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 61843131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 61943131e14SNaoya Horiguchi 620a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6211da177e4SLinus Torvalds { 6221da177e4SLinus Torvalds struct page *page; 623f96efd58SJoe Jin 624aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 625aa888a74SAndi Kleen return NULL; 626aa888a74SAndi Kleen 6276484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 628551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 629551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 630a5516438SAndi Kleen huge_page_order(h)); 6311da177e4SLinus Torvalds if (page) { 6327f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 633caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6347b8ee84dSHarvey Harrison return NULL; 6357f2e9525SGerald Schaefer } 636a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6371da177e4SLinus Torvalds } 63863b4613cSNishanth Aravamudan 63963b4613cSNishanth Aravamudan return page; 64063b4613cSNishanth Aravamudan } 64163b4613cSNishanth Aravamudan 6425ced66c9SAndi Kleen /* 6436ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 6446ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 6456ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 6466ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 6476ae11b27SLee Schermerhorn * node for alloc or free. 6489a76db09SLee Schermerhorn */ 6496ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 6509a76db09SLee Schermerhorn { 6516ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 6529a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 6536ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 6549a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 6559a76db09SLee Schermerhorn 6569a76db09SLee Schermerhorn return nid; 6579a76db09SLee Schermerhorn } 6589a76db09SLee Schermerhorn 6596ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 6605ced66c9SAndi Kleen { 6616ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 6626ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 6639a76db09SLee Schermerhorn return nid; 6645ced66c9SAndi Kleen } 6655ced66c9SAndi Kleen 6666ae11b27SLee Schermerhorn /* 6676ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 6686ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 6696ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 6706ae11b27SLee Schermerhorn * mask. 6716ae11b27SLee Schermerhorn */ 6726ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 6736ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 6746ae11b27SLee Schermerhorn { 6756ae11b27SLee Schermerhorn int nid; 6766ae11b27SLee Schermerhorn 6776ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 6786ae11b27SLee Schermerhorn 6796ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 6806ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 6816ae11b27SLee Schermerhorn 6826ae11b27SLee Schermerhorn return nid; 6836ae11b27SLee Schermerhorn } 6846ae11b27SLee Schermerhorn 6856ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 68663b4613cSNishanth Aravamudan { 68763b4613cSNishanth Aravamudan struct page *page; 68863b4613cSNishanth Aravamudan int start_nid; 68963b4613cSNishanth Aravamudan int next_nid; 69063b4613cSNishanth Aravamudan int ret = 0; 69163b4613cSNishanth Aravamudan 6926ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 693e8c5c824SLee Schermerhorn next_nid = start_nid; 69463b4613cSNishanth Aravamudan 69563b4613cSNishanth Aravamudan do { 696e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 6979a76db09SLee Schermerhorn if (page) { 69863b4613cSNishanth Aravamudan ret = 1; 6999a76db09SLee Schermerhorn break; 7009a76db09SLee Schermerhorn } 7016ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 7029a76db09SLee Schermerhorn } while (next_nid != start_nid); 70363b4613cSNishanth Aravamudan 7043b116300SAdam Litke if (ret) 7053b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7063b116300SAdam Litke else 7073b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7083b116300SAdam Litke 70963b4613cSNishanth Aravamudan return ret; 7101da177e4SLinus Torvalds } 7111da177e4SLinus Torvalds 712e8c5c824SLee Schermerhorn /* 7136ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7146ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7156ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7166ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 717e8c5c824SLee Schermerhorn */ 7186ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 719e8c5c824SLee Schermerhorn { 7206ae11b27SLee Schermerhorn int nid; 7219a76db09SLee Schermerhorn 7226ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7236ae11b27SLee Schermerhorn 7246ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7256ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7266ae11b27SLee Schermerhorn 7279a76db09SLee Schermerhorn return nid; 728e8c5c824SLee Schermerhorn } 729e8c5c824SLee Schermerhorn 730e8c5c824SLee Schermerhorn /* 731e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 732e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 733e8c5c824SLee Schermerhorn * balanced over allowed nodes. 734e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 735e8c5c824SLee Schermerhorn */ 7366ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 7376ae11b27SLee Schermerhorn bool acct_surplus) 738e8c5c824SLee Schermerhorn { 739e8c5c824SLee Schermerhorn int start_nid; 740e8c5c824SLee Schermerhorn int next_nid; 741e8c5c824SLee Schermerhorn int ret = 0; 742e8c5c824SLee Schermerhorn 7436ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 744e8c5c824SLee Schermerhorn next_nid = start_nid; 745e8c5c824SLee Schermerhorn 746e8c5c824SLee Schermerhorn do { 747685f3457SLee Schermerhorn /* 748685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 749685f3457SLee Schermerhorn * nodes with surplus pages. 750685f3457SLee Schermerhorn */ 751685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 752685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 753e8c5c824SLee Schermerhorn struct page *page = 754e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 755e8c5c824SLee Schermerhorn struct page, lru); 756e8c5c824SLee Schermerhorn list_del(&page->lru); 757e8c5c824SLee Schermerhorn h->free_huge_pages--; 758e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 759685f3457SLee Schermerhorn if (acct_surplus) { 760685f3457SLee Schermerhorn h->surplus_huge_pages--; 761685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 762685f3457SLee Schermerhorn } 763e8c5c824SLee Schermerhorn update_and_free_page(h, page); 764e8c5c824SLee Schermerhorn ret = 1; 7659a76db09SLee Schermerhorn break; 766e8c5c824SLee Schermerhorn } 7676ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 7689a76db09SLee Schermerhorn } while (next_nid != start_nid); 769e8c5c824SLee Schermerhorn 770e8c5c824SLee Schermerhorn return ret; 771e8c5c824SLee Schermerhorn } 772e8c5c824SLee Schermerhorn 773a5516438SAndi Kleen static struct page *alloc_buddy_huge_page(struct hstate *h, 774a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 7757893d1d5SAdam Litke { 7767893d1d5SAdam Litke struct page *page; 777d1c3fb1fSNishanth Aravamudan unsigned int nid; 7787893d1d5SAdam Litke 779aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 780aa888a74SAndi Kleen return NULL; 781aa888a74SAndi Kleen 782d1c3fb1fSNishanth Aravamudan /* 783d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 784d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 785d1c3fb1fSNishanth Aravamudan * overcommit 786d1c3fb1fSNishanth Aravamudan * 787d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 788d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 789d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 790d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 791d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 792d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 793d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 794d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 795d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 796d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 797d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 798d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 799d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 800d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 801d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 802d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 803d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 804d1c3fb1fSNishanth Aravamudan */ 805d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 806a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 807d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 808d1c3fb1fSNishanth Aravamudan return NULL; 809d1c3fb1fSNishanth Aravamudan } else { 810a5516438SAndi Kleen h->nr_huge_pages++; 811a5516438SAndi Kleen h->surplus_huge_pages++; 812d1c3fb1fSNishanth Aravamudan } 813d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 814d1c3fb1fSNishanth Aravamudan 815551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 816551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 817a5516438SAndi Kleen huge_page_order(h)); 818d1c3fb1fSNishanth Aravamudan 819caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 820caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 821caff3a2cSGerald Schaefer return NULL; 822caff3a2cSGerald Schaefer } 823caff3a2cSGerald Schaefer 8247893d1d5SAdam Litke spin_lock(&hugetlb_lock); 825d1c3fb1fSNishanth Aravamudan if (page) { 8262668db91SAdam Litke /* 8272668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 8282668db91SAdam Litke * no users -- drop the buddy allocator's reference. 8292668db91SAdam Litke */ 8302668db91SAdam Litke put_page_testzero(page); 8312668db91SAdam Litke VM_BUG_ON(page_count(page)); 832d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 833d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 834d1c3fb1fSNishanth Aravamudan /* 835d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 836d1c3fb1fSNishanth Aravamudan */ 837a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 838a5516438SAndi Kleen h->surplus_huge_pages_node[nid]++; 8393b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 840d1c3fb1fSNishanth Aravamudan } else { 841a5516438SAndi Kleen h->nr_huge_pages--; 842a5516438SAndi Kleen h->surplus_huge_pages--; 8433b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 8447893d1d5SAdam Litke } 845d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 8467893d1d5SAdam Litke 8477893d1d5SAdam Litke return page; 8487893d1d5SAdam Litke } 8497893d1d5SAdam Litke 850e4e574b7SAdam Litke /* 851e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 852e4e574b7SAdam Litke * of size 'delta'. 853e4e574b7SAdam Litke */ 854a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 855e4e574b7SAdam Litke { 856e4e574b7SAdam Litke struct list_head surplus_list; 857e4e574b7SAdam Litke struct page *page, *tmp; 858e4e574b7SAdam Litke int ret, i; 859e4e574b7SAdam Litke int needed, allocated; 860e4e574b7SAdam Litke 861a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 862ac09b3a1SAdam Litke if (needed <= 0) { 863a5516438SAndi Kleen h->resv_huge_pages += delta; 864e4e574b7SAdam Litke return 0; 865ac09b3a1SAdam Litke } 866e4e574b7SAdam Litke 867e4e574b7SAdam Litke allocated = 0; 868e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 869e4e574b7SAdam Litke 870e4e574b7SAdam Litke ret = -ENOMEM; 871e4e574b7SAdam Litke retry: 872e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 873e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 874a5516438SAndi Kleen page = alloc_buddy_huge_page(h, NULL, 0); 875e4e574b7SAdam Litke if (!page) { 876e4e574b7SAdam Litke /* 877e4e574b7SAdam Litke * We were not able to allocate enough pages to 878e4e574b7SAdam Litke * satisfy the entire reservation so we free what 879e4e574b7SAdam Litke * we've allocated so far. 880e4e574b7SAdam Litke */ 881e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 882e4e574b7SAdam Litke needed = 0; 883e4e574b7SAdam Litke goto free; 884e4e574b7SAdam Litke } 885e4e574b7SAdam Litke 886e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 887e4e574b7SAdam Litke } 888e4e574b7SAdam Litke allocated += needed; 889e4e574b7SAdam Litke 890e4e574b7SAdam Litke /* 891e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 892e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 893e4e574b7SAdam Litke */ 894e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 895a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 896a5516438SAndi Kleen (h->free_huge_pages + allocated); 897e4e574b7SAdam Litke if (needed > 0) 898e4e574b7SAdam Litke goto retry; 899e4e574b7SAdam Litke 900e4e574b7SAdam Litke /* 901e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 902e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 903e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 904ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 905ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 906ac09b3a1SAdam Litke * before they are reserved. 907e4e574b7SAdam Litke */ 908e4e574b7SAdam Litke needed += allocated; 909a5516438SAndi Kleen h->resv_huge_pages += delta; 910e4e574b7SAdam Litke ret = 0; 911e4e574b7SAdam Litke free: 91219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 91319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 91419fc3f0aSAdam Litke if ((--needed) < 0) 91519fc3f0aSAdam Litke break; 91619fc3f0aSAdam Litke list_del(&page->lru); 917a5516438SAndi Kleen enqueue_huge_page(h, page); 91819fc3f0aSAdam Litke } 91919fc3f0aSAdam Litke 92019fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 92119fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 92219fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 923e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 924e4e574b7SAdam Litke list_del(&page->lru); 925af767cbdSAdam Litke /* 9262668db91SAdam Litke * The page has a reference count of zero already, so 9272668db91SAdam Litke * call free_huge_page directly instead of using 9282668db91SAdam Litke * put_page. This must be done with hugetlb_lock 929af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 930af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 931af767cbdSAdam Litke */ 9322668db91SAdam Litke free_huge_page(page); 933af767cbdSAdam Litke } 93419fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 935e4e574b7SAdam Litke } 936e4e574b7SAdam Litke 937e4e574b7SAdam Litke return ret; 938e4e574b7SAdam Litke } 939e4e574b7SAdam Litke 940e4e574b7SAdam Litke /* 941e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 942e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 943e4e574b7SAdam Litke * never used. 944685f3457SLee Schermerhorn * Called with hugetlb_lock held. 945e4e574b7SAdam Litke */ 946a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 947a5516438SAndi Kleen unsigned long unused_resv_pages) 948e4e574b7SAdam Litke { 949e4e574b7SAdam Litke unsigned long nr_pages; 950e4e574b7SAdam Litke 951ac09b3a1SAdam Litke /* Uncommit the reservation */ 952a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 953ac09b3a1SAdam Litke 954aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 955aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 956aa888a74SAndi Kleen return; 957aa888a74SAndi Kleen 958a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 959e4e574b7SAdam Litke 960685f3457SLee Schermerhorn /* 961685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 9629b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 9639b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 9649b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 9659b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 9669b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 967685f3457SLee Schermerhorn */ 968685f3457SLee Schermerhorn while (nr_pages--) { 9699b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 970685f3457SLee Schermerhorn break; 971e4e574b7SAdam Litke } 972e4e574b7SAdam Litke } 973e4e574b7SAdam Litke 974c37f9fb1SAndy Whitcroft /* 975c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 976c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 977c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 978c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 979c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 980c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 981c37f9fb1SAndy Whitcroft * No action is required on failure. 982c37f9fb1SAndy Whitcroft */ 983e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 984a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 985c37f9fb1SAndy Whitcroft { 986c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 987c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 988c37f9fb1SAndy Whitcroft 989f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 990a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 991c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 992c37f9fb1SAndy Whitcroft idx, idx + 1); 993c37f9fb1SAndy Whitcroft 99484afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 995c37f9fb1SAndy Whitcroft return 1; 996c37f9fb1SAndy Whitcroft 99784afd99bSAndy Whitcroft } else { 998e2f17d94SRoel Kluin long err; 999a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 100084afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 100184afd99bSAndy Whitcroft 100284afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 100384afd99bSAndy Whitcroft if (err < 0) 100484afd99bSAndy Whitcroft return err; 1005c37f9fb1SAndy Whitcroft return 0; 1006c37f9fb1SAndy Whitcroft } 100784afd99bSAndy Whitcroft } 1008a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1009a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1010c37f9fb1SAndy Whitcroft { 1011c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1012c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1013c37f9fb1SAndy Whitcroft 1014f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1015a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1016c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 101784afd99bSAndy Whitcroft 101884afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1019a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 102084afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 102184afd99bSAndy Whitcroft 102284afd99bSAndy Whitcroft /* Mark this page used in the map. */ 102384afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1024c37f9fb1SAndy Whitcroft } 1025c37f9fb1SAndy Whitcroft } 1026c37f9fb1SAndy Whitcroft 1027348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 102804f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1029348ea204SAdam Litke { 1030a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1031348ea204SAdam Litke struct page *page; 10322fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 1033a1e78772SMel Gorman struct inode *inode = mapping->host; 1034e2f17d94SRoel Kluin long chg; 10352fc39cecSAdam Litke 1036a1e78772SMel Gorman /* 1037a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 1038a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 1039a1e78772SMel Gorman * made before satisfying the allocation 1040c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 1041c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 1042a1e78772SMel Gorman */ 1043a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1044c37f9fb1SAndy Whitcroft if (chg < 0) 1045c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 1046c37f9fb1SAndy Whitcroft if (chg) 1047a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 1048a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 104990d8b7e6SAdam Litke 1050a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1051a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1052a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1053a1e78772SMel Gorman 1054a1e78772SMel Gorman if (!page) { 1055a5516438SAndi Kleen page = alloc_buddy_huge_page(h, vma, addr); 1056a1e78772SMel Gorman if (!page) { 1057a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 10584a6018f7SMel Gorman return ERR_PTR(-VM_FAULT_SIGBUS); 1059a1e78772SMel Gorman } 1060a1e78772SMel Gorman } 1061a1e78772SMel Gorman 1062348ea204SAdam Litke set_page_refcounted(page); 10632fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 1064a1e78772SMel Gorman 1065a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1066c37f9fb1SAndy Whitcroft 10677893d1d5SAdam Litke return page; 1068b45b5bd6SDavid Gibson } 1069b45b5bd6SDavid Gibson 107091f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1071aa888a74SAndi Kleen { 1072aa888a74SAndi Kleen struct huge_bootmem_page *m; 10739b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1074aa888a74SAndi Kleen 1075aa888a74SAndi Kleen while (nr_nodes) { 1076aa888a74SAndi Kleen void *addr; 1077aa888a74SAndi Kleen 1078aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 10796ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 10809b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1081aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1082aa888a74SAndi Kleen 1083aa888a74SAndi Kleen if (addr) { 1084aa888a74SAndi Kleen /* 1085aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1086aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1087aa888a74SAndi Kleen * puts them into the mem_map). 1088aa888a74SAndi Kleen */ 1089aa888a74SAndi Kleen m = addr; 1090aa888a74SAndi Kleen goto found; 1091aa888a74SAndi Kleen } 1092aa888a74SAndi Kleen nr_nodes--; 1093aa888a74SAndi Kleen } 1094aa888a74SAndi Kleen return 0; 1095aa888a74SAndi Kleen 1096aa888a74SAndi Kleen found: 1097aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1098aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1099aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1100aa888a74SAndi Kleen m->hstate = h; 1101aa888a74SAndi Kleen return 1; 1102aa888a74SAndi Kleen } 1103aa888a74SAndi Kleen 110418229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 110518229df5SAndy Whitcroft { 110618229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 110718229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 110818229df5SAndy Whitcroft else 110918229df5SAndy Whitcroft prep_compound_page(page, order); 111018229df5SAndy Whitcroft } 111118229df5SAndy Whitcroft 1112aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1113aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1114aa888a74SAndi Kleen { 1115aa888a74SAndi Kleen struct huge_bootmem_page *m; 1116aa888a74SAndi Kleen 1117aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1118aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1119aa888a74SAndi Kleen struct hstate *h = m->hstate; 1120aa888a74SAndi Kleen __ClearPageReserved(page); 1121aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 112218229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1123aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1124aa888a74SAndi Kleen } 1125aa888a74SAndi Kleen } 1126aa888a74SAndi Kleen 11278faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 11281da177e4SLinus Torvalds { 11291da177e4SLinus Torvalds unsigned long i; 11301da177e4SLinus Torvalds 1131e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1132aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1133aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1134aa888a74SAndi Kleen break; 11359b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 11369b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 11371da177e4SLinus Torvalds break; 11381da177e4SLinus Torvalds } 11398faa8b07SAndi Kleen h->max_huge_pages = i; 1140e5ff2159SAndi Kleen } 1141e5ff2159SAndi Kleen 1142e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1143e5ff2159SAndi Kleen { 1144e5ff2159SAndi Kleen struct hstate *h; 1145e5ff2159SAndi Kleen 1146e5ff2159SAndi Kleen for_each_hstate(h) { 11478faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 11488faa8b07SAndi Kleen if (h->order < MAX_ORDER) 11498faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1150e5ff2159SAndi Kleen } 1151e5ff2159SAndi Kleen } 1152e5ff2159SAndi Kleen 11534abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 11544abd32dbSAndi Kleen { 11554abd32dbSAndi Kleen if (n >= (1UL << 30)) 11564abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 11574abd32dbSAndi Kleen else if (n >= (1UL << 20)) 11584abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 11594abd32dbSAndi Kleen else 11604abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 11614abd32dbSAndi Kleen return buf; 11624abd32dbSAndi Kleen } 11634abd32dbSAndi Kleen 1164e5ff2159SAndi Kleen static void __init report_hugepages(void) 1165e5ff2159SAndi Kleen { 1166e5ff2159SAndi Kleen struct hstate *h; 1167e5ff2159SAndi Kleen 1168e5ff2159SAndi Kleen for_each_hstate(h) { 11694abd32dbSAndi Kleen char buf[32]; 11704abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11714abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11724abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11734abd32dbSAndi Kleen h->free_huge_pages); 1174e5ff2159SAndi Kleen } 1175e5ff2159SAndi Kleen } 1176e5ff2159SAndi Kleen 11771da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 11786ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 11796ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11801da177e4SLinus Torvalds { 11814415cc8dSChristoph Lameter int i; 11824415cc8dSChristoph Lameter 1183aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1184aa888a74SAndi Kleen return; 1185aa888a74SAndi Kleen 11866ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 11871da177e4SLinus Torvalds struct page *page, *next; 1188a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1189a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1190a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11916b0c880dSAdam Litke return; 11921da177e4SLinus Torvalds if (PageHighMem(page)) 11931da177e4SLinus Torvalds continue; 11941da177e4SLinus Torvalds list_del(&page->lru); 1195e5ff2159SAndi Kleen update_and_free_page(h, page); 1196a5516438SAndi Kleen h->free_huge_pages--; 1197a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11981da177e4SLinus Torvalds } 11991da177e4SLinus Torvalds } 12001da177e4SLinus Torvalds } 12011da177e4SLinus Torvalds #else 12026ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 12036ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12041da177e4SLinus Torvalds { 12051da177e4SLinus Torvalds } 12061da177e4SLinus Torvalds #endif 12071da177e4SLinus Torvalds 120820a0307cSWu Fengguang /* 120920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 121020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 121120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 121220a0307cSWu Fengguang */ 12136ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 12146ae11b27SLee Schermerhorn int delta) 121520a0307cSWu Fengguang { 1216e8c5c824SLee Schermerhorn int start_nid, next_nid; 121720a0307cSWu Fengguang int ret = 0; 121820a0307cSWu Fengguang 121920a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 122020a0307cSWu Fengguang 1221e8c5c824SLee Schermerhorn if (delta < 0) 12226ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1223e8c5c824SLee Schermerhorn else 12246ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1225e8c5c824SLee Schermerhorn next_nid = start_nid; 1226e8c5c824SLee Schermerhorn 1227e8c5c824SLee Schermerhorn do { 1228e8c5c824SLee Schermerhorn int nid = next_nid; 1229e8c5c824SLee Schermerhorn if (delta < 0) { 1230e8c5c824SLee Schermerhorn /* 1231e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1232e8c5c824SLee Schermerhorn */ 12339a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 12346ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 12356ae11b27SLee Schermerhorn nodes_allowed); 123620a0307cSWu Fengguang continue; 1237e8c5c824SLee Schermerhorn } 12389a76db09SLee Schermerhorn } 1239e8c5c824SLee Schermerhorn if (delta > 0) { 1240e8c5c824SLee Schermerhorn /* 1241e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1242e8c5c824SLee Schermerhorn */ 1243e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 12449a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 12456ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 12466ae11b27SLee Schermerhorn nodes_allowed); 124720a0307cSWu Fengguang continue; 1248e8c5c824SLee Schermerhorn } 12499a76db09SLee Schermerhorn } 125020a0307cSWu Fengguang 125120a0307cSWu Fengguang h->surplus_huge_pages += delta; 125220a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 125320a0307cSWu Fengguang ret = 1; 125420a0307cSWu Fengguang break; 1255e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 125620a0307cSWu Fengguang 125720a0307cSWu Fengguang return ret; 125820a0307cSWu Fengguang } 125920a0307cSWu Fengguang 1260a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 12616ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 12626ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12631da177e4SLinus Torvalds { 12647893d1d5SAdam Litke unsigned long min_count, ret; 12651da177e4SLinus Torvalds 1266aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1267aa888a74SAndi Kleen return h->max_huge_pages; 1268aa888a74SAndi Kleen 12697893d1d5SAdam Litke /* 12707893d1d5SAdam Litke * Increase the pool size 12717893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 12727893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1273d1c3fb1fSNishanth Aravamudan * 1274d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1275d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1276d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1277d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1278d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 12797893d1d5SAdam Litke */ 12801da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1281a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 12826ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 12837893d1d5SAdam Litke break; 12847893d1d5SAdam Litke } 12857893d1d5SAdam Litke 1286a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 12877893d1d5SAdam Litke /* 12887893d1d5SAdam Litke * If this allocation races such that we no longer need the 12897893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12907893d1d5SAdam Litke * and reducing the surplus. 12917893d1d5SAdam Litke */ 12927893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 12936ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 12947893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12957893d1d5SAdam Litke if (!ret) 12967893d1d5SAdam Litke goto out; 12977893d1d5SAdam Litke 1298536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1299536240f2SMel Gorman if (signal_pending(current)) 1300536240f2SMel Gorman goto out; 13017893d1d5SAdam Litke } 13027893d1d5SAdam Litke 13037893d1d5SAdam Litke /* 13047893d1d5SAdam Litke * Decrease the pool size 13057893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 13067893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 13077893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 13087893d1d5SAdam Litke * to the desired size as pages become free. 1309d1c3fb1fSNishanth Aravamudan * 1310d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1311d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1312d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1313d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1314d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1315d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1316d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 13177893d1d5SAdam Litke */ 1318a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 13196b0c880dSAdam Litke min_count = max(count, min_count); 13206ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1321a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 13226ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 13231da177e4SLinus Torvalds break; 13241da177e4SLinus Torvalds } 1325a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 13266ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 13277893d1d5SAdam Litke break; 13287893d1d5SAdam Litke } 13297893d1d5SAdam Litke out: 1330a5516438SAndi Kleen ret = persistent_huge_pages(h); 13311da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 13327893d1d5SAdam Litke return ret; 13331da177e4SLinus Torvalds } 13341da177e4SLinus Torvalds 1335a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1336a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1337a3437870SNishanth Aravamudan 1338a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1339a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1340a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1341a3437870SNishanth Aravamudan 1342a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1343a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1344a3437870SNishanth Aravamudan 13459a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 13469a305230SLee Schermerhorn 13479a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1348a3437870SNishanth Aravamudan { 1349a3437870SNishanth Aravamudan int i; 13509a305230SLee Schermerhorn 1351a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 13529a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 13539a305230SLee Schermerhorn if (nidp) 13549a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1355a3437870SNishanth Aravamudan return &hstates[i]; 13569a305230SLee Schermerhorn } 13579a305230SLee Schermerhorn 13589a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1359a3437870SNishanth Aravamudan } 1360a3437870SNishanth Aravamudan 136106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1362a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1363a3437870SNishanth Aravamudan { 13649a305230SLee Schermerhorn struct hstate *h; 13659a305230SLee Schermerhorn unsigned long nr_huge_pages; 13669a305230SLee Schermerhorn int nid; 13679a305230SLee Schermerhorn 13689a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13699a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 13709a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 13719a305230SLee Schermerhorn else 13729a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 13739a305230SLee Schermerhorn 13749a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1375a3437870SNishanth Aravamudan } 137606808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 137706808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 137806808b08SLee Schermerhorn const char *buf, size_t len) 1379a3437870SNishanth Aravamudan { 1380a3437870SNishanth Aravamudan int err; 13819a305230SLee Schermerhorn int nid; 138206808b08SLee Schermerhorn unsigned long count; 13839a305230SLee Schermerhorn struct hstate *h; 1384bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1385a3437870SNishanth Aravamudan 138606808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 1387a3437870SNishanth Aravamudan if (err) 1388a3437870SNishanth Aravamudan return 0; 1389a3437870SNishanth Aravamudan 13909a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13919a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 13929a305230SLee Schermerhorn /* 13939a305230SLee Schermerhorn * global hstate attribute 13949a305230SLee Schermerhorn */ 13959a305230SLee Schermerhorn if (!(obey_mempolicy && 13969a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 139706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 13989a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 139906808b08SLee Schermerhorn } 14009a305230SLee Schermerhorn } else if (nodes_allowed) { 14019a305230SLee Schermerhorn /* 14029a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 14039a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 14049a305230SLee Schermerhorn */ 14059a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 14069a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 14079a305230SLee Schermerhorn } else 14089a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 14099a305230SLee Schermerhorn 141006808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1411a3437870SNishanth Aravamudan 14129b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 141306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 141406808b08SLee Schermerhorn 141506808b08SLee Schermerhorn return len; 141606808b08SLee Schermerhorn } 141706808b08SLee Schermerhorn 141806808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 141906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 142006808b08SLee Schermerhorn { 142106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 142206808b08SLee Schermerhorn } 142306808b08SLee Schermerhorn 142406808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 142506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 142606808b08SLee Schermerhorn { 142706808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1428a3437870SNishanth Aravamudan } 1429a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1430a3437870SNishanth Aravamudan 143106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 143206808b08SLee Schermerhorn 143306808b08SLee Schermerhorn /* 143406808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 143506808b08SLee Schermerhorn * huge page alloc/free. 143606808b08SLee Schermerhorn */ 143706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 143806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 143906808b08SLee Schermerhorn { 144006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 144106808b08SLee Schermerhorn } 144206808b08SLee Schermerhorn 144306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 144406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 144506808b08SLee Schermerhorn { 144606808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 144706808b08SLee Schermerhorn } 144806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 144906808b08SLee Schermerhorn #endif 145006808b08SLee Schermerhorn 145106808b08SLee Schermerhorn 1452a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1453a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1454a3437870SNishanth Aravamudan { 14559a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1456a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1457a3437870SNishanth Aravamudan } 1458a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1459a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1460a3437870SNishanth Aravamudan { 1461a3437870SNishanth Aravamudan int err; 1462a3437870SNishanth Aravamudan unsigned long input; 14639a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1464a3437870SNishanth Aravamudan 1465a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1466a3437870SNishanth Aravamudan if (err) 1467a3437870SNishanth Aravamudan return 0; 1468a3437870SNishanth Aravamudan 1469a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1470a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1471a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1472a3437870SNishanth Aravamudan 1473a3437870SNishanth Aravamudan return count; 1474a3437870SNishanth Aravamudan } 1475a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1476a3437870SNishanth Aravamudan 1477a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1478a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1479a3437870SNishanth Aravamudan { 14809a305230SLee Schermerhorn struct hstate *h; 14819a305230SLee Schermerhorn unsigned long free_huge_pages; 14829a305230SLee Schermerhorn int nid; 14839a305230SLee Schermerhorn 14849a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14859a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14869a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 14879a305230SLee Schermerhorn else 14889a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 14899a305230SLee Schermerhorn 14909a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1491a3437870SNishanth Aravamudan } 1492a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1493a3437870SNishanth Aravamudan 1494a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1495a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1496a3437870SNishanth Aravamudan { 14979a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1498a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1499a3437870SNishanth Aravamudan } 1500a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1501a3437870SNishanth Aravamudan 1502a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1503a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1504a3437870SNishanth Aravamudan { 15059a305230SLee Schermerhorn struct hstate *h; 15069a305230SLee Schermerhorn unsigned long surplus_huge_pages; 15079a305230SLee Schermerhorn int nid; 15089a305230SLee Schermerhorn 15099a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15109a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15119a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 15129a305230SLee Schermerhorn else 15139a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 15149a305230SLee Schermerhorn 15159a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1516a3437870SNishanth Aravamudan } 1517a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1518a3437870SNishanth Aravamudan 1519a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1520a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1521a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1522a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1523a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1524a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 152506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 152606808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 152706808b08SLee Schermerhorn #endif 1528a3437870SNishanth Aravamudan NULL, 1529a3437870SNishanth Aravamudan }; 1530a3437870SNishanth Aravamudan 1531a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1532a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1533a3437870SNishanth Aravamudan }; 1534a3437870SNishanth Aravamudan 1535094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 15369a305230SLee Schermerhorn struct kobject **hstate_kobjs, 15379a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1538a3437870SNishanth Aravamudan { 1539a3437870SNishanth Aravamudan int retval; 15409a305230SLee Schermerhorn int hi = h - hstates; 1541a3437870SNishanth Aravamudan 15429a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 15439a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1544a3437870SNishanth Aravamudan return -ENOMEM; 1545a3437870SNishanth Aravamudan 15469a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1547a3437870SNishanth Aravamudan if (retval) 15489a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1549a3437870SNishanth Aravamudan 1550a3437870SNishanth Aravamudan return retval; 1551a3437870SNishanth Aravamudan } 1552a3437870SNishanth Aravamudan 1553a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1554a3437870SNishanth Aravamudan { 1555a3437870SNishanth Aravamudan struct hstate *h; 1556a3437870SNishanth Aravamudan int err; 1557a3437870SNishanth Aravamudan 1558a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1559a3437870SNishanth Aravamudan if (!hugepages_kobj) 1560a3437870SNishanth Aravamudan return; 1561a3437870SNishanth Aravamudan 1562a3437870SNishanth Aravamudan for_each_hstate(h) { 15639a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 15649a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1565a3437870SNishanth Aravamudan if (err) 1566a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1567a3437870SNishanth Aravamudan h->name); 1568a3437870SNishanth Aravamudan } 1569a3437870SNishanth Aravamudan } 1570a3437870SNishanth Aravamudan 15719a305230SLee Schermerhorn #ifdef CONFIG_NUMA 15729a305230SLee Schermerhorn 15739a305230SLee Schermerhorn /* 15749a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 15759a305230SLee Schermerhorn * with node sysdevs in node_devices[] using a parallel array. The array 15769a305230SLee Schermerhorn * index of a node sysdev or _hstate == node id. 15779a305230SLee Schermerhorn * This is here to avoid any static dependency of the node sysdev driver, in 15789a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 15799a305230SLee Schermerhorn */ 15809a305230SLee Schermerhorn struct node_hstate { 15819a305230SLee Schermerhorn struct kobject *hugepages_kobj; 15829a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 15839a305230SLee Schermerhorn }; 15849a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 15859a305230SLee Schermerhorn 15869a305230SLee Schermerhorn /* 15879a305230SLee Schermerhorn * A subset of global hstate attributes for node sysdevs 15889a305230SLee Schermerhorn */ 15899a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 15909a305230SLee Schermerhorn &nr_hugepages_attr.attr, 15919a305230SLee Schermerhorn &free_hugepages_attr.attr, 15929a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 15939a305230SLee Schermerhorn NULL, 15949a305230SLee Schermerhorn }; 15959a305230SLee Schermerhorn 15969a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 15979a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 15989a305230SLee Schermerhorn }; 15999a305230SLee Schermerhorn 16009a305230SLee Schermerhorn /* 16019a305230SLee Schermerhorn * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. 16029a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 16039a305230SLee Schermerhorn */ 16049a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 16059a305230SLee Schermerhorn { 16069a305230SLee Schermerhorn int nid; 16079a305230SLee Schermerhorn 16089a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 16099a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 16109a305230SLee Schermerhorn int i; 16119a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 16129a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 16139a305230SLee Schermerhorn if (nidp) 16149a305230SLee Schermerhorn *nidp = nid; 16159a305230SLee Schermerhorn return &hstates[i]; 16169a305230SLee Schermerhorn } 16179a305230SLee Schermerhorn } 16189a305230SLee Schermerhorn 16199a305230SLee Schermerhorn BUG(); 16209a305230SLee Schermerhorn return NULL; 16219a305230SLee Schermerhorn } 16229a305230SLee Schermerhorn 16239a305230SLee Schermerhorn /* 16249a305230SLee Schermerhorn * Unregister hstate attributes from a single node sysdev. 16259a305230SLee Schermerhorn * No-op if no hstate attributes attached. 16269a305230SLee Schermerhorn */ 16279a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 16289a305230SLee Schermerhorn { 16299a305230SLee Schermerhorn struct hstate *h; 16309a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16319a305230SLee Schermerhorn 16329a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16339b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 16349a305230SLee Schermerhorn 16359a305230SLee Schermerhorn for_each_hstate(h) 16369a305230SLee Schermerhorn if (nhs->hstate_kobjs[h - hstates]) { 16379a305230SLee Schermerhorn kobject_put(nhs->hstate_kobjs[h - hstates]); 16389a305230SLee Schermerhorn nhs->hstate_kobjs[h - hstates] = NULL; 16399a305230SLee Schermerhorn } 16409a305230SLee Schermerhorn 16419a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 16429a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 16439a305230SLee Schermerhorn } 16449a305230SLee Schermerhorn 16459a305230SLee Schermerhorn /* 16469a305230SLee Schermerhorn * hugetlb module exit: unregister hstate attributes from node sysdevs 16479a305230SLee Schermerhorn * that have them. 16489a305230SLee Schermerhorn */ 16499a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 16509a305230SLee Schermerhorn { 16519a305230SLee Schermerhorn int nid; 16529a305230SLee Schermerhorn 16539a305230SLee Schermerhorn /* 16549a305230SLee Schermerhorn * disable node sysdev registrations. 16559a305230SLee Schermerhorn */ 16569a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 16579a305230SLee Schermerhorn 16589a305230SLee Schermerhorn /* 16599a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 16609a305230SLee Schermerhorn */ 16619a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 16629a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 16639a305230SLee Schermerhorn } 16649a305230SLee Schermerhorn 16659a305230SLee Schermerhorn /* 16669a305230SLee Schermerhorn * Register hstate attributes for a single node sysdev. 16679a305230SLee Schermerhorn * No-op if attributes already registered. 16689a305230SLee Schermerhorn */ 16699a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 16709a305230SLee Schermerhorn { 16719a305230SLee Schermerhorn struct hstate *h; 16729a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16739a305230SLee Schermerhorn int err; 16749a305230SLee Schermerhorn 16759a305230SLee Schermerhorn if (nhs->hugepages_kobj) 16769a305230SLee Schermerhorn return; /* already allocated */ 16779a305230SLee Schermerhorn 16789a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 16799a305230SLee Schermerhorn &node->sysdev.kobj); 16809a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16819a305230SLee Schermerhorn return; 16829a305230SLee Schermerhorn 16839a305230SLee Schermerhorn for_each_hstate(h) { 16849a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 16859a305230SLee Schermerhorn nhs->hstate_kobjs, 16869a305230SLee Schermerhorn &per_node_hstate_attr_group); 16879a305230SLee Schermerhorn if (err) { 16889a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 16899a305230SLee Schermerhorn " for node %d\n", 16909a305230SLee Schermerhorn h->name, node->sysdev.id); 16919a305230SLee Schermerhorn hugetlb_unregister_node(node); 16929a305230SLee Schermerhorn break; 16939a305230SLee Schermerhorn } 16949a305230SLee Schermerhorn } 16959a305230SLee Schermerhorn } 16969a305230SLee Schermerhorn 16979a305230SLee Schermerhorn /* 16989b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 16999b5e5d0fSLee Schermerhorn * sysdevs of nodes that have memory. All on-line nodes should have 17009b5e5d0fSLee Schermerhorn * registered their associated sysdev by this time. 17019a305230SLee Schermerhorn */ 17029a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 17039a305230SLee Schermerhorn { 17049a305230SLee Schermerhorn int nid; 17059a305230SLee Schermerhorn 17069b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 17079a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 17089a305230SLee Schermerhorn if (node->sysdev.id == nid) 17099a305230SLee Schermerhorn hugetlb_register_node(node); 17109a305230SLee Schermerhorn } 17119a305230SLee Schermerhorn 17129a305230SLee Schermerhorn /* 17139a305230SLee Schermerhorn * Let the node sysdev driver know we're here so it can 17149a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 17159a305230SLee Schermerhorn */ 17169a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 17179a305230SLee Schermerhorn hugetlb_unregister_node); 17189a305230SLee Schermerhorn } 17199a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 17209a305230SLee Schermerhorn 17219a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17229a305230SLee Schermerhorn { 17239a305230SLee Schermerhorn BUG(); 17249a305230SLee Schermerhorn if (nidp) 17259a305230SLee Schermerhorn *nidp = -1; 17269a305230SLee Schermerhorn return NULL; 17279a305230SLee Schermerhorn } 17289a305230SLee Schermerhorn 17299a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 17309a305230SLee Schermerhorn 17319a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 17329a305230SLee Schermerhorn 17339a305230SLee Schermerhorn #endif 17349a305230SLee Schermerhorn 1735a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1736a3437870SNishanth Aravamudan { 1737a3437870SNishanth Aravamudan struct hstate *h; 1738a3437870SNishanth Aravamudan 17399a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 17409a305230SLee Schermerhorn 1741a3437870SNishanth Aravamudan for_each_hstate(h) { 1742a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1743a3437870SNishanth Aravamudan } 1744a3437870SNishanth Aravamudan 1745a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1746a3437870SNishanth Aravamudan } 1747a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1748a3437870SNishanth Aravamudan 1749a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1750a3437870SNishanth Aravamudan { 17510ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 17520ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 17530ef89d25SBenjamin Herrenschmidt * there is no such support 17540ef89d25SBenjamin Herrenschmidt */ 17550ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 17560ef89d25SBenjamin Herrenschmidt return 0; 1757a3437870SNishanth Aravamudan 1758e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1759e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1760e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1761a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1762a3437870SNishanth Aravamudan } 1763e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1764e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1765e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1766a3437870SNishanth Aravamudan 1767a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1768a3437870SNishanth Aravamudan 1769aa888a74SAndi Kleen gather_bootmem_prealloc(); 1770aa888a74SAndi Kleen 1771a3437870SNishanth Aravamudan report_hugepages(); 1772a3437870SNishanth Aravamudan 1773a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1774a3437870SNishanth Aravamudan 17759a305230SLee Schermerhorn hugetlb_register_all_nodes(); 17769a305230SLee Schermerhorn 1777a3437870SNishanth Aravamudan return 0; 1778a3437870SNishanth Aravamudan } 1779a3437870SNishanth Aravamudan module_init(hugetlb_init); 1780a3437870SNishanth Aravamudan 1781a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1782a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1783a3437870SNishanth Aravamudan { 1784a3437870SNishanth Aravamudan struct hstate *h; 17858faa8b07SAndi Kleen unsigned long i; 17868faa8b07SAndi Kleen 1787a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1788a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1789a3437870SNishanth Aravamudan return; 1790a3437870SNishanth Aravamudan } 1791a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1792a3437870SNishanth Aravamudan BUG_ON(order == 0); 1793a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1794a3437870SNishanth Aravamudan h->order = order; 1795a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 17968faa8b07SAndi Kleen h->nr_huge_pages = 0; 17978faa8b07SAndi Kleen h->free_huge_pages = 0; 17988faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 17998faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 18009b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 18019b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1802a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1803a3437870SNishanth Aravamudan huge_page_size(h)/1024); 18048faa8b07SAndi Kleen 1805a3437870SNishanth Aravamudan parsed_hstate = h; 1806a3437870SNishanth Aravamudan } 1807a3437870SNishanth Aravamudan 1808e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1809a3437870SNishanth Aravamudan { 1810a3437870SNishanth Aravamudan unsigned long *mhp; 18118faa8b07SAndi Kleen static unsigned long *last_mhp; 1812a3437870SNishanth Aravamudan 1813a3437870SNishanth Aravamudan /* 1814a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1815a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1816a3437870SNishanth Aravamudan */ 1817a3437870SNishanth Aravamudan if (!max_hstate) 1818a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1819a3437870SNishanth Aravamudan else 1820a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1821a3437870SNishanth Aravamudan 18228faa8b07SAndi Kleen if (mhp == last_mhp) { 18238faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 18248faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 18258faa8b07SAndi Kleen return 1; 18268faa8b07SAndi Kleen } 18278faa8b07SAndi Kleen 1828a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1829a3437870SNishanth Aravamudan *mhp = 0; 1830a3437870SNishanth Aravamudan 18318faa8b07SAndi Kleen /* 18328faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 18338faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 18348faa8b07SAndi Kleen * use the bootmem allocator. 18358faa8b07SAndi Kleen */ 18368faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 18378faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 18388faa8b07SAndi Kleen 18398faa8b07SAndi Kleen last_mhp = mhp; 18408faa8b07SAndi Kleen 1841a3437870SNishanth Aravamudan return 1; 1842a3437870SNishanth Aravamudan } 1843e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1844e11bfbfcSNick Piggin 1845e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1846e11bfbfcSNick Piggin { 1847e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1848e11bfbfcSNick Piggin return 1; 1849e11bfbfcSNick Piggin } 1850e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1851a3437870SNishanth Aravamudan 18528a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 18538a213460SNishanth Aravamudan { 18548a213460SNishanth Aravamudan int node; 18558a213460SNishanth Aravamudan unsigned int nr = 0; 18568a213460SNishanth Aravamudan 18578a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 18588a213460SNishanth Aravamudan nr += array[node]; 18598a213460SNishanth Aravamudan 18608a213460SNishanth Aravamudan return nr; 18618a213460SNishanth Aravamudan } 18628a213460SNishanth Aravamudan 18638a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 186406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 186506808b08SLee Schermerhorn struct ctl_table *table, int write, 186606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 18671da177e4SLinus Torvalds { 1868e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1869e5ff2159SAndi Kleen unsigned long tmp; 1870e5ff2159SAndi Kleen 1871e5ff2159SAndi Kleen if (!write) 1872e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1873e5ff2159SAndi Kleen 1874e5ff2159SAndi Kleen table->data = &tmp; 1875e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 18768d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1877e5ff2159SAndi Kleen 187806808b08SLee Schermerhorn if (write) { 1879bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 1880bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 188106808b08SLee Schermerhorn if (!(obey_mempolicy && 188206808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 188306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 188406808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 188506808b08SLee Schermerhorn } 188606808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 188706808b08SLee Schermerhorn 188806808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 188906808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 189006808b08SLee Schermerhorn } 1891e5ff2159SAndi Kleen 18921da177e4SLinus Torvalds return 0; 18931da177e4SLinus Torvalds } 1894396faf03SMel Gorman 189506808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 189606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 189706808b08SLee Schermerhorn { 189806808b08SLee Schermerhorn 189906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 190006808b08SLee Schermerhorn buffer, length, ppos); 190106808b08SLee Schermerhorn } 190206808b08SLee Schermerhorn 190306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 190406808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 190506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 190606808b08SLee Schermerhorn { 190706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 190806808b08SLee Schermerhorn buffer, length, ppos); 190906808b08SLee Schermerhorn } 191006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 191106808b08SLee Schermerhorn 1912396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 19138d65af78SAlexey Dobriyan void __user *buffer, 1914396faf03SMel Gorman size_t *length, loff_t *ppos) 1915396faf03SMel Gorman { 19168d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 1917396faf03SMel Gorman if (hugepages_treat_as_movable) 1918396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1919396faf03SMel Gorman else 1920396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1921396faf03SMel Gorman return 0; 1922396faf03SMel Gorman } 1923396faf03SMel Gorman 1924a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 19258d65af78SAlexey Dobriyan void __user *buffer, 1926a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1927a3d0c6aaSNishanth Aravamudan { 1928a5516438SAndi Kleen struct hstate *h = &default_hstate; 1929e5ff2159SAndi Kleen unsigned long tmp; 1930e5ff2159SAndi Kleen 1931e5ff2159SAndi Kleen if (!write) 1932e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1933e5ff2159SAndi Kleen 1934e5ff2159SAndi Kleen table->data = &tmp; 1935e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 19368d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1937e5ff2159SAndi Kleen 1938e5ff2159SAndi Kleen if (write) { 1939064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1940e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1941a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1942e5ff2159SAndi Kleen } 1943e5ff2159SAndi Kleen 1944a3d0c6aaSNishanth Aravamudan return 0; 1945a3d0c6aaSNishanth Aravamudan } 1946a3d0c6aaSNishanth Aravamudan 19471da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 19481da177e4SLinus Torvalds 1949e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 19501da177e4SLinus Torvalds { 1951a5516438SAndi Kleen struct hstate *h = &default_hstate; 1952e1759c21SAlexey Dobriyan seq_printf(m, 19531da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 19541da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1955b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 19567893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 19574f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1958a5516438SAndi Kleen h->nr_huge_pages, 1959a5516438SAndi Kleen h->free_huge_pages, 1960a5516438SAndi Kleen h->resv_huge_pages, 1961a5516438SAndi Kleen h->surplus_huge_pages, 1962a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 19631da177e4SLinus Torvalds } 19641da177e4SLinus Torvalds 19651da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 19661da177e4SLinus Torvalds { 1967a5516438SAndi Kleen struct hstate *h = &default_hstate; 19681da177e4SLinus Torvalds return sprintf(buf, 19691da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1970a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1971a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1972a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1973a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1974a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 19751da177e4SLinus Torvalds } 19761da177e4SLinus Torvalds 19771da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 19781da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 19791da177e4SLinus Torvalds { 1980a5516438SAndi Kleen struct hstate *h = &default_hstate; 1981a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 19821da177e4SLinus Torvalds } 19831da177e4SLinus Torvalds 1984a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 1985fc1b8a73SMel Gorman { 1986fc1b8a73SMel Gorman int ret = -ENOMEM; 1987fc1b8a73SMel Gorman 1988fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 1989fc1b8a73SMel Gorman /* 1990fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 1991fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 1992fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 1993fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 1994fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 1995fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 1996fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 1997fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 1998fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 1999fc1b8a73SMel Gorman * 2000fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2001fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2002fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2003fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2004fc1b8a73SMel Gorman * semantics that cpuset has. 2005fc1b8a73SMel Gorman */ 2006fc1b8a73SMel Gorman if (delta > 0) { 2007a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2008fc1b8a73SMel Gorman goto out; 2009fc1b8a73SMel Gorman 2010a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2011a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2012fc1b8a73SMel Gorman goto out; 2013fc1b8a73SMel Gorman } 2014fc1b8a73SMel Gorman } 2015fc1b8a73SMel Gorman 2016fc1b8a73SMel Gorman ret = 0; 2017fc1b8a73SMel Gorman if (delta < 0) 2018a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2019fc1b8a73SMel Gorman 2020fc1b8a73SMel Gorman out: 2021fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2022fc1b8a73SMel Gorman return ret; 2023fc1b8a73SMel Gorman } 2024fc1b8a73SMel Gorman 202584afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 202684afd99bSAndy Whitcroft { 202784afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 202884afd99bSAndy Whitcroft 202984afd99bSAndy Whitcroft /* 203084afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 203184afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 203284afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 203384afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 203484afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 203584afd99bSAndy Whitcroft * new reference here without additional locking. 203684afd99bSAndy Whitcroft */ 203784afd99bSAndy Whitcroft if (reservations) 203884afd99bSAndy Whitcroft kref_get(&reservations->refs); 203984afd99bSAndy Whitcroft } 204084afd99bSAndy Whitcroft 2041a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2042a1e78772SMel Gorman { 2043a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 204484afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 204584afd99bSAndy Whitcroft unsigned long reserve; 204684afd99bSAndy Whitcroft unsigned long start; 204784afd99bSAndy Whitcroft unsigned long end; 204884afd99bSAndy Whitcroft 204984afd99bSAndy Whitcroft if (reservations) { 2050a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2051a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 205284afd99bSAndy Whitcroft 205384afd99bSAndy Whitcroft reserve = (end - start) - 205484afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 205584afd99bSAndy Whitcroft 205684afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 205784afd99bSAndy Whitcroft 20587251ff78SAdam Litke if (reserve) { 2059a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 20607251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 20617251ff78SAdam Litke } 2062a1e78772SMel Gorman } 206384afd99bSAndy Whitcroft } 2064a1e78772SMel Gorman 20651da177e4SLinus Torvalds /* 20661da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 20671da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 20681da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 20691da177e4SLinus Torvalds * this far. 20701da177e4SLinus Torvalds */ 2071d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 20721da177e4SLinus Torvalds { 20731da177e4SLinus Torvalds BUG(); 2074d0217ac0SNick Piggin return 0; 20751da177e4SLinus Torvalds } 20761da177e4SLinus Torvalds 2077f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2078d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 207984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2080a1e78772SMel Gorman .close = hugetlb_vm_op_close, 20811da177e4SLinus Torvalds }; 20821da177e4SLinus Torvalds 20831e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 20841e8f889bSDavid Gibson int writable) 208563551ae0SDavid Gibson { 208663551ae0SDavid Gibson pte_t entry; 208763551ae0SDavid Gibson 20881e8f889bSDavid Gibson if (writable) { 208963551ae0SDavid Gibson entry = 209063551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 209163551ae0SDavid Gibson } else { 20927f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 209363551ae0SDavid Gibson } 209463551ae0SDavid Gibson entry = pte_mkyoung(entry); 209563551ae0SDavid Gibson entry = pte_mkhuge(entry); 209663551ae0SDavid Gibson 209763551ae0SDavid Gibson return entry; 209863551ae0SDavid Gibson } 209963551ae0SDavid Gibson 21001e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 21011e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 21021e8f889bSDavid Gibson { 21031e8f889bSDavid Gibson pte_t entry; 21041e8f889bSDavid Gibson 21057f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 21067f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 21074b3073e1SRussell King update_mmu_cache(vma, address, ptep); 21081e8f889bSDavid Gibson } 21098dab5241SBenjamin Herrenschmidt } 21101e8f889bSDavid Gibson 21111e8f889bSDavid Gibson 211263551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 211363551ae0SDavid Gibson struct vm_area_struct *vma) 211463551ae0SDavid Gibson { 211563551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 211663551ae0SDavid Gibson struct page *ptepage; 21171c59827dSHugh Dickins unsigned long addr; 21181e8f889bSDavid Gibson int cow; 2119a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2120a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 21211e8f889bSDavid Gibson 21221e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 212363551ae0SDavid Gibson 2124a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2125c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2126c74df32cSHugh Dickins if (!src_pte) 2127c74df32cSHugh Dickins continue; 2128a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 212963551ae0SDavid Gibson if (!dst_pte) 213063551ae0SDavid Gibson goto nomem; 2131c5c99429SLarry Woodman 2132c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2133c5c99429SLarry Woodman if (dst_pte == src_pte) 2134c5c99429SLarry Woodman continue; 2135c5c99429SLarry Woodman 2136c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 213746478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 21387f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 21391e8f889bSDavid Gibson if (cow) 21407f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 21417f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 214263551ae0SDavid Gibson ptepage = pte_page(entry); 214363551ae0SDavid Gibson get_page(ptepage); 21440fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 214563551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 21461c59827dSHugh Dickins } 21471c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2148c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 214963551ae0SDavid Gibson } 215063551ae0SDavid Gibson return 0; 215163551ae0SDavid Gibson 215263551ae0SDavid Gibson nomem: 215363551ae0SDavid Gibson return -ENOMEM; 215463551ae0SDavid Gibson } 215563551ae0SDavid Gibson 2156fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2157fd6a03edSNaoya Horiguchi { 2158fd6a03edSNaoya Horiguchi swp_entry_t swp; 2159fd6a03edSNaoya Horiguchi 2160fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2161fd6a03edSNaoya Horiguchi return 0; 2162fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 2163fd6a03edSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) { 2164fd6a03edSNaoya Horiguchi return 1; 2165fd6a03edSNaoya Horiguchi } else 2166fd6a03edSNaoya Horiguchi return 0; 2167fd6a03edSNaoya Horiguchi } 2168fd6a03edSNaoya Horiguchi 2169502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 217004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 217163551ae0SDavid Gibson { 217263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 217363551ae0SDavid Gibson unsigned long address; 2174c7546f8fSDavid Gibson pte_t *ptep; 217563551ae0SDavid Gibson pte_t pte; 217663551ae0SDavid Gibson struct page *page; 2177fe1668aeSChen, Kenneth W struct page *tmp; 2178a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2179a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2180a5516438SAndi Kleen 2181c0a499c2SChen, Kenneth W /* 2182c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 2183c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 2184c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 2185c0a499c2SChen, Kenneth W */ 2186fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 218763551ae0SDavid Gibson 218863551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2189a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2190a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 219163551ae0SDavid Gibson 2192cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 2193508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2194a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2195c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2196c7546f8fSDavid Gibson if (!ptep) 2197c7546f8fSDavid Gibson continue; 2198c7546f8fSDavid Gibson 219939dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 220039dde65cSChen, Kenneth W continue; 220139dde65cSChen, Kenneth W 220204f2cbe3SMel Gorman /* 220304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 220404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 220504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 220604f2cbe3SMel Gorman */ 220704f2cbe3SMel Gorman if (ref_page) { 220804f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 220904f2cbe3SMel Gorman if (huge_pte_none(pte)) 221004f2cbe3SMel Gorman continue; 221104f2cbe3SMel Gorman page = pte_page(pte); 221204f2cbe3SMel Gorman if (page != ref_page) 221304f2cbe3SMel Gorman continue; 221404f2cbe3SMel Gorman 221504f2cbe3SMel Gorman /* 221604f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 221704f2cbe3SMel Gorman * future faults in this VMA will fail rather than 221804f2cbe3SMel Gorman * looking like data was lost 221904f2cbe3SMel Gorman */ 222004f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 222104f2cbe3SMel Gorman } 222204f2cbe3SMel Gorman 2223c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 22247f2e9525SGerald Schaefer if (huge_pte_none(pte)) 222563551ae0SDavid Gibson continue; 2226c7546f8fSDavid Gibson 2227fd6a03edSNaoya Horiguchi /* 2228fd6a03edSNaoya Horiguchi * HWPoisoned hugepage is already unmapped and dropped reference 2229fd6a03edSNaoya Horiguchi */ 2230fd6a03edSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 2231fd6a03edSNaoya Horiguchi continue; 2232fd6a03edSNaoya Horiguchi 223363551ae0SDavid Gibson page = pte_page(pte); 22346649a386SKen Chen if (pte_dirty(pte)) 22356649a386SKen Chen set_page_dirty(page); 2236fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 223763551ae0SDavid Gibson } 22381da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 2239508034a3SHugh Dickins flush_tlb_range(vma, start, end); 2240cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 2241fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 22420fe6e20bSNaoya Horiguchi page_remove_rmap(page); 2243fe1668aeSChen, Kenneth W list_del(&page->lru); 2244fe1668aeSChen, Kenneth W put_page(page); 2245fe1668aeSChen, Kenneth W } 22461da177e4SLinus Torvalds } 224763551ae0SDavid Gibson 2248502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 224904f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2250502717f4SChen, Kenneth W { 2251502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 225204f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 2253502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2254502717f4SChen, Kenneth W } 2255502717f4SChen, Kenneth W 225604f2cbe3SMel Gorman /* 225704f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 225804f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 225904f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 226004f2cbe3SMel Gorman * same region. 226104f2cbe3SMel Gorman */ 22622a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 22632a4b3dedSHarvey Harrison struct page *page, unsigned long address) 226404f2cbe3SMel Gorman { 22657526674dSAdam Litke struct hstate *h = hstate_vma(vma); 226604f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 226704f2cbe3SMel Gorman struct address_space *mapping; 226804f2cbe3SMel Gorman struct prio_tree_iter iter; 226904f2cbe3SMel Gorman pgoff_t pgoff; 227004f2cbe3SMel Gorman 227104f2cbe3SMel Gorman /* 227204f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 227304f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 227404f2cbe3SMel Gorman */ 22757526674dSAdam Litke address = address & huge_page_mask(h); 227604f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 227704f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 227804f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 227904f2cbe3SMel Gorman 22804eb2b1dcSMel Gorman /* 22814eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 22824eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 22834eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 22844eb2b1dcSMel Gorman */ 22854eb2b1dcSMel Gorman spin_lock(&mapping->i_mmap_lock); 228604f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 228704f2cbe3SMel Gorman /* Do not unmap the current VMA */ 228804f2cbe3SMel Gorman if (iter_vma == vma) 228904f2cbe3SMel Gorman continue; 229004f2cbe3SMel Gorman 229104f2cbe3SMel Gorman /* 229204f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 229304f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 229404f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 229504f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 229604f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 229704f2cbe3SMel Gorman */ 229804f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 22994eb2b1dcSMel Gorman __unmap_hugepage_range(iter_vma, 23007526674dSAdam Litke address, address + huge_page_size(h), 230104f2cbe3SMel Gorman page); 230204f2cbe3SMel Gorman } 23034eb2b1dcSMel Gorman spin_unlock(&mapping->i_mmap_lock); 230404f2cbe3SMel Gorman 230504f2cbe3SMel Gorman return 1; 230604f2cbe3SMel Gorman } 230704f2cbe3SMel Gorman 23080fe6e20bSNaoya Horiguchi /* 23090fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 23100fe6e20bSNaoya Horiguchi */ 23111e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 231204f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 231304f2cbe3SMel Gorman struct page *pagecache_page) 23141e8f889bSDavid Gibson { 2315a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 23161e8f889bSDavid Gibson struct page *old_page, *new_page; 231779ac6ba4SDavid Gibson int avoidcopy; 231804f2cbe3SMel Gorman int outside_reserve = 0; 23191e8f889bSDavid Gibson 23201e8f889bSDavid Gibson old_page = pte_page(pte); 23211e8f889bSDavid Gibson 232204f2cbe3SMel Gorman retry_avoidcopy: 23231e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 23241e8f889bSDavid Gibson * and just make the page writable */ 23250fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 23261e8f889bSDavid Gibson if (avoidcopy) { 23270fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 23280fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 23291e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 233083c54070SNick Piggin return 0; 23311e8f889bSDavid Gibson } 23321e8f889bSDavid Gibson 233304f2cbe3SMel Gorman /* 233404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 233504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 233604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 233704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 233804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 233904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 234004f2cbe3SMel Gorman * of the full address range. 234104f2cbe3SMel Gorman */ 2342f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 234304f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 234404f2cbe3SMel Gorman old_page != pagecache_page) 234504f2cbe3SMel Gorman outside_reserve = 1; 234604f2cbe3SMel Gorman 23471e8f889bSDavid Gibson page_cache_get(old_page); 2348b76c8cfbSLarry Woodman 2349b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2350b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 235104f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 23521e8f889bSDavid Gibson 23532fc39cecSAdam Litke if (IS_ERR(new_page)) { 23541e8f889bSDavid Gibson page_cache_release(old_page); 235504f2cbe3SMel Gorman 235604f2cbe3SMel Gorman /* 235704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 235804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 235904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 236004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 236104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 236204f2cbe3SMel Gorman */ 236304f2cbe3SMel Gorman if (outside_reserve) { 236404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 236504f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 236604f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 236704f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2368b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 236904f2cbe3SMel Gorman goto retry_avoidcopy; 237004f2cbe3SMel Gorman } 237104f2cbe3SMel Gorman WARN_ON_ONCE(1); 237204f2cbe3SMel Gorman } 237304f2cbe3SMel Gorman 2374b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2375b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 23762fc39cecSAdam Litke return -PTR_ERR(new_page); 23771e8f889bSDavid Gibson } 23781e8f889bSDavid Gibson 23790fe6e20bSNaoya Horiguchi /* 23800fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 23810fe6e20bSNaoya Horiguchi * anon_vma prepared. 23820fe6e20bSNaoya Horiguchi */ 23830fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) 23840fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 23850fe6e20bSNaoya Horiguchi 23869de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 23870ed361deSNick Piggin __SetPageUptodate(new_page); 23881e8f889bSDavid Gibson 2389b76c8cfbSLarry Woodman /* 2390b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2391b76c8cfbSLarry Woodman * before the page tables are altered 2392b76c8cfbSLarry Woodman */ 2393b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2394a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 23957f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 23961e8f889bSDavid Gibson /* Break COW */ 23973edd4fc9SDoug Doan mmu_notifier_invalidate_range_start(mm, 23983edd4fc9SDoug Doan address & huge_page_mask(h), 23993edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 24008fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 24011e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 24021e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 24030fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2404cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 24051e8f889bSDavid Gibson /* Make the old page be freed below */ 24061e8f889bSDavid Gibson new_page = old_page; 24073edd4fc9SDoug Doan mmu_notifier_invalidate_range_end(mm, 24083edd4fc9SDoug Doan address & huge_page_mask(h), 24093edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 24101e8f889bSDavid Gibson } 24111e8f889bSDavid Gibson page_cache_release(new_page); 24121e8f889bSDavid Gibson page_cache_release(old_page); 241383c54070SNick Piggin return 0; 24141e8f889bSDavid Gibson } 24151e8f889bSDavid Gibson 241604f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2417a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2418a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 241904f2cbe3SMel Gorman { 242004f2cbe3SMel Gorman struct address_space *mapping; 2421e7c4b0bfSAndy Whitcroft pgoff_t idx; 242204f2cbe3SMel Gorman 242304f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2424a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 242504f2cbe3SMel Gorman 242604f2cbe3SMel Gorman return find_lock_page(mapping, idx); 242704f2cbe3SMel Gorman } 242804f2cbe3SMel Gorman 24293ae77f43SHugh Dickins /* 24303ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 24313ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 24323ae77f43SHugh Dickins */ 24333ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 24342a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 24352a15efc9SHugh Dickins { 24362a15efc9SHugh Dickins struct address_space *mapping; 24372a15efc9SHugh Dickins pgoff_t idx; 24382a15efc9SHugh Dickins struct page *page; 24392a15efc9SHugh Dickins 24402a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 24412a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 24422a15efc9SHugh Dickins 24432a15efc9SHugh Dickins page = find_get_page(mapping, idx); 24442a15efc9SHugh Dickins if (page) 24452a15efc9SHugh Dickins put_page(page); 24462a15efc9SHugh Dickins return page != NULL; 24472a15efc9SHugh Dickins } 24482a15efc9SHugh Dickins 2449a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2450788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2451ac9b9c66SHugh Dickins { 2452a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2453ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2454e7c4b0bfSAndy Whitcroft pgoff_t idx; 24554c887265SAdam Litke unsigned long size; 24564c887265SAdam Litke struct page *page; 24574c887265SAdam Litke struct address_space *mapping; 24581e8f889bSDavid Gibson pte_t new_pte; 24594c887265SAdam Litke 246004f2cbe3SMel Gorman /* 246104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 246204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 246304f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 246404f2cbe3SMel Gorman */ 246504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 246604f2cbe3SMel Gorman printk(KERN_WARNING 246704f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 246804f2cbe3SMel Gorman current->pid); 246904f2cbe3SMel Gorman return ret; 247004f2cbe3SMel Gorman } 247104f2cbe3SMel Gorman 24724c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2473a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 24744c887265SAdam Litke 24754c887265SAdam Litke /* 24764c887265SAdam Litke * Use page lock to guard against racing truncation 24774c887265SAdam Litke * before we get page_table_lock. 24784c887265SAdam Litke */ 24796bda666aSChristoph Lameter retry: 24806bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 24816bda666aSChristoph Lameter if (!page) { 2482a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2483ebed4bfcSHugh Dickins if (idx >= size) 2484ebed4bfcSHugh Dickins goto out; 248504f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 24862fc39cecSAdam Litke if (IS_ERR(page)) { 24872fc39cecSAdam Litke ret = -PTR_ERR(page); 24886bda666aSChristoph Lameter goto out; 24896bda666aSChristoph Lameter } 2490a5516438SAndi Kleen clear_huge_page(page, address, huge_page_size(h)); 24910ed361deSNick Piggin __SetPageUptodate(page); 2492ac9b9c66SHugh Dickins 2493f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 24946bda666aSChristoph Lameter int err; 249545c682a6SKen Chen struct inode *inode = mapping->host; 24966bda666aSChristoph Lameter 24976bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 24986bda666aSChristoph Lameter if (err) { 24996bda666aSChristoph Lameter put_page(page); 25006bda666aSChristoph Lameter if (err == -EEXIST) 25016bda666aSChristoph Lameter goto retry; 25026bda666aSChristoph Lameter goto out; 25036bda666aSChristoph Lameter } 250445c682a6SKen Chen 250545c682a6SKen Chen spin_lock(&inode->i_lock); 2506a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 250745c682a6SKen Chen spin_unlock(&inode->i_lock); 25080fe6e20bSNaoya Horiguchi page_dup_rmap(page); 250923be7468SMel Gorman } else { 25106bda666aSChristoph Lameter lock_page(page); 25110fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 25120fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 25130fe6e20bSNaoya Horiguchi goto backout_unlocked; 251423be7468SMel Gorman } 25150fe6e20bSNaoya Horiguchi hugepage_add_new_anon_rmap(page, vma, address); 25160fe6e20bSNaoya Horiguchi } 25170fe6e20bSNaoya Horiguchi } else { 251857303d80SAndy Whitcroft /* 2519998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2520998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2521998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2522fd6a03edSNaoya Horiguchi */ 2523fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2524fd6a03edSNaoya Horiguchi ret = VM_FAULT_HWPOISON; 2525fd6a03edSNaoya Horiguchi goto backout_unlocked; 25266bda666aSChristoph Lameter } 2527998b4382SNaoya Horiguchi page_dup_rmap(page); 2528998b4382SNaoya Horiguchi } 25291e8f889bSDavid Gibson 253057303d80SAndy Whitcroft /* 253157303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 253257303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 253357303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 253457303d80SAndy Whitcroft * the spinlock. 253557303d80SAndy Whitcroft */ 2536788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 25372b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 25382b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 25392b26736cSAndy Whitcroft goto backout_unlocked; 25402b26736cSAndy Whitcroft } 254157303d80SAndy Whitcroft 2542ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2543a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 25444c887265SAdam Litke if (idx >= size) 25454c887265SAdam Litke goto backout; 25464c887265SAdam Litke 254783c54070SNick Piggin ret = 0; 25487f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 25494c887265SAdam Litke goto backout; 25504c887265SAdam Litke 25511e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 25521e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 25531e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 25541e8f889bSDavid Gibson 2555788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 25561e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 255704f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 25581e8f889bSDavid Gibson } 25591e8f889bSDavid Gibson 2560ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 25614c887265SAdam Litke unlock_page(page); 25624c887265SAdam Litke out: 2563ac9b9c66SHugh Dickins return ret; 25644c887265SAdam Litke 25654c887265SAdam Litke backout: 25664c887265SAdam Litke spin_unlock(&mm->page_table_lock); 25672b26736cSAndy Whitcroft backout_unlocked: 25684c887265SAdam Litke unlock_page(page); 25694c887265SAdam Litke put_page(page); 25704c887265SAdam Litke goto out; 2571ac9b9c66SHugh Dickins } 2572ac9b9c66SHugh Dickins 257386e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2574788c7df4SHugh Dickins unsigned long address, unsigned int flags) 257586e5216fSAdam Litke { 257686e5216fSAdam Litke pte_t *ptep; 257786e5216fSAdam Litke pte_t entry; 25781e8f889bSDavid Gibson int ret; 25790fe6e20bSNaoya Horiguchi struct page *page = NULL; 258057303d80SAndy Whitcroft struct page *pagecache_page = NULL; 25813935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2582a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 258386e5216fSAdam Litke 2584fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2585fd6a03edSNaoya Horiguchi if (ptep) { 2586fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2587fd6a03edSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2588fd6a03edSNaoya Horiguchi return VM_FAULT_HWPOISON; 2589fd6a03edSNaoya Horiguchi } 2590fd6a03edSNaoya Horiguchi 2591a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 259286e5216fSAdam Litke if (!ptep) 259386e5216fSAdam Litke return VM_FAULT_OOM; 259486e5216fSAdam Litke 25953935baa9SDavid Gibson /* 25963935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 25973935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 25983935baa9SDavid Gibson * the same page in the page cache. 25993935baa9SDavid Gibson */ 26003935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 26017f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 26027f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2603788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2604b4d1d99fSDavid Gibson goto out_mutex; 26053935baa9SDavid Gibson } 260686e5216fSAdam Litke 260783c54070SNick Piggin ret = 0; 26081e8f889bSDavid Gibson 260957303d80SAndy Whitcroft /* 261057303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 261157303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 261257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 261357303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 261457303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 261557303d80SAndy Whitcroft * consumed. 261657303d80SAndy Whitcroft */ 2617788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 26182b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 26192b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2620b4d1d99fSDavid Gibson goto out_mutex; 26212b26736cSAndy Whitcroft } 262257303d80SAndy Whitcroft 2623f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 262457303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 262557303d80SAndy Whitcroft vma, address); 262657303d80SAndy Whitcroft } 262757303d80SAndy Whitcroft 262856c9cfb1SNaoya Horiguchi /* 262956c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 263056c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 263156c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 263256c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 263356c9cfb1SNaoya Horiguchi * so no worry about deadlock. 263456c9cfb1SNaoya Horiguchi */ 26350fe6e20bSNaoya Horiguchi page = pte_page(entry); 263656c9cfb1SNaoya Horiguchi if (page != pagecache_page) 26370fe6e20bSNaoya Horiguchi lock_page(page); 26380fe6e20bSNaoya Horiguchi 26391e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 26401e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2641b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2642b4d1d99fSDavid Gibson goto out_page_table_lock; 2643b4d1d99fSDavid Gibson 2644b4d1d99fSDavid Gibson 2645788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2646b4d1d99fSDavid Gibson if (!pte_write(entry)) { 264757303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 264857303d80SAndy Whitcroft pagecache_page); 2649b4d1d99fSDavid Gibson goto out_page_table_lock; 2650b4d1d99fSDavid Gibson } 2651b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2652b4d1d99fSDavid Gibson } 2653b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2654788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2655788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 26564b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2657b4d1d99fSDavid Gibson 2658b4d1d99fSDavid Gibson out_page_table_lock: 26591e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 266057303d80SAndy Whitcroft 266157303d80SAndy Whitcroft if (pagecache_page) { 266257303d80SAndy Whitcroft unlock_page(pagecache_page); 266357303d80SAndy Whitcroft put_page(pagecache_page); 266457303d80SAndy Whitcroft } 266556c9cfb1SNaoya Horiguchi unlock_page(page); 266657303d80SAndy Whitcroft 2667b4d1d99fSDavid Gibson out_mutex: 26683935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 26691e8f889bSDavid Gibson 26701e8f889bSDavid Gibson return ret; 267186e5216fSAdam Litke } 267286e5216fSAdam Litke 2673ceb86879SAndi Kleen /* Can be overriden by architectures */ 2674ceb86879SAndi Kleen __attribute__((weak)) struct page * 2675ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2676ceb86879SAndi Kleen pud_t *pud, int write) 2677ceb86879SAndi Kleen { 2678ceb86879SAndi Kleen BUG(); 2679ceb86879SAndi Kleen return NULL; 2680ceb86879SAndi Kleen } 2681ceb86879SAndi Kleen 268263551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 268363551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 26845b23dbe8SAdam Litke unsigned long *position, int *length, int i, 26852a15efc9SHugh Dickins unsigned int flags) 268663551ae0SDavid Gibson { 2687d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2688d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 268963551ae0SDavid Gibson int remainder = *length; 2690a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 269163551ae0SDavid Gibson 26921c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 269363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 269463551ae0SDavid Gibson pte_t *pte; 26952a15efc9SHugh Dickins int absent; 269663551ae0SDavid Gibson struct page *page; 269763551ae0SDavid Gibson 26984c887265SAdam Litke /* 26994c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 27002a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 27014c887265SAdam Litke * first, for the page indexing below to work. 27024c887265SAdam Litke */ 2703a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 27042a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 270563551ae0SDavid Gibson 27062a15efc9SHugh Dickins /* 27072a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 27083ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 27093ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 27103ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 27113ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 27122a15efc9SHugh Dickins */ 27133ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 27143ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 27152a15efc9SHugh Dickins remainder = 0; 27162a15efc9SHugh Dickins break; 27172a15efc9SHugh Dickins } 27182a15efc9SHugh Dickins 27192a15efc9SHugh Dickins if (absent || 27202a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 27214c887265SAdam Litke int ret; 27224c887265SAdam Litke 27234c887265SAdam Litke spin_unlock(&mm->page_table_lock); 27242a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 27252a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 27264c887265SAdam Litke spin_lock(&mm->page_table_lock); 2727a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 27284c887265SAdam Litke continue; 27294c887265SAdam Litke 27301c59827dSHugh Dickins remainder = 0; 27311c59827dSHugh Dickins break; 27321c59827dSHugh Dickins } 273363551ae0SDavid Gibson 2734a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 27357f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2736d5d4b0aaSChen, Kenneth W same_page: 2737d6692183SChen, Kenneth W if (pages) { 273869d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 27394b2e38adSKOSAKI Motohiro get_page(pages[i]); 2740d6692183SChen, Kenneth W } 274163551ae0SDavid Gibson 274263551ae0SDavid Gibson if (vmas) 274363551ae0SDavid Gibson vmas[i] = vma; 274463551ae0SDavid Gibson 274563551ae0SDavid Gibson vaddr += PAGE_SIZE; 2746d5d4b0aaSChen, Kenneth W ++pfn_offset; 274763551ae0SDavid Gibson --remainder; 274863551ae0SDavid Gibson ++i; 2749d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2750a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2751d5d4b0aaSChen, Kenneth W /* 2752d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2753d5d4b0aaSChen, Kenneth W * of this compound page. 2754d5d4b0aaSChen, Kenneth W */ 2755d5d4b0aaSChen, Kenneth W goto same_page; 2756d5d4b0aaSChen, Kenneth W } 275763551ae0SDavid Gibson } 27581c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 275963551ae0SDavid Gibson *length = remainder; 276063551ae0SDavid Gibson *position = vaddr; 276163551ae0SDavid Gibson 27622a15efc9SHugh Dickins return i ? i : -EFAULT; 276363551ae0SDavid Gibson } 27648f860591SZhang, Yanmin 27658f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 27668f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 27678f860591SZhang, Yanmin { 27688f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 27698f860591SZhang, Yanmin unsigned long start = address; 27708f860591SZhang, Yanmin pte_t *ptep; 27718f860591SZhang, Yanmin pte_t pte; 2772a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 27738f860591SZhang, Yanmin 27748f860591SZhang, Yanmin BUG_ON(address >= end); 27758f860591SZhang, Yanmin flush_cache_range(vma, address, end); 27768f860591SZhang, Yanmin 277739dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 27788f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2779a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 27808f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 27818f860591SZhang, Yanmin if (!ptep) 27828f860591SZhang, Yanmin continue; 278339dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 278439dde65cSChen, Kenneth W continue; 27857f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 27868f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 27878f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 27888f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 27898f860591SZhang, Yanmin } 27908f860591SZhang, Yanmin } 27918f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 279239dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 27938f860591SZhang, Yanmin 27948f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 27958f860591SZhang, Yanmin } 27968f860591SZhang, Yanmin 2797a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2798a1e78772SMel Gorman long from, long to, 27995a6fe125SMel Gorman struct vm_area_struct *vma, 28005a6fe125SMel Gorman int acctflag) 2801e4e574b7SAdam Litke { 280217c9d12eSMel Gorman long ret, chg; 2803a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2804e4e574b7SAdam Litke 2805a1e78772SMel Gorman /* 280617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 280717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 280817c9d12eSMel Gorman * and filesystem quota without using reserves 280917c9d12eSMel Gorman */ 281017c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 281117c9d12eSMel Gorman return 0; 281217c9d12eSMel Gorman 281317c9d12eSMel Gorman /* 2814a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2815a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2816a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2817a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2818a1e78772SMel Gorman */ 2819f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2820e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 28215a6fe125SMel Gorman else { 28225a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 28235a6fe125SMel Gorman if (!resv_map) 28245a6fe125SMel Gorman return -ENOMEM; 28255a6fe125SMel Gorman 282617c9d12eSMel Gorman chg = to - from; 282717c9d12eSMel Gorman 28285a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 28295a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 28305a6fe125SMel Gorman } 28315a6fe125SMel Gorman 283217c9d12eSMel Gorman if (chg < 0) 283317c9d12eSMel Gorman return chg; 283417c9d12eSMel Gorman 283517c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 283617c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 283717c9d12eSMel Gorman return -ENOSPC; 283817c9d12eSMel Gorman 283917c9d12eSMel Gorman /* 284017c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 284117c9d12eSMel Gorman * Hand back the quota if there are not 284217c9d12eSMel Gorman */ 284317c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 284417c9d12eSMel Gorman if (ret < 0) { 284517c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 284617c9d12eSMel Gorman return ret; 284717c9d12eSMel Gorman } 284817c9d12eSMel Gorman 284917c9d12eSMel Gorman /* 285017c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 285117c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 285217c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 285317c9d12eSMel Gorman * the reservation was and the page cache tells how much of 285417c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 285517c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 285617c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 285717c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 285817c9d12eSMel Gorman * else has to be done for private mappings here 285917c9d12eSMel Gorman */ 2860f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 286117c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2862a43a8c39SChen, Kenneth W return 0; 2863a43a8c39SChen, Kenneth W } 2864a43a8c39SChen, Kenneth W 2865a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2866a43a8c39SChen, Kenneth W { 2867a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2868a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 286945c682a6SKen Chen 287045c682a6SKen Chen spin_lock(&inode->i_lock); 2871e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 287245c682a6SKen Chen spin_unlock(&inode->i_lock); 287345c682a6SKen Chen 287490d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2875a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2876a43a8c39SChen, Kenneth W } 287793f70f90SNaoya Horiguchi 287893f70f90SNaoya Horiguchi /* 287993f70f90SNaoya Horiguchi * This function is called from memory failure code. 288093f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 288193f70f90SNaoya Horiguchi */ 288293f70f90SNaoya Horiguchi void __isolate_hwpoisoned_huge_page(struct page *hpage) 288393f70f90SNaoya Horiguchi { 288493f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 288593f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 288693f70f90SNaoya Horiguchi 288793f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 288893f70f90SNaoya Horiguchi list_del(&hpage->lru); 288993f70f90SNaoya Horiguchi h->free_huge_pages--; 289093f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 289193f70f90SNaoya Horiguchi spin_unlock(&hugetlb_lock); 289293f70f90SNaoya Horiguchi } 2893