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> 22d6606683SLinus Torvalds 2363551ae0SDavid Gibson #include <asm/page.h> 2463551ae0SDavid Gibson #include <asm/pgtable.h> 2578a34ae2SAdrian Bunk #include <asm/io.h> 2663551ae0SDavid Gibson 2763551ae0SDavid Gibson #include <linux/hugetlb.h> 289a305230SLee Schermerhorn #include <linux/node.h> 297835e98bSNick Piggin #include "internal.h" 301da177e4SLinus Torvalds 311da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 32396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 33396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 34a5516438SAndi Kleen 35e5ff2159SAndi Kleen static int max_hstate; 36e5ff2159SAndi Kleen unsigned int default_hstate_idx; 37e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 38e5ff2159SAndi Kleen 3953ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4053ba51d2SJon Tollefson 41e5ff2159SAndi Kleen /* for command line parsing */ 42e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 43e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 44e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 45e5ff2159SAndi Kleen 46e5ff2159SAndi Kleen #define for_each_hstate(h) \ 47e5ff2159SAndi Kleen for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) 48396faf03SMel Gorman 493935baa9SDavid Gibson /* 503935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 513935baa9SDavid Gibson */ 523935baa9SDavid Gibson static DEFINE_SPINLOCK(hugetlb_lock); 530bd0f9fbSEric Paris 54e7c4b0bfSAndy Whitcroft /* 5596822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 5696822904SAndy Whitcroft * across the pages in a mapping. 5784afd99bSAndy Whitcroft * 5884afd99bSAndy Whitcroft * The region data structures are protected by a combination of the mmap_sem 5984afd99bSAndy Whitcroft * and the hugetlb_instantion_mutex. To access or modify a region the caller 6084afd99bSAndy Whitcroft * must either hold the mmap_sem for write, or the mmap_sem for read and 6184afd99bSAndy Whitcroft * the hugetlb_instantiation mutex: 6284afd99bSAndy Whitcroft * 6384afd99bSAndy Whitcroft * down_write(&mm->mmap_sem); 6484afd99bSAndy Whitcroft * or 6584afd99bSAndy Whitcroft * down_read(&mm->mmap_sem); 6684afd99bSAndy Whitcroft * mutex_lock(&hugetlb_instantiation_mutex); 6796822904SAndy Whitcroft */ 6896822904SAndy Whitcroft struct file_region { 6996822904SAndy Whitcroft struct list_head link; 7096822904SAndy Whitcroft long from; 7196822904SAndy Whitcroft long to; 7296822904SAndy Whitcroft }; 7396822904SAndy Whitcroft 7496822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 7596822904SAndy Whitcroft { 7696822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 7796822904SAndy Whitcroft 7896822904SAndy Whitcroft /* Locate the region we are either in or before. */ 7996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 8096822904SAndy Whitcroft if (f <= rg->to) 8196822904SAndy Whitcroft break; 8296822904SAndy Whitcroft 8396822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 8496822904SAndy Whitcroft if (f > rg->from) 8596822904SAndy Whitcroft f = rg->from; 8696822904SAndy Whitcroft 8796822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 8896822904SAndy Whitcroft nrg = rg; 8996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 9096822904SAndy Whitcroft if (&rg->link == head) 9196822904SAndy Whitcroft break; 9296822904SAndy Whitcroft if (rg->from > t) 9396822904SAndy Whitcroft break; 9496822904SAndy Whitcroft 9596822904SAndy Whitcroft /* If this area reaches higher then extend our area to 9696822904SAndy Whitcroft * include it completely. If this is not the first area 9796822904SAndy Whitcroft * which we intend to reuse, free it. */ 9896822904SAndy Whitcroft if (rg->to > t) 9996822904SAndy Whitcroft t = rg->to; 10096822904SAndy Whitcroft if (rg != nrg) { 10196822904SAndy Whitcroft list_del(&rg->link); 10296822904SAndy Whitcroft kfree(rg); 10396822904SAndy Whitcroft } 10496822904SAndy Whitcroft } 10596822904SAndy Whitcroft nrg->from = f; 10696822904SAndy Whitcroft nrg->to = t; 10796822904SAndy Whitcroft return 0; 10896822904SAndy Whitcroft } 10996822904SAndy Whitcroft 11096822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 11196822904SAndy Whitcroft { 11296822904SAndy Whitcroft struct file_region *rg, *nrg; 11396822904SAndy Whitcroft long chg = 0; 11496822904SAndy Whitcroft 11596822904SAndy Whitcroft /* Locate the region we are before or in. */ 11696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 11796822904SAndy Whitcroft if (f <= rg->to) 11896822904SAndy Whitcroft break; 11996822904SAndy Whitcroft 12096822904SAndy Whitcroft /* If we are below the current region then a new region is required. 12196822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 12296822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 12396822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 12496822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 12596822904SAndy Whitcroft if (!nrg) 12696822904SAndy Whitcroft return -ENOMEM; 12796822904SAndy Whitcroft nrg->from = f; 12896822904SAndy Whitcroft nrg->to = f; 12996822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 13096822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 13196822904SAndy Whitcroft 13296822904SAndy Whitcroft return t - f; 13396822904SAndy Whitcroft } 13496822904SAndy Whitcroft 13596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 13696822904SAndy Whitcroft if (f > rg->from) 13796822904SAndy Whitcroft f = rg->from; 13896822904SAndy Whitcroft chg = t - f; 13996822904SAndy Whitcroft 14096822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 14196822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 14296822904SAndy Whitcroft if (&rg->link == head) 14396822904SAndy Whitcroft break; 14496822904SAndy Whitcroft if (rg->from > t) 14596822904SAndy Whitcroft return chg; 14696822904SAndy Whitcroft 14796822904SAndy Whitcroft /* We overlap with this area, if it extends futher than 14896822904SAndy Whitcroft * us then we must extend ourselves. Account for its 14996822904SAndy Whitcroft * existing reservation. */ 15096822904SAndy Whitcroft if (rg->to > t) { 15196822904SAndy Whitcroft chg += rg->to - t; 15296822904SAndy Whitcroft t = rg->to; 15396822904SAndy Whitcroft } 15496822904SAndy Whitcroft chg -= rg->to - rg->from; 15596822904SAndy Whitcroft } 15696822904SAndy Whitcroft return chg; 15796822904SAndy Whitcroft } 15896822904SAndy Whitcroft 15996822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 16096822904SAndy Whitcroft { 16196822904SAndy Whitcroft struct file_region *rg, *trg; 16296822904SAndy Whitcroft long chg = 0; 16396822904SAndy Whitcroft 16496822904SAndy Whitcroft /* Locate the region we are either in or before. */ 16596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16696822904SAndy Whitcroft if (end <= rg->to) 16796822904SAndy Whitcroft break; 16896822904SAndy Whitcroft if (&rg->link == head) 16996822904SAndy Whitcroft return 0; 17096822904SAndy Whitcroft 17196822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 17296822904SAndy Whitcroft if (end > rg->from) { 17396822904SAndy Whitcroft chg = rg->to - end; 17496822904SAndy Whitcroft rg->to = end; 17596822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 17696822904SAndy Whitcroft } 17796822904SAndy Whitcroft 17896822904SAndy Whitcroft /* Drop any remaining regions. */ 17996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 18096822904SAndy Whitcroft if (&rg->link == head) 18196822904SAndy Whitcroft break; 18296822904SAndy Whitcroft chg += rg->to - rg->from; 18396822904SAndy Whitcroft list_del(&rg->link); 18496822904SAndy Whitcroft kfree(rg); 18596822904SAndy Whitcroft } 18696822904SAndy Whitcroft return chg; 18796822904SAndy Whitcroft } 18896822904SAndy Whitcroft 18984afd99bSAndy Whitcroft static long region_count(struct list_head *head, long f, long t) 19084afd99bSAndy Whitcroft { 19184afd99bSAndy Whitcroft struct file_region *rg; 19284afd99bSAndy Whitcroft long chg = 0; 19384afd99bSAndy Whitcroft 19484afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 19584afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 19684afd99bSAndy Whitcroft int seg_from; 19784afd99bSAndy Whitcroft int seg_to; 19884afd99bSAndy Whitcroft 19984afd99bSAndy Whitcroft if (rg->to <= f) 20084afd99bSAndy Whitcroft continue; 20184afd99bSAndy Whitcroft if (rg->from >= t) 20284afd99bSAndy Whitcroft break; 20384afd99bSAndy Whitcroft 20484afd99bSAndy Whitcroft seg_from = max(rg->from, f); 20584afd99bSAndy Whitcroft seg_to = min(rg->to, t); 20684afd99bSAndy Whitcroft 20784afd99bSAndy Whitcroft chg += seg_to - seg_from; 20884afd99bSAndy Whitcroft } 20984afd99bSAndy Whitcroft 21084afd99bSAndy Whitcroft return chg; 21184afd99bSAndy Whitcroft } 21284afd99bSAndy Whitcroft 21396822904SAndy Whitcroft /* 214e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 215e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 216e7c4b0bfSAndy Whitcroft */ 217a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 218a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 219e7c4b0bfSAndy Whitcroft { 220a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 221a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 222e7c4b0bfSAndy Whitcroft } 223e7c4b0bfSAndy Whitcroft 2240fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 2250fe6e20bSNaoya Horiguchi unsigned long address) 2260fe6e20bSNaoya Horiguchi { 2270fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 2280fe6e20bSNaoya Horiguchi } 2290fe6e20bSNaoya Horiguchi 23084afd99bSAndy Whitcroft /* 23108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 23208fba699SMel Gorman * cases this will be same size as used by the page table entries. 23308fba699SMel Gorman */ 23408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 23508fba699SMel Gorman { 23608fba699SMel Gorman struct hstate *hstate; 23708fba699SMel Gorman 23808fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 23908fba699SMel Gorman return PAGE_SIZE; 24008fba699SMel Gorman 24108fba699SMel Gorman hstate = hstate_vma(vma); 24208fba699SMel Gorman 24308fba699SMel Gorman return 1UL << (hstate->order + PAGE_SHIFT); 24408fba699SMel Gorman } 245f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 24608fba699SMel Gorman 24708fba699SMel Gorman /* 2483340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 2493340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 2503340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 2513340289dSMel Gorman * function is required. 2523340289dSMel Gorman */ 2533340289dSMel Gorman #ifndef vma_mmu_pagesize 2543340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 2553340289dSMel Gorman { 2563340289dSMel Gorman return vma_kernel_pagesize(vma); 2573340289dSMel Gorman } 2583340289dSMel Gorman #endif 2593340289dSMel Gorman 2603340289dSMel Gorman /* 26184afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 26284afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 26384afd99bSAndy Whitcroft * alignment. 26484afd99bSAndy Whitcroft */ 26584afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 26684afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 26704f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 26884afd99bSAndy Whitcroft 269a1e78772SMel Gorman /* 270a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 271a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 272a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 273a1e78772SMel Gorman * 274a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 275a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 276a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 277a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 27884afd99bSAndy Whitcroft * 27984afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 28084afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 28184afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 28284afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 28384afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 28484afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 28584afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 28684afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 287a1e78772SMel Gorman */ 288e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 289e7c4b0bfSAndy Whitcroft { 290e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 291e7c4b0bfSAndy Whitcroft } 292e7c4b0bfSAndy Whitcroft 293e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 294e7c4b0bfSAndy Whitcroft unsigned long value) 295e7c4b0bfSAndy Whitcroft { 296e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 297e7c4b0bfSAndy Whitcroft } 298e7c4b0bfSAndy Whitcroft 29984afd99bSAndy Whitcroft struct resv_map { 30084afd99bSAndy Whitcroft struct kref refs; 30184afd99bSAndy Whitcroft struct list_head regions; 30284afd99bSAndy Whitcroft }; 30384afd99bSAndy Whitcroft 3042a4b3dedSHarvey Harrison static struct resv_map *resv_map_alloc(void) 30584afd99bSAndy Whitcroft { 30684afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 30784afd99bSAndy Whitcroft if (!resv_map) 30884afd99bSAndy Whitcroft return NULL; 30984afd99bSAndy Whitcroft 31084afd99bSAndy Whitcroft kref_init(&resv_map->refs); 31184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 31284afd99bSAndy Whitcroft 31384afd99bSAndy Whitcroft return resv_map; 31484afd99bSAndy Whitcroft } 31584afd99bSAndy Whitcroft 3162a4b3dedSHarvey Harrison static void resv_map_release(struct kref *ref) 31784afd99bSAndy Whitcroft { 31884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 31984afd99bSAndy Whitcroft 32084afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 32184afd99bSAndy Whitcroft region_truncate(&resv_map->regions, 0); 32284afd99bSAndy Whitcroft kfree(resv_map); 32384afd99bSAndy Whitcroft } 32484afd99bSAndy Whitcroft 32584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 326a1e78772SMel Gorman { 327a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 328f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 32984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 33084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 3312a4b3dedSHarvey Harrison return NULL; 332a1e78772SMel Gorman } 333a1e78772SMel Gorman 33484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 335a1e78772SMel Gorman { 336a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 337f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 338a1e78772SMel Gorman 33984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 34084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 34104f2cbe3SMel Gorman } 34204f2cbe3SMel Gorman 34304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 34404f2cbe3SMel Gorman { 34504f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 346f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 347e7c4b0bfSAndy Whitcroft 348e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 34904f2cbe3SMel Gorman } 35004f2cbe3SMel Gorman 35104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 35204f2cbe3SMel Gorman { 35304f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 354e7c4b0bfSAndy Whitcroft 355e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 356a1e78772SMel Gorman } 357a1e78772SMel Gorman 358a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 359a5516438SAndi Kleen static void decrement_hugepage_resv_vma(struct hstate *h, 360a5516438SAndi Kleen struct vm_area_struct *vma) 361a1e78772SMel Gorman { 362c37f9fb1SAndy Whitcroft if (vma->vm_flags & VM_NORESERVE) 363c37f9fb1SAndy Whitcroft return; 364c37f9fb1SAndy Whitcroft 365f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 366a1e78772SMel Gorman /* Shared mappings always use reserves */ 367a5516438SAndi Kleen h->resv_huge_pages--; 36884afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 369a1e78772SMel Gorman /* 370a1e78772SMel Gorman * Only the process that called mmap() has reserves for 371a1e78772SMel Gorman * private mappings. 372a1e78772SMel Gorman */ 373a5516438SAndi Kleen h->resv_huge_pages--; 374a1e78772SMel Gorman } 375a1e78772SMel Gorman } 376a1e78772SMel Gorman 37704f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 378a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 379a1e78772SMel Gorman { 380a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 381f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 382a1e78772SMel Gorman vma->vm_private_data = (void *)0; 383a1e78772SMel Gorman } 384a1e78772SMel Gorman 385a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 3867f09ca51SMel Gorman static int vma_has_reserves(struct vm_area_struct *vma) 387a1e78772SMel Gorman { 388f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 389a1e78772SMel Gorman return 1; 3907f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3917f09ca51SMel Gorman return 1; 3927f09ca51SMel Gorman return 0; 393a1e78772SMel Gorman } 394a1e78772SMel Gorman 39569d177c2SAndy Whitcroft static void clear_gigantic_page(struct page *page, 39669d177c2SAndy Whitcroft unsigned long addr, unsigned long sz) 39769d177c2SAndy Whitcroft { 39869d177c2SAndy Whitcroft int i; 39969d177c2SAndy Whitcroft struct page *p = page; 40069d177c2SAndy Whitcroft 40169d177c2SAndy Whitcroft might_sleep(); 40269d177c2SAndy Whitcroft for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) { 40369d177c2SAndy Whitcroft cond_resched(); 40469d177c2SAndy Whitcroft clear_user_highpage(p, addr + i * PAGE_SIZE); 40569d177c2SAndy Whitcroft } 40669d177c2SAndy Whitcroft } 407a5516438SAndi Kleen static void clear_huge_page(struct page *page, 408a5516438SAndi Kleen unsigned long addr, unsigned long sz) 40979ac6ba4SDavid Gibson { 41079ac6ba4SDavid Gibson int i; 41179ac6ba4SDavid Gibson 41274dbdd23SAndrea Arcangeli if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) { 413ebdd4aeaSHannes Eder clear_gigantic_page(page, addr, sz); 414ebdd4aeaSHannes Eder return; 415ebdd4aeaSHannes Eder } 41669d177c2SAndy Whitcroft 41779ac6ba4SDavid Gibson might_sleep(); 418a5516438SAndi Kleen for (i = 0; i < sz/PAGE_SIZE; i++) { 41979ac6ba4SDavid Gibson cond_resched(); 420281e0e3bSRalf Baechle clear_user_highpage(page + i, addr + i * PAGE_SIZE); 42179ac6ba4SDavid Gibson } 42279ac6ba4SDavid Gibson } 42379ac6ba4SDavid Gibson 42469d177c2SAndy Whitcroft static void copy_gigantic_page(struct page *dst, struct page *src, 42569d177c2SAndy Whitcroft unsigned long addr, struct vm_area_struct *vma) 42669d177c2SAndy Whitcroft { 42769d177c2SAndy Whitcroft int i; 42869d177c2SAndy Whitcroft struct hstate *h = hstate_vma(vma); 42969d177c2SAndy Whitcroft struct page *dst_base = dst; 43069d177c2SAndy Whitcroft struct page *src_base = src; 43169d177c2SAndy Whitcroft might_sleep(); 43269d177c2SAndy Whitcroft for (i = 0; i < pages_per_huge_page(h); ) { 43369d177c2SAndy Whitcroft cond_resched(); 43469d177c2SAndy Whitcroft copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); 43569d177c2SAndy Whitcroft 43669d177c2SAndy Whitcroft i++; 43769d177c2SAndy Whitcroft dst = mem_map_next(dst, dst_base, i); 43869d177c2SAndy Whitcroft src = mem_map_next(src, src_base, i); 43969d177c2SAndy Whitcroft } 44069d177c2SAndy Whitcroft } 44179ac6ba4SDavid Gibson static void copy_huge_page(struct page *dst, struct page *src, 4429de455b2SAtsushi Nemoto unsigned long addr, struct vm_area_struct *vma) 44379ac6ba4SDavid Gibson { 44479ac6ba4SDavid Gibson int i; 445a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44679ac6ba4SDavid Gibson 447ebdd4aeaSHannes Eder if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 448ebdd4aeaSHannes Eder copy_gigantic_page(dst, src, addr, vma); 449ebdd4aeaSHannes Eder return; 450ebdd4aeaSHannes Eder } 45169d177c2SAndy Whitcroft 45279ac6ba4SDavid Gibson might_sleep(); 453a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 45479ac6ba4SDavid Gibson cond_resched(); 4559de455b2SAtsushi Nemoto copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 45679ac6ba4SDavid Gibson } 45779ac6ba4SDavid Gibson } 45879ac6ba4SDavid Gibson 459a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 4601da177e4SLinus Torvalds { 4611da177e4SLinus Torvalds int nid = page_to_nid(page); 462a5516438SAndi Kleen list_add(&page->lru, &h->hugepage_freelists[nid]); 463a5516438SAndi Kleen h->free_huge_pages++; 464a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 4651da177e4SLinus Torvalds } 4661da177e4SLinus Torvalds 467a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 468a5516438SAndi Kleen struct vm_area_struct *vma, 46904f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 4701da177e4SLinus Torvalds { 47131a5c6e4SNishanth Aravamudan int nid; 4721da177e4SLinus Torvalds struct page *page = NULL; 473480eccf9SLee Schermerhorn struct mempolicy *mpol; 47419770b32SMel Gorman nodemask_t *nodemask; 475c0ff7453SMiao Xie struct zonelist *zonelist; 476dd1a239fSMel Gorman struct zone *zone; 477dd1a239fSMel Gorman struct zoneref *z; 4781da177e4SLinus Torvalds 479c0ff7453SMiao Xie get_mems_allowed(); 480c0ff7453SMiao Xie zonelist = huge_zonelist(vma, address, 481c0ff7453SMiao Xie htlb_alloc_mask, &mpol, &nodemask); 482a1e78772SMel Gorman /* 483a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 484a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 485a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 486a1e78772SMel Gorman */ 4877f09ca51SMel Gorman if (!vma_has_reserves(vma) && 488a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 489c0ff7453SMiao Xie goto err; 490a1e78772SMel Gorman 49104f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 492a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 493c0ff7453SMiao Xie goto err;; 49404f2cbe3SMel Gorman 49519770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 49619770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 49754a6eb5cSMel Gorman nid = zone_to_nid(zone); 49854a6eb5cSMel Gorman if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 499a5516438SAndi Kleen !list_empty(&h->hugepage_freelists[nid])) { 500a5516438SAndi Kleen page = list_entry(h->hugepage_freelists[nid].next, 5011da177e4SLinus Torvalds struct page, lru); 5021da177e4SLinus Torvalds list_del(&page->lru); 503a5516438SAndi Kleen h->free_huge_pages--; 504a5516438SAndi Kleen h->free_huge_pages_node[nid]--; 50504f2cbe3SMel Gorman 50604f2cbe3SMel Gorman if (!avoid_reserve) 507a5516438SAndi Kleen decrement_hugepage_resv_vma(h, vma); 508a1e78772SMel Gorman 5095ab3ee7bSKen Chen break; 5101da177e4SLinus Torvalds } 5113abf7afdSAndrew Morton } 512c0ff7453SMiao Xie err: 51352cd3b07SLee Schermerhorn mpol_cond_put(mpol); 514c0ff7453SMiao Xie put_mems_allowed(); 5151da177e4SLinus Torvalds return page; 5161da177e4SLinus Torvalds } 5171da177e4SLinus Torvalds 518a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 5196af2acb6SAdam Litke { 5206af2acb6SAdam Litke int i; 521a5516438SAndi Kleen 52218229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 52318229df5SAndy Whitcroft 524a5516438SAndi Kleen h->nr_huge_pages--; 525a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 526a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 5276af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 5286af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 5296af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 5306af2acb6SAdam Litke } 5316af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 5326af2acb6SAdam Litke set_page_refcounted(page); 5337f2e9525SGerald Schaefer arch_release_hugepage(page); 534a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 5356af2acb6SAdam Litke } 5366af2acb6SAdam Litke 537e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 538e5ff2159SAndi Kleen { 539e5ff2159SAndi Kleen struct hstate *h; 540e5ff2159SAndi Kleen 541e5ff2159SAndi Kleen for_each_hstate(h) { 542e5ff2159SAndi Kleen if (huge_page_size(h) == size) 543e5ff2159SAndi Kleen return h; 544e5ff2159SAndi Kleen } 545e5ff2159SAndi Kleen return NULL; 546e5ff2159SAndi Kleen } 547e5ff2159SAndi Kleen 54827a85ef1SDavid Gibson static void free_huge_page(struct page *page) 54927a85ef1SDavid Gibson { 550a5516438SAndi Kleen /* 551a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 552a5516438SAndi Kleen * compound page destructor. 553a5516438SAndi Kleen */ 554e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 5557893d1d5SAdam Litke int nid = page_to_nid(page); 556c79fb75eSAdam Litke struct address_space *mapping; 55727a85ef1SDavid Gibson 558c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 559e5df70abSAndy Whitcroft set_page_private(page, 0); 56023be7468SMel Gorman page->mapping = NULL; 5617893d1d5SAdam Litke BUG_ON(page_count(page)); 5620fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 56327a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 56427a85ef1SDavid Gibson 56527a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 566aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 567a5516438SAndi Kleen update_and_free_page(h, page); 568a5516438SAndi Kleen h->surplus_huge_pages--; 569a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 5707893d1d5SAdam Litke } else { 571a5516438SAndi Kleen enqueue_huge_page(h, page); 5727893d1d5SAdam Litke } 57327a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 574c79fb75eSAdam Litke if (mapping) 5759a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 57627a85ef1SDavid Gibson } 57727a85ef1SDavid Gibson 578a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 579b7ba30c6SAndi Kleen { 580b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 581b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 582a5516438SAndi Kleen h->nr_huge_pages++; 583a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 584b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 585b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 586b7ba30c6SAndi Kleen } 587b7ba30c6SAndi Kleen 58820a0307cSWu Fengguang static void prep_compound_gigantic_page(struct page *page, unsigned long order) 58920a0307cSWu Fengguang { 59020a0307cSWu Fengguang int i; 59120a0307cSWu Fengguang int nr_pages = 1 << order; 59220a0307cSWu Fengguang struct page *p = page + 1; 59320a0307cSWu Fengguang 59420a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 59520a0307cSWu Fengguang set_compound_order(page, order); 59620a0307cSWu Fengguang __SetPageHead(page); 59720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 59820a0307cSWu Fengguang __SetPageTail(p); 59920a0307cSWu Fengguang p->first_page = page; 60020a0307cSWu Fengguang } 60120a0307cSWu Fengguang } 60220a0307cSWu Fengguang 60320a0307cSWu Fengguang int PageHuge(struct page *page) 60420a0307cSWu Fengguang { 60520a0307cSWu Fengguang compound_page_dtor *dtor; 60620a0307cSWu Fengguang 60720a0307cSWu Fengguang if (!PageCompound(page)) 60820a0307cSWu Fengguang return 0; 60920a0307cSWu Fengguang 61020a0307cSWu Fengguang page = compound_head(page); 61120a0307cSWu Fengguang dtor = get_compound_page_dtor(page); 61220a0307cSWu Fengguang 61320a0307cSWu Fengguang return dtor == free_huge_page; 61420a0307cSWu Fengguang } 61520a0307cSWu Fengguang 616a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 6171da177e4SLinus Torvalds { 6181da177e4SLinus Torvalds struct page *page; 619f96efd58SJoe Jin 620aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 621aa888a74SAndi Kleen return NULL; 622aa888a74SAndi Kleen 6236484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 624551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 625551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 626a5516438SAndi Kleen huge_page_order(h)); 6271da177e4SLinus Torvalds if (page) { 6287f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 629caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 6307b8ee84dSHarvey Harrison return NULL; 6317f2e9525SGerald Schaefer } 632a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 6331da177e4SLinus Torvalds } 63463b4613cSNishanth Aravamudan 63563b4613cSNishanth Aravamudan return page; 63663b4613cSNishanth Aravamudan } 63763b4613cSNishanth Aravamudan 6385ced66c9SAndi Kleen /* 6396ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 6406ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 6416ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 6426ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 6436ae11b27SLee Schermerhorn * node for alloc or free. 6449a76db09SLee Schermerhorn */ 6456ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 6469a76db09SLee Schermerhorn { 6476ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 6489a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 6496ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 6509a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 6519a76db09SLee Schermerhorn 6529a76db09SLee Schermerhorn return nid; 6539a76db09SLee Schermerhorn } 6549a76db09SLee Schermerhorn 6556ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 6565ced66c9SAndi Kleen { 6576ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 6586ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 6599a76db09SLee Schermerhorn return nid; 6605ced66c9SAndi Kleen } 6615ced66c9SAndi Kleen 6626ae11b27SLee Schermerhorn /* 6636ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 6646ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 6656ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 6666ae11b27SLee Schermerhorn * mask. 6676ae11b27SLee Schermerhorn */ 6686ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 6696ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 6706ae11b27SLee Schermerhorn { 6716ae11b27SLee Schermerhorn int nid; 6726ae11b27SLee Schermerhorn 6736ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 6746ae11b27SLee Schermerhorn 6756ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 6766ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 6776ae11b27SLee Schermerhorn 6786ae11b27SLee Schermerhorn return nid; 6796ae11b27SLee Schermerhorn } 6806ae11b27SLee Schermerhorn 6816ae11b27SLee Schermerhorn static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 68263b4613cSNishanth Aravamudan { 68363b4613cSNishanth Aravamudan struct page *page; 68463b4613cSNishanth Aravamudan int start_nid; 68563b4613cSNishanth Aravamudan int next_nid; 68663b4613cSNishanth Aravamudan int ret = 0; 68763b4613cSNishanth Aravamudan 6886ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 689e8c5c824SLee Schermerhorn next_nid = start_nid; 69063b4613cSNishanth Aravamudan 69163b4613cSNishanth Aravamudan do { 692e8c5c824SLee Schermerhorn page = alloc_fresh_huge_page_node(h, next_nid); 6939a76db09SLee Schermerhorn if (page) { 69463b4613cSNishanth Aravamudan ret = 1; 6959a76db09SLee Schermerhorn break; 6969a76db09SLee Schermerhorn } 6976ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 6989a76db09SLee Schermerhorn } while (next_nid != start_nid); 69963b4613cSNishanth Aravamudan 7003b116300SAdam Litke if (ret) 7013b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 7023b116300SAdam Litke else 7033b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 7043b116300SAdam Litke 70563b4613cSNishanth Aravamudan return ret; 7061da177e4SLinus Torvalds } 7071da177e4SLinus Torvalds 708e8c5c824SLee Schermerhorn /* 7096ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 7106ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 7116ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 7126ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 713e8c5c824SLee Schermerhorn */ 7146ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 715e8c5c824SLee Schermerhorn { 7166ae11b27SLee Schermerhorn int nid; 7179a76db09SLee Schermerhorn 7186ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 7196ae11b27SLee Schermerhorn 7206ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7216ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7226ae11b27SLee Schermerhorn 7239a76db09SLee Schermerhorn return nid; 724e8c5c824SLee Schermerhorn } 725e8c5c824SLee Schermerhorn 726e8c5c824SLee Schermerhorn /* 727e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 728e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 729e8c5c824SLee Schermerhorn * balanced over allowed nodes. 730e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 731e8c5c824SLee Schermerhorn */ 7326ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 7336ae11b27SLee Schermerhorn bool acct_surplus) 734e8c5c824SLee Schermerhorn { 735e8c5c824SLee Schermerhorn int start_nid; 736e8c5c824SLee Schermerhorn int next_nid; 737e8c5c824SLee Schermerhorn int ret = 0; 738e8c5c824SLee Schermerhorn 7396ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 740e8c5c824SLee Schermerhorn next_nid = start_nid; 741e8c5c824SLee Schermerhorn 742e8c5c824SLee Schermerhorn do { 743685f3457SLee Schermerhorn /* 744685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 745685f3457SLee Schermerhorn * nodes with surplus pages. 746685f3457SLee Schermerhorn */ 747685f3457SLee Schermerhorn if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 748685f3457SLee Schermerhorn !list_empty(&h->hugepage_freelists[next_nid])) { 749e8c5c824SLee Schermerhorn struct page *page = 750e8c5c824SLee Schermerhorn list_entry(h->hugepage_freelists[next_nid].next, 751e8c5c824SLee Schermerhorn struct page, lru); 752e8c5c824SLee Schermerhorn list_del(&page->lru); 753e8c5c824SLee Schermerhorn h->free_huge_pages--; 754e8c5c824SLee Schermerhorn h->free_huge_pages_node[next_nid]--; 755685f3457SLee Schermerhorn if (acct_surplus) { 756685f3457SLee Schermerhorn h->surplus_huge_pages--; 757685f3457SLee Schermerhorn h->surplus_huge_pages_node[next_nid]--; 758685f3457SLee Schermerhorn } 759e8c5c824SLee Schermerhorn update_and_free_page(h, page); 760e8c5c824SLee Schermerhorn ret = 1; 7619a76db09SLee Schermerhorn break; 762e8c5c824SLee Schermerhorn } 7636ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, nodes_allowed); 7649a76db09SLee Schermerhorn } while (next_nid != start_nid); 765e8c5c824SLee Schermerhorn 766e8c5c824SLee Schermerhorn return ret; 767e8c5c824SLee Schermerhorn } 768e8c5c824SLee Schermerhorn 769a5516438SAndi Kleen static struct page *alloc_buddy_huge_page(struct hstate *h, 770a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 7717893d1d5SAdam Litke { 7727893d1d5SAdam Litke struct page *page; 773d1c3fb1fSNishanth Aravamudan unsigned int nid; 7747893d1d5SAdam Litke 775aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 776aa888a74SAndi Kleen return NULL; 777aa888a74SAndi Kleen 778d1c3fb1fSNishanth Aravamudan /* 779d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 780d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 781d1c3fb1fSNishanth Aravamudan * overcommit 782d1c3fb1fSNishanth Aravamudan * 783d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 784d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 785d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 786d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 787d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 788d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 789d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 790d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 791d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 792d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 793d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 794d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 795d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 796d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 797d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 798d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 799d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 800d1c3fb1fSNishanth Aravamudan */ 801d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 802a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 803d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 804d1c3fb1fSNishanth Aravamudan return NULL; 805d1c3fb1fSNishanth Aravamudan } else { 806a5516438SAndi Kleen h->nr_huge_pages++; 807a5516438SAndi Kleen h->surplus_huge_pages++; 808d1c3fb1fSNishanth Aravamudan } 809d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 810d1c3fb1fSNishanth Aravamudan 811551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 812551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 813a5516438SAndi Kleen huge_page_order(h)); 814d1c3fb1fSNishanth Aravamudan 815caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 816caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 817caff3a2cSGerald Schaefer return NULL; 818caff3a2cSGerald Schaefer } 819caff3a2cSGerald Schaefer 8207893d1d5SAdam Litke spin_lock(&hugetlb_lock); 821d1c3fb1fSNishanth Aravamudan if (page) { 8222668db91SAdam Litke /* 8232668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 8242668db91SAdam Litke * no users -- drop the buddy allocator's reference. 8252668db91SAdam Litke */ 8262668db91SAdam Litke put_page_testzero(page); 8272668db91SAdam Litke VM_BUG_ON(page_count(page)); 828d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 829d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 830d1c3fb1fSNishanth Aravamudan /* 831d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 832d1c3fb1fSNishanth Aravamudan */ 833a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 834a5516438SAndi Kleen h->surplus_huge_pages_node[nid]++; 8353b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 836d1c3fb1fSNishanth Aravamudan } else { 837a5516438SAndi Kleen h->nr_huge_pages--; 838a5516438SAndi Kleen h->surplus_huge_pages--; 8393b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 8407893d1d5SAdam Litke } 841d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 8427893d1d5SAdam Litke 8437893d1d5SAdam Litke return page; 8447893d1d5SAdam Litke } 8457893d1d5SAdam Litke 846e4e574b7SAdam Litke /* 847e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 848e4e574b7SAdam Litke * of size 'delta'. 849e4e574b7SAdam Litke */ 850a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 851e4e574b7SAdam Litke { 852e4e574b7SAdam Litke struct list_head surplus_list; 853e4e574b7SAdam Litke struct page *page, *tmp; 854e4e574b7SAdam Litke int ret, i; 855e4e574b7SAdam Litke int needed, allocated; 856e4e574b7SAdam Litke 857a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 858ac09b3a1SAdam Litke if (needed <= 0) { 859a5516438SAndi Kleen h->resv_huge_pages += delta; 860e4e574b7SAdam Litke return 0; 861ac09b3a1SAdam Litke } 862e4e574b7SAdam Litke 863e4e574b7SAdam Litke allocated = 0; 864e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 865e4e574b7SAdam Litke 866e4e574b7SAdam Litke ret = -ENOMEM; 867e4e574b7SAdam Litke retry: 868e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 869e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 870a5516438SAndi Kleen page = alloc_buddy_huge_page(h, NULL, 0); 871e4e574b7SAdam Litke if (!page) { 872e4e574b7SAdam Litke /* 873e4e574b7SAdam Litke * We were not able to allocate enough pages to 874e4e574b7SAdam Litke * satisfy the entire reservation so we free what 875e4e574b7SAdam Litke * we've allocated so far. 876e4e574b7SAdam Litke */ 877e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 878e4e574b7SAdam Litke needed = 0; 879e4e574b7SAdam Litke goto free; 880e4e574b7SAdam Litke } 881e4e574b7SAdam Litke 882e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 883e4e574b7SAdam Litke } 884e4e574b7SAdam Litke allocated += needed; 885e4e574b7SAdam Litke 886e4e574b7SAdam Litke /* 887e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 888e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 889e4e574b7SAdam Litke */ 890e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 891a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 892a5516438SAndi Kleen (h->free_huge_pages + allocated); 893e4e574b7SAdam Litke if (needed > 0) 894e4e574b7SAdam Litke goto retry; 895e4e574b7SAdam Litke 896e4e574b7SAdam Litke /* 897e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 898e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 899e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 900ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 901ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 902ac09b3a1SAdam Litke * before they are reserved. 903e4e574b7SAdam Litke */ 904e4e574b7SAdam Litke needed += allocated; 905a5516438SAndi Kleen h->resv_huge_pages += delta; 906e4e574b7SAdam Litke ret = 0; 907e4e574b7SAdam Litke free: 90819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 90919fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 91019fc3f0aSAdam Litke if ((--needed) < 0) 91119fc3f0aSAdam Litke break; 91219fc3f0aSAdam Litke list_del(&page->lru); 913a5516438SAndi Kleen enqueue_huge_page(h, page); 91419fc3f0aSAdam Litke } 91519fc3f0aSAdam Litke 91619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 91719fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 91819fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 919e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 920e4e574b7SAdam Litke list_del(&page->lru); 921af767cbdSAdam Litke /* 9222668db91SAdam Litke * The page has a reference count of zero already, so 9232668db91SAdam Litke * call free_huge_page directly instead of using 9242668db91SAdam Litke * put_page. This must be done with hugetlb_lock 925af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 926af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 927af767cbdSAdam Litke */ 9282668db91SAdam Litke free_huge_page(page); 929af767cbdSAdam Litke } 93019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 931e4e574b7SAdam Litke } 932e4e574b7SAdam Litke 933e4e574b7SAdam Litke return ret; 934e4e574b7SAdam Litke } 935e4e574b7SAdam Litke 936e4e574b7SAdam Litke /* 937e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 938e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 939e4e574b7SAdam Litke * never used. 940685f3457SLee Schermerhorn * Called with hugetlb_lock held. 941e4e574b7SAdam Litke */ 942a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 943a5516438SAndi Kleen unsigned long unused_resv_pages) 944e4e574b7SAdam Litke { 945e4e574b7SAdam Litke unsigned long nr_pages; 946e4e574b7SAdam Litke 947ac09b3a1SAdam Litke /* Uncommit the reservation */ 948a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 949ac09b3a1SAdam Litke 950aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 951aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 952aa888a74SAndi Kleen return; 953aa888a74SAndi Kleen 954a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 955e4e574b7SAdam Litke 956685f3457SLee Schermerhorn /* 957685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 9589b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 9599b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 9609b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 9619b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 9629b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 963685f3457SLee Schermerhorn */ 964685f3457SLee Schermerhorn while (nr_pages--) { 9659b5e5d0fSLee Schermerhorn if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 966685f3457SLee Schermerhorn break; 967e4e574b7SAdam Litke } 968e4e574b7SAdam Litke } 969e4e574b7SAdam Litke 970c37f9fb1SAndy Whitcroft /* 971c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 972c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 973c37f9fb1SAndy Whitcroft * reservation and actually increase quota before an allocation can occur. 974c37f9fb1SAndy Whitcroft * Where any new reservation would be required the reservation change is 975c37f9fb1SAndy Whitcroft * prepared, but not committed. Once the page has been quota'd allocated 976c37f9fb1SAndy Whitcroft * an instantiated the change should be committed via vma_commit_reservation. 977c37f9fb1SAndy Whitcroft * No action is required on failure. 978c37f9fb1SAndy Whitcroft */ 979e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 980a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 981c37f9fb1SAndy Whitcroft { 982c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 983c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 984c37f9fb1SAndy Whitcroft 985f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 986a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 987c37f9fb1SAndy Whitcroft return region_chg(&inode->i_mapping->private_list, 988c37f9fb1SAndy Whitcroft idx, idx + 1); 989c37f9fb1SAndy Whitcroft 99084afd99bSAndy Whitcroft } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 991c37f9fb1SAndy Whitcroft return 1; 992c37f9fb1SAndy Whitcroft 99384afd99bSAndy Whitcroft } else { 994e2f17d94SRoel Kluin long err; 995a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 99684afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 99784afd99bSAndy Whitcroft 99884afd99bSAndy Whitcroft err = region_chg(&reservations->regions, idx, idx + 1); 99984afd99bSAndy Whitcroft if (err < 0) 100084afd99bSAndy Whitcroft return err; 1001c37f9fb1SAndy Whitcroft return 0; 1002c37f9fb1SAndy Whitcroft } 100384afd99bSAndy Whitcroft } 1004a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1005a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1006c37f9fb1SAndy Whitcroft { 1007c37f9fb1SAndy Whitcroft struct address_space *mapping = vma->vm_file->f_mapping; 1008c37f9fb1SAndy Whitcroft struct inode *inode = mapping->host; 1009c37f9fb1SAndy Whitcroft 1010f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 1011a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1012c37f9fb1SAndy Whitcroft region_add(&inode->i_mapping->private_list, idx, idx + 1); 101384afd99bSAndy Whitcroft 101484afd99bSAndy Whitcroft } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1015a5516438SAndi Kleen pgoff_t idx = vma_hugecache_offset(h, vma, addr); 101684afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 101784afd99bSAndy Whitcroft 101884afd99bSAndy Whitcroft /* Mark this page used in the map. */ 101984afd99bSAndy Whitcroft region_add(&reservations->regions, idx, idx + 1); 1020c37f9fb1SAndy Whitcroft } 1021c37f9fb1SAndy Whitcroft } 1022c37f9fb1SAndy Whitcroft 1023348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 102404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1025348ea204SAdam Litke { 1026a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1027348ea204SAdam Litke struct page *page; 10282fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 1029a1e78772SMel Gorman struct inode *inode = mapping->host; 1030e2f17d94SRoel Kluin long chg; 10312fc39cecSAdam Litke 1032a1e78772SMel Gorman /* 1033a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 1034a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 1035a1e78772SMel Gorman * made before satisfying the allocation 1036c37f9fb1SAndy Whitcroft * MAP_NORESERVE mappings may also need pages and quota allocated 1037c37f9fb1SAndy Whitcroft * if no reserve mapping overlaps. 1038a1e78772SMel Gorman */ 1039a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1040c37f9fb1SAndy Whitcroft if (chg < 0) 1041c37f9fb1SAndy Whitcroft return ERR_PTR(chg); 1042c37f9fb1SAndy Whitcroft if (chg) 1043a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 1044a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 104590d8b7e6SAdam Litke 1046a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1047a5516438SAndi Kleen page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1048a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 1049a1e78772SMel Gorman 1050a1e78772SMel Gorman if (!page) { 1051a5516438SAndi Kleen page = alloc_buddy_huge_page(h, vma, addr); 1052a1e78772SMel Gorman if (!page) { 1053a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 10544a6018f7SMel Gorman return ERR_PTR(-VM_FAULT_SIGBUS); 1055a1e78772SMel Gorman } 1056a1e78772SMel Gorman } 1057a1e78772SMel Gorman 1058348ea204SAdam Litke set_page_refcounted(page); 10592fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 1060a1e78772SMel Gorman 1061a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 1062c37f9fb1SAndy Whitcroft 10637893d1d5SAdam Litke return page; 1064b45b5bd6SDavid Gibson } 1065b45b5bd6SDavid Gibson 106691f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1067aa888a74SAndi Kleen { 1068aa888a74SAndi Kleen struct huge_bootmem_page *m; 10699b5e5d0fSLee Schermerhorn int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1070aa888a74SAndi Kleen 1071aa888a74SAndi Kleen while (nr_nodes) { 1072aa888a74SAndi Kleen void *addr; 1073aa888a74SAndi Kleen 1074aa888a74SAndi Kleen addr = __alloc_bootmem_node_nopanic( 10756ae11b27SLee Schermerhorn NODE_DATA(hstate_next_node_to_alloc(h, 10769b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])), 1077aa888a74SAndi Kleen huge_page_size(h), huge_page_size(h), 0); 1078aa888a74SAndi Kleen 1079aa888a74SAndi Kleen if (addr) { 1080aa888a74SAndi Kleen /* 1081aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1082aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1083aa888a74SAndi Kleen * puts them into the mem_map). 1084aa888a74SAndi Kleen */ 1085aa888a74SAndi Kleen m = addr; 1086aa888a74SAndi Kleen goto found; 1087aa888a74SAndi Kleen } 1088aa888a74SAndi Kleen nr_nodes--; 1089aa888a74SAndi Kleen } 1090aa888a74SAndi Kleen return 0; 1091aa888a74SAndi Kleen 1092aa888a74SAndi Kleen found: 1093aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1094aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1095aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1096aa888a74SAndi Kleen m->hstate = h; 1097aa888a74SAndi Kleen return 1; 1098aa888a74SAndi Kleen } 1099aa888a74SAndi Kleen 110018229df5SAndy Whitcroft static void prep_compound_huge_page(struct page *page, int order) 110118229df5SAndy Whitcroft { 110218229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 110318229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 110418229df5SAndy Whitcroft else 110518229df5SAndy Whitcroft prep_compound_page(page, order); 110618229df5SAndy Whitcroft } 110718229df5SAndy Whitcroft 1108aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1109aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1110aa888a74SAndi Kleen { 1111aa888a74SAndi Kleen struct huge_bootmem_page *m; 1112aa888a74SAndi Kleen 1113aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1114aa888a74SAndi Kleen struct page *page = virt_to_page(m); 1115aa888a74SAndi Kleen struct hstate *h = m->hstate; 1116aa888a74SAndi Kleen __ClearPageReserved(page); 1117aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 111818229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1119aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1120aa888a74SAndi Kleen } 1121aa888a74SAndi Kleen } 1122aa888a74SAndi Kleen 11238faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 11241da177e4SLinus Torvalds { 11251da177e4SLinus Torvalds unsigned long i; 11261da177e4SLinus Torvalds 1127e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1128aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1129aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1130aa888a74SAndi Kleen break; 11319b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 11329b5e5d0fSLee Schermerhorn &node_states[N_HIGH_MEMORY])) 11331da177e4SLinus Torvalds break; 11341da177e4SLinus Torvalds } 11358faa8b07SAndi Kleen h->max_huge_pages = i; 1136e5ff2159SAndi Kleen } 1137e5ff2159SAndi Kleen 1138e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1139e5ff2159SAndi Kleen { 1140e5ff2159SAndi Kleen struct hstate *h; 1141e5ff2159SAndi Kleen 1142e5ff2159SAndi Kleen for_each_hstate(h) { 11438faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 11448faa8b07SAndi Kleen if (h->order < MAX_ORDER) 11458faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1146e5ff2159SAndi Kleen } 1147e5ff2159SAndi Kleen } 1148e5ff2159SAndi Kleen 11494abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 11504abd32dbSAndi Kleen { 11514abd32dbSAndi Kleen if (n >= (1UL << 30)) 11524abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 11534abd32dbSAndi Kleen else if (n >= (1UL << 20)) 11544abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 11554abd32dbSAndi Kleen else 11564abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 11574abd32dbSAndi Kleen return buf; 11584abd32dbSAndi Kleen } 11594abd32dbSAndi Kleen 1160e5ff2159SAndi Kleen static void __init report_hugepages(void) 1161e5ff2159SAndi Kleen { 1162e5ff2159SAndi Kleen struct hstate *h; 1163e5ff2159SAndi Kleen 1164e5ff2159SAndi Kleen for_each_hstate(h) { 11654abd32dbSAndi Kleen char buf[32]; 11664abd32dbSAndi Kleen printk(KERN_INFO "HugeTLB registered %s page size, " 11674abd32dbSAndi Kleen "pre-allocated %ld pages\n", 11684abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 11694abd32dbSAndi Kleen h->free_huge_pages); 1170e5ff2159SAndi Kleen } 1171e5ff2159SAndi Kleen } 1172e5ff2159SAndi Kleen 11731da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 11746ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 11756ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 11761da177e4SLinus Torvalds { 11774415cc8dSChristoph Lameter int i; 11784415cc8dSChristoph Lameter 1179aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1180aa888a74SAndi Kleen return; 1181aa888a74SAndi Kleen 11826ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 11831da177e4SLinus Torvalds struct page *page, *next; 1184a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1185a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1186a5516438SAndi Kleen if (count >= h->nr_huge_pages) 11876b0c880dSAdam Litke return; 11881da177e4SLinus Torvalds if (PageHighMem(page)) 11891da177e4SLinus Torvalds continue; 11901da177e4SLinus Torvalds list_del(&page->lru); 1191e5ff2159SAndi Kleen update_and_free_page(h, page); 1192a5516438SAndi Kleen h->free_huge_pages--; 1193a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 11941da177e4SLinus Torvalds } 11951da177e4SLinus Torvalds } 11961da177e4SLinus Torvalds } 11971da177e4SLinus Torvalds #else 11986ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 11996ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12001da177e4SLinus Torvalds { 12011da177e4SLinus Torvalds } 12021da177e4SLinus Torvalds #endif 12031da177e4SLinus Torvalds 120420a0307cSWu Fengguang /* 120520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 120620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 120720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 120820a0307cSWu Fengguang */ 12096ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 12106ae11b27SLee Schermerhorn int delta) 121120a0307cSWu Fengguang { 1212e8c5c824SLee Schermerhorn int start_nid, next_nid; 121320a0307cSWu Fengguang int ret = 0; 121420a0307cSWu Fengguang 121520a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 121620a0307cSWu Fengguang 1217e8c5c824SLee Schermerhorn if (delta < 0) 12186ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1219e8c5c824SLee Schermerhorn else 12206ae11b27SLee Schermerhorn start_nid = hstate_next_node_to_free(h, nodes_allowed); 1221e8c5c824SLee Schermerhorn next_nid = start_nid; 1222e8c5c824SLee Schermerhorn 1223e8c5c824SLee Schermerhorn do { 1224e8c5c824SLee Schermerhorn int nid = next_nid; 1225e8c5c824SLee Schermerhorn if (delta < 0) { 1226e8c5c824SLee Schermerhorn /* 1227e8c5c824SLee Schermerhorn * To shrink on this node, there must be a surplus page 1228e8c5c824SLee Schermerhorn */ 12299a76db09SLee Schermerhorn if (!h->surplus_huge_pages_node[nid]) { 12306ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_alloc(h, 12316ae11b27SLee Schermerhorn nodes_allowed); 123220a0307cSWu Fengguang continue; 1233e8c5c824SLee Schermerhorn } 12349a76db09SLee Schermerhorn } 1235e8c5c824SLee Schermerhorn if (delta > 0) { 1236e8c5c824SLee Schermerhorn /* 1237e8c5c824SLee Schermerhorn * Surplus cannot exceed the total number of pages 1238e8c5c824SLee Schermerhorn */ 1239e8c5c824SLee Schermerhorn if (h->surplus_huge_pages_node[nid] >= 12409a76db09SLee Schermerhorn h->nr_huge_pages_node[nid]) { 12416ae11b27SLee Schermerhorn next_nid = hstate_next_node_to_free(h, 12426ae11b27SLee Schermerhorn nodes_allowed); 124320a0307cSWu Fengguang continue; 1244e8c5c824SLee Schermerhorn } 12459a76db09SLee Schermerhorn } 124620a0307cSWu Fengguang 124720a0307cSWu Fengguang h->surplus_huge_pages += delta; 124820a0307cSWu Fengguang h->surplus_huge_pages_node[nid] += delta; 124920a0307cSWu Fengguang ret = 1; 125020a0307cSWu Fengguang break; 1251e8c5c824SLee Schermerhorn } while (next_nid != start_nid); 125220a0307cSWu Fengguang 125320a0307cSWu Fengguang return ret; 125420a0307cSWu Fengguang } 125520a0307cSWu Fengguang 1256a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 12576ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 12586ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 12591da177e4SLinus Torvalds { 12607893d1d5SAdam Litke unsigned long min_count, ret; 12611da177e4SLinus Torvalds 1262aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1263aa888a74SAndi Kleen return h->max_huge_pages; 1264aa888a74SAndi Kleen 12657893d1d5SAdam Litke /* 12667893d1d5SAdam Litke * Increase the pool size 12677893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 12687893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1269d1c3fb1fSNishanth Aravamudan * 1270d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1271d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1272d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1273d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1274d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 12757893d1d5SAdam Litke */ 12761da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1277a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 12786ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 12797893d1d5SAdam Litke break; 12807893d1d5SAdam Litke } 12817893d1d5SAdam Litke 1282a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 12837893d1d5SAdam Litke /* 12847893d1d5SAdam Litke * If this allocation races such that we no longer need the 12857893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 12867893d1d5SAdam Litke * and reducing the surplus. 12877893d1d5SAdam Litke */ 12887893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 12896ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 12907893d1d5SAdam Litke spin_lock(&hugetlb_lock); 12917893d1d5SAdam Litke if (!ret) 12927893d1d5SAdam Litke goto out; 12937893d1d5SAdam Litke 1294536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1295536240f2SMel Gorman if (signal_pending(current)) 1296536240f2SMel Gorman goto out; 12977893d1d5SAdam Litke } 12987893d1d5SAdam Litke 12997893d1d5SAdam Litke /* 13007893d1d5SAdam Litke * Decrease the pool size 13017893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 13027893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 13037893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 13047893d1d5SAdam Litke * to the desired size as pages become free. 1305d1c3fb1fSNishanth Aravamudan * 1306d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1307d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1308d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1309d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1310d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1311d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1312d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 13137893d1d5SAdam Litke */ 1314a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 13156b0c880dSAdam Litke min_count = max(count, min_count); 13166ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1317a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 13186ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 13191da177e4SLinus Torvalds break; 13201da177e4SLinus Torvalds } 1321a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 13226ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 13237893d1d5SAdam Litke break; 13247893d1d5SAdam Litke } 13257893d1d5SAdam Litke out: 1326a5516438SAndi Kleen ret = persistent_huge_pages(h); 13271da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 13287893d1d5SAdam Litke return ret; 13291da177e4SLinus Torvalds } 13301da177e4SLinus Torvalds 1331a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1332a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1333a3437870SNishanth Aravamudan 1334a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1335a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1336a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1337a3437870SNishanth Aravamudan 1338a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1339a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1340a3437870SNishanth Aravamudan 13419a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 13429a305230SLee Schermerhorn 13439a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1344a3437870SNishanth Aravamudan { 1345a3437870SNishanth Aravamudan int i; 13469a305230SLee Schermerhorn 1347a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 13489a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 13499a305230SLee Schermerhorn if (nidp) 13509a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1351a3437870SNishanth Aravamudan return &hstates[i]; 13529a305230SLee Schermerhorn } 13539a305230SLee Schermerhorn 13549a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1355a3437870SNishanth Aravamudan } 1356a3437870SNishanth Aravamudan 135706808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1358a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1359a3437870SNishanth Aravamudan { 13609a305230SLee Schermerhorn struct hstate *h; 13619a305230SLee Schermerhorn unsigned long nr_huge_pages; 13629a305230SLee Schermerhorn int nid; 13639a305230SLee Schermerhorn 13649a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 13659a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 13669a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 13679a305230SLee Schermerhorn else 13689a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 13699a305230SLee Schermerhorn 13709a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1371a3437870SNishanth Aravamudan } 137206808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 137306808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 137406808b08SLee Schermerhorn const char *buf, size_t len) 1375a3437870SNishanth Aravamudan { 1376a3437870SNishanth Aravamudan int err; 13779a305230SLee Schermerhorn int nid; 137806808b08SLee Schermerhorn unsigned long count; 13799a305230SLee Schermerhorn struct hstate *h; 1380bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1381a3437870SNishanth Aravamudan 138206808b08SLee Schermerhorn err = strict_strtoul(buf, 10, &count); 1383a3437870SNishanth Aravamudan if (err) 1384a3437870SNishanth Aravamudan return 0; 1385a3437870SNishanth Aravamudan 13869a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 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; 141206808b08SLee Schermerhorn } 141306808b08SLee Schermerhorn 141406808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 141506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 141606808b08SLee Schermerhorn { 141706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 141806808b08SLee Schermerhorn } 141906808b08SLee Schermerhorn 142006808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 142106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 142206808b08SLee Schermerhorn { 142306808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1424a3437870SNishanth Aravamudan } 1425a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1426a3437870SNishanth Aravamudan 142706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 142806808b08SLee Schermerhorn 142906808b08SLee Schermerhorn /* 143006808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 143106808b08SLee Schermerhorn * huge page alloc/free. 143206808b08SLee Schermerhorn */ 143306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 143406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 143506808b08SLee Schermerhorn { 143606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 143706808b08SLee Schermerhorn } 143806808b08SLee Schermerhorn 143906808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 144006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 144106808b08SLee Schermerhorn { 144206808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 144306808b08SLee Schermerhorn } 144406808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 144506808b08SLee Schermerhorn #endif 144606808b08SLee Schermerhorn 144706808b08SLee Schermerhorn 1448a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1449a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1450a3437870SNishanth Aravamudan { 14519a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1452a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1453a3437870SNishanth Aravamudan } 1454a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1455a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1456a3437870SNishanth Aravamudan { 1457a3437870SNishanth Aravamudan int err; 1458a3437870SNishanth Aravamudan unsigned long input; 14599a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1460a3437870SNishanth Aravamudan 1461a3437870SNishanth Aravamudan err = strict_strtoul(buf, 10, &input); 1462a3437870SNishanth Aravamudan if (err) 1463a3437870SNishanth Aravamudan return 0; 1464a3437870SNishanth Aravamudan 1465a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1466a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1467a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1468a3437870SNishanth Aravamudan 1469a3437870SNishanth Aravamudan return count; 1470a3437870SNishanth Aravamudan } 1471a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1472a3437870SNishanth Aravamudan 1473a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1474a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1475a3437870SNishanth Aravamudan { 14769a305230SLee Schermerhorn struct hstate *h; 14779a305230SLee Schermerhorn unsigned long free_huge_pages; 14789a305230SLee Schermerhorn int nid; 14799a305230SLee Schermerhorn 14809a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 14819a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 14829a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 14839a305230SLee Schermerhorn else 14849a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 14859a305230SLee Schermerhorn 14869a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1487a3437870SNishanth Aravamudan } 1488a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1489a3437870SNishanth Aravamudan 1490a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1491a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1492a3437870SNishanth Aravamudan { 14939a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1494a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1495a3437870SNishanth Aravamudan } 1496a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1497a3437870SNishanth Aravamudan 1498a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1499a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1500a3437870SNishanth Aravamudan { 15019a305230SLee Schermerhorn struct hstate *h; 15029a305230SLee Schermerhorn unsigned long surplus_huge_pages; 15039a305230SLee Schermerhorn int nid; 15049a305230SLee Schermerhorn 15059a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15069a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15079a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 15089a305230SLee Schermerhorn else 15099a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 15109a305230SLee Schermerhorn 15119a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1512a3437870SNishanth Aravamudan } 1513a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1514a3437870SNishanth Aravamudan 1515a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1516a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1517a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1518a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1519a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1520a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 152106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 152206808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 152306808b08SLee Schermerhorn #endif 1524a3437870SNishanth Aravamudan NULL, 1525a3437870SNishanth Aravamudan }; 1526a3437870SNishanth Aravamudan 1527a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1528a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1529a3437870SNishanth Aravamudan }; 1530a3437870SNishanth Aravamudan 1531094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 15329a305230SLee Schermerhorn struct kobject **hstate_kobjs, 15339a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1534a3437870SNishanth Aravamudan { 1535a3437870SNishanth Aravamudan int retval; 15369a305230SLee Schermerhorn int hi = h - hstates; 1537a3437870SNishanth Aravamudan 15389a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 15399a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1540a3437870SNishanth Aravamudan return -ENOMEM; 1541a3437870SNishanth Aravamudan 15429a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1543a3437870SNishanth Aravamudan if (retval) 15449a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1545a3437870SNishanth Aravamudan 1546a3437870SNishanth Aravamudan return retval; 1547a3437870SNishanth Aravamudan } 1548a3437870SNishanth Aravamudan 1549a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1550a3437870SNishanth Aravamudan { 1551a3437870SNishanth Aravamudan struct hstate *h; 1552a3437870SNishanth Aravamudan int err; 1553a3437870SNishanth Aravamudan 1554a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1555a3437870SNishanth Aravamudan if (!hugepages_kobj) 1556a3437870SNishanth Aravamudan return; 1557a3437870SNishanth Aravamudan 1558a3437870SNishanth Aravamudan for_each_hstate(h) { 15599a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 15609a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1561a3437870SNishanth Aravamudan if (err) 1562a3437870SNishanth Aravamudan printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1563a3437870SNishanth Aravamudan h->name); 1564a3437870SNishanth Aravamudan } 1565a3437870SNishanth Aravamudan } 1566a3437870SNishanth Aravamudan 15679a305230SLee Schermerhorn #ifdef CONFIG_NUMA 15689a305230SLee Schermerhorn 15699a305230SLee Schermerhorn /* 15709a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 15719a305230SLee Schermerhorn * with node sysdevs in node_devices[] using a parallel array. The array 15729a305230SLee Schermerhorn * index of a node sysdev or _hstate == node id. 15739a305230SLee Schermerhorn * This is here to avoid any static dependency of the node sysdev driver, in 15749a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 15759a305230SLee Schermerhorn */ 15769a305230SLee Schermerhorn struct node_hstate { 15779a305230SLee Schermerhorn struct kobject *hugepages_kobj; 15789a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 15799a305230SLee Schermerhorn }; 15809a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 15819a305230SLee Schermerhorn 15829a305230SLee Schermerhorn /* 15839a305230SLee Schermerhorn * A subset of global hstate attributes for node sysdevs 15849a305230SLee Schermerhorn */ 15859a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 15869a305230SLee Schermerhorn &nr_hugepages_attr.attr, 15879a305230SLee Schermerhorn &free_hugepages_attr.attr, 15889a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 15899a305230SLee Schermerhorn NULL, 15909a305230SLee Schermerhorn }; 15919a305230SLee Schermerhorn 15929a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 15939a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 15949a305230SLee Schermerhorn }; 15959a305230SLee Schermerhorn 15969a305230SLee Schermerhorn /* 15979a305230SLee Schermerhorn * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. 15989a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 15999a305230SLee Schermerhorn */ 16009a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 16019a305230SLee Schermerhorn { 16029a305230SLee Schermerhorn int nid; 16039a305230SLee Schermerhorn 16049a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 16059a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 16069a305230SLee Schermerhorn int i; 16079a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 16089a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 16099a305230SLee Schermerhorn if (nidp) 16109a305230SLee Schermerhorn *nidp = nid; 16119a305230SLee Schermerhorn return &hstates[i]; 16129a305230SLee Schermerhorn } 16139a305230SLee Schermerhorn } 16149a305230SLee Schermerhorn 16159a305230SLee Schermerhorn BUG(); 16169a305230SLee Schermerhorn return NULL; 16179a305230SLee Schermerhorn } 16189a305230SLee Schermerhorn 16199a305230SLee Schermerhorn /* 16209a305230SLee Schermerhorn * Unregister hstate attributes from a single node sysdev. 16219a305230SLee Schermerhorn * No-op if no hstate attributes attached. 16229a305230SLee Schermerhorn */ 16239a305230SLee Schermerhorn void hugetlb_unregister_node(struct node *node) 16249a305230SLee Schermerhorn { 16259a305230SLee Schermerhorn struct hstate *h; 16269a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16279a305230SLee Schermerhorn 16289a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16299b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 16309a305230SLee Schermerhorn 16319a305230SLee Schermerhorn for_each_hstate(h) 16329a305230SLee Schermerhorn if (nhs->hstate_kobjs[h - hstates]) { 16339a305230SLee Schermerhorn kobject_put(nhs->hstate_kobjs[h - hstates]); 16349a305230SLee Schermerhorn nhs->hstate_kobjs[h - hstates] = NULL; 16359a305230SLee Schermerhorn } 16369a305230SLee Schermerhorn 16379a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 16389a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 16399a305230SLee Schermerhorn } 16409a305230SLee Schermerhorn 16419a305230SLee Schermerhorn /* 16429a305230SLee Schermerhorn * hugetlb module exit: unregister hstate attributes from node sysdevs 16439a305230SLee Schermerhorn * that have them. 16449a305230SLee Schermerhorn */ 16459a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 16469a305230SLee Schermerhorn { 16479a305230SLee Schermerhorn int nid; 16489a305230SLee Schermerhorn 16499a305230SLee Schermerhorn /* 16509a305230SLee Schermerhorn * disable node sysdev registrations. 16519a305230SLee Schermerhorn */ 16529a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 16539a305230SLee Schermerhorn 16549a305230SLee Schermerhorn /* 16559a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 16569a305230SLee Schermerhorn */ 16579a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 16589a305230SLee Schermerhorn hugetlb_unregister_node(&node_devices[nid]); 16599a305230SLee Schermerhorn } 16609a305230SLee Schermerhorn 16619a305230SLee Schermerhorn /* 16629a305230SLee Schermerhorn * Register hstate attributes for a single node sysdev. 16639a305230SLee Schermerhorn * No-op if attributes already registered. 16649a305230SLee Schermerhorn */ 16659a305230SLee Schermerhorn void hugetlb_register_node(struct node *node) 16669a305230SLee Schermerhorn { 16679a305230SLee Schermerhorn struct hstate *h; 16689a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 16699a305230SLee Schermerhorn int err; 16709a305230SLee Schermerhorn 16719a305230SLee Schermerhorn if (nhs->hugepages_kobj) 16729a305230SLee Schermerhorn return; /* already allocated */ 16739a305230SLee Schermerhorn 16749a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 16759a305230SLee Schermerhorn &node->sysdev.kobj); 16769a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 16779a305230SLee Schermerhorn return; 16789a305230SLee Schermerhorn 16799a305230SLee Schermerhorn for_each_hstate(h) { 16809a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 16819a305230SLee Schermerhorn nhs->hstate_kobjs, 16829a305230SLee Schermerhorn &per_node_hstate_attr_group); 16839a305230SLee Schermerhorn if (err) { 16849a305230SLee Schermerhorn printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 16859a305230SLee Schermerhorn " for node %d\n", 16869a305230SLee Schermerhorn h->name, node->sysdev.id); 16879a305230SLee Schermerhorn hugetlb_unregister_node(node); 16889a305230SLee Schermerhorn break; 16899a305230SLee Schermerhorn } 16909a305230SLee Schermerhorn } 16919a305230SLee Schermerhorn } 16929a305230SLee Schermerhorn 16939a305230SLee Schermerhorn /* 16949b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 16959b5e5d0fSLee Schermerhorn * sysdevs of nodes that have memory. All on-line nodes should have 16969b5e5d0fSLee Schermerhorn * registered their associated sysdev by this time. 16979a305230SLee Schermerhorn */ 16989a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 16999a305230SLee Schermerhorn { 17009a305230SLee Schermerhorn int nid; 17019a305230SLee Schermerhorn 17029b5e5d0fSLee Schermerhorn for_each_node_state(nid, N_HIGH_MEMORY) { 17039a305230SLee Schermerhorn struct node *node = &node_devices[nid]; 17049a305230SLee Schermerhorn if (node->sysdev.id == nid) 17059a305230SLee Schermerhorn hugetlb_register_node(node); 17069a305230SLee Schermerhorn } 17079a305230SLee Schermerhorn 17089a305230SLee Schermerhorn /* 17099a305230SLee Schermerhorn * Let the node sysdev driver know we're here so it can 17109a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 17119a305230SLee Schermerhorn */ 17129a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 17139a305230SLee Schermerhorn hugetlb_unregister_node); 17149a305230SLee Schermerhorn } 17159a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 17169a305230SLee Schermerhorn 17179a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 17189a305230SLee Schermerhorn { 17199a305230SLee Schermerhorn BUG(); 17209a305230SLee Schermerhorn if (nidp) 17219a305230SLee Schermerhorn *nidp = -1; 17229a305230SLee Schermerhorn return NULL; 17239a305230SLee Schermerhorn } 17249a305230SLee Schermerhorn 17259a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 17269a305230SLee Schermerhorn 17279a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 17289a305230SLee Schermerhorn 17299a305230SLee Schermerhorn #endif 17309a305230SLee Schermerhorn 1731a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1732a3437870SNishanth Aravamudan { 1733a3437870SNishanth Aravamudan struct hstate *h; 1734a3437870SNishanth Aravamudan 17359a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 17369a305230SLee Schermerhorn 1737a3437870SNishanth Aravamudan for_each_hstate(h) { 1738a3437870SNishanth Aravamudan kobject_put(hstate_kobjs[h - hstates]); 1739a3437870SNishanth Aravamudan } 1740a3437870SNishanth Aravamudan 1741a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 1742a3437870SNishanth Aravamudan } 1743a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1744a3437870SNishanth Aravamudan 1745a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1746a3437870SNishanth Aravamudan { 17470ef89d25SBenjamin Herrenschmidt /* Some platform decide whether they support huge pages at boot 17480ef89d25SBenjamin Herrenschmidt * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 17490ef89d25SBenjamin Herrenschmidt * there is no such support 17500ef89d25SBenjamin Herrenschmidt */ 17510ef89d25SBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 17520ef89d25SBenjamin Herrenschmidt return 0; 1753a3437870SNishanth Aravamudan 1754e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1755e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1756e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1757a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1758a3437870SNishanth Aravamudan } 1759e11bfbfcSNick Piggin default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1760e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1761e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1762a3437870SNishanth Aravamudan 1763a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1764a3437870SNishanth Aravamudan 1765aa888a74SAndi Kleen gather_bootmem_prealloc(); 1766aa888a74SAndi Kleen 1767a3437870SNishanth Aravamudan report_hugepages(); 1768a3437870SNishanth Aravamudan 1769a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 1770a3437870SNishanth Aravamudan 17719a305230SLee Schermerhorn hugetlb_register_all_nodes(); 17729a305230SLee Schermerhorn 1773a3437870SNishanth Aravamudan return 0; 1774a3437870SNishanth Aravamudan } 1775a3437870SNishanth Aravamudan module_init(hugetlb_init); 1776a3437870SNishanth Aravamudan 1777a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 1778a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 1779a3437870SNishanth Aravamudan { 1780a3437870SNishanth Aravamudan struct hstate *h; 17818faa8b07SAndi Kleen unsigned long i; 17828faa8b07SAndi Kleen 1783a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 1784a3437870SNishanth Aravamudan printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1785a3437870SNishanth Aravamudan return; 1786a3437870SNishanth Aravamudan } 1787a3437870SNishanth Aravamudan BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1788a3437870SNishanth Aravamudan BUG_ON(order == 0); 1789a3437870SNishanth Aravamudan h = &hstates[max_hstate++]; 1790a3437870SNishanth Aravamudan h->order = order; 1791a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 17928faa8b07SAndi Kleen h->nr_huge_pages = 0; 17938faa8b07SAndi Kleen h->free_huge_pages = 0; 17948faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 17958faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 17969b5e5d0fSLee Schermerhorn h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 17979b5e5d0fSLee Schermerhorn h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1798a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1799a3437870SNishanth Aravamudan huge_page_size(h)/1024); 18008faa8b07SAndi Kleen 1801a3437870SNishanth Aravamudan parsed_hstate = h; 1802a3437870SNishanth Aravamudan } 1803a3437870SNishanth Aravamudan 1804e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 1805a3437870SNishanth Aravamudan { 1806a3437870SNishanth Aravamudan unsigned long *mhp; 18078faa8b07SAndi Kleen static unsigned long *last_mhp; 1808a3437870SNishanth Aravamudan 1809a3437870SNishanth Aravamudan /* 1810a3437870SNishanth Aravamudan * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1811a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 1812a3437870SNishanth Aravamudan */ 1813a3437870SNishanth Aravamudan if (!max_hstate) 1814a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 1815a3437870SNishanth Aravamudan else 1816a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 1817a3437870SNishanth Aravamudan 18188faa8b07SAndi Kleen if (mhp == last_mhp) { 18198faa8b07SAndi Kleen printk(KERN_WARNING "hugepages= specified twice without " 18208faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 18218faa8b07SAndi Kleen return 1; 18228faa8b07SAndi Kleen } 18238faa8b07SAndi Kleen 1824a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 1825a3437870SNishanth Aravamudan *mhp = 0; 1826a3437870SNishanth Aravamudan 18278faa8b07SAndi Kleen /* 18288faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 18298faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 18308faa8b07SAndi Kleen * use the bootmem allocator. 18318faa8b07SAndi Kleen */ 18328faa8b07SAndi Kleen if (max_hstate && parsed_hstate->order >= MAX_ORDER) 18338faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 18348faa8b07SAndi Kleen 18358faa8b07SAndi Kleen last_mhp = mhp; 18368faa8b07SAndi Kleen 1837a3437870SNishanth Aravamudan return 1; 1838a3437870SNishanth Aravamudan } 1839e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 1840e11bfbfcSNick Piggin 1841e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 1842e11bfbfcSNick Piggin { 1843e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 1844e11bfbfcSNick Piggin return 1; 1845e11bfbfcSNick Piggin } 1846e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 1847a3437870SNishanth Aravamudan 18488a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 18498a213460SNishanth Aravamudan { 18508a213460SNishanth Aravamudan int node; 18518a213460SNishanth Aravamudan unsigned int nr = 0; 18528a213460SNishanth Aravamudan 18538a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 18548a213460SNishanth Aravamudan nr += array[node]; 18558a213460SNishanth Aravamudan 18568a213460SNishanth Aravamudan return nr; 18578a213460SNishanth Aravamudan } 18588a213460SNishanth Aravamudan 18598a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 186006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 186106808b08SLee Schermerhorn struct ctl_table *table, int write, 186206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 18631da177e4SLinus Torvalds { 1864e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 1865e5ff2159SAndi Kleen unsigned long tmp; 1866e5ff2159SAndi Kleen 1867e5ff2159SAndi Kleen if (!write) 1868e5ff2159SAndi Kleen tmp = h->max_huge_pages; 1869e5ff2159SAndi Kleen 1870e5ff2159SAndi Kleen table->data = &tmp; 1871e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 18728d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1873e5ff2159SAndi Kleen 187406808b08SLee Schermerhorn if (write) { 1875bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 1876bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 187706808b08SLee Schermerhorn if (!(obey_mempolicy && 187806808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 187906808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 188006808b08SLee Schermerhorn nodes_allowed = &node_states[N_HIGH_MEMORY]; 188106808b08SLee Schermerhorn } 188206808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 188306808b08SLee Schermerhorn 188406808b08SLee Schermerhorn if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 188506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 188606808b08SLee Schermerhorn } 1887e5ff2159SAndi Kleen 18881da177e4SLinus Torvalds return 0; 18891da177e4SLinus Torvalds } 1890396faf03SMel Gorman 189106808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 189206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 189306808b08SLee Schermerhorn { 189406808b08SLee Schermerhorn 189506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 189606808b08SLee Schermerhorn buffer, length, ppos); 189706808b08SLee Schermerhorn } 189806808b08SLee Schermerhorn 189906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 190006808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 190106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 190206808b08SLee Schermerhorn { 190306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 190406808b08SLee Schermerhorn buffer, length, ppos); 190506808b08SLee Schermerhorn } 190606808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 190706808b08SLee Schermerhorn 1908396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 19098d65af78SAlexey Dobriyan void __user *buffer, 1910396faf03SMel Gorman size_t *length, loff_t *ppos) 1911396faf03SMel Gorman { 19128d65af78SAlexey Dobriyan proc_dointvec(table, write, buffer, length, ppos); 1913396faf03SMel Gorman if (hugepages_treat_as_movable) 1914396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1915396faf03SMel Gorman else 1916396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 1917396faf03SMel Gorman return 0; 1918396faf03SMel Gorman } 1919396faf03SMel Gorman 1920a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 19218d65af78SAlexey Dobriyan void __user *buffer, 1922a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 1923a3d0c6aaSNishanth Aravamudan { 1924a5516438SAndi Kleen struct hstate *h = &default_hstate; 1925e5ff2159SAndi Kleen unsigned long tmp; 1926e5ff2159SAndi Kleen 1927e5ff2159SAndi Kleen if (!write) 1928e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 1929e5ff2159SAndi Kleen 1930e5ff2159SAndi Kleen table->data = &tmp; 1931e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 19328d65af78SAlexey Dobriyan proc_doulongvec_minmax(table, write, buffer, length, ppos); 1933e5ff2159SAndi Kleen 1934e5ff2159SAndi Kleen if (write) { 1935064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 1936e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 1937a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1938e5ff2159SAndi Kleen } 1939e5ff2159SAndi Kleen 1940a3d0c6aaSNishanth Aravamudan return 0; 1941a3d0c6aaSNishanth Aravamudan } 1942a3d0c6aaSNishanth Aravamudan 19431da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 19441da177e4SLinus Torvalds 1945e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 19461da177e4SLinus Torvalds { 1947a5516438SAndi Kleen struct hstate *h = &default_hstate; 1948e1759c21SAlexey Dobriyan seq_printf(m, 19491da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 19501da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 1951b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 19527893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 19534f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 1954a5516438SAndi Kleen h->nr_huge_pages, 1955a5516438SAndi Kleen h->free_huge_pages, 1956a5516438SAndi Kleen h->resv_huge_pages, 1957a5516438SAndi Kleen h->surplus_huge_pages, 1958a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 19591da177e4SLinus Torvalds } 19601da177e4SLinus Torvalds 19611da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 19621da177e4SLinus Torvalds { 1963a5516438SAndi Kleen struct hstate *h = &default_hstate; 19641da177e4SLinus Torvalds return sprintf(buf, 19651da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 1966a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 1967a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 1968a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 1969a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 1970a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 19711da177e4SLinus Torvalds } 19721da177e4SLinus Torvalds 19731da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 19741da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 19751da177e4SLinus Torvalds { 1976a5516438SAndi Kleen struct hstate *h = &default_hstate; 1977a5516438SAndi Kleen return h->nr_huge_pages * pages_per_huge_page(h); 19781da177e4SLinus Torvalds } 19791da177e4SLinus Torvalds 1980a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 1981fc1b8a73SMel Gorman { 1982fc1b8a73SMel Gorman int ret = -ENOMEM; 1983fc1b8a73SMel Gorman 1984fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 1985fc1b8a73SMel Gorman /* 1986fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 1987fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 1988fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 1989fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 1990fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 1991fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 1992fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 1993fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 1994fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 1995fc1b8a73SMel Gorman * 1996fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 1997fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 1998fc1b8a73SMel Gorman * we fall back to check against current free page availability as 1999fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2000fc1b8a73SMel Gorman * semantics that cpuset has. 2001fc1b8a73SMel Gorman */ 2002fc1b8a73SMel Gorman if (delta > 0) { 2003a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2004fc1b8a73SMel Gorman goto out; 2005fc1b8a73SMel Gorman 2006a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2007a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2008fc1b8a73SMel Gorman goto out; 2009fc1b8a73SMel Gorman } 2010fc1b8a73SMel Gorman } 2011fc1b8a73SMel Gorman 2012fc1b8a73SMel Gorman ret = 0; 2013fc1b8a73SMel Gorman if (delta < 0) 2014a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2015fc1b8a73SMel Gorman 2016fc1b8a73SMel Gorman out: 2017fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2018fc1b8a73SMel Gorman return ret; 2019fc1b8a73SMel Gorman } 2020fc1b8a73SMel Gorman 202184afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 202284afd99bSAndy Whitcroft { 202384afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 202484afd99bSAndy Whitcroft 202584afd99bSAndy Whitcroft /* 202684afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 202784afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 202884afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 202984afd99bSAndy Whitcroft * has a reference to the reservation map it cannot dissappear until 203084afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 203184afd99bSAndy Whitcroft * new reference here without additional locking. 203284afd99bSAndy Whitcroft */ 203384afd99bSAndy Whitcroft if (reservations) 203484afd99bSAndy Whitcroft kref_get(&reservations->refs); 203584afd99bSAndy Whitcroft } 203684afd99bSAndy Whitcroft 2037a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2038a1e78772SMel Gorman { 2039a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 204084afd99bSAndy Whitcroft struct resv_map *reservations = vma_resv_map(vma); 204184afd99bSAndy Whitcroft unsigned long reserve; 204284afd99bSAndy Whitcroft unsigned long start; 204384afd99bSAndy Whitcroft unsigned long end; 204484afd99bSAndy Whitcroft 204584afd99bSAndy Whitcroft if (reservations) { 2046a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2047a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 204884afd99bSAndy Whitcroft 204984afd99bSAndy Whitcroft reserve = (end - start) - 205084afd99bSAndy Whitcroft region_count(&reservations->regions, start, end); 205184afd99bSAndy Whitcroft 205284afd99bSAndy Whitcroft kref_put(&reservations->refs, resv_map_release); 205384afd99bSAndy Whitcroft 20547251ff78SAdam Litke if (reserve) { 2055a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 20567251ff78SAdam Litke hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 20577251ff78SAdam Litke } 2058a1e78772SMel Gorman } 205984afd99bSAndy Whitcroft } 2060a1e78772SMel Gorman 20611da177e4SLinus Torvalds /* 20621da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 20631da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 20641da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 20651da177e4SLinus Torvalds * this far. 20661da177e4SLinus Torvalds */ 2067d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 20681da177e4SLinus Torvalds { 20691da177e4SLinus Torvalds BUG(); 2070d0217ac0SNick Piggin return 0; 20711da177e4SLinus Torvalds } 20721da177e4SLinus Torvalds 2073f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2074d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 207584afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2076a1e78772SMel Gorman .close = hugetlb_vm_op_close, 20771da177e4SLinus Torvalds }; 20781da177e4SLinus Torvalds 20791e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 20801e8f889bSDavid Gibson int writable) 208163551ae0SDavid Gibson { 208263551ae0SDavid Gibson pte_t entry; 208363551ae0SDavid Gibson 20841e8f889bSDavid Gibson if (writable) { 208563551ae0SDavid Gibson entry = 208663551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 208763551ae0SDavid Gibson } else { 20887f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 208963551ae0SDavid Gibson } 209063551ae0SDavid Gibson entry = pte_mkyoung(entry); 209163551ae0SDavid Gibson entry = pte_mkhuge(entry); 209263551ae0SDavid Gibson 209363551ae0SDavid Gibson return entry; 209463551ae0SDavid Gibson } 209563551ae0SDavid Gibson 20961e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 20971e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 20981e8f889bSDavid Gibson { 20991e8f889bSDavid Gibson pte_t entry; 21001e8f889bSDavid Gibson 21017f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 21027f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 21034b3073e1SRussell King update_mmu_cache(vma, address, ptep); 21041e8f889bSDavid Gibson } 21058dab5241SBenjamin Herrenschmidt } 21061e8f889bSDavid Gibson 21071e8f889bSDavid Gibson 210863551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 210963551ae0SDavid Gibson struct vm_area_struct *vma) 211063551ae0SDavid Gibson { 211163551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 211263551ae0SDavid Gibson struct page *ptepage; 21131c59827dSHugh Dickins unsigned long addr; 21141e8f889bSDavid Gibson int cow; 2115a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2116a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 21171e8f889bSDavid Gibson 21181e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 211963551ae0SDavid Gibson 2120a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2121c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2122c74df32cSHugh Dickins if (!src_pte) 2123c74df32cSHugh Dickins continue; 2124a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 212563551ae0SDavid Gibson if (!dst_pte) 212663551ae0SDavid Gibson goto nomem; 2127c5c99429SLarry Woodman 2128c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2129c5c99429SLarry Woodman if (dst_pte == src_pte) 2130c5c99429SLarry Woodman continue; 2131c5c99429SLarry Woodman 2132c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 213346478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 21347f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 21351e8f889bSDavid Gibson if (cow) 21367f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 21377f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 213863551ae0SDavid Gibson ptepage = pte_page(entry); 213963551ae0SDavid Gibson get_page(ptepage); 21400fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 214163551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 21421c59827dSHugh Dickins } 21431c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 2144c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 214563551ae0SDavid Gibson } 214663551ae0SDavid Gibson return 0; 214763551ae0SDavid Gibson 214863551ae0SDavid Gibson nomem: 214963551ae0SDavid Gibson return -ENOMEM; 215063551ae0SDavid Gibson } 215163551ae0SDavid Gibson 2152502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 215304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 215463551ae0SDavid Gibson { 215563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 215663551ae0SDavid Gibson unsigned long address; 2157c7546f8fSDavid Gibson pte_t *ptep; 215863551ae0SDavid Gibson pte_t pte; 215963551ae0SDavid Gibson struct page *page; 2160fe1668aeSChen, Kenneth W struct page *tmp; 2161a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2162a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2163a5516438SAndi Kleen 2164c0a499c2SChen, Kenneth W /* 2165c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 2166c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 2167c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 2168c0a499c2SChen, Kenneth W */ 2169fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 217063551ae0SDavid Gibson 217163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2172a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2173a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 217463551ae0SDavid Gibson 2175cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_start(mm, start, end); 2176508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 2177a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2178c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2179c7546f8fSDavid Gibson if (!ptep) 2180c7546f8fSDavid Gibson continue; 2181c7546f8fSDavid Gibson 218239dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 218339dde65cSChen, Kenneth W continue; 218439dde65cSChen, Kenneth W 218504f2cbe3SMel Gorman /* 218604f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 218704f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 218804f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 218904f2cbe3SMel Gorman */ 219004f2cbe3SMel Gorman if (ref_page) { 219104f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 219204f2cbe3SMel Gorman if (huge_pte_none(pte)) 219304f2cbe3SMel Gorman continue; 219404f2cbe3SMel Gorman page = pte_page(pte); 219504f2cbe3SMel Gorman if (page != ref_page) 219604f2cbe3SMel Gorman continue; 219704f2cbe3SMel Gorman 219804f2cbe3SMel Gorman /* 219904f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 220004f2cbe3SMel Gorman * future faults in this VMA will fail rather than 220104f2cbe3SMel Gorman * looking like data was lost 220204f2cbe3SMel Gorman */ 220304f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 220404f2cbe3SMel Gorman } 220504f2cbe3SMel Gorman 2206c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 22077f2e9525SGerald Schaefer if (huge_pte_none(pte)) 220863551ae0SDavid Gibson continue; 2209c7546f8fSDavid Gibson 221063551ae0SDavid Gibson page = pte_page(pte); 22116649a386SKen Chen if (pte_dirty(pte)) 22126649a386SKen Chen set_page_dirty(page); 2213fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 221463551ae0SDavid Gibson } 22151da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 2216508034a3SHugh Dickins flush_tlb_range(vma, start, end); 2217cddb8a5cSAndrea Arcangeli mmu_notifier_invalidate_range_end(mm, start, end); 2218fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 22190fe6e20bSNaoya Horiguchi page_remove_rmap(page); 2220fe1668aeSChen, Kenneth W list_del(&page->lru); 2221fe1668aeSChen, Kenneth W put_page(page); 2222fe1668aeSChen, Kenneth W } 22231da177e4SLinus Torvalds } 222463551ae0SDavid Gibson 2225502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 222604f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2227502717f4SChen, Kenneth W { 2228502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 222904f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 2230502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2231502717f4SChen, Kenneth W } 2232502717f4SChen, Kenneth W 223304f2cbe3SMel Gorman /* 223404f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 223504f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 223604f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 223704f2cbe3SMel Gorman * same region. 223804f2cbe3SMel Gorman */ 22392a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 22402a4b3dedSHarvey Harrison struct page *page, unsigned long address) 224104f2cbe3SMel Gorman { 22427526674dSAdam Litke struct hstate *h = hstate_vma(vma); 224304f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 224404f2cbe3SMel Gorman struct address_space *mapping; 224504f2cbe3SMel Gorman struct prio_tree_iter iter; 224604f2cbe3SMel Gorman pgoff_t pgoff; 224704f2cbe3SMel Gorman 224804f2cbe3SMel Gorman /* 224904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 225004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 225104f2cbe3SMel Gorman */ 22527526674dSAdam Litke address = address & huge_page_mask(h); 225304f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 225404f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 225504f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 225604f2cbe3SMel Gorman 22574eb2b1dcSMel Gorman /* 22584eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 22594eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 22604eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 22614eb2b1dcSMel Gorman */ 22624eb2b1dcSMel Gorman spin_lock(&mapping->i_mmap_lock); 226304f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 226404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 226504f2cbe3SMel Gorman if (iter_vma == vma) 226604f2cbe3SMel Gorman continue; 226704f2cbe3SMel Gorman 226804f2cbe3SMel Gorman /* 226904f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 227004f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 227104f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 227204f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 227304f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 227404f2cbe3SMel Gorman */ 227504f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 22764eb2b1dcSMel Gorman __unmap_hugepage_range(iter_vma, 22777526674dSAdam Litke address, address + huge_page_size(h), 227804f2cbe3SMel Gorman page); 227904f2cbe3SMel Gorman } 22804eb2b1dcSMel Gorman spin_unlock(&mapping->i_mmap_lock); 228104f2cbe3SMel Gorman 228204f2cbe3SMel Gorman return 1; 228304f2cbe3SMel Gorman } 228404f2cbe3SMel Gorman 22850fe6e20bSNaoya Horiguchi /* 22860fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 22870fe6e20bSNaoya Horiguchi */ 22881e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 228904f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 229004f2cbe3SMel Gorman struct page *pagecache_page) 22911e8f889bSDavid Gibson { 2292a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 22931e8f889bSDavid Gibson struct page *old_page, *new_page; 229479ac6ba4SDavid Gibson int avoidcopy; 229504f2cbe3SMel Gorman int outside_reserve = 0; 22961e8f889bSDavid Gibson 22971e8f889bSDavid Gibson old_page = pte_page(pte); 22981e8f889bSDavid Gibson 229904f2cbe3SMel Gorman retry_avoidcopy: 23001e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 23011e8f889bSDavid Gibson * and just make the page writable */ 23020fe6e20bSNaoya Horiguchi avoidcopy = (page_mapcount(old_page) == 1); 23031e8f889bSDavid Gibson if (avoidcopy) { 23040fe6e20bSNaoya Horiguchi if (!trylock_page(old_page)) 23050fe6e20bSNaoya Horiguchi if (PageAnon(old_page)) 23060fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 23071e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 230883c54070SNick Piggin return 0; 23091e8f889bSDavid Gibson } 23101e8f889bSDavid Gibson 231104f2cbe3SMel Gorman /* 231204f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 231304f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 231404f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 231504f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 231604f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 231704f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 231804f2cbe3SMel Gorman * of the full address range. 231904f2cbe3SMel Gorman */ 2320f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE) && 232104f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 232204f2cbe3SMel Gorman old_page != pagecache_page) 232304f2cbe3SMel Gorman outside_reserve = 1; 232404f2cbe3SMel Gorman 23251e8f889bSDavid Gibson page_cache_get(old_page); 2326b76c8cfbSLarry Woodman 2327b76c8cfbSLarry Woodman /* Drop page_table_lock as buddy allocator may be called */ 2328b76c8cfbSLarry Woodman spin_unlock(&mm->page_table_lock); 232904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 23301e8f889bSDavid Gibson 23312fc39cecSAdam Litke if (IS_ERR(new_page)) { 23321e8f889bSDavid Gibson page_cache_release(old_page); 233304f2cbe3SMel Gorman 233404f2cbe3SMel Gorman /* 233504f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 233604f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 233704f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 233804f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 233904f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 234004f2cbe3SMel Gorman */ 234104f2cbe3SMel Gorman if (outside_reserve) { 234204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 234304f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 234404f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 234504f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2346b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 234704f2cbe3SMel Gorman goto retry_avoidcopy; 234804f2cbe3SMel Gorman } 234904f2cbe3SMel Gorman WARN_ON_ONCE(1); 235004f2cbe3SMel Gorman } 235104f2cbe3SMel Gorman 2352b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2353b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 23542fc39cecSAdam Litke return -PTR_ERR(new_page); 23551e8f889bSDavid Gibson } 23561e8f889bSDavid Gibson 23570fe6e20bSNaoya Horiguchi /* 23580fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 23590fe6e20bSNaoya Horiguchi * anon_vma prepared. 23600fe6e20bSNaoya Horiguchi */ 23610fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) 23620fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 23630fe6e20bSNaoya Horiguchi 23649de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 23650ed361deSNick Piggin __SetPageUptodate(new_page); 23661e8f889bSDavid Gibson 2367b76c8cfbSLarry Woodman /* 2368b76c8cfbSLarry Woodman * Retake the page_table_lock to check for racing updates 2369b76c8cfbSLarry Woodman * before the page tables are altered 2370b76c8cfbSLarry Woodman */ 2371b76c8cfbSLarry Woodman spin_lock(&mm->page_table_lock); 2372a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 23737f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 23741e8f889bSDavid Gibson /* Break COW */ 23758fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 23761e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 23771e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 23780fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 23790fe6e20bSNaoya Horiguchi hugepage_add_anon_rmap(new_page, vma, address); 23801e8f889bSDavid Gibson /* Make the old page be freed below */ 23811e8f889bSDavid Gibson new_page = old_page; 23821e8f889bSDavid Gibson } 23831e8f889bSDavid Gibson page_cache_release(new_page); 23841e8f889bSDavid Gibson page_cache_release(old_page); 238583c54070SNick Piggin return 0; 23861e8f889bSDavid Gibson } 23871e8f889bSDavid Gibson 238804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2389a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2390a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 239104f2cbe3SMel Gorman { 239204f2cbe3SMel Gorman struct address_space *mapping; 2393e7c4b0bfSAndy Whitcroft pgoff_t idx; 239404f2cbe3SMel Gorman 239504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2396a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 239704f2cbe3SMel Gorman 239804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 239904f2cbe3SMel Gorman } 240004f2cbe3SMel Gorman 24013ae77f43SHugh Dickins /* 24023ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 24033ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 24043ae77f43SHugh Dickins */ 24053ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 24062a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 24072a15efc9SHugh Dickins { 24082a15efc9SHugh Dickins struct address_space *mapping; 24092a15efc9SHugh Dickins pgoff_t idx; 24102a15efc9SHugh Dickins struct page *page; 24112a15efc9SHugh Dickins 24122a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 24132a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 24142a15efc9SHugh Dickins 24152a15efc9SHugh Dickins page = find_get_page(mapping, idx); 24162a15efc9SHugh Dickins if (page) 24172a15efc9SHugh Dickins put_page(page); 24182a15efc9SHugh Dickins return page != NULL; 24192a15efc9SHugh Dickins } 24202a15efc9SHugh Dickins 2421a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2422788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2423ac9b9c66SHugh Dickins { 2424a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2425ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2426e7c4b0bfSAndy Whitcroft pgoff_t idx; 24274c887265SAdam Litke unsigned long size; 24284c887265SAdam Litke struct page *page; 24294c887265SAdam Litke struct address_space *mapping; 24301e8f889bSDavid Gibson pte_t new_pte; 24314c887265SAdam Litke 243204f2cbe3SMel Gorman /* 243304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 243404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 243504f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 243604f2cbe3SMel Gorman */ 243704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 243804f2cbe3SMel Gorman printk(KERN_WARNING 243904f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 244004f2cbe3SMel Gorman current->pid); 244104f2cbe3SMel Gorman return ret; 244204f2cbe3SMel Gorman } 244304f2cbe3SMel Gorman 24444c887265SAdam Litke mapping = vma->vm_file->f_mapping; 2445a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 24464c887265SAdam Litke 24474c887265SAdam Litke /* 24484c887265SAdam Litke * Use page lock to guard against racing truncation 24494c887265SAdam Litke * before we get page_table_lock. 24504c887265SAdam Litke */ 24516bda666aSChristoph Lameter retry: 24526bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 24536bda666aSChristoph Lameter if (!page) { 2454a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2455ebed4bfcSHugh Dickins if (idx >= size) 2456ebed4bfcSHugh Dickins goto out; 245704f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 24582fc39cecSAdam Litke if (IS_ERR(page)) { 24592fc39cecSAdam Litke ret = -PTR_ERR(page); 24606bda666aSChristoph Lameter goto out; 24616bda666aSChristoph Lameter } 2462a5516438SAndi Kleen clear_huge_page(page, address, huge_page_size(h)); 24630ed361deSNick Piggin __SetPageUptodate(page); 2464ac9b9c66SHugh Dickins 2465f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 24666bda666aSChristoph Lameter int err; 246745c682a6SKen Chen struct inode *inode = mapping->host; 24686bda666aSChristoph Lameter 24696bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 24706bda666aSChristoph Lameter if (err) { 24716bda666aSChristoph Lameter put_page(page); 24726bda666aSChristoph Lameter if (err == -EEXIST) 24736bda666aSChristoph Lameter goto retry; 24746bda666aSChristoph Lameter goto out; 24756bda666aSChristoph Lameter } 247645c682a6SKen Chen 247745c682a6SKen Chen spin_lock(&inode->i_lock); 2478a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 247945c682a6SKen Chen spin_unlock(&inode->i_lock); 24800fe6e20bSNaoya Horiguchi page_dup_rmap(page); 248123be7468SMel Gorman } else { 24826bda666aSChristoph Lameter lock_page(page); 24830fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 24840fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 24850fe6e20bSNaoya Horiguchi goto backout_unlocked; 248623be7468SMel Gorman } 24870fe6e20bSNaoya Horiguchi hugepage_add_new_anon_rmap(page, vma, address); 24880fe6e20bSNaoya Horiguchi } 24890fe6e20bSNaoya Horiguchi } else { 24900fe6e20bSNaoya Horiguchi page_dup_rmap(page); 24916bda666aSChristoph Lameter } 24921e8f889bSDavid Gibson 249357303d80SAndy Whitcroft /* 249457303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 249557303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 249657303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 249757303d80SAndy Whitcroft * the spinlock. 249857303d80SAndy Whitcroft */ 2499788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 25002b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 25012b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 25022b26736cSAndy Whitcroft goto backout_unlocked; 25032b26736cSAndy Whitcroft } 250457303d80SAndy Whitcroft 2505ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 2506a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 25074c887265SAdam Litke if (idx >= size) 25084c887265SAdam Litke goto backout; 25094c887265SAdam Litke 251083c54070SNick Piggin ret = 0; 25117f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 25124c887265SAdam Litke goto backout; 25134c887265SAdam Litke 25141e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 25151e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 25161e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 25171e8f889bSDavid Gibson 2518788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 25191e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 252004f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 25211e8f889bSDavid Gibson } 25221e8f889bSDavid Gibson 2523ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 25244c887265SAdam Litke unlock_page(page); 25254c887265SAdam Litke out: 2526ac9b9c66SHugh Dickins return ret; 25274c887265SAdam Litke 25284c887265SAdam Litke backout: 25294c887265SAdam Litke spin_unlock(&mm->page_table_lock); 25302b26736cSAndy Whitcroft backout_unlocked: 25314c887265SAdam Litke unlock_page(page); 25324c887265SAdam Litke put_page(page); 25334c887265SAdam Litke goto out; 2534ac9b9c66SHugh Dickins } 2535ac9b9c66SHugh Dickins 253686e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2537788c7df4SHugh Dickins unsigned long address, unsigned int flags) 253886e5216fSAdam Litke { 253986e5216fSAdam Litke pte_t *ptep; 254086e5216fSAdam Litke pte_t entry; 25411e8f889bSDavid Gibson int ret; 25420fe6e20bSNaoya Horiguchi struct page *page = NULL; 254357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 25443935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2545a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 254686e5216fSAdam Litke 2547a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 254886e5216fSAdam Litke if (!ptep) 254986e5216fSAdam Litke return VM_FAULT_OOM; 255086e5216fSAdam Litke 25513935baa9SDavid Gibson /* 25523935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 25533935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 25543935baa9SDavid Gibson * the same page in the page cache. 25553935baa9SDavid Gibson */ 25563935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 25577f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 25587f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 2559788c7df4SHugh Dickins ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2560b4d1d99fSDavid Gibson goto out_mutex; 25613935baa9SDavid Gibson } 256286e5216fSAdam Litke 256383c54070SNick Piggin ret = 0; 25641e8f889bSDavid Gibson 256557303d80SAndy Whitcroft /* 256657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 256757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 256857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 256957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 257057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 257157303d80SAndy Whitcroft * consumed. 257257303d80SAndy Whitcroft */ 2573788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 25742b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 25752b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 2576b4d1d99fSDavid Gibson goto out_mutex; 25772b26736cSAndy Whitcroft } 257857303d80SAndy Whitcroft 2579f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 258057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 258157303d80SAndy Whitcroft vma, address); 258257303d80SAndy Whitcroft } 258357303d80SAndy Whitcroft 25840fe6e20bSNaoya Horiguchi if (!pagecache_page) { 25850fe6e20bSNaoya Horiguchi page = pte_page(entry); 25860fe6e20bSNaoya Horiguchi lock_page(page); 25870fe6e20bSNaoya Horiguchi } 25880fe6e20bSNaoya Horiguchi 25891e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 25901e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 2591b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2592b4d1d99fSDavid Gibson goto out_page_table_lock; 2593b4d1d99fSDavid Gibson 2594b4d1d99fSDavid Gibson 2595788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 2596b4d1d99fSDavid Gibson if (!pte_write(entry)) { 259757303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 259857303d80SAndy Whitcroft pagecache_page); 2599b4d1d99fSDavid Gibson goto out_page_table_lock; 2600b4d1d99fSDavid Gibson } 2601b4d1d99fSDavid Gibson entry = pte_mkdirty(entry); 2602b4d1d99fSDavid Gibson } 2603b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 2604788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2605788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 26064b3073e1SRussell King update_mmu_cache(vma, address, ptep); 2607b4d1d99fSDavid Gibson 2608b4d1d99fSDavid Gibson out_page_table_lock: 26091e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 261057303d80SAndy Whitcroft 261157303d80SAndy Whitcroft if (pagecache_page) { 261257303d80SAndy Whitcroft unlock_page(pagecache_page); 261357303d80SAndy Whitcroft put_page(pagecache_page); 26140fe6e20bSNaoya Horiguchi } else { 26150fe6e20bSNaoya Horiguchi unlock_page(page); 261657303d80SAndy Whitcroft } 261757303d80SAndy Whitcroft 2618b4d1d99fSDavid Gibson out_mutex: 26193935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 26201e8f889bSDavid Gibson 26211e8f889bSDavid Gibson return ret; 262286e5216fSAdam Litke } 262386e5216fSAdam Litke 2624ceb86879SAndi Kleen /* Can be overriden by architectures */ 2625ceb86879SAndi Kleen __attribute__((weak)) struct page * 2626ceb86879SAndi Kleen follow_huge_pud(struct mm_struct *mm, unsigned long address, 2627ceb86879SAndi Kleen pud_t *pud, int write) 2628ceb86879SAndi Kleen { 2629ceb86879SAndi Kleen BUG(); 2630ceb86879SAndi Kleen return NULL; 2631ceb86879SAndi Kleen } 2632ceb86879SAndi Kleen 263363551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 263463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 26355b23dbe8SAdam Litke unsigned long *position, int *length, int i, 26362a15efc9SHugh Dickins unsigned int flags) 263763551ae0SDavid Gibson { 2638d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 2639d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 264063551ae0SDavid Gibson int remainder = *length; 2641a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 264263551ae0SDavid Gibson 26431c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 264463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 264563551ae0SDavid Gibson pte_t *pte; 26462a15efc9SHugh Dickins int absent; 264763551ae0SDavid Gibson struct page *page; 264863551ae0SDavid Gibson 26494c887265SAdam Litke /* 26504c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 26512a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 26524c887265SAdam Litke * first, for the page indexing below to work. 26534c887265SAdam Litke */ 2654a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 26552a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 265663551ae0SDavid Gibson 26572a15efc9SHugh Dickins /* 26582a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 26593ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 26603ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 26613ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 26623ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 26632a15efc9SHugh Dickins */ 26643ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 26653ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 26662a15efc9SHugh Dickins remainder = 0; 26672a15efc9SHugh Dickins break; 26682a15efc9SHugh Dickins } 26692a15efc9SHugh Dickins 26702a15efc9SHugh Dickins if (absent || 26712a15efc9SHugh Dickins ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 26724c887265SAdam Litke int ret; 26734c887265SAdam Litke 26744c887265SAdam Litke spin_unlock(&mm->page_table_lock); 26752a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 26762a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 26774c887265SAdam Litke spin_lock(&mm->page_table_lock); 2678a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 26794c887265SAdam Litke continue; 26804c887265SAdam Litke 26811c59827dSHugh Dickins remainder = 0; 26821c59827dSHugh Dickins break; 26831c59827dSHugh Dickins } 268463551ae0SDavid Gibson 2685a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 26867f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 2687d5d4b0aaSChen, Kenneth W same_page: 2688d6692183SChen, Kenneth W if (pages) { 268969d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 26904b2e38adSKOSAKI Motohiro get_page(pages[i]); 2691d6692183SChen, Kenneth W } 269263551ae0SDavid Gibson 269363551ae0SDavid Gibson if (vmas) 269463551ae0SDavid Gibson vmas[i] = vma; 269563551ae0SDavid Gibson 269663551ae0SDavid Gibson vaddr += PAGE_SIZE; 2697d5d4b0aaSChen, Kenneth W ++pfn_offset; 269863551ae0SDavid Gibson --remainder; 269963551ae0SDavid Gibson ++i; 2700d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 2701a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 2702d5d4b0aaSChen, Kenneth W /* 2703d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 2704d5d4b0aaSChen, Kenneth W * of this compound page. 2705d5d4b0aaSChen, Kenneth W */ 2706d5d4b0aaSChen, Kenneth W goto same_page; 2707d5d4b0aaSChen, Kenneth W } 270863551ae0SDavid Gibson } 27091c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 271063551ae0SDavid Gibson *length = remainder; 271163551ae0SDavid Gibson *position = vaddr; 271263551ae0SDavid Gibson 27132a15efc9SHugh Dickins return i ? i : -EFAULT; 271463551ae0SDavid Gibson } 27158f860591SZhang, Yanmin 27168f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 27178f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 27188f860591SZhang, Yanmin { 27198f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 27208f860591SZhang, Yanmin unsigned long start = address; 27218f860591SZhang, Yanmin pte_t *ptep; 27228f860591SZhang, Yanmin pte_t pte; 2723a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 27248f860591SZhang, Yanmin 27258f860591SZhang, Yanmin BUG_ON(address >= end); 27268f860591SZhang, Yanmin flush_cache_range(vma, address, end); 27278f860591SZhang, Yanmin 272839dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 27298f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 2730a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 27318f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 27328f860591SZhang, Yanmin if (!ptep) 27338f860591SZhang, Yanmin continue; 273439dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 273539dde65cSChen, Kenneth W continue; 27367f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 27378f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 27388f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 27398f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 27408f860591SZhang, Yanmin } 27418f860591SZhang, Yanmin } 27428f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 274339dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 27448f860591SZhang, Yanmin 27458f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 27468f860591SZhang, Yanmin } 27478f860591SZhang, Yanmin 2748a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 2749a1e78772SMel Gorman long from, long to, 27505a6fe125SMel Gorman struct vm_area_struct *vma, 27515a6fe125SMel Gorman int acctflag) 2752e4e574b7SAdam Litke { 275317c9d12eSMel Gorman long ret, chg; 2754a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2755e4e574b7SAdam Litke 2756a1e78772SMel Gorman /* 275717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 275817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 275917c9d12eSMel Gorman * and filesystem quota without using reserves 276017c9d12eSMel Gorman */ 276117c9d12eSMel Gorman if (acctflag & VM_NORESERVE) 276217c9d12eSMel Gorman return 0; 276317c9d12eSMel Gorman 276417c9d12eSMel Gorman /* 2765a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 2766a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 2767a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 2768a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 2769a1e78772SMel Gorman */ 2770f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 2771e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 27725a6fe125SMel Gorman else { 27735a6fe125SMel Gorman struct resv_map *resv_map = resv_map_alloc(); 27745a6fe125SMel Gorman if (!resv_map) 27755a6fe125SMel Gorman return -ENOMEM; 27765a6fe125SMel Gorman 277717c9d12eSMel Gorman chg = to - from; 277817c9d12eSMel Gorman 27795a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 27805a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 27815a6fe125SMel Gorman } 27825a6fe125SMel Gorman 278317c9d12eSMel Gorman if (chg < 0) 278417c9d12eSMel Gorman return chg; 278517c9d12eSMel Gorman 278617c9d12eSMel Gorman /* There must be enough filesystem quota for the mapping */ 278717c9d12eSMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 278817c9d12eSMel Gorman return -ENOSPC; 278917c9d12eSMel Gorman 279017c9d12eSMel Gorman /* 279117c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 279217c9d12eSMel Gorman * Hand back the quota if there are not 279317c9d12eSMel Gorman */ 279417c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 279517c9d12eSMel Gorman if (ret < 0) { 279617c9d12eSMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 279717c9d12eSMel Gorman return ret; 279817c9d12eSMel Gorman } 279917c9d12eSMel Gorman 280017c9d12eSMel Gorman /* 280117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 280217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 280317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 280417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 280517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 280617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 280717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 280817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 280917c9d12eSMel Gorman * else has to be done for private mappings here 281017c9d12eSMel Gorman */ 2811f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 281217c9d12eSMel Gorman region_add(&inode->i_mapping->private_list, from, to); 2813a43a8c39SChen, Kenneth W return 0; 2814a43a8c39SChen, Kenneth W } 2815a43a8c39SChen, Kenneth W 2816a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2817a43a8c39SChen, Kenneth W { 2818a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 2819a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 282045c682a6SKen Chen 282145c682a6SKen Chen spin_lock(&inode->i_lock); 2822e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 282345c682a6SKen Chen spin_unlock(&inode->i_lock); 282445c682a6SKen Chen 282590d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2826a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 2827a43a8c39SChen, Kenneth W } 2828