11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 31da177e4SLinus Torvalds * (C) William Irwin, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/gfp.h> 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/module.h> 91da177e4SLinus Torvalds #include <linux/mm.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 121da177e4SLinus Torvalds #include <linux/nodemask.h> 1363551ae0SDavid Gibson #include <linux/pagemap.h> 145da7ca86SChristoph Lameter #include <linux/mempolicy.h> 15aea47ff3SChristoph Lameter #include <linux/cpuset.h> 163935baa9SDavid Gibson #include <linux/mutex.h> 175da7ca86SChristoph Lameter 1863551ae0SDavid Gibson #include <asm/page.h> 1963551ae0SDavid Gibson #include <asm/pgtable.h> 2063551ae0SDavid Gibson 2163551ae0SDavid Gibson #include <linux/hugetlb.h> 227835e98bSNick Piggin #include "internal.h" 231da177e4SLinus Torvalds 241da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 25a43a8c39SChen, Kenneth W static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages; 267893d1d5SAdam Litke static unsigned long surplus_huge_pages; 27064d9efeSNishanth Aravamudan static unsigned long nr_overcommit_huge_pages; 281da177e4SLinus Torvalds unsigned long max_huge_pages; 29064d9efeSNishanth Aravamudan unsigned long sysctl_overcommit_huge_pages; 301da177e4SLinus Torvalds static struct list_head hugepage_freelists[MAX_NUMNODES]; 311da177e4SLinus Torvalds static unsigned int nr_huge_pages_node[MAX_NUMNODES]; 321da177e4SLinus Torvalds static unsigned int free_huge_pages_node[MAX_NUMNODES]; 337893d1d5SAdam Litke static unsigned int surplus_huge_pages_node[MAX_NUMNODES]; 34396faf03SMel Gorman static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 35396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 3663b4613cSNishanth Aravamudan static int hugetlb_next_nid; 37396faf03SMel Gorman 383935baa9SDavid Gibson /* 393935baa9SDavid Gibson * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 403935baa9SDavid Gibson */ 413935baa9SDavid Gibson static DEFINE_SPINLOCK(hugetlb_lock); 420bd0f9fbSEric Paris 43e7c4b0bfSAndy Whitcroft /* 4496822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 4596822904SAndy Whitcroft * across the pages in a mapping. 4696822904SAndy Whitcroft */ 4796822904SAndy Whitcroft struct file_region { 4896822904SAndy Whitcroft struct list_head link; 4996822904SAndy Whitcroft long from; 5096822904SAndy Whitcroft long to; 5196822904SAndy Whitcroft }; 5296822904SAndy Whitcroft 5396822904SAndy Whitcroft static long region_add(struct list_head *head, long f, long t) 5496822904SAndy Whitcroft { 5596822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 5696822904SAndy Whitcroft 5796822904SAndy Whitcroft /* Locate the region we are either in or before. */ 5896822904SAndy Whitcroft list_for_each_entry(rg, head, link) 5996822904SAndy Whitcroft if (f <= rg->to) 6096822904SAndy Whitcroft break; 6196822904SAndy Whitcroft 6296822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 6396822904SAndy Whitcroft if (f > rg->from) 6496822904SAndy Whitcroft f = rg->from; 6596822904SAndy Whitcroft 6696822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 6796822904SAndy Whitcroft nrg = rg; 6896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 6996822904SAndy Whitcroft if (&rg->link == head) 7096822904SAndy Whitcroft break; 7196822904SAndy Whitcroft if (rg->from > t) 7296822904SAndy Whitcroft break; 7396822904SAndy Whitcroft 7496822904SAndy Whitcroft /* If this area reaches higher then extend our area to 7596822904SAndy Whitcroft * include it completely. If this is not the first area 7696822904SAndy Whitcroft * which we intend to reuse, free it. */ 7796822904SAndy Whitcroft if (rg->to > t) 7896822904SAndy Whitcroft t = rg->to; 7996822904SAndy Whitcroft if (rg != nrg) { 8096822904SAndy Whitcroft list_del(&rg->link); 8196822904SAndy Whitcroft kfree(rg); 8296822904SAndy Whitcroft } 8396822904SAndy Whitcroft } 8496822904SAndy Whitcroft nrg->from = f; 8596822904SAndy Whitcroft nrg->to = t; 8696822904SAndy Whitcroft return 0; 8796822904SAndy Whitcroft } 8896822904SAndy Whitcroft 8996822904SAndy Whitcroft static long region_chg(struct list_head *head, long f, long t) 9096822904SAndy Whitcroft { 9196822904SAndy Whitcroft struct file_region *rg, *nrg; 9296822904SAndy Whitcroft long chg = 0; 9396822904SAndy Whitcroft 9496822904SAndy Whitcroft /* Locate the region we are before or in. */ 9596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 9696822904SAndy Whitcroft if (f <= rg->to) 9796822904SAndy Whitcroft break; 9896822904SAndy Whitcroft 9996822904SAndy Whitcroft /* If we are below the current region then a new region is required. 10096822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 10196822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 10296822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 10396822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 10496822904SAndy Whitcroft if (!nrg) 10596822904SAndy Whitcroft return -ENOMEM; 10696822904SAndy Whitcroft nrg->from = f; 10796822904SAndy Whitcroft nrg->to = f; 10896822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 10996822904SAndy Whitcroft list_add(&nrg->link, rg->link.prev); 11096822904SAndy Whitcroft 11196822904SAndy Whitcroft return t - f; 11296822904SAndy Whitcroft } 11396822904SAndy Whitcroft 11496822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 11596822904SAndy Whitcroft if (f > rg->from) 11696822904SAndy Whitcroft f = rg->from; 11796822904SAndy Whitcroft chg = t - f; 11896822904SAndy Whitcroft 11996822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 12096822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 12196822904SAndy Whitcroft if (&rg->link == head) 12296822904SAndy Whitcroft break; 12396822904SAndy Whitcroft if (rg->from > t) 12496822904SAndy Whitcroft return chg; 12596822904SAndy Whitcroft 12696822904SAndy Whitcroft /* We overlap with this area, if it extends futher than 12796822904SAndy Whitcroft * us then we must extend ourselves. Account for its 12896822904SAndy Whitcroft * existing reservation. */ 12996822904SAndy Whitcroft if (rg->to > t) { 13096822904SAndy Whitcroft chg += rg->to - t; 13196822904SAndy Whitcroft t = rg->to; 13296822904SAndy Whitcroft } 13396822904SAndy Whitcroft chg -= rg->to - rg->from; 13496822904SAndy Whitcroft } 13596822904SAndy Whitcroft return chg; 13696822904SAndy Whitcroft } 13796822904SAndy Whitcroft 13896822904SAndy Whitcroft static long region_truncate(struct list_head *head, long end) 13996822904SAndy Whitcroft { 14096822904SAndy Whitcroft struct file_region *rg, *trg; 14196822904SAndy Whitcroft long chg = 0; 14296822904SAndy Whitcroft 14396822904SAndy Whitcroft /* Locate the region we are either in or before. */ 14496822904SAndy Whitcroft list_for_each_entry(rg, head, link) 14596822904SAndy Whitcroft if (end <= rg->to) 14696822904SAndy Whitcroft break; 14796822904SAndy Whitcroft if (&rg->link == head) 14896822904SAndy Whitcroft return 0; 14996822904SAndy Whitcroft 15096822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 15196822904SAndy Whitcroft if (end > rg->from) { 15296822904SAndy Whitcroft chg = rg->to - end; 15396822904SAndy Whitcroft rg->to = end; 15496822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 15596822904SAndy Whitcroft } 15696822904SAndy Whitcroft 15796822904SAndy Whitcroft /* Drop any remaining regions. */ 15896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 15996822904SAndy Whitcroft if (&rg->link == head) 16096822904SAndy Whitcroft break; 16196822904SAndy Whitcroft chg += rg->to - rg->from; 16296822904SAndy Whitcroft list_del(&rg->link); 16396822904SAndy Whitcroft kfree(rg); 16496822904SAndy Whitcroft } 16596822904SAndy Whitcroft return chg; 16696822904SAndy Whitcroft } 16796822904SAndy Whitcroft 16896822904SAndy Whitcroft /* 169e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 170e7c4b0bfSAndy Whitcroft * the mapping, in base page units. 171e7c4b0bfSAndy Whitcroft */ 172e7c4b0bfSAndy Whitcroft static pgoff_t vma_page_offset(struct vm_area_struct *vma, 173e7c4b0bfSAndy Whitcroft unsigned long address) 174e7c4b0bfSAndy Whitcroft { 175e7c4b0bfSAndy Whitcroft return ((address - vma->vm_start) >> PAGE_SHIFT) + 176e7c4b0bfSAndy Whitcroft (vma->vm_pgoff >> PAGE_SHIFT); 177e7c4b0bfSAndy Whitcroft } 178e7c4b0bfSAndy Whitcroft 179e7c4b0bfSAndy Whitcroft /* 180e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 181e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 182e7c4b0bfSAndy Whitcroft */ 183e7c4b0bfSAndy Whitcroft static pgoff_t vma_pagecache_offset(struct vm_area_struct *vma, 184e7c4b0bfSAndy Whitcroft unsigned long address) 185e7c4b0bfSAndy Whitcroft { 186e7c4b0bfSAndy Whitcroft return ((address - vma->vm_start) >> HPAGE_SHIFT) + 187e7c4b0bfSAndy Whitcroft (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); 188e7c4b0bfSAndy Whitcroft } 189e7c4b0bfSAndy Whitcroft 19004f2cbe3SMel Gorman #define HPAGE_RESV_OWNER (1UL << (BITS_PER_LONG - 1)) 19104f2cbe3SMel Gorman #define HPAGE_RESV_UNMAPPED (1UL << (BITS_PER_LONG - 2)) 19204f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 193a1e78772SMel Gorman /* 194a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 195a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 196a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 197a1e78772SMel Gorman * 198a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 199a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 200a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 201a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 202a1e78772SMel Gorman */ 203e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 204e7c4b0bfSAndy Whitcroft { 205e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 206e7c4b0bfSAndy Whitcroft } 207e7c4b0bfSAndy Whitcroft 208e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 209e7c4b0bfSAndy Whitcroft unsigned long value) 210e7c4b0bfSAndy Whitcroft { 211e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 212e7c4b0bfSAndy Whitcroft } 213e7c4b0bfSAndy Whitcroft 214a1e78772SMel Gorman static unsigned long vma_resv_huge_pages(struct vm_area_struct *vma) 215a1e78772SMel Gorman { 216a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 217a1e78772SMel Gorman if (!(vma->vm_flags & VM_SHARED)) 218e7c4b0bfSAndy Whitcroft return get_vma_private_data(vma) & ~HPAGE_RESV_MASK; 219a1e78772SMel Gorman return 0; 220a1e78772SMel Gorman } 221a1e78772SMel Gorman 222a1e78772SMel Gorman static void set_vma_resv_huge_pages(struct vm_area_struct *vma, 223a1e78772SMel Gorman unsigned long reserve) 224a1e78772SMel Gorman { 225a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 226a1e78772SMel Gorman VM_BUG_ON(vma->vm_flags & VM_SHARED); 227a1e78772SMel Gorman 228e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, 229e7c4b0bfSAndy Whitcroft (get_vma_private_data(vma) & HPAGE_RESV_MASK) | reserve); 23004f2cbe3SMel Gorman } 23104f2cbe3SMel Gorman 23204f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 23304f2cbe3SMel Gorman { 23404f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 235e7c4b0bfSAndy Whitcroft VM_BUG_ON(vma->vm_flags & VM_SHARED); 236e7c4b0bfSAndy Whitcroft 237e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 23804f2cbe3SMel Gorman } 23904f2cbe3SMel Gorman 24004f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 24104f2cbe3SMel Gorman { 24204f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 243e7c4b0bfSAndy Whitcroft 244e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 245a1e78772SMel Gorman } 246a1e78772SMel Gorman 247a1e78772SMel Gorman /* Decrement the reserved pages in the hugepage pool by one */ 248a1e78772SMel Gorman static void decrement_hugepage_resv_vma(struct vm_area_struct *vma) 249a1e78772SMel Gorman { 250a1e78772SMel Gorman if (vma->vm_flags & VM_SHARED) { 251a1e78772SMel Gorman /* Shared mappings always use reserves */ 252a1e78772SMel Gorman resv_huge_pages--; 253a1e78772SMel Gorman } else { 254a1e78772SMel Gorman /* 255a1e78772SMel Gorman * Only the process that called mmap() has reserves for 256a1e78772SMel Gorman * private mappings. 257a1e78772SMel Gorman */ 25804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 25904f2cbe3SMel Gorman unsigned long flags, reserve; 260a1e78772SMel Gorman resv_huge_pages--; 26104f2cbe3SMel Gorman flags = (unsigned long)vma->vm_private_data & 26204f2cbe3SMel Gorman HPAGE_RESV_MASK; 263a1e78772SMel Gorman reserve = (unsigned long)vma->vm_private_data - 1; 26404f2cbe3SMel Gorman vma->vm_private_data = (void *)(reserve | flags); 265a1e78772SMel Gorman } 266a1e78772SMel Gorman } 267a1e78772SMel Gorman } 268a1e78772SMel Gorman 26904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 270a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 271a1e78772SMel Gorman { 272a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 273a1e78772SMel Gorman if (!(vma->vm_flags & VM_SHARED)) 274a1e78772SMel Gorman vma->vm_private_data = (void *)0; 275a1e78772SMel Gorman } 276a1e78772SMel Gorman 277a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 278a1e78772SMel Gorman static int vma_has_private_reserves(struct vm_area_struct *vma) 279a1e78772SMel Gorman { 280a1e78772SMel Gorman if (vma->vm_flags & VM_SHARED) 281a1e78772SMel Gorman return 0; 282a1e78772SMel Gorman if (!vma_resv_huge_pages(vma)) 283a1e78772SMel Gorman return 0; 284a1e78772SMel Gorman return 1; 285a1e78772SMel Gorman } 286a1e78772SMel Gorman 28779ac6ba4SDavid Gibson static void clear_huge_page(struct page *page, unsigned long addr) 28879ac6ba4SDavid Gibson { 28979ac6ba4SDavid Gibson int i; 29079ac6ba4SDavid Gibson 29179ac6ba4SDavid Gibson might_sleep(); 29279ac6ba4SDavid Gibson for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) { 29379ac6ba4SDavid Gibson cond_resched(); 294281e0e3bSRalf Baechle clear_user_highpage(page + i, addr + i * PAGE_SIZE); 29579ac6ba4SDavid Gibson } 29679ac6ba4SDavid Gibson } 29779ac6ba4SDavid Gibson 29879ac6ba4SDavid Gibson static void copy_huge_page(struct page *dst, struct page *src, 2999de455b2SAtsushi Nemoto unsigned long addr, struct vm_area_struct *vma) 30079ac6ba4SDavid Gibson { 30179ac6ba4SDavid Gibson int i; 30279ac6ba4SDavid Gibson 30379ac6ba4SDavid Gibson might_sleep(); 30479ac6ba4SDavid Gibson for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) { 30579ac6ba4SDavid Gibson cond_resched(); 3069de455b2SAtsushi Nemoto copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 30779ac6ba4SDavid Gibson } 30879ac6ba4SDavid Gibson } 30979ac6ba4SDavid Gibson 3101da177e4SLinus Torvalds static void enqueue_huge_page(struct page *page) 3111da177e4SLinus Torvalds { 3121da177e4SLinus Torvalds int nid = page_to_nid(page); 3131da177e4SLinus Torvalds list_add(&page->lru, &hugepage_freelists[nid]); 3141da177e4SLinus Torvalds free_huge_pages++; 3151da177e4SLinus Torvalds free_huge_pages_node[nid]++; 3161da177e4SLinus Torvalds } 3171da177e4SLinus Torvalds 318348e1e04SNishanth Aravamudan static struct page *dequeue_huge_page(void) 319348e1e04SNishanth Aravamudan { 320348e1e04SNishanth Aravamudan int nid; 321348e1e04SNishanth Aravamudan struct page *page = NULL; 322348e1e04SNishanth Aravamudan 323348e1e04SNishanth Aravamudan for (nid = 0; nid < MAX_NUMNODES; ++nid) { 324348e1e04SNishanth Aravamudan if (!list_empty(&hugepage_freelists[nid])) { 325348e1e04SNishanth Aravamudan page = list_entry(hugepage_freelists[nid].next, 326348e1e04SNishanth Aravamudan struct page, lru); 327348e1e04SNishanth Aravamudan list_del(&page->lru); 328348e1e04SNishanth Aravamudan free_huge_pages--; 329348e1e04SNishanth Aravamudan free_huge_pages_node[nid]--; 330348e1e04SNishanth Aravamudan break; 331348e1e04SNishanth Aravamudan } 332348e1e04SNishanth Aravamudan } 333348e1e04SNishanth Aravamudan return page; 334348e1e04SNishanth Aravamudan } 335348e1e04SNishanth Aravamudan 336348e1e04SNishanth Aravamudan static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma, 33704f2cbe3SMel Gorman unsigned long address, int avoid_reserve) 3381da177e4SLinus Torvalds { 33931a5c6e4SNishanth Aravamudan int nid; 3401da177e4SLinus Torvalds struct page *page = NULL; 341480eccf9SLee Schermerhorn struct mempolicy *mpol; 34219770b32SMel Gorman nodemask_t *nodemask; 343396faf03SMel Gorman struct zonelist *zonelist = huge_zonelist(vma, address, 34419770b32SMel Gorman htlb_alloc_mask, &mpol, &nodemask); 345dd1a239fSMel Gorman struct zone *zone; 346dd1a239fSMel Gorman struct zoneref *z; 3471da177e4SLinus Torvalds 348a1e78772SMel Gorman /* 349a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 350a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 351a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 352a1e78772SMel Gorman */ 353a1e78772SMel Gorman if (!vma_has_private_reserves(vma) && 354a1e78772SMel Gorman free_huge_pages - resv_huge_pages == 0) 355a1e78772SMel Gorman return NULL; 356a1e78772SMel Gorman 35704f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 35804f2cbe3SMel Gorman if (avoid_reserve && free_huge_pages - resv_huge_pages == 0) 35904f2cbe3SMel Gorman return NULL; 36004f2cbe3SMel Gorman 36119770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 36219770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 36354a6eb5cSMel Gorman nid = zone_to_nid(zone); 36454a6eb5cSMel Gorman if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 3653abf7afdSAndrew Morton !list_empty(&hugepage_freelists[nid])) { 3661da177e4SLinus Torvalds page = list_entry(hugepage_freelists[nid].next, 3671da177e4SLinus Torvalds struct page, lru); 3681da177e4SLinus Torvalds list_del(&page->lru); 3691da177e4SLinus Torvalds free_huge_pages--; 3701da177e4SLinus Torvalds free_huge_pages_node[nid]--; 37104f2cbe3SMel Gorman 37204f2cbe3SMel Gorman if (!avoid_reserve) 373a1e78772SMel Gorman decrement_hugepage_resv_vma(vma); 374a1e78772SMel Gorman 3755ab3ee7bSKen Chen break; 3761da177e4SLinus Torvalds } 3773abf7afdSAndrew Morton } 37852cd3b07SLee Schermerhorn mpol_cond_put(mpol); 3791da177e4SLinus Torvalds return page; 3801da177e4SLinus Torvalds } 3811da177e4SLinus Torvalds 3826af2acb6SAdam Litke static void update_and_free_page(struct page *page) 3836af2acb6SAdam Litke { 3846af2acb6SAdam Litke int i; 3856af2acb6SAdam Litke nr_huge_pages--; 3866af2acb6SAdam Litke nr_huge_pages_node[page_to_nid(page)]--; 3876af2acb6SAdam Litke for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { 3886af2acb6SAdam Litke page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 3896af2acb6SAdam Litke 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 3906af2acb6SAdam Litke 1 << PG_private | 1<< PG_writeback); 3916af2acb6SAdam Litke } 3926af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 3936af2acb6SAdam Litke set_page_refcounted(page); 3947f2e9525SGerald Schaefer arch_release_hugepage(page); 3956af2acb6SAdam Litke __free_pages(page, HUGETLB_PAGE_ORDER); 3966af2acb6SAdam Litke } 3976af2acb6SAdam Litke 39827a85ef1SDavid Gibson static void free_huge_page(struct page *page) 39927a85ef1SDavid Gibson { 4007893d1d5SAdam Litke int nid = page_to_nid(page); 401c79fb75eSAdam Litke struct address_space *mapping; 40227a85ef1SDavid Gibson 403c79fb75eSAdam Litke mapping = (struct address_space *) page_private(page); 404e5df70abSAndy Whitcroft set_page_private(page, 0); 4057893d1d5SAdam Litke BUG_ON(page_count(page)); 40627a85ef1SDavid Gibson INIT_LIST_HEAD(&page->lru); 40727a85ef1SDavid Gibson 40827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 4097893d1d5SAdam Litke if (surplus_huge_pages_node[nid]) { 4107893d1d5SAdam Litke update_and_free_page(page); 4117893d1d5SAdam Litke surplus_huge_pages--; 4127893d1d5SAdam Litke surplus_huge_pages_node[nid]--; 4137893d1d5SAdam Litke } else { 41427a85ef1SDavid Gibson enqueue_huge_page(page); 4157893d1d5SAdam Litke } 41627a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 417c79fb75eSAdam Litke if (mapping) 4189a119c05SAdam Litke hugetlb_put_quota(mapping, 1); 41927a85ef1SDavid Gibson } 42027a85ef1SDavid Gibson 4217893d1d5SAdam Litke /* 4227893d1d5SAdam Litke * Increment or decrement surplus_huge_pages. Keep node-specific counters 4237893d1d5SAdam Litke * balanced by operating on them in a round-robin fashion. 4247893d1d5SAdam Litke * Returns 1 if an adjustment was made. 4257893d1d5SAdam Litke */ 4267893d1d5SAdam Litke static int adjust_pool_surplus(int delta) 4277893d1d5SAdam Litke { 4287893d1d5SAdam Litke static int prev_nid; 4297893d1d5SAdam Litke int nid = prev_nid; 4307893d1d5SAdam Litke int ret = 0; 4317893d1d5SAdam Litke 4327893d1d5SAdam Litke VM_BUG_ON(delta != -1 && delta != 1); 4337893d1d5SAdam Litke do { 4347893d1d5SAdam Litke nid = next_node(nid, node_online_map); 4357893d1d5SAdam Litke if (nid == MAX_NUMNODES) 4367893d1d5SAdam Litke nid = first_node(node_online_map); 4377893d1d5SAdam Litke 4387893d1d5SAdam Litke /* To shrink on this node, there must be a surplus page */ 4397893d1d5SAdam Litke if (delta < 0 && !surplus_huge_pages_node[nid]) 4407893d1d5SAdam Litke continue; 4417893d1d5SAdam Litke /* Surplus cannot exceed the total number of pages */ 4427893d1d5SAdam Litke if (delta > 0 && surplus_huge_pages_node[nid] >= 4437893d1d5SAdam Litke nr_huge_pages_node[nid]) 4447893d1d5SAdam Litke continue; 4457893d1d5SAdam Litke 4467893d1d5SAdam Litke surplus_huge_pages += delta; 4477893d1d5SAdam Litke surplus_huge_pages_node[nid] += delta; 4487893d1d5SAdam Litke ret = 1; 4497893d1d5SAdam Litke break; 4507893d1d5SAdam Litke } while (nid != prev_nid); 4517893d1d5SAdam Litke 4527893d1d5SAdam Litke prev_nid = nid; 4537893d1d5SAdam Litke return ret; 4547893d1d5SAdam Litke } 4557893d1d5SAdam Litke 45663b4613cSNishanth Aravamudan static struct page *alloc_fresh_huge_page_node(int nid) 4571da177e4SLinus Torvalds { 4581da177e4SLinus Torvalds struct page *page; 459f96efd58SJoe Jin 46063b4613cSNishanth Aravamudan page = alloc_pages_node(nid, 461551883aeSNishanth Aravamudan htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 462551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 463f96efd58SJoe Jin HUGETLB_PAGE_ORDER); 4641da177e4SLinus Torvalds if (page) { 4657f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 4667f2e9525SGerald Schaefer __free_pages(page, HUGETLB_PAGE_ORDER); 4677b8ee84dSHarvey Harrison return NULL; 4687f2e9525SGerald Schaefer } 46933f2ef89SAndy Whitcroft set_compound_page_dtor(page, free_huge_page); 4700bd0f9fbSEric Paris spin_lock(&hugetlb_lock); 4711da177e4SLinus Torvalds nr_huge_pages++; 47263b4613cSNishanth Aravamudan nr_huge_pages_node[nid]++; 4730bd0f9fbSEric Paris spin_unlock(&hugetlb_lock); 474a482289dSNick Piggin put_page(page); /* free it into the hugepage allocator */ 4751da177e4SLinus Torvalds } 47663b4613cSNishanth Aravamudan 47763b4613cSNishanth Aravamudan return page; 47863b4613cSNishanth Aravamudan } 47963b4613cSNishanth Aravamudan 48063b4613cSNishanth Aravamudan static int alloc_fresh_huge_page(void) 48163b4613cSNishanth Aravamudan { 48263b4613cSNishanth Aravamudan struct page *page; 48363b4613cSNishanth Aravamudan int start_nid; 48463b4613cSNishanth Aravamudan int next_nid; 48563b4613cSNishanth Aravamudan int ret = 0; 48663b4613cSNishanth Aravamudan 48763b4613cSNishanth Aravamudan start_nid = hugetlb_next_nid; 48863b4613cSNishanth Aravamudan 48963b4613cSNishanth Aravamudan do { 49063b4613cSNishanth Aravamudan page = alloc_fresh_huge_page_node(hugetlb_next_nid); 49163b4613cSNishanth Aravamudan if (page) 49263b4613cSNishanth Aravamudan ret = 1; 49363b4613cSNishanth Aravamudan /* 49463b4613cSNishanth Aravamudan * Use a helper variable to find the next node and then 49563b4613cSNishanth Aravamudan * copy it back to hugetlb_next_nid afterwards: 49663b4613cSNishanth Aravamudan * otherwise there's a window in which a racer might 49763b4613cSNishanth Aravamudan * pass invalid nid MAX_NUMNODES to alloc_pages_node. 49863b4613cSNishanth Aravamudan * But we don't need to use a spin_lock here: it really 49963b4613cSNishanth Aravamudan * doesn't matter if occasionally a racer chooses the 50063b4613cSNishanth Aravamudan * same nid as we do. Move nid forward in the mask even 50163b4613cSNishanth Aravamudan * if we just successfully allocated a hugepage so that 50263b4613cSNishanth Aravamudan * the next caller gets hugepages on the next node. 50363b4613cSNishanth Aravamudan */ 50463b4613cSNishanth Aravamudan next_nid = next_node(hugetlb_next_nid, node_online_map); 50563b4613cSNishanth Aravamudan if (next_nid == MAX_NUMNODES) 50663b4613cSNishanth Aravamudan next_nid = first_node(node_online_map); 50763b4613cSNishanth Aravamudan hugetlb_next_nid = next_nid; 50863b4613cSNishanth Aravamudan } while (!page && hugetlb_next_nid != start_nid); 50963b4613cSNishanth Aravamudan 5103b116300SAdam Litke if (ret) 5113b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC); 5123b116300SAdam Litke else 5133b116300SAdam Litke count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 5143b116300SAdam Litke 51563b4613cSNishanth Aravamudan return ret; 5161da177e4SLinus Torvalds } 5171da177e4SLinus Torvalds 5187893d1d5SAdam Litke static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma, 5197893d1d5SAdam Litke unsigned long address) 5207893d1d5SAdam Litke { 5217893d1d5SAdam Litke struct page *page; 522d1c3fb1fSNishanth Aravamudan unsigned int nid; 5237893d1d5SAdam Litke 524d1c3fb1fSNishanth Aravamudan /* 525d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 526d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 527d1c3fb1fSNishanth Aravamudan * overcommit 528d1c3fb1fSNishanth Aravamudan * 529d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 530d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 531d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 532d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 533d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 534d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 535d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 536d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 537d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 538d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 539d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 540d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 541d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 542d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 543d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 544d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 545d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 546d1c3fb1fSNishanth Aravamudan */ 547d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 548d1c3fb1fSNishanth Aravamudan if (surplus_huge_pages >= nr_overcommit_huge_pages) { 549d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 550d1c3fb1fSNishanth Aravamudan return NULL; 551d1c3fb1fSNishanth Aravamudan } else { 552d1c3fb1fSNishanth Aravamudan nr_huge_pages++; 553d1c3fb1fSNishanth Aravamudan surplus_huge_pages++; 554d1c3fb1fSNishanth Aravamudan } 555d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 556d1c3fb1fSNishanth Aravamudan 557551883aeSNishanth Aravamudan page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 558551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 5597893d1d5SAdam Litke HUGETLB_PAGE_ORDER); 560d1c3fb1fSNishanth Aravamudan 5617893d1d5SAdam Litke spin_lock(&hugetlb_lock); 562d1c3fb1fSNishanth Aravamudan if (page) { 5632668db91SAdam Litke /* 5642668db91SAdam Litke * This page is now managed by the hugetlb allocator and has 5652668db91SAdam Litke * no users -- drop the buddy allocator's reference. 5662668db91SAdam Litke */ 5672668db91SAdam Litke put_page_testzero(page); 5682668db91SAdam Litke VM_BUG_ON(page_count(page)); 569d1c3fb1fSNishanth Aravamudan nid = page_to_nid(page); 570d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 571d1c3fb1fSNishanth Aravamudan /* 572d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 573d1c3fb1fSNishanth Aravamudan */ 574d1c3fb1fSNishanth Aravamudan nr_huge_pages_node[nid]++; 575d1c3fb1fSNishanth Aravamudan surplus_huge_pages_node[nid]++; 5763b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 577d1c3fb1fSNishanth Aravamudan } else { 578d1c3fb1fSNishanth Aravamudan nr_huge_pages--; 579d1c3fb1fSNishanth Aravamudan surplus_huge_pages--; 5803b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 5817893d1d5SAdam Litke } 582d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 5837893d1d5SAdam Litke 5847893d1d5SAdam Litke return page; 5857893d1d5SAdam Litke } 5867893d1d5SAdam Litke 587e4e574b7SAdam Litke /* 588e4e574b7SAdam Litke * Increase the hugetlb pool such that it can accomodate a reservation 589e4e574b7SAdam Litke * of size 'delta'. 590e4e574b7SAdam Litke */ 591e4e574b7SAdam Litke static int gather_surplus_pages(int delta) 592e4e574b7SAdam Litke { 593e4e574b7SAdam Litke struct list_head surplus_list; 594e4e574b7SAdam Litke struct page *page, *tmp; 595e4e574b7SAdam Litke int ret, i; 596e4e574b7SAdam Litke int needed, allocated; 597e4e574b7SAdam Litke 598e4e574b7SAdam Litke needed = (resv_huge_pages + delta) - free_huge_pages; 599ac09b3a1SAdam Litke if (needed <= 0) { 600ac09b3a1SAdam Litke resv_huge_pages += delta; 601e4e574b7SAdam Litke return 0; 602ac09b3a1SAdam Litke } 603e4e574b7SAdam Litke 604e4e574b7SAdam Litke allocated = 0; 605e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 606e4e574b7SAdam Litke 607e4e574b7SAdam Litke ret = -ENOMEM; 608e4e574b7SAdam Litke retry: 609e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 610e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 611e4e574b7SAdam Litke page = alloc_buddy_huge_page(NULL, 0); 612e4e574b7SAdam Litke if (!page) { 613e4e574b7SAdam Litke /* 614e4e574b7SAdam Litke * We were not able to allocate enough pages to 615e4e574b7SAdam Litke * satisfy the entire reservation so we free what 616e4e574b7SAdam Litke * we've allocated so far. 617e4e574b7SAdam Litke */ 618e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 619e4e574b7SAdam Litke needed = 0; 620e4e574b7SAdam Litke goto free; 621e4e574b7SAdam Litke } 622e4e574b7SAdam Litke 623e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 624e4e574b7SAdam Litke } 625e4e574b7SAdam Litke allocated += needed; 626e4e574b7SAdam Litke 627e4e574b7SAdam Litke /* 628e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 629e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 630e4e574b7SAdam Litke */ 631e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 632e4e574b7SAdam Litke needed = (resv_huge_pages + delta) - (free_huge_pages + allocated); 633e4e574b7SAdam Litke if (needed > 0) 634e4e574b7SAdam Litke goto retry; 635e4e574b7SAdam Litke 636e4e574b7SAdam Litke /* 637e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 638e4e574b7SAdam Litke * needed to accomodate the reservation. Add the appropriate number 639e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 640ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 641ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 642ac09b3a1SAdam Litke * before they are reserved. 643e4e574b7SAdam Litke */ 644e4e574b7SAdam Litke needed += allocated; 645ac09b3a1SAdam Litke resv_huge_pages += delta; 646e4e574b7SAdam Litke ret = 0; 647e4e574b7SAdam Litke free: 64819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 64919fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 65019fc3f0aSAdam Litke if ((--needed) < 0) 65119fc3f0aSAdam Litke break; 65219fc3f0aSAdam Litke list_del(&page->lru); 65319fc3f0aSAdam Litke enqueue_huge_page(page); 65419fc3f0aSAdam Litke } 65519fc3f0aSAdam Litke 65619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 65719fc3f0aSAdam Litke if (!list_empty(&surplus_list)) { 65819fc3f0aSAdam Litke spin_unlock(&hugetlb_lock); 659e4e574b7SAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 660e4e574b7SAdam Litke list_del(&page->lru); 661af767cbdSAdam Litke /* 6622668db91SAdam Litke * The page has a reference count of zero already, so 6632668db91SAdam Litke * call free_huge_page directly instead of using 6642668db91SAdam Litke * put_page. This must be done with hugetlb_lock 665af767cbdSAdam Litke * unlocked which is safe because free_huge_page takes 666af767cbdSAdam Litke * hugetlb_lock before deciding how to free the page. 667af767cbdSAdam Litke */ 6682668db91SAdam Litke free_huge_page(page); 669af767cbdSAdam Litke } 67019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 671e4e574b7SAdam Litke } 672e4e574b7SAdam Litke 673e4e574b7SAdam Litke return ret; 674e4e574b7SAdam Litke } 675e4e574b7SAdam Litke 676e4e574b7SAdam Litke /* 677e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 678e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 679e4e574b7SAdam Litke * never used. 680e4e574b7SAdam Litke */ 6818cde045cSAdrian Bunk static void return_unused_surplus_pages(unsigned long unused_resv_pages) 682e4e574b7SAdam Litke { 683e4e574b7SAdam Litke static int nid = -1; 684e4e574b7SAdam Litke struct page *page; 685e4e574b7SAdam Litke unsigned long nr_pages; 686e4e574b7SAdam Litke 68711320d17SNishanth Aravamudan /* 68811320d17SNishanth Aravamudan * We want to release as many surplus pages as possible, spread 68911320d17SNishanth Aravamudan * evenly across all nodes. Iterate across all nodes until we 69011320d17SNishanth Aravamudan * can no longer free unreserved surplus pages. This occurs when 69111320d17SNishanth Aravamudan * the nodes with surplus pages have no free pages. 69211320d17SNishanth Aravamudan */ 69311320d17SNishanth Aravamudan unsigned long remaining_iterations = num_online_nodes(); 69411320d17SNishanth Aravamudan 695ac09b3a1SAdam Litke /* Uncommit the reservation */ 696ac09b3a1SAdam Litke resv_huge_pages -= unused_resv_pages; 697ac09b3a1SAdam Litke 698e4e574b7SAdam Litke nr_pages = min(unused_resv_pages, surplus_huge_pages); 699e4e574b7SAdam Litke 70011320d17SNishanth Aravamudan while (remaining_iterations-- && nr_pages) { 701e4e574b7SAdam Litke nid = next_node(nid, node_online_map); 702e4e574b7SAdam Litke if (nid == MAX_NUMNODES) 703e4e574b7SAdam Litke nid = first_node(node_online_map); 704e4e574b7SAdam Litke 705e4e574b7SAdam Litke if (!surplus_huge_pages_node[nid]) 706e4e574b7SAdam Litke continue; 707e4e574b7SAdam Litke 708e4e574b7SAdam Litke if (!list_empty(&hugepage_freelists[nid])) { 709e4e574b7SAdam Litke page = list_entry(hugepage_freelists[nid].next, 710e4e574b7SAdam Litke struct page, lru); 711e4e574b7SAdam Litke list_del(&page->lru); 712e4e574b7SAdam Litke update_and_free_page(page); 713e4e574b7SAdam Litke free_huge_pages--; 714e4e574b7SAdam Litke free_huge_pages_node[nid]--; 715e4e574b7SAdam Litke surplus_huge_pages--; 716e4e574b7SAdam Litke surplus_huge_pages_node[nid]--; 717e4e574b7SAdam Litke nr_pages--; 71811320d17SNishanth Aravamudan remaining_iterations = num_online_nodes(); 719e4e574b7SAdam Litke } 720e4e574b7SAdam Litke } 721e4e574b7SAdam Litke } 722e4e574b7SAdam Litke 723348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 72404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 725348ea204SAdam Litke { 726348ea204SAdam Litke struct page *page; 7272fc39cecSAdam Litke struct address_space *mapping = vma->vm_file->f_mapping; 728a1e78772SMel Gorman struct inode *inode = mapping->host; 729a1e78772SMel Gorman unsigned int chg = 0; 7302fc39cecSAdam Litke 731a1e78772SMel Gorman /* 732a1e78772SMel Gorman * Processes that did not create the mapping will have no reserves and 733a1e78772SMel Gorman * will not have accounted against quota. Check that the quota can be 734a1e78772SMel Gorman * made before satisfying the allocation 735a1e78772SMel Gorman */ 73604f2cbe3SMel Gorman if (!(vma->vm_flags & VM_SHARED) && 73704f2cbe3SMel Gorman !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 738a1e78772SMel Gorman chg = 1; 739a1e78772SMel Gorman if (hugetlb_get_quota(inode->i_mapping, chg)) 740a1e78772SMel Gorman return ERR_PTR(-ENOSPC); 741a1e78772SMel Gorman } 74290d8b7e6SAdam Litke 743a1e78772SMel Gorman spin_lock(&hugetlb_lock); 74404f2cbe3SMel Gorman page = dequeue_huge_page_vma(vma, addr, avoid_reserve); 745a1e78772SMel Gorman spin_unlock(&hugetlb_lock); 746a1e78772SMel Gorman 747a1e78772SMel Gorman if (!page) { 748a1e78772SMel Gorman page = alloc_buddy_huge_page(vma, addr); 749a1e78772SMel Gorman if (!page) { 750a1e78772SMel Gorman hugetlb_put_quota(inode->i_mapping, chg); 751a1e78772SMel Gorman return ERR_PTR(-VM_FAULT_OOM); 752a1e78772SMel Gorman } 753a1e78772SMel Gorman } 754a1e78772SMel Gorman 755348ea204SAdam Litke set_page_refcounted(page); 7562fc39cecSAdam Litke set_page_private(page, (unsigned long) mapping); 757a1e78772SMel Gorman 7587893d1d5SAdam Litke return page; 759b45b5bd6SDavid Gibson } 760b45b5bd6SDavid Gibson 7611da177e4SLinus Torvalds static int __init hugetlb_init(void) 7621da177e4SLinus Torvalds { 7631da177e4SLinus Torvalds unsigned long i; 7641da177e4SLinus Torvalds 7653c726f8dSBenjamin Herrenschmidt if (HPAGE_SHIFT == 0) 7663c726f8dSBenjamin Herrenschmidt return 0; 7673c726f8dSBenjamin Herrenschmidt 7681da177e4SLinus Torvalds for (i = 0; i < MAX_NUMNODES; ++i) 7691da177e4SLinus Torvalds INIT_LIST_HEAD(&hugepage_freelists[i]); 7701da177e4SLinus Torvalds 77163b4613cSNishanth Aravamudan hugetlb_next_nid = first_node(node_online_map); 77263b4613cSNishanth Aravamudan 7731da177e4SLinus Torvalds for (i = 0; i < max_huge_pages; ++i) { 774a482289dSNick Piggin if (!alloc_fresh_huge_page()) 7751da177e4SLinus Torvalds break; 7761da177e4SLinus Torvalds } 7771da177e4SLinus Torvalds max_huge_pages = free_huge_pages = nr_huge_pages = i; 7781da177e4SLinus Torvalds printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); 7791da177e4SLinus Torvalds return 0; 7801da177e4SLinus Torvalds } 7811da177e4SLinus Torvalds module_init(hugetlb_init); 7821da177e4SLinus Torvalds 7831da177e4SLinus Torvalds static int __init hugetlb_setup(char *s) 7841da177e4SLinus Torvalds { 7851da177e4SLinus Torvalds if (sscanf(s, "%lu", &max_huge_pages) <= 0) 7861da177e4SLinus Torvalds max_huge_pages = 0; 7871da177e4SLinus Torvalds return 1; 7881da177e4SLinus Torvalds } 7891da177e4SLinus Torvalds __setup("hugepages=", hugetlb_setup); 7901da177e4SLinus Torvalds 7918a630112SKen Chen static unsigned int cpuset_mems_nr(unsigned int *array) 7928a630112SKen Chen { 7938a630112SKen Chen int node; 7948a630112SKen Chen unsigned int nr = 0; 7958a630112SKen Chen 7968a630112SKen Chen for_each_node_mask(node, cpuset_current_mems_allowed) 7978a630112SKen Chen nr += array[node]; 7988a630112SKen Chen 7998a630112SKen Chen return nr; 8008a630112SKen Chen } 8018a630112SKen Chen 8021da177e4SLinus Torvalds #ifdef CONFIG_SYSCTL 8031da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 8041da177e4SLinus Torvalds static void try_to_free_low(unsigned long count) 8051da177e4SLinus Torvalds { 8064415cc8dSChristoph Lameter int i; 8074415cc8dSChristoph Lameter 8081da177e4SLinus Torvalds for (i = 0; i < MAX_NUMNODES; ++i) { 8091da177e4SLinus Torvalds struct page *page, *next; 8101da177e4SLinus Torvalds list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { 8116b0c880dSAdam Litke if (count >= nr_huge_pages) 8126b0c880dSAdam Litke return; 8131da177e4SLinus Torvalds if (PageHighMem(page)) 8141da177e4SLinus Torvalds continue; 8151da177e4SLinus Torvalds list_del(&page->lru); 8161da177e4SLinus Torvalds update_and_free_page(page); 8171da177e4SLinus Torvalds free_huge_pages--; 8184415cc8dSChristoph Lameter free_huge_pages_node[page_to_nid(page)]--; 8191da177e4SLinus Torvalds } 8201da177e4SLinus Torvalds } 8211da177e4SLinus Torvalds } 8221da177e4SLinus Torvalds #else 8231da177e4SLinus Torvalds static inline void try_to_free_low(unsigned long count) 8241da177e4SLinus Torvalds { 8251da177e4SLinus Torvalds } 8261da177e4SLinus Torvalds #endif 8271da177e4SLinus Torvalds 8287893d1d5SAdam Litke #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages) 8291da177e4SLinus Torvalds static unsigned long set_max_huge_pages(unsigned long count) 8301da177e4SLinus Torvalds { 8317893d1d5SAdam Litke unsigned long min_count, ret; 8321da177e4SLinus Torvalds 8337893d1d5SAdam Litke /* 8347893d1d5SAdam Litke * Increase the pool size 8357893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 8367893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 837d1c3fb1fSNishanth Aravamudan * 838d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 839d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 840d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 841d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 842d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 8437893d1d5SAdam Litke */ 8441da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 8457893d1d5SAdam Litke while (surplus_huge_pages && count > persistent_huge_pages) { 8467893d1d5SAdam Litke if (!adjust_pool_surplus(-1)) 8477893d1d5SAdam Litke break; 8487893d1d5SAdam Litke } 8497893d1d5SAdam Litke 8507893d1d5SAdam Litke while (count > persistent_huge_pages) { 8517893d1d5SAdam Litke /* 8527893d1d5SAdam Litke * If this allocation races such that we no longer need the 8537893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 8547893d1d5SAdam Litke * and reducing the surplus. 8557893d1d5SAdam Litke */ 8567893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 8577893d1d5SAdam Litke ret = alloc_fresh_huge_page(); 8587893d1d5SAdam Litke spin_lock(&hugetlb_lock); 8597893d1d5SAdam Litke if (!ret) 8607893d1d5SAdam Litke goto out; 8617893d1d5SAdam Litke 8627893d1d5SAdam Litke } 8637893d1d5SAdam Litke 8647893d1d5SAdam Litke /* 8657893d1d5SAdam Litke * Decrease the pool size 8667893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 8677893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 8687893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 8697893d1d5SAdam Litke * to the desired size as pages become free. 870d1c3fb1fSNishanth Aravamudan * 871d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 872d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 873d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 874d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 875d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 876d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 877d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 8787893d1d5SAdam Litke */ 8796b0c880dSAdam Litke min_count = resv_huge_pages + nr_huge_pages - free_huge_pages; 8806b0c880dSAdam Litke min_count = max(count, min_count); 8817893d1d5SAdam Litke try_to_free_low(min_count); 8827893d1d5SAdam Litke while (min_count < persistent_huge_pages) { 883348e1e04SNishanth Aravamudan struct page *page = dequeue_huge_page(); 8841da177e4SLinus Torvalds if (!page) 8851da177e4SLinus Torvalds break; 8861da177e4SLinus Torvalds update_and_free_page(page); 8871da177e4SLinus Torvalds } 8887893d1d5SAdam Litke while (count < persistent_huge_pages) { 8897893d1d5SAdam Litke if (!adjust_pool_surplus(1)) 8907893d1d5SAdam Litke break; 8917893d1d5SAdam Litke } 8927893d1d5SAdam Litke out: 8937893d1d5SAdam Litke ret = persistent_huge_pages; 8941da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 8957893d1d5SAdam Litke return ret; 8961da177e4SLinus Torvalds } 8971da177e4SLinus Torvalds 8981da177e4SLinus Torvalds int hugetlb_sysctl_handler(struct ctl_table *table, int write, 8991da177e4SLinus Torvalds struct file *file, void __user *buffer, 9001da177e4SLinus Torvalds size_t *length, loff_t *ppos) 9011da177e4SLinus Torvalds { 9021da177e4SLinus Torvalds proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 9031da177e4SLinus Torvalds max_huge_pages = set_max_huge_pages(max_huge_pages); 9041da177e4SLinus Torvalds return 0; 9051da177e4SLinus Torvalds } 906396faf03SMel Gorman 907396faf03SMel Gorman int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 908396faf03SMel Gorman struct file *file, void __user *buffer, 909396faf03SMel Gorman size_t *length, loff_t *ppos) 910396faf03SMel Gorman { 911396faf03SMel Gorman proc_dointvec(table, write, file, buffer, length, ppos); 912396faf03SMel Gorman if (hugepages_treat_as_movable) 913396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 914396faf03SMel Gorman else 915396faf03SMel Gorman htlb_alloc_mask = GFP_HIGHUSER; 916396faf03SMel Gorman return 0; 917396faf03SMel Gorman } 918396faf03SMel Gorman 919a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 920a3d0c6aaSNishanth Aravamudan struct file *file, void __user *buffer, 921a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 922a3d0c6aaSNishanth Aravamudan { 923a3d0c6aaSNishanth Aravamudan proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 924064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 925064d9efeSNishanth Aravamudan nr_overcommit_huge_pages = sysctl_overcommit_huge_pages; 926a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 927a3d0c6aaSNishanth Aravamudan return 0; 928a3d0c6aaSNishanth Aravamudan } 929a3d0c6aaSNishanth Aravamudan 9301da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 9311da177e4SLinus Torvalds 9321da177e4SLinus Torvalds int hugetlb_report_meminfo(char *buf) 9331da177e4SLinus Torvalds { 9341da177e4SLinus Torvalds return sprintf(buf, 9351da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 9361da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 937b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 9387893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 9391da177e4SLinus Torvalds "Hugepagesize: %5lu kB\n", 9401da177e4SLinus Torvalds nr_huge_pages, 9411da177e4SLinus Torvalds free_huge_pages, 942a43a8c39SChen, Kenneth W resv_huge_pages, 9437893d1d5SAdam Litke surplus_huge_pages, 9441da177e4SLinus Torvalds HPAGE_SIZE/1024); 9451da177e4SLinus Torvalds } 9461da177e4SLinus Torvalds 9471da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 9481da177e4SLinus Torvalds { 9491da177e4SLinus Torvalds return sprintf(buf, 9501da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 951a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 952a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 9531da177e4SLinus Torvalds nid, nr_huge_pages_node[nid], 954a1de0919SNishanth Aravamudan nid, free_huge_pages_node[nid], 955a1de0919SNishanth Aravamudan nid, surplus_huge_pages_node[nid]); 9561da177e4SLinus Torvalds } 9571da177e4SLinus Torvalds 9581da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 9591da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 9601da177e4SLinus Torvalds { 9611da177e4SLinus Torvalds return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); 9621da177e4SLinus Torvalds } 9631da177e4SLinus Torvalds 964fc1b8a73SMel Gorman static int hugetlb_acct_memory(long delta) 965fc1b8a73SMel Gorman { 966fc1b8a73SMel Gorman int ret = -ENOMEM; 967fc1b8a73SMel Gorman 968fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 969fc1b8a73SMel Gorman /* 970fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 971fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 972fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 973fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 974fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 975fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 976fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 977fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 978fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 979fc1b8a73SMel Gorman * 980fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 981fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 982fc1b8a73SMel Gorman * we fall back to check against current free page availability as 983fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 984fc1b8a73SMel Gorman * semantics that cpuset has. 985fc1b8a73SMel Gorman */ 986fc1b8a73SMel Gorman if (delta > 0) { 987fc1b8a73SMel Gorman if (gather_surplus_pages(delta) < 0) 988fc1b8a73SMel Gorman goto out; 989fc1b8a73SMel Gorman 990fc1b8a73SMel Gorman if (delta > cpuset_mems_nr(free_huge_pages_node)) { 991fc1b8a73SMel Gorman return_unused_surplus_pages(delta); 992fc1b8a73SMel Gorman goto out; 993fc1b8a73SMel Gorman } 994fc1b8a73SMel Gorman } 995fc1b8a73SMel Gorman 996fc1b8a73SMel Gorman ret = 0; 997fc1b8a73SMel Gorman if (delta < 0) 998fc1b8a73SMel Gorman return_unused_surplus_pages((unsigned long) -delta); 999fc1b8a73SMel Gorman 1000fc1b8a73SMel Gorman out: 1001fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 1002fc1b8a73SMel Gorman return ret; 1003fc1b8a73SMel Gorman } 1004fc1b8a73SMel Gorman 1005a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 1006a1e78772SMel Gorman { 1007a1e78772SMel Gorman unsigned long reserve = vma_resv_huge_pages(vma); 1008a1e78772SMel Gorman if (reserve) 1009a1e78772SMel Gorman hugetlb_acct_memory(-reserve); 1010a1e78772SMel Gorman } 1011a1e78772SMel Gorman 10121da177e4SLinus Torvalds /* 10131da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 10141da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 10151da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 10161da177e4SLinus Torvalds * this far. 10171da177e4SLinus Torvalds */ 1018d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 10191da177e4SLinus Torvalds { 10201da177e4SLinus Torvalds BUG(); 1021d0217ac0SNick Piggin return 0; 10221da177e4SLinus Torvalds } 10231da177e4SLinus Torvalds 10241da177e4SLinus Torvalds struct vm_operations_struct hugetlb_vm_ops = { 1025d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 1026a1e78772SMel Gorman .close = hugetlb_vm_op_close, 10271da177e4SLinus Torvalds }; 10281da177e4SLinus Torvalds 10291e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 10301e8f889bSDavid Gibson int writable) 103163551ae0SDavid Gibson { 103263551ae0SDavid Gibson pte_t entry; 103363551ae0SDavid Gibson 10341e8f889bSDavid Gibson if (writable) { 103563551ae0SDavid Gibson entry = 103663551ae0SDavid Gibson pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 103763551ae0SDavid Gibson } else { 10387f2e9525SGerald Schaefer entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 103963551ae0SDavid Gibson } 104063551ae0SDavid Gibson entry = pte_mkyoung(entry); 104163551ae0SDavid Gibson entry = pte_mkhuge(entry); 104263551ae0SDavid Gibson 104363551ae0SDavid Gibson return entry; 104463551ae0SDavid Gibson } 104563551ae0SDavid Gibson 10461e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 10471e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 10481e8f889bSDavid Gibson { 10491e8f889bSDavid Gibson pte_t entry; 10501e8f889bSDavid Gibson 10517f2e9525SGerald Schaefer entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 10527f2e9525SGerald Schaefer if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 10531e8f889bSDavid Gibson update_mmu_cache(vma, address, entry); 10541e8f889bSDavid Gibson } 10558dab5241SBenjamin Herrenschmidt } 10561e8f889bSDavid Gibson 10571e8f889bSDavid Gibson 105863551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 105963551ae0SDavid Gibson struct vm_area_struct *vma) 106063551ae0SDavid Gibson { 106163551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 106263551ae0SDavid Gibson struct page *ptepage; 10631c59827dSHugh Dickins unsigned long addr; 10641e8f889bSDavid Gibson int cow; 10651e8f889bSDavid Gibson 10661e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 106763551ae0SDavid Gibson 10681c59827dSHugh Dickins for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { 1069c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 1070c74df32cSHugh Dickins if (!src_pte) 1071c74df32cSHugh Dickins continue; 107263551ae0SDavid Gibson dst_pte = huge_pte_alloc(dst, addr); 107363551ae0SDavid Gibson if (!dst_pte) 107463551ae0SDavid Gibson goto nomem; 1075c5c99429SLarry Woodman 1076c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 1077c5c99429SLarry Woodman if (dst_pte == src_pte) 1078c5c99429SLarry Woodman continue; 1079c5c99429SLarry Woodman 1080c74df32cSHugh Dickins spin_lock(&dst->page_table_lock); 108146478758SNick Piggin spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 10827f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 10831e8f889bSDavid Gibson if (cow) 10847f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 10857f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 108663551ae0SDavid Gibson ptepage = pte_page(entry); 108763551ae0SDavid Gibson get_page(ptepage); 108863551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 10891c59827dSHugh Dickins } 10901c59827dSHugh Dickins spin_unlock(&src->page_table_lock); 1091c74df32cSHugh Dickins spin_unlock(&dst->page_table_lock); 109263551ae0SDavid Gibson } 109363551ae0SDavid Gibson return 0; 109463551ae0SDavid Gibson 109563551ae0SDavid Gibson nomem: 109663551ae0SDavid Gibson return -ENOMEM; 109763551ae0SDavid Gibson } 109863551ae0SDavid Gibson 1099502717f4SChen, Kenneth W void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 110004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 110163551ae0SDavid Gibson { 110263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 110363551ae0SDavid Gibson unsigned long address; 1104c7546f8fSDavid Gibson pte_t *ptep; 110563551ae0SDavid Gibson pte_t pte; 110663551ae0SDavid Gibson struct page *page; 1107fe1668aeSChen, Kenneth W struct page *tmp; 1108c0a499c2SChen, Kenneth W /* 1109c0a499c2SChen, Kenneth W * A page gathering list, protected by per file i_mmap_lock. The 1110c0a499c2SChen, Kenneth W * lock is used to avoid list corruption from multiple unmapping 1111c0a499c2SChen, Kenneth W * of the same page since we are using page->lru. 1112c0a499c2SChen, Kenneth W */ 1113fe1668aeSChen, Kenneth W LIST_HEAD(page_list); 111463551ae0SDavid Gibson 111563551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 111663551ae0SDavid Gibson BUG_ON(start & ~HPAGE_MASK); 111763551ae0SDavid Gibson BUG_ON(end & ~HPAGE_MASK); 111863551ae0SDavid Gibson 1119508034a3SHugh Dickins spin_lock(&mm->page_table_lock); 112063551ae0SDavid Gibson for (address = start; address < end; address += HPAGE_SIZE) { 1121c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 1122c7546f8fSDavid Gibson if (!ptep) 1123c7546f8fSDavid Gibson continue; 1124c7546f8fSDavid Gibson 112539dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 112639dde65cSChen, Kenneth W continue; 112739dde65cSChen, Kenneth W 112804f2cbe3SMel Gorman /* 112904f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 113004f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 113104f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 113204f2cbe3SMel Gorman */ 113304f2cbe3SMel Gorman if (ref_page) { 113404f2cbe3SMel Gorman pte = huge_ptep_get(ptep); 113504f2cbe3SMel Gorman if (huge_pte_none(pte)) 113604f2cbe3SMel Gorman continue; 113704f2cbe3SMel Gorman page = pte_page(pte); 113804f2cbe3SMel Gorman if (page != ref_page) 113904f2cbe3SMel Gorman continue; 114004f2cbe3SMel Gorman 114104f2cbe3SMel Gorman /* 114204f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 114304f2cbe3SMel Gorman * future faults in this VMA will fail rather than 114404f2cbe3SMel Gorman * looking like data was lost 114504f2cbe3SMel Gorman */ 114604f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 114704f2cbe3SMel Gorman } 114804f2cbe3SMel Gorman 1149c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 11507f2e9525SGerald Schaefer if (huge_pte_none(pte)) 115163551ae0SDavid Gibson continue; 1152c7546f8fSDavid Gibson 115363551ae0SDavid Gibson page = pte_page(pte); 11546649a386SKen Chen if (pte_dirty(pte)) 11556649a386SKen Chen set_page_dirty(page); 1156fe1668aeSChen, Kenneth W list_add(&page->lru, &page_list); 115763551ae0SDavid Gibson } 11581da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 1159508034a3SHugh Dickins flush_tlb_range(vma, start, end); 1160fe1668aeSChen, Kenneth W list_for_each_entry_safe(page, tmp, &page_list, lru) { 1161fe1668aeSChen, Kenneth W list_del(&page->lru); 1162fe1668aeSChen, Kenneth W put_page(page); 1163fe1668aeSChen, Kenneth W } 11641da177e4SLinus Torvalds } 116563551ae0SDavid Gibson 1166502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 116704f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 1168502717f4SChen, Kenneth W { 1169502717f4SChen, Kenneth W /* 1170502717f4SChen, Kenneth W * It is undesirable to test vma->vm_file as it should be non-null 1171502717f4SChen, Kenneth W * for valid hugetlb area. However, vm_file will be NULL in the error 1172502717f4SChen, Kenneth W * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, 1173502717f4SChen, Kenneth W * do_mmap_pgoff() nullifies vma->vm_file before calling this function 1174502717f4SChen, Kenneth W * to clean up. Since no pte has actually been setup, it is safe to 1175502717f4SChen, Kenneth W * do nothing in this case. 1176502717f4SChen, Kenneth W */ 1177502717f4SChen, Kenneth W if (vma->vm_file) { 1178502717f4SChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 117904f2cbe3SMel Gorman __unmap_hugepage_range(vma, start, end, ref_page); 1180502717f4SChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 1181502717f4SChen, Kenneth W } 1182502717f4SChen, Kenneth W } 1183502717f4SChen, Kenneth W 118404f2cbe3SMel Gorman /* 118504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 118604f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 118704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 118804f2cbe3SMel Gorman * same region. 118904f2cbe3SMel Gorman */ 119004f2cbe3SMel Gorman int unmap_ref_private(struct mm_struct *mm, 119104f2cbe3SMel Gorman struct vm_area_struct *vma, 119204f2cbe3SMel Gorman struct page *page, 119304f2cbe3SMel Gorman unsigned long address) 119404f2cbe3SMel Gorman { 119504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 119604f2cbe3SMel Gorman struct address_space *mapping; 119704f2cbe3SMel Gorman struct prio_tree_iter iter; 119804f2cbe3SMel Gorman pgoff_t pgoff; 119904f2cbe3SMel Gorman 120004f2cbe3SMel Gorman /* 120104f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 120204f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 120304f2cbe3SMel Gorman */ 120404f2cbe3SMel Gorman address = address & huge_page_mask(hstate_vma(vma)); 120504f2cbe3SMel Gorman pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 120604f2cbe3SMel Gorman + (vma->vm_pgoff >> PAGE_SHIFT); 120704f2cbe3SMel Gorman mapping = (struct address_space *)page_private(page); 120804f2cbe3SMel Gorman 120904f2cbe3SMel Gorman vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 121004f2cbe3SMel Gorman /* Do not unmap the current VMA */ 121104f2cbe3SMel Gorman if (iter_vma == vma) 121204f2cbe3SMel Gorman continue; 121304f2cbe3SMel Gorman 121404f2cbe3SMel Gorman /* 121504f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 121604f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 121704f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 121804f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 121904f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 122004f2cbe3SMel Gorman */ 122104f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 122204f2cbe3SMel Gorman unmap_hugepage_range(iter_vma, 122304f2cbe3SMel Gorman address, address + HPAGE_SIZE, 122404f2cbe3SMel Gorman page); 122504f2cbe3SMel Gorman } 122604f2cbe3SMel Gorman 122704f2cbe3SMel Gorman return 1; 122804f2cbe3SMel Gorman } 122904f2cbe3SMel Gorman 12301e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 123104f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 123204f2cbe3SMel Gorman struct page *pagecache_page) 12331e8f889bSDavid Gibson { 12341e8f889bSDavid Gibson struct page *old_page, *new_page; 123579ac6ba4SDavid Gibson int avoidcopy; 123604f2cbe3SMel Gorman int outside_reserve = 0; 12371e8f889bSDavid Gibson 12381e8f889bSDavid Gibson old_page = pte_page(pte); 12391e8f889bSDavid Gibson 124004f2cbe3SMel Gorman retry_avoidcopy: 12411e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 12421e8f889bSDavid Gibson * and just make the page writable */ 12431e8f889bSDavid Gibson avoidcopy = (page_count(old_page) == 1); 12441e8f889bSDavid Gibson if (avoidcopy) { 12451e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 124683c54070SNick Piggin return 0; 12471e8f889bSDavid Gibson } 12481e8f889bSDavid Gibson 124904f2cbe3SMel Gorman /* 125004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 125104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 125204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 125304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 125404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 125504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 125604f2cbe3SMel Gorman * of the full address range. 125704f2cbe3SMel Gorman */ 125804f2cbe3SMel Gorman if (!(vma->vm_flags & VM_SHARED) && 125904f2cbe3SMel Gorman is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 126004f2cbe3SMel Gorman old_page != pagecache_page) 126104f2cbe3SMel Gorman outside_reserve = 1; 126204f2cbe3SMel Gorman 12631e8f889bSDavid Gibson page_cache_get(old_page); 126404f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 12651e8f889bSDavid Gibson 12662fc39cecSAdam Litke if (IS_ERR(new_page)) { 12671e8f889bSDavid Gibson page_cache_release(old_page); 126804f2cbe3SMel Gorman 126904f2cbe3SMel Gorman /* 127004f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 127104f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 127204f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 127304f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 127404f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 127504f2cbe3SMel Gorman */ 127604f2cbe3SMel Gorman if (outside_reserve) { 127704f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 127804f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 127904f2cbe3SMel Gorman BUG_ON(page_count(old_page) != 1); 128004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 128104f2cbe3SMel Gorman goto retry_avoidcopy; 128204f2cbe3SMel Gorman } 128304f2cbe3SMel Gorman WARN_ON_ONCE(1); 128404f2cbe3SMel Gorman } 128504f2cbe3SMel Gorman 12862fc39cecSAdam Litke return -PTR_ERR(new_page); 12871e8f889bSDavid Gibson } 12881e8f889bSDavid Gibson 12891e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 12909de455b2SAtsushi Nemoto copy_huge_page(new_page, old_page, address, vma); 12910ed361deSNick Piggin __SetPageUptodate(new_page); 12921e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 12931e8f889bSDavid Gibson 12941e8f889bSDavid Gibson ptep = huge_pte_offset(mm, address & HPAGE_MASK); 12957f2e9525SGerald Schaefer if (likely(pte_same(huge_ptep_get(ptep), pte))) { 12961e8f889bSDavid Gibson /* Break COW */ 12978fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 12981e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 12991e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 13001e8f889bSDavid Gibson /* Make the old page be freed below */ 13011e8f889bSDavid Gibson new_page = old_page; 13021e8f889bSDavid Gibson } 13031e8f889bSDavid Gibson page_cache_release(new_page); 13041e8f889bSDavid Gibson page_cache_release(old_page); 130583c54070SNick Piggin return 0; 13061e8f889bSDavid Gibson } 13071e8f889bSDavid Gibson 130804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 130904f2cbe3SMel Gorman static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma, 131004f2cbe3SMel Gorman unsigned long address) 131104f2cbe3SMel Gorman { 131204f2cbe3SMel Gorman struct address_space *mapping; 1313e7c4b0bfSAndy Whitcroft pgoff_t idx; 131404f2cbe3SMel Gorman 131504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 1316e7c4b0bfSAndy Whitcroft idx = vma_pagecache_offset(vma, address); 131704f2cbe3SMel Gorman 131804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 131904f2cbe3SMel Gorman } 132004f2cbe3SMel Gorman 1321a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 13221e8f889bSDavid Gibson unsigned long address, pte_t *ptep, int write_access) 1323ac9b9c66SHugh Dickins { 1324ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 1325e7c4b0bfSAndy Whitcroft pgoff_t idx; 13264c887265SAdam Litke unsigned long size; 13274c887265SAdam Litke struct page *page; 13284c887265SAdam Litke struct address_space *mapping; 13291e8f889bSDavid Gibson pte_t new_pte; 13304c887265SAdam Litke 133104f2cbe3SMel Gorman /* 133204f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 133304f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 133404f2cbe3SMel Gorman * COW. Warn that such a situation has occured as it may not be obvious 133504f2cbe3SMel Gorman */ 133604f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 133704f2cbe3SMel Gorman printk(KERN_WARNING 133804f2cbe3SMel Gorman "PID %d killed due to inadequate hugepage pool\n", 133904f2cbe3SMel Gorman current->pid); 134004f2cbe3SMel Gorman return ret; 134104f2cbe3SMel Gorman } 134204f2cbe3SMel Gorman 13434c887265SAdam Litke mapping = vma->vm_file->f_mapping; 1344e7c4b0bfSAndy Whitcroft idx = vma_pagecache_offset(vma, address); 13454c887265SAdam Litke 13464c887265SAdam Litke /* 13474c887265SAdam Litke * Use page lock to guard against racing truncation 13484c887265SAdam Litke * before we get page_table_lock. 13494c887265SAdam Litke */ 13506bda666aSChristoph Lameter retry: 13516bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 13526bda666aSChristoph Lameter if (!page) { 1353ebed4bfcSHugh Dickins size = i_size_read(mapping->host) >> HPAGE_SHIFT; 1354ebed4bfcSHugh Dickins if (idx >= size) 1355ebed4bfcSHugh Dickins goto out; 135604f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 13572fc39cecSAdam Litke if (IS_ERR(page)) { 13582fc39cecSAdam Litke ret = -PTR_ERR(page); 13596bda666aSChristoph Lameter goto out; 13606bda666aSChristoph Lameter } 136179ac6ba4SDavid Gibson clear_huge_page(page, address); 13620ed361deSNick Piggin __SetPageUptodate(page); 1363ac9b9c66SHugh Dickins 13646bda666aSChristoph Lameter if (vma->vm_flags & VM_SHARED) { 13656bda666aSChristoph Lameter int err; 136645c682a6SKen Chen struct inode *inode = mapping->host; 13676bda666aSChristoph Lameter 13686bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 13696bda666aSChristoph Lameter if (err) { 13706bda666aSChristoph Lameter put_page(page); 13716bda666aSChristoph Lameter if (err == -EEXIST) 13726bda666aSChristoph Lameter goto retry; 13736bda666aSChristoph Lameter goto out; 13746bda666aSChristoph Lameter } 137545c682a6SKen Chen 137645c682a6SKen Chen spin_lock(&inode->i_lock); 137745c682a6SKen Chen inode->i_blocks += BLOCKS_PER_HUGEPAGE; 137845c682a6SKen Chen spin_unlock(&inode->i_lock); 13796bda666aSChristoph Lameter } else 13806bda666aSChristoph Lameter lock_page(page); 13816bda666aSChristoph Lameter } 13821e8f889bSDavid Gibson 1383ac9b9c66SHugh Dickins spin_lock(&mm->page_table_lock); 13844c887265SAdam Litke size = i_size_read(mapping->host) >> HPAGE_SHIFT; 13854c887265SAdam Litke if (idx >= size) 13864c887265SAdam Litke goto backout; 13874c887265SAdam Litke 138883c54070SNick Piggin ret = 0; 13897f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 13904c887265SAdam Litke goto backout; 13914c887265SAdam Litke 13921e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 13931e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 13941e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 13951e8f889bSDavid Gibson 13961e8f889bSDavid Gibson if (write_access && !(vma->vm_flags & VM_SHARED)) { 13971e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 139804f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 13991e8f889bSDavid Gibson } 14001e8f889bSDavid Gibson 1401ac9b9c66SHugh Dickins spin_unlock(&mm->page_table_lock); 14024c887265SAdam Litke unlock_page(page); 14034c887265SAdam Litke out: 1404ac9b9c66SHugh Dickins return ret; 14054c887265SAdam Litke 14064c887265SAdam Litke backout: 14074c887265SAdam Litke spin_unlock(&mm->page_table_lock); 14084c887265SAdam Litke unlock_page(page); 14094c887265SAdam Litke put_page(page); 14104c887265SAdam Litke goto out; 1411ac9b9c66SHugh Dickins } 1412ac9b9c66SHugh Dickins 141386e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 141486e5216fSAdam Litke unsigned long address, int write_access) 141586e5216fSAdam Litke { 141686e5216fSAdam Litke pte_t *ptep; 141786e5216fSAdam Litke pte_t entry; 14181e8f889bSDavid Gibson int ret; 14193935baa9SDavid Gibson static DEFINE_MUTEX(hugetlb_instantiation_mutex); 142086e5216fSAdam Litke 142186e5216fSAdam Litke ptep = huge_pte_alloc(mm, address); 142286e5216fSAdam Litke if (!ptep) 142386e5216fSAdam Litke return VM_FAULT_OOM; 142486e5216fSAdam Litke 14253935baa9SDavid Gibson /* 14263935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 14273935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 14283935baa9SDavid Gibson * the same page in the page cache. 14293935baa9SDavid Gibson */ 14303935baa9SDavid Gibson mutex_lock(&hugetlb_instantiation_mutex); 14317f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 14327f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 14333935baa9SDavid Gibson ret = hugetlb_no_page(mm, vma, address, ptep, write_access); 14343935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 14353935baa9SDavid Gibson return ret; 14363935baa9SDavid Gibson } 143786e5216fSAdam Litke 143883c54070SNick Piggin ret = 0; 14391e8f889bSDavid Gibson 14401e8f889bSDavid Gibson spin_lock(&mm->page_table_lock); 14411e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 14427f2e9525SGerald Schaefer if (likely(pte_same(entry, huge_ptep_get(ptep)))) 144304f2cbe3SMel Gorman if (write_access && !pte_write(entry)) { 144404f2cbe3SMel Gorman struct page *page; 144504f2cbe3SMel Gorman page = hugetlbfs_pagecache_page(vma, address); 144604f2cbe3SMel Gorman ret = hugetlb_cow(mm, vma, address, ptep, entry, page); 144704f2cbe3SMel Gorman if (page) { 144804f2cbe3SMel Gorman unlock_page(page); 144904f2cbe3SMel Gorman put_page(page); 145004f2cbe3SMel Gorman } 145104f2cbe3SMel Gorman } 14521e8f889bSDavid Gibson spin_unlock(&mm->page_table_lock); 14533935baa9SDavid Gibson mutex_unlock(&hugetlb_instantiation_mutex); 14541e8f889bSDavid Gibson 14551e8f889bSDavid Gibson return ret; 145686e5216fSAdam Litke } 145786e5216fSAdam Litke 145863551ae0SDavid Gibson int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 145963551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 14605b23dbe8SAdam Litke unsigned long *position, int *length, int i, 14615b23dbe8SAdam Litke int write) 146263551ae0SDavid Gibson { 1463d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 1464d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 146563551ae0SDavid Gibson int remainder = *length; 146663551ae0SDavid Gibson 14671c59827dSHugh Dickins spin_lock(&mm->page_table_lock); 146863551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 146963551ae0SDavid Gibson pte_t *pte; 147063551ae0SDavid Gibson struct page *page; 147163551ae0SDavid Gibson 14724c887265SAdam Litke /* 14734c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 14744c887265SAdam Litke * each hugepage. We have to make * sure we get the 14754c887265SAdam Litke * first, for the page indexing below to work. 14764c887265SAdam Litke */ 147763551ae0SDavid Gibson pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); 147863551ae0SDavid Gibson 14797f2e9525SGerald Schaefer if (!pte || huge_pte_none(huge_ptep_get(pte)) || 14807f2e9525SGerald Schaefer (write && !pte_write(huge_ptep_get(pte)))) { 14814c887265SAdam Litke int ret; 14824c887265SAdam Litke 14834c887265SAdam Litke spin_unlock(&mm->page_table_lock); 14845b23dbe8SAdam Litke ret = hugetlb_fault(mm, vma, vaddr, write); 14854c887265SAdam Litke spin_lock(&mm->page_table_lock); 1486a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 14874c887265SAdam Litke continue; 14884c887265SAdam Litke 14891c59827dSHugh Dickins remainder = 0; 14901c59827dSHugh Dickins if (!i) 14911c59827dSHugh Dickins i = -EFAULT; 14921c59827dSHugh Dickins break; 14931c59827dSHugh Dickins } 149463551ae0SDavid Gibson 1495d5d4b0aaSChen, Kenneth W pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; 14967f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 1497d5d4b0aaSChen, Kenneth W same_page: 1498d6692183SChen, Kenneth W if (pages) { 149963551ae0SDavid Gibson get_page(page); 1500d5d4b0aaSChen, Kenneth W pages[i] = page + pfn_offset; 1501d6692183SChen, Kenneth W } 150263551ae0SDavid Gibson 150363551ae0SDavid Gibson if (vmas) 150463551ae0SDavid Gibson vmas[i] = vma; 150563551ae0SDavid Gibson 150663551ae0SDavid Gibson vaddr += PAGE_SIZE; 1507d5d4b0aaSChen, Kenneth W ++pfn_offset; 150863551ae0SDavid Gibson --remainder; 150963551ae0SDavid Gibson ++i; 1510d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 1511d5d4b0aaSChen, Kenneth W pfn_offset < HPAGE_SIZE/PAGE_SIZE) { 1512d5d4b0aaSChen, Kenneth W /* 1513d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 1514d5d4b0aaSChen, Kenneth W * of this compound page. 1515d5d4b0aaSChen, Kenneth W */ 1516d5d4b0aaSChen, Kenneth W goto same_page; 1517d5d4b0aaSChen, Kenneth W } 151863551ae0SDavid Gibson } 15191c59827dSHugh Dickins spin_unlock(&mm->page_table_lock); 152063551ae0SDavid Gibson *length = remainder; 152163551ae0SDavid Gibson *position = vaddr; 152263551ae0SDavid Gibson 152363551ae0SDavid Gibson return i; 152463551ae0SDavid Gibson } 15258f860591SZhang, Yanmin 15268f860591SZhang, Yanmin void hugetlb_change_protection(struct vm_area_struct *vma, 15278f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 15288f860591SZhang, Yanmin { 15298f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 15308f860591SZhang, Yanmin unsigned long start = address; 15318f860591SZhang, Yanmin pte_t *ptep; 15328f860591SZhang, Yanmin pte_t pte; 15338f860591SZhang, Yanmin 15348f860591SZhang, Yanmin BUG_ON(address >= end); 15358f860591SZhang, Yanmin flush_cache_range(vma, address, end); 15368f860591SZhang, Yanmin 153739dde65cSChen, Kenneth W spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 15388f860591SZhang, Yanmin spin_lock(&mm->page_table_lock); 15398f860591SZhang, Yanmin for (; address < end; address += HPAGE_SIZE) { 15408f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 15418f860591SZhang, Yanmin if (!ptep) 15428f860591SZhang, Yanmin continue; 154339dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 154439dde65cSChen, Kenneth W continue; 15457f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 15468f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 15478f860591SZhang, Yanmin pte = pte_mkhuge(pte_modify(pte, newprot)); 15488f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 15498f860591SZhang, Yanmin } 15508f860591SZhang, Yanmin } 15518f860591SZhang, Yanmin spin_unlock(&mm->page_table_lock); 155239dde65cSChen, Kenneth W spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 15538f860591SZhang, Yanmin 15548f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 15558f860591SZhang, Yanmin } 15568f860591SZhang, Yanmin 1557a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 1558a1e78772SMel Gorman long from, long to, 1559a1e78772SMel Gorman struct vm_area_struct *vma) 1560e4e574b7SAdam Litke { 1561e4e574b7SAdam Litke long ret, chg; 1562e4e574b7SAdam Litke 1563a1e78772SMel Gorman /* 1564a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 1565a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 1566a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 1567a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 1568a1e78772SMel Gorman */ 1569a1e78772SMel Gorman if (!vma || vma->vm_flags & VM_SHARED) 1570e4e574b7SAdam Litke chg = region_chg(&inode->i_mapping->private_list, from, to); 1571a1e78772SMel Gorman else { 1572a1e78772SMel Gorman chg = to - from; 1573a1e78772SMel Gorman set_vma_resv_huge_pages(vma, chg); 157404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 1575a1e78772SMel Gorman } 1576a1e78772SMel Gorman 1577e4e574b7SAdam Litke if (chg < 0) 1578e4e574b7SAdam Litke return chg; 15798a630112SKen Chen 158090d8b7e6SAdam Litke if (hugetlb_get_quota(inode->i_mapping, chg)) 158190d8b7e6SAdam Litke return -ENOSPC; 1582a43a8c39SChen, Kenneth W ret = hugetlb_acct_memory(chg); 158368842c9bSKen Chen if (ret < 0) { 158468842c9bSKen Chen hugetlb_put_quota(inode->i_mapping, chg); 1585a43a8c39SChen, Kenneth W return ret; 158668842c9bSKen Chen } 1587a1e78772SMel Gorman if (!vma || vma->vm_flags & VM_SHARED) 1588a43a8c39SChen, Kenneth W region_add(&inode->i_mapping->private_list, from, to); 1589a43a8c39SChen, Kenneth W return 0; 1590a43a8c39SChen, Kenneth W } 1591a43a8c39SChen, Kenneth W 1592a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 1593a43a8c39SChen, Kenneth W { 1594a43a8c39SChen, Kenneth W long chg = region_truncate(&inode->i_mapping->private_list, offset); 159545c682a6SKen Chen 159645c682a6SKen Chen spin_lock(&inode->i_lock); 159745c682a6SKen Chen inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed; 159845c682a6SKen Chen spin_unlock(&inode->i_lock); 159945c682a6SKen Chen 160090d8b7e6SAdam Litke hugetlb_put_quota(inode->i_mapping, (chg - freed)); 160190d8b7e6SAdam Litke hugetlb_acct_memory(-(chg - freed)); 1602a43a8c39SChen, Kenneth W } 1603