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 3970ebabb41SNaoya Horiguchi static void copy_gigantic_page(struct page *dst, struct page *src) 3980ebabb41SNaoya Horiguchi { 3990ebabb41SNaoya Horiguchi int i; 4000ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4010ebabb41SNaoya Horiguchi struct page *dst_base = dst; 4020ebabb41SNaoya Horiguchi struct page *src_base = src; 4030ebabb41SNaoya Horiguchi 4040ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); ) { 4050ebabb41SNaoya Horiguchi cond_resched(); 4060ebabb41SNaoya Horiguchi copy_highpage(dst, src); 4070ebabb41SNaoya Horiguchi 4080ebabb41SNaoya Horiguchi i++; 4090ebabb41SNaoya Horiguchi dst = mem_map_next(dst, dst_base, i); 4100ebabb41SNaoya Horiguchi src = mem_map_next(src, src_base, i); 4110ebabb41SNaoya Horiguchi } 4120ebabb41SNaoya Horiguchi } 4130ebabb41SNaoya Horiguchi 4140ebabb41SNaoya Horiguchi void copy_huge_page(struct page *dst, struct page *src) 4150ebabb41SNaoya Horiguchi { 4160ebabb41SNaoya Horiguchi int i; 4170ebabb41SNaoya Horiguchi struct hstate *h = page_hstate(src); 4180ebabb41SNaoya Horiguchi 4190ebabb41SNaoya Horiguchi if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 4200ebabb41SNaoya Horiguchi copy_gigantic_page(dst, src); 4210ebabb41SNaoya Horiguchi return; 4220ebabb41SNaoya Horiguchi } 4230ebabb41SNaoya Horiguchi 4240ebabb41SNaoya Horiguchi might_sleep(); 4250ebabb41SNaoya Horiguchi for (i = 0; i < pages_per_huge_page(h); i++) { 4260ebabb41SNaoya Horiguchi cond_resched(); 4270ebabb41SNaoya Horiguchi copy_highpage(dst + i, src + i); 4280ebabb41SNaoya Horiguchi } 4290ebabb41SNaoya Horiguchi } 4300ebabb41SNaoya Horiguchi 431a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 4321da177e4SLinus Torvalds { 4331da177e4SLinus Torvalds int nid = page_to_nid(page); 434a5516438SAndi Kleen list_add(&page->lru, &h->hugepage_freelists[nid]); 435a5516438SAndi Kleen h->free_huge_pages++; 436a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 4371da177e4SLinus Torvalds } 4381da177e4SLinus Torvalds 439bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 440bf50bab2SNaoya Horiguchi { 441bf50bab2SNaoya Horiguchi struct page *page; 442bf50bab2SNaoya Horiguchi 443bf50bab2SNaoya Horiguchi if (list_empty(&h->hugepage_freelists[nid])) 444bf50bab2SNaoya Horiguchi return NULL; 445bf50bab2SNaoya Horiguchi page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); 446bf50bab2SNaoya Horiguchi list_del(&page->lru); 447a9869b83SNaoya Horiguchi set_page_refcounted(page); 448bf50bab2SNaoya Horiguchi h->free_huge_pages--; 449bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 450bf50bab2SNaoya Horiguchi return page; 451bf50bab2SNaoya Horiguchi } 452bf50bab2SNaoya Horiguchi 453a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 454a5516438SAndi Kleen struct vm_area_struct *vma, 45504f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 4561da177e4SLinus Torvalds { 4571da177e4SLinus Torvalds struct page *page = NULL; 458480eccf9SLee Schermerhorn struct mempolicy *mpol; 45919770b32SMel Gorman nodemask_t *nodemask; 460c0ff7453SMiao Xie struct zonelist *zonelist; 461dd1a239fSMel Gorman struct zone *zone; 462dd1a239fSMel Gorman struct zoneref *z; 4631da177e4SLinus Torvalds 464c0ff7453SMiao Xie get_mems_allowed(); 465c0ff7453SMiao Xie zonelist = huge_zonelist(vma, address, 466c0ff7453SMiao Xie htlb_alloc_mask, &mpol, &nodemask); 467a1e78772SMel Gorman /* 468a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 469a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 470a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 471a1e78772SMel Gorman */ 4727f09ca51SMel Gorman if (!vma_has_reserves(vma) && 473a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 474c0ff7453SMiao Xie goto err; 475a1e78772SMel Gorman 47604f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 477a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 478c0ff7453SMiao Xie goto err;; 47904f2cbe3SMel Gorman 48019770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 48119770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 482bf50bab2SNaoya Horiguchi if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { 483bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 484bf50bab2SNaoya Horiguchi if (page) { 48504f2cbe3SMel Gorman if (!avoid_reserve) 486a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 4875ab3ee7bSKen Chen break; 4881da177e4SLinus Torvalds } 4893abf7afdSAndrew Morton } 490bf50bab2SNaoya Horiguchi } 491c0ff7453SMiao Xie err: 49252cd3b07SLee Schermerhorn mpol_cond_put(mpol); 493c0ff7453SMiao Xie put_mems_allowed(); 4941da177e4SLinus Torvalds return page; 4951da177e4SLinus Torvalds } 4961da177e4SLinus Torvalds 497a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 4986af2acb6SAdam Litke { 4996af2acb6SAdam Litke int i; 500a5516438SAndi Kleen 50118229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 50218229df5SAndy Whitcroft 503a5516438SAndi Kleen h->nr_huge_pages--; 504a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 505a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 5066af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 5076af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 5086af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 5096af2acb6SAdam Litke } 5106af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5116af2acb6SAdam Litke set_page_refcounted(page); 5127f2e9525SGerald Schaefer arch_release_hugepage(page); 513a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 5146af2acb6SAdam Litke } 5156af2acb6SAdam Litke 516e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 517e5ff2159SAndi Kleen { 518e5ff2159SAndi Kleen struct hstate *h; 519e5ff2159SAndi Kleen 520e5ff2159SAndi Kleen for_each_hstate(h) { 521e5ff2159SAndi Kleen if (huge_page_size(h) == size) 522e5ff2159SAndi Kleen return h; 523e5ff2159SAndi Kleen } 524e5ff2159SAndi Kleen return NULL; 525e5ff2159SAndi Kleen } 526e5ff2159SAndi Kleen 52727a85ef1SDavid Gibson static void free_huge_page(struct page *page) 52827a85ef1SDavid Gibson { 529a5516438SAndi Kleen /* 530a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 531a5516438SAndi Kleen * compound page destructor. 532a5516438SAndi Kleen */ 533e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 5347893d1d5SAdam Litke int nid = page_to_nid(page); 535c79fb75eSAdam Litke struct address_space *mapping; 53627a85ef1SDavid Gibson 537c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 538e5df70abSAndy Whitcroft set_page_private(page, 0); 53923be7468SMel Gorman page->mapping = NULL; 5407893d1d5SAdam Litke BUG_ON(page_count(page)); 5410fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 54227a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 54327a85ef1SDavid Gibson 54427a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 545aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 546a5516438SAndi Kleen update_and_free_page(h, page); 547a5516438SAndi Kleen h->surplus_huge_pages--; 548a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 5497893d1d5SAdam Litke } else { 550a5516438SAndi Kleen enqueue_huge_page(h, page); 5517893d1d5SAdam Litke } 55227a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 553c79fb75eSAdam Litke if (mapping) 5549a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 55527a85ef1SDavid Gibson } 55627a85ef1SDavid Gibson 557a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 558b7ba30c6SAndi Kleen { 559b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 560b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 561a5516438SAndi Kleen h->nr_huge_pages++; 562a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 563b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 564b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 565b7ba30c6SAndi Kleen } 566b7ba30c6SAndi Kleen 56720a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 56820a0307cSWu Fengguang { 56920a0307cSWu Fengguang int i; 57020a0307cSWu Fengguang int nr_pages = 1 << order; 57120a0307cSWu Fengguang struct page *p = page + 1; 57220a0307cSWu Fengguang 57320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 57420a0307cSWu Fengguang set_compound_order(page, order); 57520a0307cSWu Fengguang __SetPageHead(page); 57620a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 57720a0307cSWu Fengguang __SetPageTail(p); 57820a0307cSWu Fengguang p->first_page = page; 57920a0307cSWu Fengguang } 58020a0307cSWu Fengguang } 58120a0307cSWu Fengguang 58220a0307cSWu Fengguang int PageHuge(struct page *page) 58320a0307cSWu Fengguang { 58420a0307cSWu Fengguang compound_page_dtor *dtor; 58520a0307cSWu Fengguang 58620a0307cSWu Fengguang if (!PageCompound(page)) 58720a0307cSWu Fengguang return 0; 58820a0307cSWu Fengguang 58920a0307cSWu Fengguang page = compound_head(page); 59020a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 59120a0307cSWu Fengguang 59220a0307cSWu Fengguang return dtor == free_huge_page; 59320a0307cSWu Fengguang } 59420a0307cSWu Fengguang 59543131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 59643131e14SNaoya Horiguchi 597a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 5981da177e4SLinus Torvalds { 5991da177e4SLinus Torvalds struct page *page; 600f96efd58SJoe Jin 601aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 602aa888a74SAndi Kleen return NULL; 603aa888a74SAndi Kleen 6046484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 605551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 606551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 607a5516438SAndi Kleen huge_page_order(h)); 6081da177e4SLinus Torvalds if (page) { 6097f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 610caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6117b8ee84dSHarvey Harrison return NULL; 6127f2e9525SGerald Schaefer } 613a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6141da177e4SLinus Torvalds } 61563b4613cSNishanth Aravamudan 61663b4613cSNishanth Aravamudan return page; 61763b4613cSNishanth Aravamudan } 61863b4613cSNishanth Aravamudan 6195ced66c9SAndi Kleen /* 6206ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 6216ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 6226ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 6236ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 6246ae11b27SLee Schermerhorn * node for alloc or free. 6259a76db09SLee Schermerhorn */ 6266ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 6279a76db09SLee Schermerhorn { 6286ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 6299a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 6306ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 6319a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 6329a76db09SLee Schermerhorn 6339a76db09SLee Schermerhorn return nid; 6349a76db09SLee Schermerhorn } 6359a76db09SLee Schermerhorn 6366ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 6375ced66c9SAndi Kleen { 6386ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 6396ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 6409a76db09SLee Schermerhorn return nid; 6415ced66c9SAndi Kleen } 6425ced66c9SAndi Kleen 6436ae11b27SLee Schermerhorn /* 6446ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 6456ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 6466ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 6476ae11b27SLee Schermerhorn * mask. 6486ae11b27SLee Schermerhorn */ 6496ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 6506ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 6516ae11b27SLee Schermerhorn { 6526ae11b27SLee Schermerhorn int nid; 6536ae11b27SLee Schermerhorn 6546ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 6556ae11b27SLee Schermerhorn 6566ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 6576ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 6586ae11b27SLee Schermerhorn 6596ae11b27SLee Schermerhorn return nid; 6606ae11b27SLee Schermerhorn } 6616ae11b27SLee Schermerhorn 6626ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 66363b4613cSNishanth Aravamudan { 66463b4613cSNishanth Aravamudan struct page *page; 66563b4613cSNishanth Aravamudan int start_nid; 66663b4613cSNishanth Aravamudan int next_nid; 66763b4613cSNishanth Aravamudan int ret = 0; 66863b4613cSNishanth Aravamudan 6696ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 670e8c5c824SLee Schermerhorn next_nid = start_nid; 67163b4613cSNishanth Aravamudan 67263b4613cSNishanth Aravamudan do { 673e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 6749a76db09SLee Schermerhorn if (page) { 67563b4613cSNishanth Aravamudan ret = 1; 6769a76db09SLee Schermerhorn break; 6779a76db09SLee Schermerhorn } 6786ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 6799a76db09SLee Schermerhorn } while (next_nid != start_nid); 68063b4613cSNishanth Aravamudan 6813b116300SAdam Litke if (ret) 6823b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 6833b116300SAdam Litke else 6843b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 6853b116300SAdam Litke 68663b4613cSNishanth Aravamudan return ret; 6871da177e4SLinus Torvalds } 6881da177e4SLinus Torvalds 689e8c5c824SLee Schermerhorn /* 6906ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 6916ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 6926ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 6936ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 694e8c5c824SLee Schermerhorn */ 6956ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 696e8c5c824SLee Schermerhorn { 6976ae11b27SLee Schermerhorn int nid; 6989a76db09SLee Schermerhorn 6996ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7006ae11b27SLee Schermerhorn 7016ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7026ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7036ae11b27SLee Schermerhorn 7049a76db09SLee Schermerhorn return nid; 705e8c5c824SLee Schermerhorn } 706e8c5c824SLee Schermerhorn 707e8c5c824SLee Schermerhorn /* 708e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 709e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 710e8c5c824SLee Schermerhorn * balanced over allowed nodes. 711e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 712e8c5c824SLee Schermerhorn */ 7136ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 7146ae11b27SLee Schermerhorn bool acct_surplus) 715e8c5c824SLee Schermerhorn { 716e8c5c824SLee Schermerhorn int start_nid; 717e8c5c824SLee Schermerhorn int next_nid; 718e8c5c824SLee Schermerhorn int ret = 0; 719e8c5c824SLee Schermerhorn 7206ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 721e8c5c824SLee Schermerhorn next_nid = start_nid; 722e8c5c824SLee Schermerhorn 723e8c5c824SLee Schermerhorn do { 724685f3457SLee Schermerhorn /* 725685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 726685f3457SLee Schermerhorn * nodes with surplus pages. 727685f3457SLee Schermerhorn */ 728685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 729685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 730e8c5c824SLee Schermerhorn struct page *page = 731e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 732e8c5c824SLee Schermerhorn struct page, lru); 733e8c5c824SLee Schermerhorn list_del(&page->lru); 734e8c5c824SLee Schermerhorn h->free_huge_pages--; 735e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 736685f3457SLee Schermerhorn if (acct_surplus) { 737685f3457SLee Schermerhorn h->surplus_huge_pages--; 738685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 739685f3457SLee Schermerhorn } 740e8c5c824SLee Schermerhorn update_and_free_page(h, page); 741e8c5c824SLee Schermerhorn ret = 1; 7429a76db09SLee Schermerhorn break; 743e8c5c824SLee Schermerhorn } 7446ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 7459a76db09SLee Schermerhorn } while (next_nid != start_nid); 746e8c5c824SLee Schermerhorn 747e8c5c824SLee Schermerhorn return ret; 748e8c5c824SLee Schermerhorn } 749e8c5c824SLee Schermerhorn 750bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 7517893d1d5SAdam Litke { 7527893d1d5SAdam Litke struct page *page; 753bf50bab2SNaoya Horiguchi unsigned int r_nid; 7547893d1d5SAdam Litke 755aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 756aa888a74SAndi Kleen return NULL; 757aa888a74SAndi Kleen 758d1c3fb1fSNishanth Aravamudan /* 759d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 760d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 761d1c3fb1fSNishanth Aravamudan * overcommit 762d1c3fb1fSNishanth Aravamudan * 763d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 764d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 765d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 766d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 767d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 768d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 769d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 770d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 771d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 772d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 773d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 774d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 775d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 776d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 777d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 778d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 779d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 780d1c3fb1fSNishanth Aravamudan */ 781d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 782a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 783d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 784d1c3fb1fSNishanth Aravamudan return NULL; 785d1c3fb1fSNishanth Aravamudan } else { 786a5516438SAndi Kleen h->nr_huge_pages++; 787a5516438SAndi Kleen h->surplus_huge_pages++; 788d1c3fb1fSNishanth Aravamudan } 789d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 790d1c3fb1fSNishanth Aravamudan 791bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 792551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 793551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 794a5516438SAndi Kleen huge_page_order(h)); 795bf50bab2SNaoya Horiguchi else 796bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 797bf50bab2SNaoya Horiguchi htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 798bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 799d1c3fb1fSNishanth Aravamudan 800caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 801caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 802caff3a2cSGerald Schaefer return NULL; 803caff3a2cSGerald Schaefer } 804caff3a2cSGerald Schaefer 8057893d1d5SAdam Litke spin_lock(&hugetlb_lock); 806d1c3fb1fSNishanth Aravamudan if (page) { 807bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 808d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 809d1c3fb1fSNishanth Aravamudan /* 810d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 811d1c3fb1fSNishanth Aravamudan */ 812bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 813bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 8143b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 815d1c3fb1fSNishanth Aravamudan } else { 816a5516438SAndi Kleen h->nr_huge_pages--; 817a5516438SAndi Kleen h->surplus_huge_pages--; 8183b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 8197893d1d5SAdam Litke } 820d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 8217893d1d5SAdam Litke 8227893d1d5SAdam Litke return page; 8237893d1d5SAdam Litke } 8247893d1d5SAdam Litke 825e4e574b7SAdam Litke /* 826bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 827bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 828bf50bab2SNaoya Horiguchi * address of error page. 829bf50bab2SNaoya Horiguchi */ 830bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 831bf50bab2SNaoya Horiguchi { 832bf50bab2SNaoya Horiguchi struct page *page; 833bf50bab2SNaoya Horiguchi 834bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 835bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 836bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 837bf50bab2SNaoya Horiguchi 838bf50bab2SNaoya Horiguchi if (!page) 839bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 840bf50bab2SNaoya Horiguchi 841bf50bab2SNaoya Horiguchi return page; 842bf50bab2SNaoya Horiguchi } 843bf50bab2SNaoya Horiguchi 844bf50bab2SNaoya Horiguchi /* 845e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 846e4e574b7SAdam Litke * of size 'delta'. 847e4e574b7SAdam Litke */ 848a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 849e4e574b7SAdam Litke { 850e4e574b7SAdam Litke struct list_head surplus_list; 851e4e574b7SAdam Litke struct page *page, *tmp; 852e4e574b7SAdam Litke int ret, i; 853e4e574b7SAdam Litke int needed, allocated; 854e4e574b7SAdam Litke 855a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 856ac09b3a1SAdam Litke if (needed <= 0) { 857a5516438SAndi Kleen h->resv_huge_pages += delta; 858e4e574b7SAdam Litke return 0; 859ac09b3a1SAdam Litke } 860e4e574b7SAdam Litke 861e4e574b7SAdam Litke allocated = 0; 862e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 863e4e574b7SAdam Litke 864e4e574b7SAdam Litke ret = -ENOMEM; 865e4e574b7SAdam Litke retry: 866e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 867e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 868bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 869a9869b83SNaoya Horiguchi if (!page) 870e4e574b7SAdam Litke /* 871e4e574b7SAdam Litke * We were not able to allocate enough pages to 872e4e574b7SAdam Litke * satisfy the entire reservation so we free what 873e4e574b7SAdam Litke * we've allocated so far. 874e4e574b7SAdam Litke */ 875e4e574b7SAdam Litke goto free; 876e4e574b7SAdam Litke 877e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 878e4e574b7SAdam Litke } 879e4e574b7SAdam Litke allocated += needed; 880e4e574b7SAdam Litke 881e4e574b7SAdam Litke /* 882e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 883e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 884e4e574b7SAdam Litke */ 885e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 886a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 887a5516438SAndi Kleen (h->free_huge_pages + allocated); 888e4e574b7SAdam Litke if (needed > 0) 889e4e574b7SAdam Litke goto retry; 890e4e574b7SAdam Litke 891e4e574b7SAdam Litke /* 892e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 893e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 894e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 895ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 896ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 897ac09b3a1SAdam Litke * before they are reserved. 898e4e574b7SAdam Litke */ 899e4e574b7SAdam Litke needed += allocated; 900a5516438SAndi Kleen h->resv_huge_pages += delta; 901e4e574b7SAdam Litke ret = 0; 902a9869b83SNaoya Horiguchi 903a9869b83SNaoya Horiguchi spin_unlock(&hugetlb_lock); 90419fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 90519fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 90619fc3f0aSAdam Litke if ((--needed) < 0) 90719fc3f0aSAdam Litke break; 90819fc3f0aSAdam Litke list_del(&page->lru); 909a9869b83SNaoya Horiguchi /* 910a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 911a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 912a9869b83SNaoya Horiguchi */ 913a9869b83SNaoya Horiguchi put_page_testzero(page); 914a9869b83SNaoya Horiguchi VM_BUG_ON(page_count(page)); 915a5516438SAndi Kleen enqueue_huge_page(h, page); 91619fc3f0aSAdam Litke } 91719fc3f0aSAdam Litke 91819fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 919a9869b83SNaoya Horiguchi free: 92019fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 921e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 922e4e574b7SAdam Litke list_del(&page->lru); 923a9869b83SNaoya Horiguchi put_page(page); 924a9869b83SNaoya Horiguchi } 925af767cbdSAdam Litke } 92619fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 927e4e574b7SAdam Litke 928e4e574b7SAdam Litke return ret; 929e4e574b7SAdam Litke } 930e4e574b7SAdam Litke 931e4e574b7SAdam Litke /* 932e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 933e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 934e4e574b7SAdam Litke * never used. 935685f3457SLee Schermerhorn * Called with hugetlb_lock held. 936e4e574b7SAdam Litke */ 937a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 938a5516438SAndi Kleen unsigned long unused_resv_pages) 939e4e574b7SAdam Litke { 940e4e574b7SAdam Litke unsigned long nr_pages; 941e4e574b7SAdam Litke 942ac09b3a1SAdam Litke /* Uncommit the reservation */ 943a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 944ac09b3a1SAdam Litke 945aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 946aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 947aa888a74SAndi Kleen return; 948aa888a74SAndi Kleen 949a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 950e4e574b7SAdam Litke 951685f3457SLee Schermerhorn /* 952685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 9539b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 9549b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 9559b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 9569b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 9579b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 958685f3457SLee Schermerhorn */ 959685f3457SLee Schermerhorn while (nr_pages--) { 9609b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 961685f3457SLee Schermerhorn break; 962e4e574b7SAdam Litke } 963e4e574b7SAdam Litke } 964e4e574b7SAdam Litke 965c37f9fb1SAndy Whitcroft /* 966c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 967c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 968c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 969c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 970c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 971c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 972c37f9fb1SAndy Whitcroft * No action is required on failure. 973c37f9fb1SAndy Whitcroft */ 974e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 975a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 976c37f9fb1SAndy Whitcroft { 977c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 978c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 979c37f9fb1SAndy Whitcroft 980f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 981a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 982c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 983c37f9fb1SAndy Whitcroft idx, idx + 1); 984c37f9fb1SAndy Whitcroft 98584afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 986c37f9fb1SAndy Whitcroft return 1; 987c37f9fb1SAndy Whitcroft 98884afd99bSAndy Whitcroft } else { 989e2f17d94SRoel Kluin long err; 990a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 99184afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 99284afd99bSAndy Whitcroft 99384afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 99484afd99bSAndy Whitcroft if (err < 0) 99584afd99bSAndy Whitcroft return err; 996c37f9fb1SAndy Whitcroft return 0; 997c37f9fb1SAndy Whitcroft } 99884afd99bSAndy Whitcroft } 999a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1000a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1001c37f9fb1SAndy Whitcroft { 1002c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1003c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1004c37f9fb1SAndy Whitcroft 1005f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1006a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1007c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 100884afd99bSAndy Whitcroft 100984afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1010a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 101184afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 101284afd99bSAndy Whitcroft 101384afd99bSAndy Whitcroft /* Mark this page used in the map. */ 101484afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1015c37f9fb1SAndy Whitcroft } 1016c37f9fb1SAndy Whitcroft } 1017c37f9fb1SAndy Whitcroft 1018348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 101904f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1020348ea204SAdam Litke { 1021a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1022348ea204SAdam Litke struct page *page; 10232fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 1024a1e78772SMel Gorman struct inode *inode = mapping->host; 1025e2f17d94SRoel Kluin long chg; 10262fc39cecSAdam Litke 1027a1e78772SMel Gorman /* 1028a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 1029a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 1030a1e78772SMel Gorman * made before satisfying the allocation 1031c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 1032c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 1033a1e78772SMel Gorman */ 1034a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1035c37f9fb1SAndy Whitcroft if (chg < 0) 1036c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 1037c37f9fb1SAndy Whitcroft if (chg) 1038a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 1039a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 104090d8b7e6SAdam Litke 1041a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1042a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1043a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1044a1e78772SMel Gorman 1045a1e78772SMel Gorman if (!page) { 1046bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1047a1e78772SMel Gorman if (!page) { 1048a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 10494a6018f7SMel Gorman return ERR_PTR(-VM_FAULT_SIGBUS); 1050a1e78772SMel Gorman } 1051a1e78772SMel Gorman } 1052a1e78772SMel Gorman 10532fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 1054a1e78772SMel Gorman 1055a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1056c37f9fb1SAndy Whitcroft 10577893d1d5SAdam Litke return page; 1058b45b5bd6SDavid Gibson } 1059b45b5bd6SDavid Gibson 106091f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1061aa888a74SAndi Kleen { 1062aa888a74SAndi Kleen struct huge_bootmem_page *m; 10639b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1064aa888a74SAndi Kleen 1065aa888a74SAndi Kleen while (nr_nodes) { 1066aa888a74SAndi Kleen void *addr; 1067aa888a74SAndi Kleen 1068aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 10696ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 10709b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1071aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1072aa888a74SAndi Kleen 1073aa888a74SAndi Kleen if (addr) { 1074aa888a74SAndi Kleen /* 1075aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1076aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1077aa888a74SAndi Kleen * puts them into the mem_map). 1078aa888a74SAndi Kleen */ 1079aa888a74SAndi Kleen m = addr; 1080aa888a74SAndi Kleen goto found; 1081aa888a74SAndi Kleen } 1082aa888a74SAndi Kleen nr_nodes--; 1083aa888a74SAndi Kleen } 1084aa888a74SAndi Kleen return 0; 1085aa888a74SAndi Kleen 1086aa888a74SAndi Kleen found: 1087aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1088aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1089aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1090aa888a74SAndi Kleen m->hstate = h; 1091aa888a74SAndi Kleen return 1; 1092aa888a74SAndi Kleen } 1093aa888a74SAndi Kleen 109418229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 109518229df5SAndy Whitcroft { 109618229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 109718229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 109818229df5SAndy Whitcroft else 109918229df5SAndy Whitcroft prep_compound_page(page, order); 110018229df5SAndy Whitcroft } 110118229df5SAndy Whitcroft 1102aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1103aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1104aa888a74SAndi Kleen { 1105aa888a74SAndi Kleen struct huge_bootmem_page *m; 1106aa888a74SAndi Kleen 1107aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1108aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1109aa888a74SAndi Kleen struct hstate *h = m->hstate; 1110aa888a74SAndi Kleen __ClearPageReserved(page); 1111aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 111218229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1113aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1114aa888a74SAndi Kleen } 1115aa888a74SAndi Kleen } 1116aa888a74SAndi Kleen 11178faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 11181da177e4SLinus Torvalds { 11191da177e4SLinus Torvalds unsigned long i; 11201da177e4SLinus Torvalds 1121e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1122aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1123aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1124aa888a74SAndi Kleen break; 11259b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 11269b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 11271da177e4SLinus Torvalds break; 11281da177e4SLinus Torvalds } 11298faa8b07SAndi Kleen h->max_huge_pages = i; 1130e5ff2159SAndi Kleen } 1131e5ff2159SAndi Kleen 1132e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1133e5ff2159SAndi Kleen { 1134e5ff2159SAndi Kleen struct hstate *h; 1135e5ff2159SAndi Kleen 1136e5ff2159SAndi Kleen for_each_hstate(h) { 11378faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 11388faa8b07SAndi Kleen if (h->order < MAX_ORDER) 11398faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1140e5ff2159SAndi Kleen } 1141e5ff2159SAndi Kleen } 1142e5ff2159SAndi Kleen 11434abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 11444abd32dbSAndi Kleen { 11454abd32dbSAndi Kleen if (n >= (1UL << 30)) 11464abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 11474abd32dbSAndi Kleen else if (n >= (1UL << 20)) 11484abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 11494abd32dbSAndi Kleen else 11504abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 11514abd32dbSAndi Kleen return buf; 11524abd32dbSAndi Kleen } 11534abd32dbSAndi Kleen 1154e5ff2159SAndi Kleen static void __init report_hugepages(void) 1155e5ff2159SAndi Kleen { 1156e5ff2159SAndi Kleen struct hstate *h; 1157e5ff2159SAndi Kleen 1158e5ff2159SAndi Kleen for_each_hstate(h) { 11594abd32dbSAndi Kleen char buf[32]; 11604abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11614abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11624abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11634abd32dbSAndi Kleen h->free_huge_pages); 1164e5ff2159SAndi Kleen } 1165e5ff2159SAndi Kleen } 1166e5ff2159SAndi Kleen 11671da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 11686ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 11696ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11701da177e4SLinus Torvalds { 11714415cc8dSChristoph Lameter int i; 11724415cc8dSChristoph Lameter 1173aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1174aa888a74SAndi Kleen return; 1175aa888a74SAndi Kleen 11766ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 11771da177e4SLinus Torvalds struct page *page, *next; 1178a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1179a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1180a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11816b0c880dSAdam Litke return; 11821da177e4SLinus Torvalds if (PageHighMem(page)) 11831da177e4SLinus Torvalds continue; 11841da177e4SLinus Torvalds list_del(&page->lru); 1185e5ff2159SAndi Kleen update_and_free_page(h, page); 1186a5516438SAndi Kleen h->free_huge_pages--; 1187a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11881da177e4SLinus Torvalds } 11891da177e4SLinus Torvalds } 11901da177e4SLinus Torvalds } 11911da177e4SLinus Torvalds #else 11926ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 11936ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11941da177e4SLinus Torvalds { 11951da177e4SLinus Torvalds } 11961da177e4SLinus Torvalds #endif 11971da177e4SLinus Torvalds 119820a0307cSWu Fengguang /* 119920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 120020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 120120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 120220a0307cSWu Fengguang */ 12036ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 12046ae11b27SLee Schermerhorn int delta) 120520a0307cSWu Fengguang { 1206e8c5c824SLee Schermerhorn int start_nid, next_nid; 120720a0307cSWu Fengguang int ret = 0; 120820a0307cSWu Fengguang 120920a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 121020a0307cSWu Fengguang 1211e8c5c824SLee Schermerhorn if (delta < 0) 12126ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1213e8c5c824SLee Schermerhorn else 12146ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1215e8c5c824SLee Schermerhorn next_nid = start_nid; 1216e8c5c824SLee Schermerhorn 1217e8c5c824SLee Schermerhorn do { 1218e8c5c824SLee Schermerhorn int nid = next_nid; 1219e8c5c824SLee Schermerhorn if (delta < 0) { 1220e8c5c824SLee Schermerhorn /* 1221e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1222e8c5c824SLee Schermerhorn */ 12239a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 12246ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 12256ae11b27SLee Schermerhorn nodes_allowed); 122620a0307cSWu Fengguang continue; 1227e8c5c824SLee Schermerhorn } 12289a76db09SLee Schermerhorn } 1229e8c5c824SLee Schermerhorn if (delta > 0) { 1230e8c5c824SLee Schermerhorn /* 1231e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1232e8c5c824SLee Schermerhorn */ 1233e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 12349a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 12356ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 12366ae11b27SLee Schermerhorn nodes_allowed); 123720a0307cSWu Fengguang continue; 1238e8c5c824SLee Schermerhorn } 12399a76db09SLee Schermerhorn } 124020a0307cSWu Fengguang 124120a0307cSWu Fengguang h->surplus_huge_pages += delta; 124220a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 124320a0307cSWu Fengguang ret = 1; 124420a0307cSWu Fengguang break; 1245e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 124620a0307cSWu Fengguang 124720a0307cSWu Fengguang return ret; 124820a0307cSWu Fengguang } 124920a0307cSWu Fengguang 1250a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 12516ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 12526ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12531da177e4SLinus Torvalds { 12547893d1d5SAdam Litke unsigned long min_count, ret; 12551da177e4SLinus Torvalds 1256aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1257aa888a74SAndi Kleen return h->max_huge_pages; 1258aa888a74SAndi Kleen 12597893d1d5SAdam Litke /* 12607893d1d5SAdam Litke * Increase the pool size 12617893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 12627893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1263d1c3fb1fSNishanth Aravamudan * 1264d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1265d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1266d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1267d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1268d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 12697893d1d5SAdam Litke */ 12701da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1271a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 12726ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 12737893d1d5SAdam Litke break; 12747893d1d5SAdam Litke } 12757893d1d5SAdam Litke 1276a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 12777893d1d5SAdam Litke /* 12787893d1d5SAdam Litke * If this allocation races such that we no longer need the 12797893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12807893d1d5SAdam Litke * and reducing the surplus. 12817893d1d5SAdam Litke */ 12827893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 12836ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 12847893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12857893d1d5SAdam Litke if (!ret) 12867893d1d5SAdam Litke goto out; 12877893d1d5SAdam Litke 1288536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1289536240f2SMel Gorman if (signal_pending(current)) 1290536240f2SMel Gorman goto out; 12917893d1d5SAdam Litke } 12927893d1d5SAdam Litke 12937893d1d5SAdam Litke /* 12947893d1d5SAdam Litke * Decrease the pool size 12957893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 12967893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 12977893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 12987893d1d5SAdam Litke * to the desired size as pages become free. 1299d1c3fb1fSNishanth Aravamudan * 1300d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1301d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1302d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1303d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1304d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1305d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1306d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 13077893d1d5SAdam Litke */ 1308a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 13096b0c880dSAdam Litke min_count = max(count, min_count); 13106ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1311a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 13126ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 13131da177e4SLinus Torvalds break; 13141da177e4SLinus Torvalds } 1315a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 13166ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 13177893d1d5SAdam Litke break; 13187893d1d5SAdam Litke } 13197893d1d5SAdam Litke out: 1320a5516438SAndi Kleen ret = persistent_huge_pages(h); 13211da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 13227893d1d5SAdam Litke return ret; 13231da177e4SLinus Torvalds } 13241da177e4SLinus Torvalds 1325a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1326a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1327a3437870SNishanth Aravamudan 1328a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1329a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1330a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1331a3437870SNishanth Aravamudan 1332a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1333a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1334a3437870SNishanth Aravamudan 13359a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 13369a305230SLee Schermerhorn 13379a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1338a3437870SNishanth Aravamudan { 1339a3437870SNishanth Aravamudan int i; 13409a305230SLee Schermerhorn 1341a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 13429a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 13439a305230SLee Schermerhorn if (nidp) 13449a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1345a3437870SNishanth Aravamudan return &hstates[i]; 13469a305230SLee Schermerhorn } 13479a305230SLee Schermerhorn 13489a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1349a3437870SNishanth Aravamudan } 1350a3437870SNishanth Aravamudan 135106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1352a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1353a3437870SNishanth Aravamudan { 13549a305230SLee Schermerhorn struct hstate *h; 13559a305230SLee Schermerhorn unsigned long nr_huge_pages; 13569a305230SLee Schermerhorn int nid; 13579a305230SLee Schermerhorn 13589a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13599a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 13609a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 13619a305230SLee Schermerhorn else 13629a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 13639a305230SLee Schermerhorn 13649a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1365a3437870SNishanth Aravamudan } 1366adbe8726SEric B Munson 136706808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 136806808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 136906808b08SLee Schermerhorn const char *buf, size_t len) 1370a3437870SNishanth Aravamudan { 1371a3437870SNishanth Aravamudan int err; 13729a305230SLee Schermerhorn int nid; 137306808b08SLee Schermerhorn unsigned long count; 13749a305230SLee Schermerhorn struct hstate *h; 1375bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1376a3437870SNishanth Aravamudan 137706808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 137873ae31e5SEric B Munson if (err) 1379adbe8726SEric B Munson goto out; 1380a3437870SNishanth Aravamudan 13819a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1382adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1383adbe8726SEric B Munson err = -EINVAL; 1384adbe8726SEric B Munson goto out; 1385adbe8726SEric B Munson } 1386adbe8726SEric B Munson 13879a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 13889a305230SLee Schermerhorn /* 13899a305230SLee Schermerhorn * global hstate attribute 13909a305230SLee Schermerhorn */ 13919a305230SLee Schermerhorn if (!(obey_mempolicy && 13929a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 139306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 13949a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 139506808b08SLee Schermerhorn } 13969a305230SLee Schermerhorn } else if (nodes_allowed) { 13979a305230SLee Schermerhorn /* 13989a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 13999a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 14009a305230SLee Schermerhorn */ 14019a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 14029a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 14039a305230SLee Schermerhorn } else 14049a305230SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 14059a305230SLee Schermerhorn 140606808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1407a3437870SNishanth Aravamudan 14089b5e5d0fSLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 140906808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 141006808b08SLee Schermerhorn 141106808b08SLee Schermerhorn return len; 1412adbe8726SEric B Munson out: 1413adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1414adbe8726SEric B Munson return err; 141506808b08SLee Schermerhorn } 141606808b08SLee Schermerhorn 141706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 141806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 141906808b08SLee Schermerhorn { 142006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 142106808b08SLee Schermerhorn } 142206808b08SLee Schermerhorn 142306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 142406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 142506808b08SLee Schermerhorn { 142606808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1427a3437870SNishanth Aravamudan } 1428a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1429a3437870SNishanth Aravamudan 143006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 143106808b08SLee Schermerhorn 143206808b08SLee Schermerhorn /* 143306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 143406808b08SLee Schermerhorn * huge page alloc/free. 143506808b08SLee Schermerhorn */ 143606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 143706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 143806808b08SLee Schermerhorn { 143906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 144006808b08SLee Schermerhorn } 144106808b08SLee Schermerhorn 144206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 144306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 144406808b08SLee Schermerhorn { 144506808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 144606808b08SLee Schermerhorn } 144706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 144806808b08SLee Schermerhorn #endif 144906808b08SLee Schermerhorn 145006808b08SLee Schermerhorn 1451a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1452a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1453a3437870SNishanth Aravamudan { 14549a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1455a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1456a3437870SNishanth Aravamudan } 1457adbe8726SEric B Munson 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 1465adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1466adbe8726SEric B Munson return -EINVAL; 1467adbe8726SEric B Munson 1468a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1469a3437870SNishanth Aravamudan if (err) 147073ae31e5SEric B Munson return err; 1471a3437870SNishanth Aravamudan 1472a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1473a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1474a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1475a3437870SNishanth Aravamudan 1476a3437870SNishanth Aravamudan return count; 1477a3437870SNishanth Aravamudan } 1478a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1479a3437870SNishanth Aravamudan 1480a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1481a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1482a3437870SNishanth Aravamudan { 14839a305230SLee Schermerhorn struct hstate *h; 14849a305230SLee Schermerhorn unsigned long free_huge_pages; 14859a305230SLee Schermerhorn int nid; 14869a305230SLee Schermerhorn 14879a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14889a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14899a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 14909a305230SLee Schermerhorn else 14919a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 14929a305230SLee Schermerhorn 14939a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1494a3437870SNishanth Aravamudan } 1495a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1496a3437870SNishanth Aravamudan 1497a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1498a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1499a3437870SNishanth Aravamudan { 15009a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1501a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1502a3437870SNishanth Aravamudan } 1503a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1504a3437870SNishanth Aravamudan 1505a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1506a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1507a3437870SNishanth Aravamudan { 15089a305230SLee Schermerhorn struct hstate *h; 15099a305230SLee Schermerhorn unsigned long surplus_huge_pages; 15109a305230SLee Schermerhorn int nid; 15119a305230SLee Schermerhorn 15129a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15139a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15149a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 15159a305230SLee Schermerhorn else 15169a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 15179a305230SLee Schermerhorn 15189a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1519a3437870SNishanth Aravamudan } 1520a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1521a3437870SNishanth Aravamudan 1522a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1523a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1524a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1525a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1526a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1527a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 152806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 152906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 153006808b08SLee Schermerhorn #endif 1531a3437870SNishanth Aravamudan NULL, 1532a3437870SNishanth Aravamudan }; 1533a3437870SNishanth Aravamudan 1534a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1535a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1536a3437870SNishanth Aravamudan }; 1537a3437870SNishanth Aravamudan 1538094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 15399a305230SLee Schermerhorn struct kobject **hstate_kobjs, 15409a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1541a3437870SNishanth Aravamudan { 1542a3437870SNishanth Aravamudan int retval; 15439a305230SLee Schermerhorn int hi = h - hstates; 1544a3437870SNishanth Aravamudan 15459a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 15469a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1547a3437870SNishanth Aravamudan return -ENOMEM; 1548a3437870SNishanth Aravamudan 15499a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1550a3437870SNishanth Aravamudan if (retval) 15519a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1552a3437870SNishanth Aravamudan 1553a3437870SNishanth Aravamudan return retval; 1554a3437870SNishanth Aravamudan } 1555a3437870SNishanth Aravamudan 1556a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1557a3437870SNishanth Aravamudan { 1558a3437870SNishanth Aravamudan struct hstate *h; 1559a3437870SNishanth Aravamudan int err; 1560a3437870SNishanth Aravamudan 1561a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1562a3437870SNishanth Aravamudan if (!hugepages_kobj) 1563a3437870SNishanth Aravamudan return; 1564a3437870SNishanth Aravamudan 1565a3437870SNishanth Aravamudan for_each_hstate(h) { 15669a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 15679a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1568a3437870SNishanth Aravamudan if (err) 1569a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1570a3437870SNishanth Aravamudan h->name); 1571a3437870SNishanth Aravamudan } 1572a3437870SNishanth Aravamudan } 1573a3437870SNishanth Aravamudan 15749a305230SLee Schermerhorn #ifdef CONFIG_NUMA 15759a305230SLee Schermerhorn 15769a305230SLee Schermerhorn /* 15779a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 15789a305230SLee Schermerhorn * with node sysdevs in node_devices[] using a parallel array. The array 15799a305230SLee Schermerhorn * index of a node sysdev or _hstate == node id. 15809a305230SLee Schermerhorn * This is here to avoid any static dependency of the node sysdev driver, in 15819a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 15829a305230SLee Schermerhorn */ 15839a305230SLee Schermerhorn struct node_hstate { 15849a305230SLee Schermerhorn struct kobject *hugepages_kobj; 15859a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 15869a305230SLee Schermerhorn }; 15879a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 15889a305230SLee Schermerhorn 15899a305230SLee Schermerhorn /* 15909a305230SLee Schermerhorn * A subset of global hstate attributes for node sysdevs 15919a305230SLee Schermerhorn */ 15929a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 15939a305230SLee Schermerhorn &nr_hugepages_attr.attr, 15949a305230SLee Schermerhorn &free_hugepages_attr.attr, 15959a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 15969a305230SLee Schermerhorn NULL, 15979a305230SLee Schermerhorn }; 15989a305230SLee Schermerhorn 15999a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 16009a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 16019a305230SLee Schermerhorn }; 16029a305230SLee Schermerhorn 16039a305230SLee Schermerhorn /* 16049a305230SLee Schermerhorn * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. 16059a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 16069a305230SLee Schermerhorn */ 16079a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 16089a305230SLee Schermerhorn { 16099a305230SLee Schermerhorn int nid; 16109a305230SLee Schermerhorn 16119a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 16129a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 16139a305230SLee Schermerhorn int i; 16149a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 16159a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 16169a305230SLee Schermerhorn if (nidp) 16179a305230SLee Schermerhorn *nidp = nid; 16189a305230SLee Schermerhorn return &hstates[i]; 16199a305230SLee Schermerhorn } 16209a305230SLee Schermerhorn } 16219a305230SLee Schermerhorn 16229a305230SLee Schermerhorn BUG(); 16239a305230SLee Schermerhorn return NULL; 16249a305230SLee Schermerhorn } 16259a305230SLee Schermerhorn 16269a305230SLee Schermerhorn /* 16279a305230SLee Schermerhorn * Unregister hstate attributes from a single node sysdev. 16289a305230SLee Schermerhorn * No-op if no hstate attributes attached. 16299a305230SLee Schermerhorn */ 16309a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 16319a305230SLee Schermerhorn { 16329a305230SLee Schermerhorn struct hstate *h; 16339a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16349a305230SLee Schermerhorn 16359a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16369b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 16379a305230SLee Schermerhorn 16389a305230SLee Schermerhorn for_each_hstate(h) 16399a305230SLee Schermerhorn if (nhs->hstate_kobjs[h - hstates]) { 16409a305230SLee Schermerhorn kobject_put(nhs->hstate_kobjs[h - hstates]); 16419a305230SLee Schermerhorn nhs->hstate_kobjs[h - hstates] = NULL; 16429a305230SLee Schermerhorn } 16439a305230SLee Schermerhorn 16449a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 16459a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 16469a305230SLee Schermerhorn } 16479a305230SLee Schermerhorn 16489a305230SLee Schermerhorn /* 16499a305230SLee Schermerhorn * hugetlb module exit: unregister hstate attributes from node sysdevs 16509a305230SLee Schermerhorn * that have them. 16519a305230SLee Schermerhorn */ 16529a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 16539a305230SLee Schermerhorn { 16549a305230SLee Schermerhorn int nid; 16559a305230SLee Schermerhorn 16569a305230SLee Schermerhorn /* 16579a305230SLee Schermerhorn * disable node sysdev registrations. 16589a305230SLee Schermerhorn */ 16599a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 16609a305230SLee Schermerhorn 16619a305230SLee Schermerhorn /* 16629a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 16639a305230SLee Schermerhorn */ 16649a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 16659a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 16669a305230SLee Schermerhorn } 16679a305230SLee Schermerhorn 16689a305230SLee Schermerhorn /* 16699a305230SLee Schermerhorn * Register hstate attributes for a single node sysdev. 16709a305230SLee Schermerhorn * No-op if attributes already registered. 16719a305230SLee Schermerhorn */ 16729a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 16739a305230SLee Schermerhorn { 16749a305230SLee Schermerhorn struct hstate *h; 16759a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16769a305230SLee Schermerhorn int err; 16779a305230SLee Schermerhorn 16789a305230SLee Schermerhorn if (nhs->hugepages_kobj) 16799a305230SLee Schermerhorn return; /* already allocated */ 16809a305230SLee Schermerhorn 16819a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 16829a305230SLee Schermerhorn &node->sysdev.kobj); 16839a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16849a305230SLee Schermerhorn return; 16859a305230SLee Schermerhorn 16869a305230SLee Schermerhorn for_each_hstate(h) { 16879a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 16889a305230SLee Schermerhorn nhs->hstate_kobjs, 16899a305230SLee Schermerhorn &per_node_hstate_attr_group); 16909a305230SLee Schermerhorn if (err) { 16919a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 16929a305230SLee Schermerhorn " for node %d\n", 16939a305230SLee Schermerhorn h->name, node->sysdev.id); 16949a305230SLee Schermerhorn hugetlb_unregister_node(node); 16959a305230SLee Schermerhorn break; 16969a305230SLee Schermerhorn } 16979a305230SLee Schermerhorn } 16989a305230SLee Schermerhorn } 16999a305230SLee Schermerhorn 17009a305230SLee Schermerhorn /* 17019b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 17029b5e5d0fSLee Schermerhorn * sysdevs of nodes that have memory. All on-line nodes should have 17039b5e5d0fSLee Schermerhorn * registered their associated sysdev by this time. 17049a305230SLee Schermerhorn */ 17059a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 17069a305230SLee Schermerhorn { 17079a305230SLee Schermerhorn int nid; 17089a305230SLee Schermerhorn 17099b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 17109a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 17119a305230SLee Schermerhorn if (node->sysdev.id == nid) 17129a305230SLee Schermerhorn hugetlb_register_node(node); 17139a305230SLee Schermerhorn } 17149a305230SLee Schermerhorn 17159a305230SLee Schermerhorn /* 17169a305230SLee Schermerhorn * Let the node sysdev driver know we're here so it can 17179a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 17189a305230SLee Schermerhorn */ 17199a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 17209a305230SLee Schermerhorn hugetlb_unregister_node); 17219a305230SLee Schermerhorn } 17229a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 17239a305230SLee Schermerhorn 17249a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17259a305230SLee Schermerhorn { 17269a305230SLee Schermerhorn BUG(); 17279a305230SLee Schermerhorn if (nidp) 17289a305230SLee Schermerhorn *nidp = -1; 17299a305230SLee Schermerhorn return NULL; 17309a305230SLee Schermerhorn } 17319a305230SLee Schermerhorn 17329a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 17339a305230SLee Schermerhorn 17349a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 17359a305230SLee Schermerhorn 17369a305230SLee Schermerhorn #endif 17379a305230SLee Schermerhorn 1738a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1739a3437870SNishanth Aravamudan { 1740a3437870SNishanth Aravamudan struct hstate *h; 1741a3437870SNishanth Aravamudan 17429a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 17439a305230SLee Schermerhorn 1744a3437870SNishanth Aravamudan for_each_hstate(h) { 1745a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1746a3437870SNishanth Aravamudan } 1747a3437870SNishanth Aravamudan 1748a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1749a3437870SNishanth Aravamudan } 1750a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1751a3437870SNishanth Aravamudan 1752a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1753a3437870SNishanth Aravamudan { 17540ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 17550ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 17560ef89d25SBenjamin Herrenschmidt * there is no such support 17570ef89d25SBenjamin Herrenschmidt */ 17580ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 17590ef89d25SBenjamin Herrenschmidt return 0; 1760a3437870SNishanth Aravamudan 1761e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1762e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1763e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1764a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1765a3437870SNishanth Aravamudan } 1766e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1767e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1768e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1769a3437870SNishanth Aravamudan 1770a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1771a3437870SNishanth Aravamudan 1772aa888a74SAndi Kleen gather_bootmem_prealloc(); 1773aa888a74SAndi Kleen 1774a3437870SNishanth Aravamudan report_hugepages(); 1775a3437870SNishanth Aravamudan 1776a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1777a3437870SNishanth Aravamudan 17789a305230SLee Schermerhorn hugetlb_register_all_nodes(); 17799a305230SLee Schermerhorn 1780a3437870SNishanth Aravamudan return 0; 1781a3437870SNishanth Aravamudan } 1782a3437870SNishanth Aravamudan module_init(hugetlb_init); 1783a3437870SNishanth Aravamudan 1784a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1785a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1786a3437870SNishanth Aravamudan { 1787a3437870SNishanth Aravamudan struct hstate *h; 17888faa8b07SAndi Kleen unsigned long i; 17898faa8b07SAndi Kleen 1790a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1791a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1792a3437870SNishanth Aravamudan return; 1793a3437870SNishanth Aravamudan } 1794a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1795a3437870SNishanth Aravamudan BUG_ON(order == 0); 1796a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1797a3437870SNishanth Aravamudan h->order = order; 1798a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 17998faa8b07SAndi Kleen h->nr_huge_pages = 0; 18008faa8b07SAndi Kleen h->free_huge_pages = 0; 18018faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 18028faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 18039b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 18049b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1805a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1806a3437870SNishanth Aravamudan huge_page_size(h)/1024); 18078faa8b07SAndi Kleen 1808a3437870SNishanth Aravamudan parsed_hstate = h; 1809a3437870SNishanth Aravamudan } 1810a3437870SNishanth Aravamudan 1811e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1812a3437870SNishanth Aravamudan { 1813a3437870SNishanth Aravamudan unsigned long *mhp; 18148faa8b07SAndi Kleen static unsigned long *last_mhp; 1815a3437870SNishanth Aravamudan 1816a3437870SNishanth Aravamudan /* 1817a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1818a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1819a3437870SNishanth Aravamudan */ 1820a3437870SNishanth Aravamudan if (!max_hstate) 1821a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1822a3437870SNishanth Aravamudan else 1823a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1824a3437870SNishanth Aravamudan 18258faa8b07SAndi Kleen if (mhp == last_mhp) { 18268faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 18278faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 18288faa8b07SAndi Kleen return 1; 18298faa8b07SAndi Kleen } 18308faa8b07SAndi Kleen 1831a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1832a3437870SNishanth Aravamudan *mhp = 0; 1833a3437870SNishanth Aravamudan 18348faa8b07SAndi Kleen /* 18358faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 18368faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 18378faa8b07SAndi Kleen * use the bootmem allocator. 18388faa8b07SAndi Kleen */ 18398faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 18408faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 18418faa8b07SAndi Kleen 18428faa8b07SAndi Kleen last_mhp = mhp; 18438faa8b07SAndi Kleen 1844a3437870SNishanth Aravamudan return 1; 1845a3437870SNishanth Aravamudan } 1846e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1847e11bfbfcSNick Piggin 1848e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1849e11bfbfcSNick Piggin { 1850e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1851e11bfbfcSNick Piggin return 1; 1852e11bfbfcSNick Piggin } 1853e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1854a3437870SNishanth Aravamudan 18558a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 18568a213460SNishanth Aravamudan { 18578a213460SNishanth Aravamudan int node; 18588a213460SNishanth Aravamudan unsigned int nr = 0; 18598a213460SNishanth Aravamudan 18608a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 18618a213460SNishanth Aravamudan nr += array[node]; 18628a213460SNishanth Aravamudan 18638a213460SNishanth Aravamudan return nr; 18648a213460SNishanth Aravamudan } 18658a213460SNishanth Aravamudan 18668a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 186706808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 186806808b08SLee Schermerhorn struct ctl_table *table, int write, 186906808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 18701da177e4SLinus Torvalds { 1871e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1872e5ff2159SAndi Kleen unsigned long tmp; 187308d4a246SMichal Hocko int ret; 1874e5ff2159SAndi Kleen 1875e5ff2159SAndi Kleen if (!write) 1876e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1877e5ff2159SAndi Kleen 1878adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 1879adbe8726SEric B Munson return -EINVAL; 1880adbe8726SEric B Munson 1881e5ff2159SAndi Kleen table->data = &tmp; 1882e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 188308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 188408d4a246SMichal Hocko if (ret) 188508d4a246SMichal Hocko goto out; 1886e5ff2159SAndi Kleen 188706808b08SLee Schermerhorn if (write) { 1888bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 1889bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 189006808b08SLee Schermerhorn if (!(obey_mempolicy && 189106808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 189206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 189306808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 189406808b08SLee Schermerhorn } 189506808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 189606808b08SLee Schermerhorn 189706808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 189806808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 189906808b08SLee Schermerhorn } 190008d4a246SMichal Hocko out: 190108d4a246SMichal Hocko return ret; 19021da177e4SLinus Torvalds } 1903396faf03SMel Gorman 190406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 190506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 190606808b08SLee Schermerhorn { 190706808b08SLee Schermerhorn 190806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 190906808b08SLee Schermerhorn buffer, length, ppos); 191006808b08SLee Schermerhorn } 191106808b08SLee Schermerhorn 191206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 191306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 191406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 191506808b08SLee Schermerhorn { 191606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 191706808b08SLee Schermerhorn buffer, length, ppos); 191806808b08SLee Schermerhorn } 191906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 192006808b08SLee Schermerhorn 1921396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 19228d65af78SAlexey Dobriyan void __user *buffer, 1923396faf03SMel Gorman size_t *length, loff_t *ppos) 1924396faf03SMel Gorman { 19258d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 1926396faf03SMel Gorman if (hugepages_treat_as_movable) 1927396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1928396faf03SMel Gorman else 1929396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1930396faf03SMel Gorman return 0; 1931396faf03SMel Gorman } 1932396faf03SMel Gorman 1933a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 19348d65af78SAlexey Dobriyan void __user *buffer, 1935a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1936a3d0c6aaSNishanth Aravamudan { 1937a5516438SAndi Kleen struct hstate *h = &default_hstate; 1938e5ff2159SAndi Kleen unsigned long tmp; 193908d4a246SMichal Hocko int ret; 1940e5ff2159SAndi Kleen 1941e5ff2159SAndi Kleen if (!write) 1942e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1943e5ff2159SAndi Kleen 1944adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 1945adbe8726SEric B Munson return -EINVAL; 1946adbe8726SEric B Munson 1947e5ff2159SAndi Kleen table->data = &tmp; 1948e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 194908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 195008d4a246SMichal Hocko if (ret) 195108d4a246SMichal Hocko goto out; 1952e5ff2159SAndi Kleen 1953e5ff2159SAndi Kleen if (write) { 1954064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1955e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1956a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1957e5ff2159SAndi Kleen } 195808d4a246SMichal Hocko out: 195908d4a246SMichal Hocko return ret; 1960a3d0c6aaSNishanth Aravamudan } 1961a3d0c6aaSNishanth Aravamudan 19621da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 19631da177e4SLinus Torvalds 1964e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 19651da177e4SLinus Torvalds { 1966a5516438SAndi Kleen struct hstate *h = &default_hstate; 1967e1759c21SAlexey Dobriyan seq_printf(m, 19681da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 19691da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1970b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 19717893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 19724f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1973a5516438SAndi Kleen h->nr_huge_pages, 1974a5516438SAndi Kleen h->free_huge_pages, 1975a5516438SAndi Kleen h->resv_huge_pages, 1976a5516438SAndi Kleen h->surplus_huge_pages, 1977a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 19781da177e4SLinus Torvalds } 19791da177e4SLinus Torvalds 19801da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 19811da177e4SLinus Torvalds { 1982a5516438SAndi Kleen struct hstate *h = &default_hstate; 19831da177e4SLinus Torvalds return sprintf(buf, 19841da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1985a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1986a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1987a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1988a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1989a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 19901da177e4SLinus Torvalds } 19911da177e4SLinus Torvalds 19921da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 19931da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 19941da177e4SLinus Torvalds { 1995a5516438SAndi Kleen struct hstate *h = &default_hstate; 1996a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 19971da177e4SLinus Torvalds } 19981da177e4SLinus Torvalds 1999a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2000fc1b8a73SMel Gorman { 2001fc1b8a73SMel Gorman int ret = -ENOMEM; 2002fc1b8a73SMel Gorman 2003fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2004fc1b8a73SMel Gorman /* 2005fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2006fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2007fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2008fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2009fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2010fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2011fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2012fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2013fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2014fc1b8a73SMel Gorman * 2015fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2016fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2017fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2018fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2019fc1b8a73SMel Gorman * semantics that cpuset has. 2020fc1b8a73SMel Gorman */ 2021fc1b8a73SMel Gorman if (delta > 0) { 2022a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2023fc1b8a73SMel Gorman goto out; 2024fc1b8a73SMel Gorman 2025a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2026a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2027fc1b8a73SMel Gorman goto out; 2028fc1b8a73SMel Gorman } 2029fc1b8a73SMel Gorman } 2030fc1b8a73SMel Gorman 2031fc1b8a73SMel Gorman ret = 0; 2032fc1b8a73SMel Gorman if (delta < 0) 2033a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2034fc1b8a73SMel Gorman 2035fc1b8a73SMel Gorman out: 2036fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2037fc1b8a73SMel Gorman return ret; 2038fc1b8a73SMel Gorman } 2039fc1b8a73SMel Gorman 204084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 204184afd99bSAndy Whitcroft { 204284afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 204384afd99bSAndy Whitcroft 204484afd99bSAndy Whitcroft /* 204584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 204684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 204784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 204884afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 204984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 205084afd99bSAndy Whitcroft * new reference here without additional locking. 205184afd99bSAndy Whitcroft */ 205284afd99bSAndy Whitcroft if (reservations) 205384afd99bSAndy Whitcroft kref_get(&reservations->refs); 205484afd99bSAndy Whitcroft } 205584afd99bSAndy Whitcroft 2056a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2057a1e78772SMel Gorman { 2058a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 205984afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 206084afd99bSAndy Whitcroft unsigned long reserve; 206184afd99bSAndy Whitcroft unsigned long start; 206284afd99bSAndy Whitcroft unsigned long end; 206384afd99bSAndy Whitcroft 206484afd99bSAndy Whitcroft if (reservations) { 2065a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2066a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 206784afd99bSAndy Whitcroft 206884afd99bSAndy Whitcroft reserve = (end - start) - 206984afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 207084afd99bSAndy Whitcroft 207184afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 207284afd99bSAndy Whitcroft 20737251ff78SAdam Litke if (reserve) { 2074a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 20757251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 20767251ff78SAdam Litke } 2077a1e78772SMel Gorman } 207884afd99bSAndy Whitcroft } 2079a1e78772SMel Gorman 20801da177e4SLinus Torvalds /* 20811da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 20821da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 20831da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 20841da177e4SLinus Torvalds * this far. 20851da177e4SLinus Torvalds */ 2086d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 20871da177e4SLinus Torvalds { 20881da177e4SLinus Torvalds BUG(); 2089d0217ac0SNick Piggin return 0; 20901da177e4SLinus Torvalds } 20911da177e4SLinus Torvalds 2092f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2093d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 209484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2095a1e78772SMel Gorman .close = hugetlb_vm_op_close, 20961da177e4SLinus Torvalds }; 20971da177e4SLinus Torvalds 20981e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 20991e8f889bSDavid Gibson int writable) 210063551ae0SDavid Gibson { 210163551ae0SDavid Gibson pte_t entry; 210263551ae0SDavid Gibson 21031e8f889bSDavid Gibson if (writable) { 210463551ae0SDavid Gibson entry = 210563551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 210663551ae0SDavid Gibson } else { 21077f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 210863551ae0SDavid Gibson } 210963551ae0SDavid Gibson entry = pte_mkyoung(entry); 211063551ae0SDavid Gibson entry = pte_mkhuge(entry); 211163551ae0SDavid Gibson 211263551ae0SDavid Gibson return entry; 211363551ae0SDavid Gibson } 211463551ae0SDavid Gibson 21151e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 21161e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 21171e8f889bSDavid Gibson { 21181e8f889bSDavid Gibson pte_t entry; 21191e8f889bSDavid Gibson 21207f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 21217f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 21224b3073e1SRussell King update_mmu_cache(vma, address, ptep); 21231e8f889bSDavid Gibson } 21248dab5241SBenjamin Herrenschmidt } 21251e8f889bSDavid Gibson 21261e8f889bSDavid Gibson 212763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 212863551ae0SDavid Gibson struct vm_area_struct *vma) 212963551ae0SDavid Gibson { 213063551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 213163551ae0SDavid Gibson struct page *ptepage; 21321c59827dSHugh Dickins unsigned long addr; 21331e8f889bSDavid Gibson int cow; 2134a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2135a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 21361e8f889bSDavid Gibson 21371e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 213863551ae0SDavid Gibson 2139a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2140c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2141c74df32cSHugh Dickins if (!src_pte) 2142c74df32cSHugh Dickins continue; 2143a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 214463551ae0SDavid Gibson if (!dst_pte) 214563551ae0SDavid Gibson goto nomem; 2146c5c99429SLarry Woodman 2147c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2148c5c99429SLarry Woodman if (dst_pte == src_pte) 2149c5c99429SLarry Woodman continue; 2150c5c99429SLarry Woodman 2151c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 215246478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 21537f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 21541e8f889bSDavid Gibson if (cow) 21557f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 21567f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 215763551ae0SDavid Gibson ptepage = pte_page(entry); 215863551ae0SDavid Gibson get_page(ptepage); 21590fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 216063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 21611c59827dSHugh Dickins } 21621c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2163c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 216463551ae0SDavid Gibson } 216563551ae0SDavid Gibson return 0; 216663551ae0SDavid Gibson 216763551ae0SDavid Gibson nomem: 216863551ae0SDavid Gibson return -ENOMEM; 216963551ae0SDavid Gibson } 217063551ae0SDavid Gibson 2171290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2172290408d4SNaoya Horiguchi { 2173290408d4SNaoya Horiguchi swp_entry_t swp; 2174290408d4SNaoya Horiguchi 2175290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2176290408d4SNaoya Horiguchi return 0; 2177290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 2178290408d4SNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) { 2179290408d4SNaoya Horiguchi return 1; 2180290408d4SNaoya Horiguchi } else 2181290408d4SNaoya Horiguchi return 0; 2182290408d4SNaoya Horiguchi } 2183290408d4SNaoya Horiguchi 2184fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2185fd6a03edSNaoya Horiguchi { 2186fd6a03edSNaoya Horiguchi swp_entry_t swp; 2187fd6a03edSNaoya Horiguchi 2188fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2189fd6a03edSNaoya Horiguchi return 0; 2190fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 2191fd6a03edSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) { 2192fd6a03edSNaoya Horiguchi return 1; 2193fd6a03edSNaoya Horiguchi } else 2194fd6a03edSNaoya Horiguchi return 0; 2195fd6a03edSNaoya Horiguchi } 2196fd6a03edSNaoya Horiguchi 2197502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 219804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 219963551ae0SDavid Gibson { 220063551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 220163551ae0SDavid Gibson unsigned long address; 2202c7546f8fSDavid Gibson pte_t *ptep; 220363551ae0SDavid Gibson pte_t pte; 220463551ae0SDavid Gibson struct page *page; 2205fe1668aeSChen, Kenneth W struct page *tmp; 2206a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2207a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2208a5516438SAndi Kleen 2209c0a499c2SChen, Kenneth W /* 2210c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 2211c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 2212c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 2213c0a499c2SChen, Kenneth W */ 2214fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 221563551ae0SDavid Gibson 221663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2217a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2218a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 221963551ae0SDavid Gibson 2220cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 2221508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2222a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2223c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2224c7546f8fSDavid Gibson if (!ptep) 2225c7546f8fSDavid Gibson continue; 2226c7546f8fSDavid Gibson 222739dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 222839dde65cSChen, Kenneth W continue; 222939dde65cSChen, Kenneth W 223004f2cbe3SMel Gorman /* 223104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 223204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 223304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 223404f2cbe3SMel Gorman */ 223504f2cbe3SMel Gorman if (ref_page) { 223604f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 223704f2cbe3SMel Gorman if (huge_pte_none(pte)) 223804f2cbe3SMel Gorman continue; 223904f2cbe3SMel Gorman page = pte_page(pte); 224004f2cbe3SMel Gorman if (page != ref_page) 224104f2cbe3SMel Gorman continue; 224204f2cbe3SMel Gorman 224304f2cbe3SMel Gorman /* 224404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 224504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 224604f2cbe3SMel Gorman * looking like data was lost 224704f2cbe3SMel Gorman */ 224804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 224904f2cbe3SMel Gorman } 225004f2cbe3SMel Gorman 2251c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 22527f2e9525SGerald Schaefer if (huge_pte_none(pte)) 225363551ae0SDavid Gibson continue; 2254c7546f8fSDavid Gibson 2255fd6a03edSNaoya Horiguchi /* 2256fd6a03edSNaoya Horiguchi * HWPoisoned hugepage is already unmapped and dropped reference 2257fd6a03edSNaoya Horiguchi */ 2258fd6a03edSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 2259fd6a03edSNaoya Horiguchi continue; 2260fd6a03edSNaoya Horiguchi 226163551ae0SDavid Gibson page = pte_page(pte); 22626649a386SKen Chen if (pte_dirty(pte)) 22636649a386SKen Chen set_page_dirty(page); 2264fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 226563551ae0SDavid Gibson } 22661da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 2267508034a3SHugh Dickins flush_tlb_range(vma, start, end); 2268cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 2269fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 22700fe6e20bSNaoya Horiguchi page_remove_rmap(page); 2271fe1668aeSChen, Kenneth W list_del(&page->lru); 2272fe1668aeSChen, Kenneth W put_page(page); 2273fe1668aeSChen, Kenneth W } 22741da177e4SLinus Torvalds } 227563551ae0SDavid Gibson 2276502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 227704f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2278502717f4SChen, Kenneth W { 2279502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 228004f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 2281502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2282502717f4SChen, Kenneth W } 2283502717f4SChen, Kenneth W 228404f2cbe3SMel Gorman /* 228504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 228604f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 228704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 228804f2cbe3SMel Gorman * same region. 228904f2cbe3SMel Gorman */ 22902a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 22912a4b3dedSHarvey Harrison struct page *page, unsigned long address) 229204f2cbe3SMel Gorman { 22937526674dSAdam Litke struct hstate *h = hstate_vma(vma); 229404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 229504f2cbe3SMel Gorman struct address_space *mapping; 229604f2cbe3SMel Gorman struct prio_tree_iter iter; 229704f2cbe3SMel Gorman pgoff_t pgoff; 229804f2cbe3SMel Gorman 229904f2cbe3SMel Gorman /* 230004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 230104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 230204f2cbe3SMel Gorman */ 23037526674dSAdam Litke address = address & huge_page_mask(h); 230404f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 230504f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 230604f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 230704f2cbe3SMel Gorman 23084eb2b1dcSMel Gorman /* 23094eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 23104eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 23114eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 23124eb2b1dcSMel Gorman */ 23134eb2b1dcSMel Gorman spin_lock(&mapping->i_mmap_lock); 231404f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 231504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 231604f2cbe3SMel Gorman if (iter_vma == vma) 231704f2cbe3SMel Gorman continue; 231804f2cbe3SMel Gorman 231904f2cbe3SMel Gorman /* 232004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 232104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 232204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 232304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 232404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 232504f2cbe3SMel Gorman */ 232604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 23274eb2b1dcSMel Gorman __unmap_hugepage_range(iter_vma, 23287526674dSAdam Litke address, address + huge_page_size(h), 232904f2cbe3SMel Gorman page); 233004f2cbe3SMel Gorman } 23314eb2b1dcSMel Gorman spin_unlock(&mapping->i_mmap_lock); 233204f2cbe3SMel Gorman 233304f2cbe3SMel Gorman return 1; 233404f2cbe3SMel Gorman } 233504f2cbe3SMel Gorman 23360fe6e20bSNaoya Horiguchi /* 23370fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 23380fe6e20bSNaoya Horiguchi */ 23391e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 234004f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 234104f2cbe3SMel Gorman struct page *pagecache_page) 23421e8f889bSDavid Gibson { 2343a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 23441e8f889bSDavid Gibson struct page *old_page, *new_page; 234579ac6ba4SDavid Gibson int avoidcopy; 234604f2cbe3SMel Gorman int outside_reserve = 0; 23471e8f889bSDavid Gibson 23481e8f889bSDavid Gibson old_page = pte_page(pte); 23491e8f889bSDavid Gibson 235004f2cbe3SMel Gorman retry_avoidcopy: 23511e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 23521e8f889bSDavid Gibson * and just make the page writable */ 23530fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 23541e8f889bSDavid Gibson if (avoidcopy) { 23550fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 23560fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 23571e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 235883c54070SNick Piggin return 0; 23591e8f889bSDavid Gibson } 23601e8f889bSDavid Gibson 236104f2cbe3SMel Gorman /* 236204f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 236304f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 236404f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 236504f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 236604f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 236704f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 236804f2cbe3SMel Gorman * of the full address range. 236904f2cbe3SMel Gorman */ 2370f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 237104f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 237204f2cbe3SMel Gorman old_page != pagecache_page) 237304f2cbe3SMel Gorman outside_reserve = 1; 237404f2cbe3SMel Gorman 23751e8f889bSDavid Gibson page_cache_get(old_page); 2376b76c8cfbSLarry Woodman 2377b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2378b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 237904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 23801e8f889bSDavid Gibson 23812fc39cecSAdam Litke if (IS_ERR(new_page)) { 23821e8f889bSDavid Gibson page_cache_release(old_page); 238304f2cbe3SMel Gorman 238404f2cbe3SMel Gorman /* 238504f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 238604f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 238704f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 238804f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 238904f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 239004f2cbe3SMel Gorman */ 239104f2cbe3SMel Gorman if (outside_reserve) { 239204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 239304f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 239404f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 239504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2396b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 239704f2cbe3SMel Gorman goto retry_avoidcopy; 239804f2cbe3SMel Gorman } 239904f2cbe3SMel Gorman WARN_ON_ONCE(1); 240004f2cbe3SMel Gorman } 240104f2cbe3SMel Gorman 2402b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2403b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 24042fc39cecSAdam Litke return -PTR_ERR(new_page); 24051e8f889bSDavid Gibson } 24061e8f889bSDavid Gibson 24070fe6e20bSNaoya Horiguchi /* 24080fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 24090fe6e20bSNaoya Horiguchi * anon_vma prepared. 24100fe6e20bSNaoya Horiguchi */ 241144e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 241244e2aa93SDean Nelson /* Caller expects lock to be held */ 241344e2aa93SDean Nelson spin_lock(&mm->page_table_lock); 24140fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 241544e2aa93SDean Nelson } 24160fe6e20bSNaoya Horiguchi 241747ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 241847ad8475SAndrea Arcangeli pages_per_huge_page(h)); 24190ed361deSNick Piggin __SetPageUptodate(new_page); 24201e8f889bSDavid Gibson 2421b76c8cfbSLarry Woodman /* 2422b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2423b76c8cfbSLarry Woodman * before the page tables are altered 2424b76c8cfbSLarry Woodman */ 2425b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2426a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 24277f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 24281e8f889bSDavid Gibson /* Break COW */ 24293edd4fc9SDoug Doan mmu_notifier_invalidate_range_start(mm, 24303edd4fc9SDoug Doan address & huge_page_mask(h), 24313edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 24328fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 24331e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 24341e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 24350fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2436cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 24371e8f889bSDavid Gibson /* Make the old page be freed below */ 24381e8f889bSDavid Gibson new_page = old_page; 24393edd4fc9SDoug Doan mmu_notifier_invalidate_range_end(mm, 24403edd4fc9SDoug Doan address & huge_page_mask(h), 24413edd4fc9SDoug Doan (address & huge_page_mask(h)) + huge_page_size(h)); 24421e8f889bSDavid Gibson } 24431e8f889bSDavid Gibson page_cache_release(new_page); 24441e8f889bSDavid Gibson page_cache_release(old_page); 244583c54070SNick Piggin return 0; 24461e8f889bSDavid Gibson } 24471e8f889bSDavid Gibson 244804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2449a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2450a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 245104f2cbe3SMel Gorman { 245204f2cbe3SMel Gorman struct address_space *mapping; 2453e7c4b0bfSAndy Whitcroft pgoff_t idx; 245404f2cbe3SMel Gorman 245504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2456a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 245704f2cbe3SMel Gorman 245804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 245904f2cbe3SMel Gorman } 246004f2cbe3SMel Gorman 24613ae77f43SHugh Dickins /* 24623ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 24633ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 24643ae77f43SHugh Dickins */ 24653ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 24662a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 24672a15efc9SHugh Dickins { 24682a15efc9SHugh Dickins struct address_space *mapping; 24692a15efc9SHugh Dickins pgoff_t idx; 24702a15efc9SHugh Dickins struct page *page; 24712a15efc9SHugh Dickins 24722a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 24732a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 24742a15efc9SHugh Dickins 24752a15efc9SHugh Dickins page = find_get_page(mapping, idx); 24762a15efc9SHugh Dickins if (page) 24772a15efc9SHugh Dickins put_page(page); 24782a15efc9SHugh Dickins return page != NULL; 24792a15efc9SHugh Dickins } 24802a15efc9SHugh Dickins 2481a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2482788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2483ac9b9c66SHugh Dickins { 2484a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2485ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2486e7c4b0bfSAndy Whitcroft pgoff_t idx; 24874c887265SAdam Litke unsigned long size; 24884c887265SAdam Litke struct page *page; 24894c887265SAdam Litke struct address_space *mapping; 24901e8f889bSDavid Gibson pte_t new_pte; 24914c887265SAdam Litke 249204f2cbe3SMel Gorman /* 249304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 249404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 249504f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 249604f2cbe3SMel Gorman */ 249704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 249804f2cbe3SMel Gorman printk(KERN_WARNING 249904f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 250004f2cbe3SMel Gorman current->pid); 250104f2cbe3SMel Gorman return ret; 250204f2cbe3SMel Gorman } 250304f2cbe3SMel Gorman 25044c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2505a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 25064c887265SAdam Litke 25074c887265SAdam Litke /* 25084c887265SAdam Litke * Use page lock to guard against racing truncation 25094c887265SAdam Litke * before we get page_table_lock. 25104c887265SAdam Litke */ 25116bda666aSChristoph Lameter retry: 25126bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 25136bda666aSChristoph Lameter if (!page) { 2514a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2515ebed4bfcSHugh Dickins if (idx >= size) 2516ebed4bfcSHugh Dickins goto out; 251704f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 25182fc39cecSAdam Litke if (IS_ERR(page)) { 25192fc39cecSAdam Litke ret = -PTR_ERR(page); 25206bda666aSChristoph Lameter goto out; 25216bda666aSChristoph Lameter } 252247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 25230ed361deSNick Piggin __SetPageUptodate(page); 2524ac9b9c66SHugh Dickins 2525f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 25266bda666aSChristoph Lameter int err; 252745c682a6SKen Chen struct inode *inode = mapping->host; 25286bda666aSChristoph Lameter 25296bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 25306bda666aSChristoph Lameter if (err) { 25316bda666aSChristoph Lameter put_page(page); 25326bda666aSChristoph Lameter if (err == -EEXIST) 25336bda666aSChristoph Lameter goto retry; 25346bda666aSChristoph Lameter goto out; 25356bda666aSChristoph Lameter } 253645c682a6SKen Chen 253745c682a6SKen Chen spin_lock(&inode->i_lock); 2538a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 253945c682a6SKen Chen spin_unlock(&inode->i_lock); 25400fe6e20bSNaoya Horiguchi page_dup_rmap(page); 254123be7468SMel Gorman } else { 25426bda666aSChristoph Lameter lock_page(page); 25430fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 25440fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 25450fe6e20bSNaoya Horiguchi goto backout_unlocked; 254623be7468SMel Gorman } 25470fe6e20bSNaoya Horiguchi hugepage_add_new_anon_rmap(page, vma, address); 25480fe6e20bSNaoya Horiguchi } 25490fe6e20bSNaoya Horiguchi } else { 255057303d80SAndy Whitcroft /* 2551998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2552998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2553998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2554fd6a03edSNaoya Horiguchi */ 2555fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2556aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2557aa50d3a7SAndi Kleen VM_FAULT_SET_HINDEX(h - hstates); 2558fd6a03edSNaoya Horiguchi goto backout_unlocked; 25596bda666aSChristoph Lameter } 2560998b4382SNaoya Horiguchi page_dup_rmap(page); 2561998b4382SNaoya Horiguchi } 25621e8f889bSDavid Gibson 256357303d80SAndy Whitcroft /* 256457303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 256557303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 256657303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 256757303d80SAndy Whitcroft * the spinlock. 256857303d80SAndy Whitcroft */ 2569788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 25702b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 25712b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 25722b26736cSAndy Whitcroft goto backout_unlocked; 25732b26736cSAndy Whitcroft } 257457303d80SAndy Whitcroft 2575ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2576a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 25774c887265SAdam Litke if (idx >= size) 25784c887265SAdam Litke goto backout; 25794c887265SAdam Litke 258083c54070SNick Piggin ret = 0; 25817f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 25824c887265SAdam Litke goto backout; 25834c887265SAdam Litke 25841e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 25851e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 25861e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 25871e8f889bSDavid Gibson 2588788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 25891e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 259004f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 25911e8f889bSDavid Gibson } 25921e8f889bSDavid Gibson 2593ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 25944c887265SAdam Litke unlock_page(page); 25954c887265SAdam Litke out: 2596ac9b9c66SHugh Dickins return ret; 25974c887265SAdam Litke 25984c887265SAdam Litke backout: 25994c887265SAdam Litke spin_unlock(&mm->page_table_lock); 26002b26736cSAndy Whitcroft backout_unlocked: 26014c887265SAdam Litke unlock_page(page); 26024c887265SAdam Litke put_page(page); 26034c887265SAdam Litke goto out; 2604ac9b9c66SHugh Dickins } 2605ac9b9c66SHugh Dickins 260686e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2607788c7df4SHugh Dickins unsigned long address, unsigned int flags) 260886e5216fSAdam Litke { 260986e5216fSAdam Litke pte_t *ptep; 261086e5216fSAdam Litke pte_t entry; 26111e8f889bSDavid Gibson int ret; 26120fe6e20bSNaoya Horiguchi struct page *page = NULL; 261357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 26143935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2615a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 261686e5216fSAdam Litke 2617fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2618fd6a03edSNaoya Horiguchi if (ptep) { 2619fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2620290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 2621290408d4SNaoya Horiguchi migration_entry_wait(mm, (pmd_t *)ptep, address); 2622290408d4SNaoya Horiguchi return 0; 2623290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2624aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2625aa50d3a7SAndi Kleen VM_FAULT_SET_HINDEX(h - hstates); 2626fd6a03edSNaoya Horiguchi } 2627fd6a03edSNaoya Horiguchi 2628a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 262986e5216fSAdam Litke if (!ptep) 263086e5216fSAdam Litke return VM_FAULT_OOM; 263186e5216fSAdam Litke 26323935baa9SDavid Gibson /* 26333935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 26343935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 26353935baa9SDavid Gibson * the same page in the page cache. 26363935baa9SDavid Gibson */ 26373935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 26387f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 26397f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2640788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2641b4d1d99fSDavid Gibson goto out_mutex; 26423935baa9SDavid Gibson } 264386e5216fSAdam Litke 264483c54070SNick Piggin ret = 0; 26451e8f889bSDavid Gibson 264657303d80SAndy Whitcroft /* 264757303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 264857303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 264957303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 265057303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 265157303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 265257303d80SAndy Whitcroft * consumed. 265357303d80SAndy Whitcroft */ 2654788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 26552b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 26562b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2657b4d1d99fSDavid Gibson goto out_mutex; 26582b26736cSAndy Whitcroft } 265957303d80SAndy Whitcroft 2660f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 266157303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 266257303d80SAndy Whitcroft vma, address); 266357303d80SAndy Whitcroft } 266457303d80SAndy Whitcroft 266556c9cfb1SNaoya Horiguchi /* 266656c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 266756c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 266856c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 266956c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 267056c9cfb1SNaoya Horiguchi * so no worry about deadlock. 267156c9cfb1SNaoya Horiguchi */ 26720fe6e20bSNaoya Horiguchi page = pte_page(entry); 267356c9cfb1SNaoya Horiguchi if (page != pagecache_page) 26740fe6e20bSNaoya Horiguchi lock_page(page); 26750fe6e20bSNaoya Horiguchi 26761e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 26771e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2678b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2679b4d1d99fSDavid Gibson goto out_page_table_lock; 2680b4d1d99fSDavid Gibson 2681b4d1d99fSDavid Gibson 2682788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2683b4d1d99fSDavid Gibson if (!pte_write(entry)) { 268457303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 268557303d80SAndy Whitcroft pagecache_page); 2686b4d1d99fSDavid Gibson goto out_page_table_lock; 2687b4d1d99fSDavid Gibson } 2688b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2689b4d1d99fSDavid Gibson } 2690b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2691788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2692788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 26934b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2694b4d1d99fSDavid Gibson 2695b4d1d99fSDavid Gibson out_page_table_lock: 26961e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 269757303d80SAndy Whitcroft 269857303d80SAndy Whitcroft if (pagecache_page) { 269957303d80SAndy Whitcroft unlock_page(pagecache_page); 270057303d80SAndy Whitcroft put_page(pagecache_page); 270157303d80SAndy Whitcroft } 27021f64d69cSDean Nelson if (page != pagecache_page) 270356c9cfb1SNaoya Horiguchi unlock_page(page); 270457303d80SAndy Whitcroft 2705b4d1d99fSDavid Gibson out_mutex: 27063935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 27071e8f889bSDavid Gibson 27081e8f889bSDavid Gibson return ret; 270986e5216fSAdam Litke } 271086e5216fSAdam Litke 2711ceb86879SAndi Kleen /* Can be overriden by architectures */ 2712ceb86879SAndi Kleen __attribute__((weak)) struct page * 2713ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2714ceb86879SAndi Kleen pud_t *pud, int write) 2715ceb86879SAndi Kleen { 2716ceb86879SAndi Kleen BUG(); 2717ceb86879SAndi Kleen return NULL; 2718ceb86879SAndi Kleen } 2719ceb86879SAndi Kleen 272063551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 272163551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 27225b23dbe8SAdam Litke unsigned long *position, int *length, int i, 27232a15efc9SHugh Dickins unsigned int flags) 272463551ae0SDavid Gibson { 2725d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2726d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 272763551ae0SDavid Gibson int remainder = *length; 2728a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 272963551ae0SDavid Gibson 27301c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 273163551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 273263551ae0SDavid Gibson pte_t *pte; 27332a15efc9SHugh Dickins int absent; 273463551ae0SDavid Gibson struct page *page; 273563551ae0SDavid Gibson 27364c887265SAdam Litke /* 27374c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 27382a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 27394c887265SAdam Litke * first, for the page indexing below to work. 27404c887265SAdam Litke */ 2741a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 27422a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 274363551ae0SDavid Gibson 27442a15efc9SHugh Dickins /* 27452a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 27463ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 27473ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 27483ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 27493ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 27502a15efc9SHugh Dickins */ 27513ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 27523ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 27532a15efc9SHugh Dickins remainder = 0; 27542a15efc9SHugh Dickins break; 27552a15efc9SHugh Dickins } 27562a15efc9SHugh Dickins 27572a15efc9SHugh Dickins if (absent || 27582a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 27594c887265SAdam Litke int ret; 27604c887265SAdam Litke 27614c887265SAdam Litke spin_unlock(&mm->page_table_lock); 27622a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 27632a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 27644c887265SAdam Litke spin_lock(&mm->page_table_lock); 2765a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 27664c887265SAdam Litke continue; 27674c887265SAdam Litke 27681c59827dSHugh Dickins remainder = 0; 27691c59827dSHugh Dickins break; 27701c59827dSHugh Dickins } 277163551ae0SDavid Gibson 2772a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 27737f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2774d5d4b0aaSChen, Kenneth W same_page: 2775d6692183SChen, Kenneth W if (pages) { 277669d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 27774b2e38adSKOSAKI Motohiro get_page(pages[i]); 2778d6692183SChen, Kenneth W } 277963551ae0SDavid Gibson 278063551ae0SDavid Gibson if (vmas) 278163551ae0SDavid Gibson vmas[i] = vma; 278263551ae0SDavid Gibson 278363551ae0SDavid Gibson vaddr += PAGE_SIZE; 2784d5d4b0aaSChen, Kenneth W ++pfn_offset; 278563551ae0SDavid Gibson --remainder; 278663551ae0SDavid Gibson ++i; 2787d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2788a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2789d5d4b0aaSChen, Kenneth W /* 2790d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2791d5d4b0aaSChen, Kenneth W * of this compound page. 2792d5d4b0aaSChen, Kenneth W */ 2793d5d4b0aaSChen, Kenneth W goto same_page; 2794d5d4b0aaSChen, Kenneth W } 279563551ae0SDavid Gibson } 27961c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 279763551ae0SDavid Gibson *length = remainder; 279863551ae0SDavid Gibson *position = vaddr; 279963551ae0SDavid Gibson 28002a15efc9SHugh Dickins return i ? i : -EFAULT; 280163551ae0SDavid Gibson } 28028f860591SZhang, Yanmin 28038f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 28048f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 28058f860591SZhang, Yanmin { 28068f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 28078f860591SZhang, Yanmin unsigned long start = address; 28088f860591SZhang, Yanmin pte_t *ptep; 28098f860591SZhang, Yanmin pte_t pte; 2810a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 28118f860591SZhang, Yanmin 28128f860591SZhang, Yanmin BUG_ON(address >= end); 28138f860591SZhang, Yanmin flush_cache_range(vma, address, end); 28148f860591SZhang, Yanmin 281539dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 28168f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2817a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 28188f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 28198f860591SZhang, Yanmin if (!ptep) 28208f860591SZhang, Yanmin continue; 282139dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 282239dde65cSChen, Kenneth W continue; 28237f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 28248f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 28258f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 28268f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 28278f860591SZhang, Yanmin } 28288f860591SZhang, Yanmin } 28298f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 283039dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 28318f860591SZhang, Yanmin 28328f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 28338f860591SZhang, Yanmin } 28348f860591SZhang, Yanmin 2835a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2836a1e78772SMel Gorman long from, long to, 28375a6fe125SMel Gorman struct vm_area_struct *vma, 28385a6fe125SMel Gorman int acctflag) 2839e4e574b7SAdam Litke { 284017c9d12eSMel Gorman long ret, chg; 2841a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2842e4e574b7SAdam Litke 2843a1e78772SMel Gorman /* 284417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 284517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 284617c9d12eSMel Gorman * and filesystem quota without using reserves 284717c9d12eSMel Gorman */ 284817c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 284917c9d12eSMel Gorman return 0; 285017c9d12eSMel Gorman 285117c9d12eSMel Gorman /* 2852a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2853a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2854a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2855a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2856a1e78772SMel Gorman */ 2857f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2858e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 28595a6fe125SMel Gorman else { 28605a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 28615a6fe125SMel Gorman if (!resv_map) 28625a6fe125SMel Gorman return -ENOMEM; 28635a6fe125SMel Gorman 286417c9d12eSMel Gorman chg = to - from; 286517c9d12eSMel Gorman 28665a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 28675a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 28685a6fe125SMel Gorman } 28695a6fe125SMel Gorman 287017c9d12eSMel Gorman if (chg < 0) 287117c9d12eSMel Gorman return chg; 287217c9d12eSMel Gorman 287317c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 287417c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 287517c9d12eSMel Gorman return -ENOSPC; 287617c9d12eSMel Gorman 287717c9d12eSMel Gorman /* 287817c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 287917c9d12eSMel Gorman * Hand back the quota if there are not 288017c9d12eSMel Gorman */ 288117c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 288217c9d12eSMel Gorman if (ret < 0) { 288317c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 288417c9d12eSMel Gorman return ret; 288517c9d12eSMel Gorman } 288617c9d12eSMel Gorman 288717c9d12eSMel Gorman /* 288817c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 288917c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 289017c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 289117c9d12eSMel Gorman * the reservation was and the page cache tells how much of 289217c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 289317c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 289417c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 289517c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 289617c9d12eSMel Gorman * else has to be done for private mappings here 289717c9d12eSMel Gorman */ 2898f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 289917c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2900a43a8c39SChen, Kenneth W return 0; 2901a43a8c39SChen, Kenneth W } 2902a43a8c39SChen, Kenneth W 2903a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2904a43a8c39SChen, Kenneth W { 2905a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2906a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 290745c682a6SKen Chen 290845c682a6SKen Chen spin_lock(&inode->i_lock); 2909e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 291045c682a6SKen Chen spin_unlock(&inode->i_lock); 291145c682a6SKen Chen 291290d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2913a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2914a43a8c39SChen, Kenneth W } 291593f70f90SNaoya Horiguchi 2916d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 2917d5bd9106SAndi Kleen 29186de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 29196de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 29206de2b1aaSNaoya Horiguchi { 29216de2b1aaSNaoya Horiguchi struct page *page; 29226de2b1aaSNaoya Horiguchi struct page *tmp; 29236de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 29246de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 29256de2b1aaSNaoya Horiguchi 29266de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 29276de2b1aaSNaoya Horiguchi if (page == hpage) 29286de2b1aaSNaoya Horiguchi return 1; 29296de2b1aaSNaoya Horiguchi return 0; 29306de2b1aaSNaoya Horiguchi } 29316de2b1aaSNaoya Horiguchi 293293f70f90SNaoya Horiguchi /* 293393f70f90SNaoya Horiguchi * This function is called from memory failure code. 293493f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 293593f70f90SNaoya Horiguchi */ 29366de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 293793f70f90SNaoya Horiguchi { 293893f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 293993f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 29406de2b1aaSNaoya Horiguchi int ret = -EBUSY; 294193f70f90SNaoya Horiguchi 294293f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 29436de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 294493f70f90SNaoya Horiguchi list_del(&hpage->lru); 29458c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 294693f70f90SNaoya Horiguchi h->free_huge_pages--; 294793f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 29486de2b1aaSNaoya Horiguchi ret = 0; 294993f70f90SNaoya Horiguchi } 29506de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 29516de2b1aaSNaoya Horiguchi return ret; 29526de2b1aaSNaoya Horiguchi } 29536de2b1aaSNaoya Horiguchi #endif 2954