen table->data = &tmp; 1933e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 19348d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1935e5ff2159SAndi Kleen 1936e5ff2159SAndi Kleen if (write) { 1937064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1938e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1939a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1940e5ff2159SAndi Kleen } 1941e5ff2159SAndi Kleen 1942a3d0c6aaSNishanth Aravamudan return 0; 1943a3d0c6aaSNishanth Aravamudan } 1944a3d0c6aaSNishanth Aravamudan 19451da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 19461da177e4SLinus Torvalds 1947e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 19481da177e4SLinus Torvalds { 1949a5516438SAndi Kleen struct hstate *h = &default_hstate; 1950e1759c21SAlexey Dobriyan seq_printf(m, 19511da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 19521da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1953b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 19547893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 19554f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1956a5516438SAndi Kleen h->nr_huge_pages, 1957a5516438SAndi Kleen h->free_huge_pages, 1958a5516438SAndi Kleen h->resv_huge_pages, 1959a5516438SAndi Kleen h->surplus_huge_pages, 1960a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 19611da177e4SLinus Torvalds } 19621da177e4SLinus Torvalds 19631da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 19641da177e4SLinus Torvalds { 1965a5516438SAndi Kleen struct hstate *h = &default_hstate; 19661da177e4SLinus Torvalds return sprintf(buf, 19671da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1968a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1969a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1970a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1971a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1972a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 19731da177e4SLinus Torvalds } 19741da177e4SLinus Torvalds 19751da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 19761da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 19771da177e4SLinus Torvalds { 1978a5516438SAndi Kleen struct hstate *h = &default_hstate; 1979a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 19801da177e4SLinus Torvalds } 19811da177e4SLinus Torvalds 1982a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 1983fc1b8a73SMel Gorman { 1984fc1b8a73SMel Gorman int ret = -ENOMEM; 1985fc1b8a73SMel Gorman 1986fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 1987fc1b8a73SMel Gorman /* 1988fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 1989fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 1990fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 1991fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 1992fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 1993fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 1994fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 1995fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 1996fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 1997fc1b8a73SMel Gorman * 1998fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 1999fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2000fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2001fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2002fc1b8a73SMel Gorman * semantics that cpuset has. 2003fc1b8a73SMel Gorman */ 2004fc1b8a73SMel Gorman if (delta > 0) { 2005a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2006fc1b8a73SMel Gorman goto out; 2007fc1b8a73SMel Gorman 2008a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2009a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2010fc1b8a73SMel Gorman goto out; 2011fc1b8a73SMel Gorman } 2012fc1b8a73SMel Gorman } 2013fc1b8a73SMel Gorman 2014fc1b8a73SMel Gorman ret = 0; 2015fc1b8a73SMel Gorman if (delta < 0) 2016a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2017fc1b8a73SMel Gorman 2018fc1b8a73SMel Gorman out: 2019fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2020fc1b8a73SMel Gorman return ret; 2021fc1b8a73SMel Gorman } 2022fc1b8a73SMel Gorman 202384afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 202484afd99bSAndy Whitcroft { 202584afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 202684afd99bSAndy Whitcroft 202784afd99bSAndy Whitcroft /* 202884afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 202984afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 203084afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 203184afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 203284afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 203384afd99bSAndy Whitcroft * new reference here without additional locking. 203484afd99bSAndy Whitcroft */ 203584afd99bSAndy Whitcroft if (reservations) 203684afd99bSAndy Whitcroft kref_get(&reservations->refs); 203784afd99bSAndy Whitcroft } 203884afd99bSAndy Whitcroft 2039a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2040a1e78772SMel Gorman { 2041a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 204284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 204384afd99bSAndy Whitcroft unsigned long reserve; 204484afd99bSAndy Whitcroft unsigned long start; 204584afd99bSAndy Whitcroft unsigned long end; 204684afd99bSAndy Whitcroft 204784afd99bSAndy Whitcroft if (reservations) { 2048a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2049a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 205084afd99bSAndy Whitcroft 205184afd99bSAndy Whitcroft reserve = (end - start) - 205284afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 205384afd99bSAndy Whitcroft 205484afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 205584afd99bSAndy Whitcroft 20567251ff78SAdam Litke if (reserve) { 2057a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 20587251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 20597251ff78SAdam Litke } 2060a1e78772SMel Gorman } 206184afd99bSAndy Whitcroft } 2062a1e78772SMel Gorman 20631da177e4SLinus Torvalds /* 20641da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 20651da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 20661da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 20671da177e4SLinus Torvalds * this far. 20681da177e4SLinus Torvalds */ 2069d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 20701da177e4SLinus Torvalds { 20711da177e4SLinus Torvalds BUG(); 2072d0217ac0SNick Piggin return 0; 20731da177e4SLinus Torvalds } 20741da177e4SLinus Torvalds 2075f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2076d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 207784afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2078a1e78772SMel Gorman .close = hugetlb_vm_op_close, 20791da177e4SLinus Torvalds }; 20801da177e4SLinus Torvalds 20811e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 20821e8f889bSDavid Gibson int writable) 208363551ae0SDavid Gibson { 208463551ae0SDavid Gibson pte_t entry; 208563551ae0SDavid Gibson 20861e8f889bSDavid Gibson if (writable) { 208763551ae0SDavid Gibson entry = 208863551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 208963551ae0SDavid Gibson } else { 20907f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 209163551ae0SDavid Gibson } 209263551ae0SDavid Gibson entry = pte_mkyoung(entry); 209363551ae0SDavid Gibson entry = pte_mkhuge(entry); 209463551ae0SDavid Gibson 209563551ae0SDavid Gibson return entry; 209663551ae0SDavid Gibson } 209763551ae0SDavid Gibson 20981e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 20991e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 21001e8f889bSDavid Gibson { 21011e8f889bSDavid Gibson pte_t entry; 21021e8f889bSDavid Gibson 21037f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 21047f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 21054b3073e1SRussell King update_mmu_cache(vma, address, ptep); 21061e8f889bSDavid Gibson } 21078dab5241SBenjamin Herrenschmidt } 21081e8f889bSDavid Gibson 21091e8f889bSDavid Gibson 211063551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 211163551ae0SDavid Gibson struct vm_area_struct *vma) 211263551ae0SDavid Gibson { 211363551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 211463551ae0SDavid Gibson struct page *ptepage; 21151c59827dSHugh Dickins unsigned long addr; 21161e8f889bSDavid Gibson int cow; 2117a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2118a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 21191e8f889bSDavid Gibson 21201e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 212163551ae0SDavid Gibson 2122a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2123c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2124c74df32cSHugh Dickins if (!src_pte) 2125c74df32cSHugh Dickins continue; 2126a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 212763551ae0SDavid Gibson if (!dst_pte) 212863551ae0SDavid Gibson goto nomem; 2129c5c99429SLarry Woodman 2130c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2131c5c99429SLarry Woodman if (dst_pte == src_pte) 2132c5c99429SLarry Woodman continue; 2133c5c99429SLarry Woodman 2134c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 213546478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 21367f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 21371e8f889bSDavid Gibson if (cow) 21387f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 21397f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 214063551ae0SDavid Gibson ptepage = pte_page(entry); 214163551ae0SDavid Gibson get_page(ptepage); 21420fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 214363551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 21441c59827dSHugh Dickins } 21451c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2146c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 214763551ae0SDavid Gibson } 214863551ae0SDavid Gibson return 0; 214963551ae0SDavid Gibson 215063551ae0SDavid Gibson nomem: 215163551ae0SDavid Gibson return -ENOMEM; 215263551ae0SDavid Gibson } 215363551ae0SDavid Gibson 2154fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2155fd6a03edSNaoya Horiguchi { 2156fd6a03edSNaoya Horiguchi swp_entry_t swp; 2157fd6a03edSNaoya Horiguchi 2158fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2159fd6a03edSNaoya Horiguchi return 0; 2160fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 2161fd6a03edSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) { 2162fd6a03edSNaoya Horiguchi return 1; 2163fd6a03edSNaoya Horiguchi } else 2164fd6a03edSNaoya Horiguchi return 0; 2165fd6a03edSNaoya Horiguchi } 2166fd6a03edSNaoya Horiguchi 2167502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 216804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 216963551ae0SDavid Gibson { 217063551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 217163551ae0SDavid Gibson unsigned long address; 2172c7546f8fSDavid Gibson pte_t *ptep; 217363551ae0SDavid Gibson pte_t pte; 217463551ae0SDavid Gibson struct page *page; 2175fe1668aeSChen, Kenneth W struct page *tmp; 2176a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2177a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2178a5516438SAndi Kleen 2179c0a499c2SChen, Kenneth W /* 2180c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 2181c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 2182c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 2183c0a499c2SChen, Kenneth W */ 2184fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 218563551ae0SDavid Gibson 218663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2187a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2188a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 218963551ae0SDavid Gibson 2190cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 2191508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2192a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2193c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2194c7546f8fSDavid Gibson if (!ptep) 2195c7546f8fSDavid Gibson continue; 2196c7546f8fSDavid Gibson 219739dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 219839dde65cSChen, Kenneth W continue; 219939dde65cSChen, Kenneth W 220004f2cbe3SMel Gorman /* 220104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 220204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 220304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 220404f2cbe3SMel Gorman */ 220504f2cbe3SMel Gorman if (ref_page) { 220604f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 220704f2cbe3SMel Gorman if (huge_pte_none(pte)) 220804f2cbe3SMel Gorman continue; 220904f2cbe3SMel Gorman page = pte_page(pte); 221004f2cbe3SMel Gorman if (page != ref_page) 221104f2cbe3SMel Gorman continue; 221204f2cbe3SMel Gorman 221304f2cbe3SMel Gorman /* 221404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 221504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 221604f2cbe3SMel Gorman * looking like data was lost 221704f2cbe3SMel Gorman */ 221804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 221904f2cbe3SMel Gorman } 222004f2cbe3SMel Gorman 2221c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 22227f2e9525SGerald Schaefer if (huge_pte_none(pte)) 222363551ae0SDavid Gibson continue; 2224c7546f8fSDavid Gibson 2225fd6a03edSNaoya Horiguchi /* 2226fd6a03edSNaoya Horiguchi * HWPoisoned hugepage is already unmapped and dropped reference 2227fd6a03edSNaoya Horiguchi */ 2228fd6a03edSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 2229fd6a03edSNaoya Horiguchi continue; 2230fd6a03edSNaoya Horiguchi 223163551ae0SDavid Gibson page = pte_page(pte); 22326649a386SKen Chen if (pte_dirty(pte)) 22336649a386SKen Chen set_page_dirty(page); 2234fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 223563551ae0SDavid Gibson } 22361da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 2237508034a3SHugh Dickins flush_tlb_range(vma, start, end); 2238cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 2239fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 22400fe6e20bSNaoya Horiguchi page_remove_rmap(page); 2241fe1668aeSChen, Kenneth W list_del(&page->lru); 2242fe1668aeSChen, Kenneth W put_page(page); 2243fe1668aeSChen, Kenneth W } 22441da177e4SLinus Torvalds } 224563551ae0SDavid Gibson 2246502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 224704f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2248502717f4SChen, Kenneth W { 2249502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 225004f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 2251502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2252502717f4SChen, Kenneth W } 2253502717f4SChen, Kenneth W 225404f2cbe3SMel Gorman /* 225504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 225604f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 225704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 225804f2cbe3SMel Gorman * same region. 225904f2cbe3SMel Gorman */ 22602a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 22612a4b3dedSHarvey Harrison struct page *page, unsigned long address) 226204f2cbe3SMel Gorman { 22637526674dSAdam Litke struct hstate *h = hstate_vma(vma); 226404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 226504f2cbe3SMel Gorman struct address_space *mapping; 226604f2cbe3SMel Gorman struct prio_tree_iter iter; 226704f2cbe3SMel Gorman pgoff_t pgoff; 226804f2cbe3SMel Gorman 226904f2cbe3SMel Gorman /* 227004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 227104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 227204f2cbe3SMel Gorman */ 22737526674dSAdam Litke address = address & huge_page_mask(h); 227404f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 227504f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 227604f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 227704f2cbe3SMel Gorman 22784eb2b1dcSMel Gorman /* 22794eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 22804eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 22814eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 22824eb2b1dcSMel Gorman */ 22834eb2b1dcSMel Gorman spin_lock(&mapping->i_mmap_lock); 228404f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 228504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 228604f2cbe3SMel Gorman if (iter_vma == vma) 228704f2cbe3SMel Gorman continue; 228804f2cbe3SMel Gorman 228904f2cbe3SMel Gorman /* 229004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 229104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 229204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 229304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 229404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 229504f2cbe3SMel Gorman */ 229604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 22974eb2b1dcSMel Gorman __unmap_hugepage_range(iter_vma, 22987526674dSAdam Litke address, address + huge_page_size(h), 229904f2cbe3SMel Gorman page); 230004f2cbe3SMel Gorman } 23014eb2b1dcSMel Gorman spin_unlock(&mapping->i_mmap_lock); 230204f2cbe3SMel Gorman 230304f2cbe3SMel Gorman return 1; 230404f2cbe3SMel Gorman } 230504f2cbe3SMel Gorman 23060fe6e20bSNaoya Horiguchi /* 23070fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 23080fe6e20bSNaoya Horiguchi */ 23091e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 231004f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 231104f2cbe3SMel Gorman struct page *pagecache_page) 23121e8f889bSDavid Gibson { 2313a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 23141e8f889bSDavid Gibson struct page *old_page, *new_page; 231579ac6ba4SDavid Gibson int avoidcopy; 231604f2cbe3SMel Gorman int outside_reserve = 0; 23171e8f889bSDavid Gibson 23181e8f889bSDavid Gibson old_page = pte_page(pte); 23191e8f889bSDavid Gibson 232004f2cbe3SMel Gorman retry_avoidcopy: 23211e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 23221e8f889bSDavid Gibson * and just make the page writable */ 23230fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 23241e8f889bSDavid Gibson if (avoidcopy) { 23250fe6e20bSNaoya Horiguchi if (!trylock_page(old_page)) 23260fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 23270fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 23281e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 232983c54070SNick Piggin return 0; 23301e8f889bSDavid Gibson } 23311e8f889bSDavid Gibson 233204f2cbe3SMel Gorman /* 233304f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 233404f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 233504f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 233604f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 233704f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 233804f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 233904f2cbe3SMel Gorman * of the full address range. 234004f2cbe3SMel Gorman */ 2341f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 234204f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 234304f2cbe3SMel Gorman old_page != pagecache_page) 234404f2cbe3SMel Gorman outside_reserve = 1; 234504f2cbe3SMel Gorman 23461e8f889bSDavid Gibson page_cache_get(old_page); 2347b76c8cfbSLarry Woodman 2348b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2349b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 235004f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 23511e8f889bSDavid Gibson 23522fc39cecSAdam Litke if (IS_ERR(new_page)) { 23531e8f889bSDavid Gibson page_cache_release(old_page); 235404f2cbe3SMel Gorman 235504f2cbe3SMel Gorman /* 235604f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 235704f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 235804f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 235904f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 236004f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 236104f2cbe3SMel Gorman */ 236204f2cbe3SMel Gorman if (outside_reserve) { 236304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 236404f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 236504f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 236604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2367b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 236804f2cbe3SMel Gorman goto retry_avoidcopy; 236904f2cbe3SMel Gorman } 237004f2cbe3SMel Gorman WARN_ON_ONCE(1); 237104f2cbe3SMel Gorman } 237204f2cbe3SMel Gorman 2373b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2374b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 23752fc39cecSAdam Litke return -PTR_ERR(new_page); 23761e8f889bSDavid Gibson } 23771e8f889bSDavid Gibson 23780fe6e20bSNaoya Horiguchi /* 23790fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 23800fe6e20bSNaoya Horiguchi * anon_vma prepared. 23810fe6e20bSNaoya Horiguchi */ 23820fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) 23830fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 23840fe6e20bSNaoya Horiguchi 23859de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 23860ed361deSNick Piggin __SetPageUptodate(new_page); 23871e8f889bSDavid Gibson 2388b76c8cfbSLarry Woodman /* 2389b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2390b76c8cfbSLarry Woodman * before the page tables are altered 2391b76c8cfbSLarry Woodman */ 2392b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2393a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 23947f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 23951e8f889bSDavid Gibson /* Break COW */ 23968fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 23971e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 23981e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 23990fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 24000fe6e20bSNaoya Horiguchi hugepage_add_anon_rmap(new_page, vma, address); 24011e8f889bSDavid Gibson /* Make the old page be freed below */ 24021e8f889bSDavid Gibson new_page = old_page; 24031e8f889bSDavid Gibson } 24041e8f889bSDavid Gibson page_cache_release(new_page); 24051e8f889bSDavid Gibson page_cache_release(old_page); 240683c54070SNick Piggin return 0; 24071e8f889bSDavid Gibson } 24081e8f889bSDavid Gibson 240904f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2410a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2411a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 241204f2cbe3SMel Gorman { 241304f2cbe3SMel Gorman struct address_space *mapping; 2414e7c4b0bfSAndy Whitcroft pgoff_t idx; 241504f2cbe3SMel Gorman 241604f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2417a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 241804f2cbe3SMel Gorman 241904f2cbe3SMel Gorman return find_lock_page(mapping, idx); 242004f2cbe3SMel Gorman } 242104f2cbe3SMel Gorman 24223ae77f43SHugh Dickins /* 24233ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 24243ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 24253ae77f43SHugh Dickins */ 24263ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 24272a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 24282a15efc9SHugh Dickins { 24292a15efc9SHugh Dickins struct address_space *mapping; 24302a15efc9SHugh Dickins pgoff_t idx; 24312a15efc9SHugh Dickins struct page *page; 24322a15efc9SHugh Dickins 24332a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 24342a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 24352a15efc9SHugh Dickins 24362a15efc9SHugh Dickins page = find_get_page(mapping, idx); 24372a15efc9SHugh Dickins if (page) 24382a15efc9SHugh Dickins put_page(page); 24392a15efc9SHugh Dickins return page != NULL; 24402a15efc9SHugh Dickins } 24412a15efc9SHugh Dickins 2442a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2443788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2444ac9b9c66SHugh Dickins { 2445a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2446ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2447e7c4b0bfSAndy Whitcroft pgoff_t idx; 24484c887265SAdam Litke unsigned long size; 24494c887265SAdam Litke struct page *page; 24504c887265SAdam Litke struct address_space *mapping; 24511e8f889bSDavid Gibson pte_t new_pte; 24524c887265SAdam Litke 245304f2cbe3SMel Gorman /* 245404f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 245504f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 245604f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 245704f2cbe3SMel Gorman */ 245804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 245904f2cbe3SMel Gorman printk(KERN_WARNING 246004f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 246104f2cbe3SMel Gorman current->pid); 246204f2cbe3SMel Gorman return ret; 246304f2cbe3SMel Gorman } 246404f2cbe3SMel Gorman 24654c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2466a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 24674c887265SAdam Litke 24684c887265SAdam Litke /* 24694c887265SAdam Litke * Use page lock to guard against racing truncation 24704c887265SAdam Litke * before we get page_table_lock. 24714c887265SAdam Litke */ 24726bda666aSChristoph Lameter retry: 24736bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 24746bda666aSChristoph Lameter if (!page) { 2475a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2476ebed4bfcSHugh Dickins if (idx >= size) 2477ebed4bfcSHugh Dickins goto out; 247804f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 24792fc39cecSAdam Litke if (IS_ERR(page)) { 24802fc39cecSAdam Litke ret = -PTR_ERR(page); 24816bda666aSChristoph Lameter goto out; 24826bda666aSChristoph Lameter } 2483a5516438SAndi Kleen clear_huge_page(page, address, huge_page_size(h)); 24840ed361deSNick Piggin __SetPageUptodate(page); 2485ac9b9c66SHugh Dickins 2486f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 24876bda666aSChristoph Lameter int err; 248845c682a6SKen Chen struct inode *inode = mapping->host; 24896bda666aSChristoph Lameter 24906bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 24916bda666aSChristoph Lameter if (err) { 24926bda666aSChristoph Lameter put_page(page); 24936bda666aSChristoph Lameter if (err == -EEXIST) 24946bda666aSChristoph Lameter goto retry; 24956bda666aSChristoph Lameter goto out; 24966bda666aSChristoph Lameter } 249745c682a6SKen Chen 249845c682a6SKen Chen spin_lock(&inode->i_lock); 2499a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 250045c682a6SKen Chen spin_unlock(&inode->i_lock); 25010fe6e20bSNaoya Horiguchi page_dup_rmap(page); 250223be7468SMel Gorman } else { 25036bda666aSChristoph Lameter lock_page(page); 25040fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 25050fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 25060fe6e20bSNaoya Horiguchi goto backout_unlocked; 250723be7468SMel Gorman } 25080fe6e20bSNaoya Horiguchi hugepage_add_new_anon_rmap(page, vma, address); 25090fe6e20bSNaoya Horiguchi } 25100fe6e20bSNaoya Horiguchi } else { 25110fe6e20bSNaoya Horiguchi page_dup_rmap(page); 25126bda666aSChristoph Lameter } 25131e8f889bSDavid Gibson 251457303d80SAndy Whitcroft /* 2515fd6a03edSNaoya Horiguchi * Since memory error handler replaces pte into hwpoison swap entry 2516fd6a03edSNaoya Horiguchi * at the time of error handling, a process which reserved but not have 2517fd6a03edSNaoya Horiguchi * the mapping to the error hugepage does not have hwpoison swap entry. 2518fd6a03edSNaoya Horiguchi * So we need to block accesses from such a process by checking 2519fd6a03edSNaoya Horiguchi * PG_hwpoison bit here. 2520fd6a03edSNaoya Horiguchi */ 2521fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2522fd6a03edSNaoya Horiguchi ret = VM_FAULT_HWPOISON; 2523fd6a03edSNaoya Horiguchi goto backout_unlocked; 2524fd6a03edSNaoya Horiguchi } 2525fd6a03edSNaoya Horiguchi 2526fd6a03edSNaoya Horiguchi /* 252757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 252857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 252957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 253057303d80SAndy Whitcroft * the spinlock. 253157303d80SAndy Whitcroft */ 2532788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 25332b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 25342b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 25352b26736cSAndy Whitcroft goto backout_unlocked; 25362b26736cSAndy Whitcroft } 253757303d80SAndy Whitcroft 2538ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2539a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 25404c887265SAdam Litke if (idx >= size) 25414c887265SAdam Litke goto backout; 25424c887265SAdam Litke 254383c54070SNick Piggin ret = 0; 25447f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 25454c887265SAdam Litke goto backout; 25464c887265SAdam Litke 25471e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 25481e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 25491e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 25501e8f889bSDavid Gibson 2551788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 25521e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 255304f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 25541e8f889bSDavid Gibson } 25551e8f889bSDavid Gibson 2556ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 25574c887265SAdam Litke unlock_page(page); 25584c887265SAdam Litke out: 2559ac9b9c66SHugh Dickins return ret; 25604c887265SAdam Litke 25614c887265SAdam Litke backout: 25624c887265SAdam Litke spin_unlock(&mm->page_table_lock); 25632b26736cSAndy Whitcroft backout_unlocked: 25644c887265SAdam Litke unlock_page(page); 25654c887265SAdam Litke put_page(page); 25664c887265SAdam Litke goto out; 2567ac9b9c66SHugh Dickins } 2568ac9b9c66SHugh Dickins 256986e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2570788c7df4SHugh Dickins unsigned long address, unsigned int flags) 257186e5216fSAdam Litke { 257286e5216fSAdam Litke pte_t *ptep; 257386e5216fSAdam Litke pte_t entry; 25741e8f889bSDavid Gibson int ret; 25750fe6e20bSNaoya Horiguchi struct page *page = NULL; 257657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 25773935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2578a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 257986e5216fSAdam Litke 2580fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2581fd6a03edSNaoya Horiguchi if (ptep) { 2582fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2583fd6a03edSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2584fd6a03edSNaoya Horiguchi return VM_FAULT_HWPOISON; 2585fd6a03edSNaoya Horiguchi } 2586fd6a03edSNaoya Horiguchi 2587a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 258886e5216fSAdam Litke if (!ptep) 258986e5216fSAdam Litke return VM_FAULT_OOM; 259086e5216fSAdam Litke 25913935baa9SDavid Gibson /* 25923935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 25933935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 25943935baa9SDavid Gibson * the same page in the page cache. 25953935baa9SDavid Gibson */ 25963935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 25977f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 25987f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2599788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2600b4d1d99fSDavid Gibson goto out_mutex; 26013935baa9SDavid Gibson } 260286e5216fSAdam Litke 260383c54070SNick Piggin ret = 0; 26041e8f889bSDavid Gibson 260557303d80SAndy Whitcroft /* 260657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 260757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 260857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 260957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 261057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 261157303d80SAndy Whitcroft * consumed. 261257303d80SAndy Whitcroft */ 2613788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 26142b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 26152b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2616b4d1d99fSDavid Gibson goto out_mutex; 26172b26736cSAndy Whitcroft } 261857303d80SAndy Whitcroft 2619f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 262057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 262157303d80SAndy Whitcroft vma, address); 262257303d80SAndy Whitcroft } 262357303d80SAndy Whitcroft 26240fe6e20bSNaoya Horiguchi if (!pagecache_page) { 26250fe6e20bSNaoya Horiguchi page = pte_page(entry); 26260fe6e20bSNaoya Horiguchi lock_page(page); 26270fe6e20bSNaoya Horiguchi } 26280fe6e20bSNaoya Horiguchi 26291e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 26301e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2631b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2632b4d1d99fSDavid Gibson goto out_page_table_lock; 2633b4d1d99fSDavid Gibson 2634b4d1d99fSDavid Gibson 2635788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2636b4d1d99fSDavid Gibson if (!pte_write(entry)) { 263757303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 263857303d80SAndy Whitcroft pagecache_page); 2639b4d1d99fSDavid Gibson goto out_page_table_lock; 2640b4d1d99fSDavid Gibson } 2641b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2642b4d1d99fSDavid Gibson } 2643b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2644788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2645788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 26464b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2647b4d1d99fSDavid Gibson 2648b4d1d99fSDavid Gibson out_page_table_lock: 26491e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 265057303d80SAndy Whitcroft 265157303d80SAndy Whitcroft if (pagecache_page) { 265257303d80SAndy Whitcroft unlock_page(pagecache_page); 265357303d80SAndy Whitcroft put_page(pagecache_page); 26540fe6e20bSNaoya Horiguchi } else { 26550fe6e20bSNaoya Horiguchi unlock_page(page); 265657303d80SAndy Whitcroft } 265757303d80SAndy Whitcroft 2658b4d1d99fSDavid Gibson out_mutex: 26593935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 26601e8f889bSDavid Gibson 26611e8f889bSDavid Gibson return ret; 266286e5216fSAdam Litke } 266386e5216fSAdam Litke 2664ceb86879SAndi Kleen /* Can be overriden by architectures */ 2665ceb86879SAndi Kleen __attribute__((weak)) struct page * 2666ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2667ceb86879SAndi Kleen pud_t *pud, int write) 2668ceb86879SAndi Kleen { 2669ceb86879SAndi Kleen BUG(); 2670ceb86879SAndi Kleen return NULL; 2671ceb86879SAndi Kleen } 2672ceb86879SAndi Kleen 267363551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 267463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 26755b23dbe8SAdam Litke unsigned long *position, int *length, int i, 26762a15efc9SHugh Dickins unsigned int flags) 267763551ae0SDavid Gibson { 2678d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2679d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 268063551ae0SDavid Gibson int remainder = *length; 2681a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 268263551ae0SDavid Gibson 26831c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 268463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 268563551ae0SDavid Gibson pte_t *pte; 26862a15efc9SHugh Dickins int absent; 268763551ae0SDavid Gibson struct page *page; 268863551ae0SDavid Gibson 26894c887265SAdam Litke /* 26904c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 26912a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 26924c887265SAdam Litke * first, for the page indexing below to work. 26934c887265SAdam Litke */ 2694a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 26952a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 269663551ae0SDavid Gibson 26972a15efc9SHugh Dickins /* 26982a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 26993ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 27003ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 27013ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 27023ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 27032a15efc9SHugh Dickins */ 27043ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 27053ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 27062a15efc9SHugh Dickins remainder = 0; 27072a15efc9SHugh Dickins break; 27082a15efc9SHugh Dickins } 27092a15efc9SHugh Dickins 27102a15efc9SHugh Dickins if (absent || 27112a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 27124c887265SAdam Litke int ret; 27134c887265SAdam Litke 27144c887265SAdam Litke spin_unlock(&mm->page_table_lock); 27152a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 27162a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 27174c887265SAdam Litke spin_lock(&mm->page_table_lock); 2718a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 27194c887265SAdam Litke continue; 27204c887265SAdam Litke 27211c59827dSHugh Dickins remainder = 0; 27221c59827dSHugh Dickins break; 27231c59827dSHugh Dickins } 272463551ae0SDavid Gibson 2725a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 27267f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2727d5d4b0aaSChen, Kenneth W same_page: 2728d6692183SChen, Kenneth W if (pages) { 272969d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 27304b2e38adSKOSAKI Motohiro get_page(pages[i]); 2731d6692183SChen, Kenneth W } 273263551ae0SDavid Gibson 273363551ae0SDavid Gibson if (vmas) 273463551ae0SDavid Gibson vmas[i] = vma; 273563551ae0SDavid Gibson 273663551ae0SDavid Gibson vaddr += PAGE_SIZE; 2737d5d4b0aaSChen, Kenneth W ++pfn_offset; 273863551ae0SDavid Gibson --remainder; 273963551ae0SDavid Gibson ++i; 2740d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2741a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2742d5d4b0aaSChen, Kenneth W /* 2743d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2744d5d4b0aaSChen, Kenneth W * of this compound page. 2745d5d4b0aaSChen, Kenneth W */ 2746d5d4b0aaSChen, Kenneth W goto same_page; 2747d5d4b0aaSChen, Kenneth W } 274863551ae0SDavid Gibson } 27491c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 275063551ae0SDavid Gibson *length = remainder; 275163551ae0SDavid Gibson *position = vaddr; 275263551ae0SDavid Gibson 27532a15efc9SHugh Dickins return i ? i : -EFAULT; 275463551ae0SDavid Gibson } 27558f860591SZhang, Yanmin 27568f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 27578f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 27588f860591SZhang, Yanmin { 27598f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 27608f860591SZhang, Yanmin unsigned long start = address; 27618f860591SZhang, Yanmin pte_t *ptep; 27628f860591SZhang, Yanmin pte_t pte; 2763a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 27648f860591SZhang, Yanmin 27658f860591SZhang, Yanmin BUG_ON(address >= end); 27668f860591SZhang, Yanmin flush_cache_range(vma, address, end); 27678f860591SZhang, Yanmin 276839dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 27698f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2770a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 27718f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 27728f860591SZhang, Yanmin if (!ptep) 27738f860591SZhang, Yanmin continue; 277439dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 277539dde65cSChen, Kenneth W continue; 27767f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 27778f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 27788f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 27798f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 27808f860591SZhang, Yanmin } 27818f860591SZhang, Yanmin } 27828f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 278339dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 27848f860591SZhang, Yanmin 27858f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 27868f860591SZhang, Yanmin } 27878f860591SZhang, Yanmin 2788a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2789a1e78772SMel Gorman long from, long to, 27905a6fe125SMel Gorman struct vm_area_struct *vma, 27915a6fe125SMel Gorman int acctflag) 2792e4e574b7SAdam Litke { 279317c9d12eSMel Gorman long ret, chg; 2794a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2795e4e574b7SAdam Litke 2796a1e78772SMel Gorman /* 279717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 279817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 279917c9d12eSMel Gorman * and filesystem quota without using reserves 280017c9d12eSMel Gorman */ 280117c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 280217c9d12eSMel Gorman return 0; 280317c9d12eSMel Gorman 280417c9d12eSMel Gorman /* 2805a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2806a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2807a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2808a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2809a1e78772SMel Gorman */ 2810f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2811e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 28125a6fe125SMel Gorman else { 28135a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 28145a6fe125SMel Gorman if (!resv_map) 28155a6fe125SMel Gorman return -ENOMEM; 28165a6fe125SMel Gorman 281717c9d12eSMel Gorman chg = to - from; 281817c9d12eSMel Gorman 28195a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 28205a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 28215a6fe125SMel Gorman } 28225a6fe125SMel Gorman 282317c9d12eSMel Gorman if (chg < 0) 282417c9d12eSMel Gorman return chg; 282517c9d12eSMel Gorman 282617c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 282717c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 282817c9d12eSMel Gorman return -ENOSPC; 282917c9d12eSMel Gorman 283017c9d12eSMel Gorman /* 283117c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 283217c9d12eSMel Gorman * Hand back the quota if there are not 283317c9d12eSMel Gorman */ 283417c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 283517c9d12eSMel Gorman if (ret < 0) { 283617c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 283717c9d12eSMel Gorman return ret; 283817c9d12eSMel Gorman } 283917c9d12eSMel Gorman 284017c9d12eSMel Gorman /* 284117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 284217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 284317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 284417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 284517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 284617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 284717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 284817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 284917c9d12eSMel Gorman * else has to be done for private mappings here 285017c9d12eSMel Gorman */ 2851f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 285217c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2853a43a8c39SChen, Kenneth W return 0; 2854a43a8c39SChen, Kenneth W } 2855a43a8c39SChen, Kenneth W 2856a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2857a43a8c39SChen, Kenneth W { 2858a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2859a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 286045c682a6SKen Chen 286145c682a6SKen Chen spin_lock(&inode->i_lock); 2862e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 286345c682a6SKen Chen spin_unlock(&inode->i_lock); 286445c682a6SKen Chen 286590d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2866a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2867a43a8c39SChen, Kenneth W } 286893f70f90SNaoya Horiguchi 286993f70f90SNaoya Horiguchi /* 287093f70f90SNaoya Horiguchi * This function is called from memory failure code. 287193f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 287293f70f90SNaoya Horiguchi */ 287393f70f90SNaoya Horiguchi void __isolate_hwpoisoned_huge_page(struct page *hpage) 287493f70f90SNaoya Horiguchi { 287593f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 287693f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 287793f70f90SNaoya Horiguchi 287893f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 287993f70f90SNaoya Horiguchi list_del(&hpage->lru); 288093f70f90SNaoya Horiguchi h->free_huge_pages--; 288193f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 288293f70f90SNaoya Horiguchi spin_unlock(&hugetlb_lock); 288393f70f90SNaoya Horiguchi } 2884