mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/sysctl.h>
  12 #include <linux/highmem.h>
  13 #include <linux/mmu_notifier.h>
  14 #include <linux/nodemask.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/mempolicy.h>
  17 #include <linux/cpuset.h>
  18 #include <linux/mutex.h>
  19 #include <linux/bootmem.h>
  20 #include <linux/sysfs.h>
  21
  22 #include <asm/page.h>
  23 #include <asm/pgtable.h>
  24 #include <asm/io.h>
  25
  26 #include <linux/hugetlb.h>
  27 #include "internal.h"
  28
  29 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  30 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  31 unsigned long hugepages_treat_as_movable;
  32
  33 static int max_hstate;
  34 unsigned int default_hstate_idx;
  35 struct hstate hstates[HUGE_MAX_HSTATE];
  36
  37 __initdata LIST_HEAD(huge_boot_pages);
  38
  39 /* for command line parsing */
  40 static struct hstate * __initdata parsed_hstate;
  41 static unsigned long __initdata default_hstate_max_huge_pages;
  42 static unsigned long __initdata default_hstate_size;
  43
  44 #define for_each_hstate(h) \
  45         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
  46
  47 /*
  48  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  49  */
  50 static DEFINE_SPINLOCK(hugetlb_lock);
  51
  52 /*
  53  * Region tracking -- allows tracking of reservations and instantiated pages
  54  *                    across the pages in a mapping.
  55  *
  56  * The region data structures are protected by a combination of the mmap_sem
  57  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
  58  * must either hold the mmap_sem for write, or the mmap_sem for read and
  59  * the hugetlb_instantiation mutex:
  60  *
  61  *      down_write(&mm->mmap_sem);
  62  * or
  63  *      down_read(&mm->mmap_sem);
  64  *      mutex_lock(&hugetlb_instantiation_mutex);
  65  */
  66 struct file_region {
  67         struct list_head link;
  68         long from;
  69         long to;
  70 };
  71
  72 static long region_add(struct list_head *head, long f, long t)
  73 {
  74         struct file_region *rg, *nrg, *trg;
  75
  76         /* Locate the region we are either in or before. */
  77         list_for_each_entry(rg, head, link)
  78                 if (f <= rg->to)
  79                         break;
  80
  81         /* Round our left edge to the current segment if it encloses us. */
  82         if (f > rg->from)
  83                 f = rg->from;
  84
  85         /* Check for and consume any regions we now overlap with. */
  86         nrg = rg;
  87         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
  88                 if (&rg->link == head)
  89                         break;
  90                 if (rg->from > t)
  91                         break;
  92
  93                 /* If this area reaches higher then extend our area to
  94                  * include it completely.  If this is not the first area
  95                  * which we intend to reuse, free it. */
  96                 if (rg->to > t)
  97                         t = rg->to;
  98                 if (rg != nrg) {
  99                         list_del(&rg->link);
 100                         kfree(rg);
 101                 }
 102         }
 103         nrg->from = f;
 104         nrg->to = t;
 105         return 0;
 106 }
 107
 108 static long region_chg(struct list_head *head, long f, long t)
 109 {
 110         struct file_region *rg, *nrg;
 111         long chg = 0;
 112
 113         /* Locate the region we are before or in. */
 114         list_for_each_entry(rg, head, link)
 115                 if (f <= rg->to)
 116                         break;
 117
 118         /* If we are below the current region then a new region is required.
 119          * Subtle, allocate a new region at the position but make it zero
 120          * size such that we can guarantee to record the reservation. */
 121         if (&rg->link == head || t < rg->from) {
 122                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 123                 if (!nrg)
 124                         return -ENOMEM;
 125                 nrg->from = f;
 126                 nrg->to   = f;
 127                 INIT_LIST_HEAD(&nrg->link);
 128                 list_add(&nrg->link, rg->link.prev);
 129
 130                 return t - f;
 131         }
 132
 133         /* Round our left edge to the current segment if it encloses us. */
 134         if (f > rg->from)
 135                 f = rg->from;
 136         chg = t - f;
 137
 138         /* Check for and consume any regions we now overlap with. */
 139         list_for_each_entry(rg, rg->link.prev, link) {
 140                 if (&rg->link == head)
 141                         break;
 142                 if (rg->from > t)
 143                         return chg;
 144
 145                 /* We overlap with this area, if it extends futher than
 146                  * us then we must extend ourselves.  Account for its
 147                  * existing reservation. */
 148                 if (rg->to > t) {
 149                         chg += rg->to - t;
 150                         t = rg->to;
 151                 }
 152                 chg -= rg->to - rg->from;
 153         }
 154         return chg;
 155 }
 156
 157 static long region_truncate(struct list_head *head, long end)
 158 {
 159         struct file_region *rg, *trg;
 160         long chg = 0;
 161
 162         /* Locate the region we are either in or before. */
 163         list_for_each_entry(rg, head, link)
 164                 if (end <= rg->to)
 165                         break;
 166         if (&rg->link == head)
 167                 return 0;
 168
 169         /* If we are in the middle of a region then adjust it. */
 170         if (end > rg->from) {
 171                 chg = rg->to - end;
 172                 rg->to = end;
 173                 rg = list_entry(rg->link.next, typeof(*rg), link);
 174         }
 175
 176         /* Drop any remaining regions. */
 177         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 178                 if (&rg->link == head)
 179                         break;
 180                 chg += rg->to - rg->from;
 181                 list_del(&rg->link);
 182                 kfree(rg);
 183         }
 184         return chg;
 185 }
 186
 187 static long region_count(struct list_head *head, long f, long t)
 188 {
 189         struct file_region *rg;
 190         long chg = 0;
 191
 192         /* Locate each segment we overlap with, and count that overlap. */
 193         list_for_each_entry(rg, head, link) {
 194                 int seg_from;
 195                 int seg_to;
 196
 197                 if (rg->to <= f)
 198                         continue;
 199                 if (rg->from >= t)
 200                         break;
 201
 202                 seg_from = max(rg->from, f);
 203                 seg_to = min(rg->to, t);
 204
 205                 chg += seg_to - seg_from;
 206         }
 207
 208         return chg;
 209 }
 210
 211 /*
 212  * Convert the address within this vma to the page offset within
 213  * the mapping, in pagecache page units; huge pages here.
 214  */
 215 static pgoff_t vma_hugecache_offset(struct hstate *h,
 216                         struct vm_area_struct *vma, unsigned long address)
 217 {
 218         return ((address - vma->vm_start) >> huge_page_shift(h)) +
 219                         (vma->vm_pgoff >> huge_page_order(h));
 220 }
 221
 222 /*
 223  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 224  * bits of the reservation map pointer, which are always clear due to
 225  * alignment.
 226  */
 227 #define HPAGE_RESV_OWNER    (1UL << 0)
 228 #define HPAGE_RESV_UNMAPPED (1UL << 1)
 229 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 230
 231 /*
 232  * These helpers are used to track how many pages are reserved for
 233  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 234  * is guaranteed to have their future faults succeed.
 235  *
 236  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 237  * the reserve counters are updated with the hugetlb_lock held. It is safe
 238  * to reset the VMA at fork() time as it is not in use yet and there is no
 239  * chance of the global counters getting corrupted as a result of the values.
 240  *
 241  * The private mapping reservation is represented in a subtly different
 242  * manner to a shared mapping.  A shared mapping has a region map associated
 243  * with the underlying file, this region map represents the backing file
 244  * pages which have ever had a reservation assigned which this persists even
 245  * after the page is instantiated.  A private mapping has a region map
 246  * associated with the original mmap which is attached to all VMAs which
 247  * reference it, this region map represents those offsets which have consumed
 248  * reservation ie. where pages have been instantiated.
 249  */
 250 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 251 {
 252         return (unsigned long)vma->vm_private_data;
 253 }
 254
 255 static void set_vma_private_data(struct vm_area_struct *vma,
 256                                                         unsigned long value)
 257 {
 258         vma->vm_private_data = (void *)value;
 259 }
 260
 261 struct resv_map {
 262         struct kref refs;
 263         struct list_head regions;
 264 };
 265
 266 static struct resv_map *resv_map_alloc(void)
 267 {
 268         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 269         if (!resv_map)
 270                 return NULL;
 271
 272         kref_init(&resv_map->refs);
 273         INIT_LIST_HEAD(&resv_map->regions);
 274
 275         return resv_map;
 276 }
 277
 278 static void resv_map_release(struct kref *ref)
 279 {
 280         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 281
 282         /* Clear out any active regions before we release the map. */
 283         region_truncate(&resv_map->regions, 0);
 284         kfree(resv_map);
 285 }
 286
 287 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 288 {
 289         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 290         if (!(vma->vm_flags & VM_SHARED))
 291                 return (struct resv_map *)(get_vma_private_data(vma) &
 292                                                         ~HPAGE_RESV_MASK);
 293         return NULL;
 294 }
 295
 296 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 297 {
 298         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 299         VM_BUG_ON(vma->vm_flags & VM_SHARED);
 300
 301         set_vma_private_data(vma, (get_vma_private_data(vma) &
 302                                 HPAGE_RESV_MASK) | (unsigned long)map);
 303 }
 304
 305 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 306 {
 307         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 308         VM_BUG_ON(vma->vm_flags & VM_SHARED);
 309
 310         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 311 }
 312
 313 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 314 {
 315         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 316
 317         return (get_vma_private_data(vma) & flag) != 0;
 318 }
 319
 320 /* Decrement the reserved pages in the hugepage pool by one */
 321 static void decrement_hugepage_resv_vma(struct hstate *h,
 322                         struct vm_area_struct *vma)
 323 {
 324         if (vma->vm_flags & VM_NORESERVE)
 325                 return;
 326
 327         if (vma->vm_flags & VM_SHARED) {
 328                 /* Shared mappings always use reserves */
 329                 h->resv_huge_pages--;
 330         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 331                 /*
 332                  * Only the process that called mmap() has reserves for
 333                  * private mappings.
 334                  */
 335                 h->resv_huge_pages--;
 336         }
 337 }
 338
 339 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 340 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 341 {
 342         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 343         if (!(vma->vm_flags & VM_SHARED))
 344                 vma->vm_private_data = (void *)0;
 345 }
 346
 347 /* Returns true if the VMA has associated reserve pages */
 348 static int vma_has_reserves(struct vm_area_struct *vma)
 349 {
 350         if (vma->vm_flags & VM_SHARED)
 351                 return 1;
 352         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 353                 return 1;
 354         return 0;
 355 }
 356
 357 static void clear_gigantic_page(struct page *page,
 358                         unsigned long addr, unsigned long sz)
 359 {
 360         int i;
 361         struct page *p = page;
 362
 363         might_sleep();
 364         for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
 365                 cond_resched();
 366                 clear_user_highpage(p, addr + i * PAGE_SIZE);
 367         }
 368 }
 369 static void clear_huge_page(struct page *page,
 370                         unsigned long addr, unsigned long sz)
 371 {
 372         int i;
 373
 374         if (unlikely(sz > MAX_ORDER_NR_PAGES))
 375                 return clear_gigantic_page(page, addr, sz);
 376
 377         might_sleep();
 378         for (i = 0; i < sz/PAGE_SIZE; i++) {
 379                 cond_resched();
 380                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
 381         }
 382 }
 383
 384 static void copy_gigantic_page(struct page *dst, struct page *src,
 385                            unsigned long addr, struct vm_area_struct *vma)
 386 {
 387         int i;
 388         struct hstate *h = hstate_vma(vma);
 389         struct page *dst_base = dst;
 390         struct page *src_base = src;
 391         might_sleep();
 392         for (i = 0; i < pages_per_huge_page(h); ) {
 393                 cond_resched();
 394                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
 395
 396                 i++;
 397                 dst = mem_map_next(dst, dst_base, i);
 398                 src = mem_map_next(src, src_base, i);
 399         }
 400 }
 401 static void copy_huge_page(struct page *dst, struct page *src,
 402                            unsigned long addr, struct vm_area_struct *vma)
 403 {
 404         int i;
 405         struct hstate *h = hstate_vma(vma);
 406
 407         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES))
 408                 return copy_gigantic_page(dst, src, addr, vma);
 409
 410         might_sleep();
 411         for (i = 0; i < pages_per_huge_page(h); i++) {
 412                 cond_resched();
 413                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
 414         }
 415 }
 416
 417 static void enqueue_huge_page(struct hstate *h, struct page *page)
 418 {
 419         int nid = page_to_nid(page);
 420         list_add(&page->lru, &h->hugepage_freelists[nid]);
 421         h->free_huge_pages++;
 422         h->free_huge_pages_node[nid]++;
 423 }
 424
 425 static struct page *dequeue_huge_page(struct hstate *h)
 426 {
 427         int nid;
 428         struct page *page = NULL;
 429
 430         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
 431                 if (!list_empty(&h->hugepage_freelists[nid])) {
 432                         page = list_entry(h->hugepage_freelists[nid].next,
 433                                           struct page, lru);
 434                         list_del(&page->lru);
 435                         h->free_huge_pages--;
 436                         h->free_huge_pages_node[nid]--;
 437                         break;
 438                 }
 439         }
 440         return page;
 441 }
 442
 443 static struct page *dequeue_huge_page_vma(struct hstate *h,
 444                                 struct vm_area_struct *vma,
 445                                 unsigned long address, int avoid_reserve)
 446 {
 447         int nid;
 448         struct page *page = NULL;
 449         struct mempolicy *mpol;
 450         nodemask_t *nodemask;
 451         struct zonelist *zonelist = huge_zonelist(vma, address,
 452                                         htlb_alloc_mask, &mpol, &nodemask);
 453         struct zone *zone;
 454         struct zoneref *z;
 455
 456         /*
 457          * A child process with MAP_PRIVATE mappings created by their parent
 458          * have no page reserves. This check ensures that reservations are
 459          * not "stolen". The child may still get SIGKILLed
 460          */
 461         if (!vma_has_reserves(vma) &&
 462                         h->free_huge_pages - h->resv_huge_pages == 0)
 463                 return NULL;
 464
 465         /* If reserves cannot be used, ensure enough pages are in the pool */
 466         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 467                 return NULL;
 468
 469         for_each_zone_zonelist_nodemask(zone, z, zonelist,
 470                                                 MAX_NR_ZONES - 1, nodemask) {
 471                 nid = zone_to_nid(zone);
 472                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
 473                     !list_empty(&h->hugepage_freelists[nid])) {
 474                         page = list_entry(h->hugepage_freelists[nid].next,
 475                                           struct page, lru);
 476                         list_del(&page->lru);
 477                         h->free_huge_pages--;
 478                         h->free_huge_pages_node[nid]--;
 479
 480                         if (!avoid_reserve)
 481                                 decrement_hugepage_resv_vma(h, vma);
 482
 483                         break;
 484                 }
 485         }
 486         mpol_cond_put(mpol);
 487         return page;
 488 }
 489
 490 static void update_and_free_page(struct hstate *h, struct page *page)
 491 {
 492         int i;
 493
 494         VM_BUG_ON(h->order >= MAX_ORDER);
 495
 496         h->nr_huge_pages--;
 497         h->nr_huge_pages_node[page_to_nid(page)]--;
 498         for (i = 0; i < pages_per_huge_page(h); i++) {
 499                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 500                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 501                                 1 << PG_private | 1<< PG_writeback);
 502         }
 503         set_compound_page_dtor(page, NULL);
 504         set_page_refcounted(page);
 505         arch_release_hugepage(page);
 506         __free_pages(page, huge_page_order(h));
 507 }
 508
 509 struct hstate *size_to_hstate(unsigned long size)
 510 {
 511         struct hstate *h;
 512
 513         for_each_hstate(h) {
 514                 if (huge_page_size(h) == size)
 515                         return h;
 516         }
 517         return NULL;
 518 }
 519
 520 static void free_huge_page(struct page *page)
 521 {
 522         /*
 523          * Can't pass hstate in here because it is called from the
 524          * compound page destructor.
 525          */
 526         struct hstate *h = page_hstate(page);
 527         int nid = page_to_nid(page);
 528         struct address_space *mapping;
 529
 530         mapping = (struct address_space *) page_private(page);
 531         set_page_private(page, 0);
 532         BUG_ON(page_count(page));
 533         INIT_LIST_HEAD(&page->lru);
 534
 535         spin_lock(&hugetlb_lock);
 536         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 537                 update_and_free_page(h, page);
 538                 h->surplus_huge_pages--;
 539                 h->surplus_huge_pages_node[nid]--;
 540         } else {
 541                 enqueue_huge_page(h, page);
 542         }
 543         spin_unlock(&hugetlb_lock);
 544         if (mapping)
 545                 hugetlb_put_quota(mapping, 1);
 546 }
 547
 548 /*
 549  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 550  * balanced by operating on them in a round-robin fashion.
 551  * Returns 1 if an adjustment was made.
 552  */
 553 static int adjust_pool_surplus(struct hstate *h, int delta)
 554 {
 555         static int prev_nid;
 556         int nid = prev_nid;
 557         int ret = 0;
 558
 559         VM_BUG_ON(delta != -1 && delta != 1);
 560         do {
 561                 nid = next_node(nid, node_online_map);
 562                 if (nid == MAX_NUMNODES)
 563                         nid = first_node(node_online_map);
 564
 565                 /* To shrink on this node, there must be a surplus page */
 566                 if (delta < 0 && !h->surplus_huge_pages_node[nid])
 567                         continue;
 568                 /* Surplus cannot exceed the total number of pages */
 569                 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
 570                                                 h->nr_huge_pages_node[nid])
 571                         continue;
 572
 573                 h->surplus_huge_pages += delta;
 574                 h->surplus_huge_pages_node[nid] += delta;
 575                 ret = 1;
 576                 break;
 577         } while (nid != prev_nid);
 578
 579         prev_nid = nid;
 580         return ret;
 581 }
 582
 583 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 584 {
 585         set_compound_page_dtor(page, free_huge_page);
 586         spin_lock(&hugetlb_lock);
 587         h->nr_huge_pages++;
 588         h->nr_huge_pages_node[nid]++;
 589         spin_unlock(&hugetlb_lock);
 590         put_page(page); /* free it into the hugepage allocator */
 591 }
 592
 593 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 594 {
 595         struct page *page;
 596
 597         if (h->order >= MAX_ORDER)
 598                 return NULL;
 599
 600         page = alloc_pages_node(nid,
 601                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 602                                                 __GFP_REPEAT|__GFP_NOWARN,
 603                 huge_page_order(h));
 604         if (page) {
 605                 if (arch_prepare_hugepage(page)) {
 606                         __free_pages(page, huge_page_order(h));
 607                         return NULL;
 608                 }
 609                 prep_new_huge_page(h, page, nid);
 610         }
 611
 612         return page;
 613 }
 614
 615 /*
 616  * Use a helper variable to find the next node and then
 617  * copy it back to hugetlb_next_nid afterwards:
 618  * otherwise there's a window in which a racer might
 619  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 620  * But we don't need to use a spin_lock here: it really
 621  * doesn't matter if occasionally a racer chooses the
 622  * same nid as we do.  Move nid forward in the mask even
 623  * if we just successfully allocated a hugepage so that
 624  * the next caller gets hugepages on the next node.
 625  */
 626 static int hstate_next_node(struct hstate *h)
 627 {
 628         int next_nid;
 629         next_nid = next_node(h->hugetlb_next_nid, node_online_map);
 630         if (next_nid == MAX_NUMNODES)
 631                 next_nid = first_node(node_online_map);
 632         h->hugetlb_next_nid = next_nid;
 633         return next_nid;
 634 }
 635
 636 static int alloc_fresh_huge_page(struct hstate *h)
 637 {
 638         struct page *page;
 639         int start_nid;
 640         int next_nid;
 641         int ret = 0;
 642
 643         start_nid = h->hugetlb_next_nid;
 644
 645         do {
 646                 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
 647                 if (page)
 648                         ret = 1;
 649                 next_nid = hstate_next_node(h);
 650         } while (!page && h->hugetlb_next_nid != start_nid);
 651
 652         if (ret)
 653                 count_vm_event(HTLB_BUDDY_PGALLOC);
 654         else
 655                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 656
 657         return ret;
 658 }
 659
 660 static struct page *alloc_buddy_huge_page(struct hstate *h,
 661                         struct vm_area_struct *vma, unsigned long address)
 662 {
 663         struct page *page;
 664         unsigned int nid;
 665
 666         if (h->order >= MAX_ORDER)
 667                 return NULL;
 668
 669         /*
 670          * Assume we will successfully allocate the surplus page to
 671          * prevent racing processes from causing the surplus to exceed
 672          * overcommit
 673          *
 674          * This however introduces a different race, where a process B
 675          * tries to grow the static hugepage pool while alloc_pages() is
 676          * called by process A. B will only examine the per-node
 677          * counters in determining if surplus huge pages can be
 678          * converted to normal huge pages in adjust_pool_surplus(). A
 679          * won't be able to increment the per-node counter, until the
 680          * lock is dropped by B, but B doesn't drop hugetlb_lock until
 681          * no more huge pages can be converted from surplus to normal
 682          * state (and doesn't try to convert again). Thus, we have a
 683          * case where a surplus huge page exists, the pool is grown, and
 684          * the surplus huge page still exists after, even though it
 685          * should just have been converted to a normal huge page. This
 686          * does not leak memory, though, as the hugepage will be freed
 687          * once it is out of use. It also does not allow the counters to
 688          * go out of whack in adjust_pool_surplus() as we don't modify
 689          * the node values until we've gotten the hugepage and only the
 690          * per-node value is checked there.
 691          */
 692         spin_lock(&hugetlb_lock);
 693         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 694                 spin_unlock(&hugetlb_lock);
 695                 return NULL;
 696         } else {
 697                 h->nr_huge_pages++;
 698                 h->surplus_huge_pages++;
 699         }
 700         spin_unlock(&hugetlb_lock);
 701
 702         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 703                                         __GFP_REPEAT|__GFP_NOWARN,
 704                                         huge_page_order(h));
 705
 706         if (page && arch_prepare_hugepage(page)) {
 707                 __free_pages(page, huge_page_order(h));
 708                 return NULL;
 709         }
 710
 711         spin_lock(&hugetlb_lock);
 712         if (page) {
 713                 /*
 714                  * This page is now managed by the hugetlb allocator and has
 715                  * no users -- drop the buddy allocator's reference.
 716                  */
 717                 put_page_testzero(page);
 718                 VM_BUG_ON(page_count(page));
 719                 nid = page_to_nid(page);
 720                 set_compound_page_dtor(page, free_huge_page);
 721                 /*
 722                  * We incremented the global counters already
 723                  */
 724                 h->nr_huge_pages_node[nid]++;
 725                 h->surplus_huge_pages_node[nid]++;
 726                 __count_vm_event(HTLB_BUDDY_PGALLOC);
 727         } else {
 728                 h->nr_huge_pages--;
 729                 h->surplus_huge_pages--;
 730                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 731         }
 732         spin_unlock(&hugetlb_lock);
 733
 734         return page;
 735 }
 736
 737 /*
 738  * Increase the hugetlb pool such that it can accomodate a reservation
 739  * of size 'delta'.
 740  */
 741 static int gather_surplus_pages(struct hstate *h, int delta)
 742 {
 743         struct list_head surplus_list;
 744         struct page *page, *tmp;
 745         int ret, i;
 746         int needed, allocated;
 747
 748         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
 749         if (needed <= 0) {
 750                 h->resv_huge_pages += delta;
 751                 return 0;
 752         }
 753
 754         allocated = 0;
 755         INIT_LIST_HEAD(&surplus_list);
 756
 757         ret = -ENOMEM;
 758 retry:
 759         spin_unlock(&hugetlb_lock);
 760         for (i = 0; i < needed; i++) {
 761                 page = alloc_buddy_huge_page(h, NULL, 0);
 762                 if (!page) {
 763                         /*
 764                          * We were not able to allocate enough pages to
 765                          * satisfy the entire reservation so we free what
 766                          * we've allocated so far.
 767                          */
 768                         spin_lock(&hugetlb_lock);
 769                         needed = 0;
 770                         goto free;
 771                 }
 772
 773                 list_add(&page->lru, &surplus_list);
 774         }
 775         allocated += needed;
 776
 777         /*
 778          * After retaking hugetlb_lock, we need to recalculate 'needed'
 779          * because either resv_huge_pages or free_huge_pages may have changed.
 780          */
 781         spin_lock(&hugetlb_lock);
 782         needed = (h->resv_huge_pages + delta) -
 783                         (h->free_huge_pages + allocated);
 784         if (needed > 0)
 785                 goto retry;
 786
 787         /*
 788          * The surplus_list now contains _at_least_ the number of extra pages
 789          * needed to accomodate the reservation.  Add the appropriate number
 790          * of pages to the hugetlb pool and free the extras back to the buddy
 791          * allocator.  Commit the entire reservation here to prevent another
 792          * process from stealing the pages as they are added to the pool but
 793          * before they are reserved.
 794          */
 795         needed += allocated;
 796         h->resv_huge_pages += delta;
 797         ret = 0;
 798 free:
 799         /* Free the needed pages to the hugetlb pool */
 800         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 801                 if ((--needed) < 0)
 802                         break;
 803                 list_del(&page->lru);
 804                 enqueue_huge_page(h, page);
 805         }
 806
 807         /* Free unnecessary surplus pages to the buddy allocator */
 808         if (!list_empty(&surplus_list)) {
 809                 spin_unlock(&hugetlb_lock);
 810                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 811                         list_del(&page->lru);
 812                         /*
 813                          * The page has a reference count of zero already, so
 814                          * call free_huge_page directly instead of using
 815                          * put_page.  This must be done with hugetlb_lock
 816                          * unlocked which is safe because free_huge_page takes
 817                          * hugetlb_lock before deciding how to free the page.
 818                          */
 819                         free_huge_page(page);
 820                 }
 821                 spin_lock(&hugetlb_lock);
 822         }
 823
 824         return ret;
 825 }
 826
 827 /*
 828  * When releasing a hugetlb pool reservation, any surplus pages that were
 829  * allocated to satisfy the reservation must be explicitly freed if they were
 830  * never used.
 831  */
 832 static void return_unused_surplus_pages(struct hstate *h,
 833                                         unsigned long unused_resv_pages)
 834 {
 835         static int nid = -1;
 836         struct page *page;
 837         unsigned long nr_pages;
 838
 839         /*
 840          * We want to release as many surplus pages as possible, spread
 841          * evenly across all nodes. Iterate across all nodes until we
 842          * can no longer free unreserved surplus pages. This occurs when
 843          * the nodes with surplus pages have no free pages.
 844          */
 845         unsigned long remaining_iterations = num_online_nodes();
 846
 847         /* Uncommit the reservation */
 848         h->resv_huge_pages -= unused_resv_pages;
 849
 850         /* Cannot return gigantic pages currently */
 851         if (h->order >= MAX_ORDER)
 852                 return;
 853
 854         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
 855
 856         while (remaining_iterations-- && nr_pages) {
 857                 nid = next_node(nid, node_online_map);
 858                 if (nid == MAX_NUMNODES)
 859                         nid = first_node(node_online_map);
 860
 861                 if (!h->surplus_huge_pages_node[nid])
 862                         continue;
 863
 864                 if (!list_empty(&h->hugepage_freelists[nid])) {
 865                         page = list_entry(h->hugepage_freelists[nid].next,
 866                                           struct page, lru);
 867                         list_del(&page->lru);
 868                         update_and_free_page(h, page);
 869                         h->free_huge_pages--;
 870                         h->free_huge_pages_node[nid]--;
 871                         h->surplus_huge_pages--;
 872                         h->surplus_huge_pages_node[nid]--;
 873                         nr_pages--;
 874                         remaining_iterations = num_online_nodes();
 875                 }
 876         }
 877 }
 878
 879 /*
 880  * Determine if the huge page at addr within the vma has an associated
 881  * reservation.  Where it does not we will need to logically increase
 882  * reservation and actually increase quota before an allocation can occur.
 883  * Where any new reservation would be required the reservation change is
 884  * prepared, but not committed.  Once the page has been quota'd allocated
 885  * an instantiated the change should be committed via vma_commit_reservation.
 886  * No action is required on failure.
 887  */
 888 static int vma_needs_reservation(struct hstate *h,
 889                         struct vm_area_struct *vma, unsigned long addr)
 890 {
 891         struct address_space *mapping = vma->vm_file->f_mapping;
 892         struct inode *inode = mapping->host;
 893
 894         if (vma->vm_flags & VM_SHARED) {
 895                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 896                 return region_chg(&inode->i_mapping->private_list,
 897                                                         idx, idx + 1);
 898
 899         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 900                 return 1;
 901
 902         } else  {
 903                 int err;
 904                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 905                 struct resv_map *reservations = vma_resv_map(vma);
 906
 907                 err = region_chg(&reservations->regions, idx, idx + 1);
 908                 if (err < 0)
 909                         return err;
 910                 return 0;
 911         }
 912 }
 913 static void vma_commit_reservation(struct hstate *h,
 914                         struct vm_area_struct *vma, unsigned long addr)
 915 {
 916         struct address_space *mapping = vma->vm_file->f_mapping;
 917         struct inode *inode = mapping->host;
 918
 919         if (vma->vm_flags & VM_SHARED) {
 920                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 921                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
 922
 923         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 924                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 925                 struct resv_map *reservations = vma_resv_map(vma);
 926
 927                 /* Mark this page used in the map. */
 928                 region_add(&reservations->regions, idx, idx + 1);
 929         }
 930 }
 931
 932 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 933                                     unsigned long addr, int avoid_reserve)
 934 {
 935         struct hstate *h = hstate_vma(vma);
 936         struct page *page;
 937         struct address_space *mapping = vma->vm_file->f_mapping;
 938         struct inode *inode = mapping->host;
 939         unsigned int chg;
 940
 941         /*
 942          * Processes that did not create the mapping will have no reserves and
 943          * will not have accounted against quota. Check that the quota can be
 944          * made before satisfying the allocation
 945          * MAP_NORESERVE mappings may also need pages and quota allocated
 946          * if no reserve mapping overlaps.
 947          */
 948         chg = vma_needs_reservation(h, vma, addr);
 949         if (chg < 0)
 950                 return ERR_PTR(chg);
 951         if (chg)
 952                 if (hugetlb_get_quota(inode->i_mapping, chg))
 953                         return ERR_PTR(-ENOSPC);
 954
 955         spin_lock(&hugetlb_lock);
 956         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
 957         spin_unlock(&hugetlb_lock);
 958
 959         if (!page) {
 960                 page = alloc_buddy_huge_page(h, vma, addr);
 961                 if (!page) {
 962                         hugetlb_put_quota(inode->i_mapping, chg);
 963                         return ERR_PTR(-VM_FAULT_OOM);
 964                 }
 965         }
 966
 967         set_page_refcounted(page);
 968         set_page_private(page, (unsigned long) mapping);
 969
 970         vma_commit_reservation(h, vma, addr);
 971
 972         return page;
 973 }
 974
 975 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
 976 {
 977         struct huge_bootmem_page *m;
 978         int nr_nodes = nodes_weight(node_online_map);
 979
 980         while (nr_nodes) {
 981                 void *addr;
 982
 983                 addr = __alloc_bootmem_node_nopanic(
 984                                 NODE_DATA(h->hugetlb_next_nid),
 985                                 huge_page_size(h), huge_page_size(h), 0);
 986
 987                 if (addr) {
 988                         /*
 989                          * Use the beginning of the huge page to store the
 990                          * huge_bootmem_page struct (until gather_bootmem
 991                          * puts them into the mem_map).
 992                          */
 993                         m = addr;
 994                         if (m)
 995                                 goto found;
 996                 }
 997                 hstate_next_node(h);
 998                 nr_nodes--;
 999         }
1000         return 0;
1001
1002 found:
1003         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1004         /* Put them into a private list first because mem_map is not up yet */
1005         list_add(&m->list, &huge_boot_pages);
1006         m->hstate = h;
1007         return 1;
1008 }
1009
1010 static void prep_compound_huge_page(struct page *page, int order)
1011 {
1012         if (unlikely(order > (MAX_ORDER - 1)))
1013                 prep_compound_gigantic_page(page, order);
1014         else
1015                 prep_compound_page(page, order);
1016 }
1017
1018 /* Put bootmem huge pages into the standard lists after mem_map is up */
1019 static void __init gather_bootmem_prealloc(void)
1020 {
1021         struct huge_bootmem_page *m;
1022
1023         list_for_each_entry(m, &huge_boot_pages, list) {
1024                 struct page *page = virt_to_page(m);
1025                 struct hstate *h = m->hstate;
1026                 __ClearPageReserved(page);
1027                 WARN_ON(page_count(page) != 1);
1028                 prep_compound_huge_page(page, h->order);
1029                 prep_new_huge_page(h, page, page_to_nid(page));
1030         }
1031 }
1032
1033 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1034 {
1035         unsigned long i;
1036
1037         for (i = 0; i < h->max_huge_pages; ++i) {
1038                 if (h->order >= MAX_ORDER) {
1039                         if (!alloc_bootmem_huge_page(h))
1040                                 break;
1041                 } else if (!alloc_fresh_huge_page(h))
1042                         break;
1043         }
1044         h->max_huge_pages = i;
1045 }
1046
1047 static void __init hugetlb_init_hstates(void)
1048 {
1049         struct hstate *h;
1050
1051         for_each_hstate(h) {
1052                 /* oversize hugepages were init'ed in early boot */
1053                 if (h->order < MAX_ORDER)
1054                         hugetlb_hstate_alloc_pages(h);
1055         }
1056 }
1057
1058 static char * __init memfmt(char *buf, unsigned long n)
1059 {
1060         if (n >= (1UL << 30))
1061                 sprintf(buf, "%lu GB", n >> 30);
1062         else if (n >= (1UL << 20))
1063                 sprintf(buf, "%lu MB", n >> 20);
1064         else
1065                 sprintf(buf, "%lu KB", n >> 10);
1066         return buf;
1067 }
1068
1069 static void __init report_hugepages(void)
1070 {
1071         struct hstate *h;
1072
1073         for_each_hstate(h) {
1074                 char buf[32];
1075                 printk(KERN_INFO "HugeTLB registered %s page size, "
1076                                  "pre-allocated %ld pages\n",
1077                         memfmt(buf, huge_page_size(h)),
1078                         h->free_huge_pages);
1079         }
1080 }
1081
1082 #ifdef CONFIG_HIGHMEM
1083 static void try_to_free_low(struct hstate *h, unsigned long count)
1084 {
1085         int i;
1086
1087         if (h->order >= MAX_ORDER)
1088                 return;
1089
1090         for (i = 0; i < MAX_NUMNODES; ++i) {
1091                 struct page *page, *next;
1092                 struct list_head *freel = &h->hugepage_freelists[i];
1093                 list_for_each_entry_safe(page, next, freel, lru) {
1094                         if (count >= h->nr_huge_pages)
1095                                 return;
1096                         if (PageHighMem(page))
1097                                 continue;
1098                         list_del(&page->lru);
1099                         update_and_free_page(h, page);
1100                         h->free_huge_pages--;
1101                         h->free_huge_pages_node[page_to_nid(page)]--;
1102                 }
1103         }
1104 }
1105 #else
1106 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1107 {
1108 }
1109 #endif
1110
1111 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1112 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1113 {
1114         unsigned long min_count, ret;
1115
1116         if (h->order >= MAX_ORDER)
1117                 return h->max_huge_pages;
1118
1119         /*
1120          * Increase the pool size
1121          * First take pages out of surplus state.  Then make up the
1122          * remaining difference by allocating fresh huge pages.
1123          *
1124          * We might race with alloc_buddy_huge_page() here and be unable
1125          * to convert a surplus huge page to a normal huge page. That is
1126          * not critical, though, it just means the overall size of the
1127          * pool might be one hugepage larger than it needs to be, but
1128          * within all the constraints specified by the sysctls.
1129          */
1130         spin_lock(&hugetlb_lock);
1131         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1132                 if (!adjust_pool_surplus(h, -1))
1133                         break;
1134         }
1135
1136         while (count > persistent_huge_pages(h)) {
1137                 /*
1138                  * If this allocation races such that we no longer need the
1139                  * page, free_huge_page will handle it by freeing the page
1140                  * and reducing the surplus.
1141                  */
1142                 spin_unlock(&hugetlb_lock);
1143                 ret = alloc_fresh_huge_page(h);
1144                 spin_lock(&hugetlb_lock);
1145                 if (!ret)
1146                         goto out;
1147
1148         }
1149
1150         /*
1151          * Decrease the pool size
1152          * First return free pages to the buddy allocator (being careful
1153          * to keep enough around to satisfy reservations).  Then place
1154          * pages into surplus state as needed so the pool will shrink
1155          * to the desired size as pages become free.
1156          *
1157          * By placing pages into the surplus state independent of the
1158          * overcommit value, we are allowing the surplus pool size to
1159          * exceed overcommit. There are few sane options here. Since
1160          * alloc_buddy_huge_page() is checking the global counter,
1161          * though, we'll note that we're not allowed to exceed surplus
1162          * and won't grow the pool anywhere else. Not until one of the
1163          * sysctls are changed, or the surplus pages go out of use.
1164          */
1165         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1166         min_count = max(count, min_count);
1167         try_to_free_low(h, min_count);
1168         while (min_count < persistent_huge_pages(h)) {
1169                 struct page *page = dequeue_huge_page(h);
1170                 if (!page)
1171                         break;
1172                 update_and_free_page(h, page);
1173         }
1174         while (count < persistent_huge_pages(h)) {
1175                 if (!adjust_pool_surplus(h, 1))
1176                         break;
1177         }
1178 out:
1179         ret = persistent_huge_pages(h);
1180         spin_unlock(&hugetlb_lock);
1181         return ret;
1182 }
1183
1184 #define HSTATE_ATTR_RO(_name) \
1185         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1186
1187 #define HSTATE_ATTR(_name) \
1188         static struct kobj_attribute _name##_attr = \
1189                 __ATTR(_name, 0644, _name##_show, _name##_store)
1190
1191 static struct kobject *hugepages_kobj;
1192 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1193
1194 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1195 {
1196         int i;
1197         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1198                 if (hstate_kobjs[i] == kobj)
1199                         return &hstates[i];
1200         BUG();
1201         return NULL;
1202 }
1203
1204 static ssize_t nr_hugepages_show(struct kobject *kobj,
1205                                         struct kobj_attribute *attr, char *buf)
1206 {
1207         struct hstate *h = kobj_to_hstate(kobj);
1208         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1209 }
1210 static ssize_t nr_hugepages_store(struct kobject *kobj,
1211                 struct kobj_attribute *attr, const char *buf, size_t count)
1212 {
1213         int err;
1214         unsigned long input;
1215         struct hstate *h = kobj_to_hstate(kobj);
1216
1217         err = strict_strtoul(buf, 10, &input);
1218         if (err)
1219                 return 0;
1220
1221         h->max_huge_pages = set_max_huge_pages(h, input);
1222
1223         return count;
1224 }
1225 HSTATE_ATTR(nr_hugepages);
1226
1227 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1228                                         struct kobj_attribute *attr, char *buf)
1229 {
1230         struct hstate *h = kobj_to_hstate(kobj);
1231         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1232 }
1233 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1234                 struct kobj_attribute *attr, const char *buf, size_t count)
1235 {
1236         int err;
1237         unsigned long input;
1238         struct hstate *h = kobj_to_hstate(kobj);
1239
1240         err = strict_strtoul(buf, 10, &input);
1241         if (err)
1242                 return 0;
1243
1244         spin_lock(&hugetlb_lock);
1245         h->nr_overcommit_huge_pages = input;
1246         spin_unlock(&hugetlb_lock);
1247
1248         return count;
1249 }
1250 HSTATE_ATTR(nr_overcommit_hugepages);
1251
1252 static ssize_t free_hugepages_show(struct kobject *kobj,
1253                                         struct kobj_attribute *attr, char *buf)
1254 {
1255         struct hstate *h = kobj_to_hstate(kobj);
1256         return sprintf(buf, "%lu\n", h->free_huge_pages);
1257 }
1258 HSTATE_ATTR_RO(free_hugepages);
1259
1260 static ssize_t resv_hugepages_show(struct kobject *kobj,
1261                                         struct kobj_attribute *attr, char *buf)
1262 {
1263         struct hstate *h = kobj_to_hstate(kobj);
1264         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1265 }
1266 HSTATE_ATTR_RO(resv_hugepages);
1267
1268 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1269                                         struct kobj_attribute *attr, char *buf)
1270 {
1271         struct hstate *h = kobj_to_hstate(kobj);
1272         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1273 }
1274 HSTATE_ATTR_RO(surplus_hugepages);
1275
1276 static struct attribute *hstate_attrs[] = {
1277         &nr_hugepages_attr.attr,
1278         &nr_overcommit_hugepages_attr.attr,
1279         &free_hugepages_attr.attr,
1280         &resv_hugepages_attr.attr,
1281         &surplus_hugepages_attr.attr,
1282         NULL,
1283 };
1284
1285 static struct attribute_group hstate_attr_group = {
1286         .attrs = hstate_attrs,
1287 };
1288
1289 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1290 {
1291         int retval;
1292
1293         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1294                                                         hugepages_kobj);
1295         if (!hstate_kobjs[h - hstates])
1296                 return -ENOMEM;
1297
1298         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1299                                                         &hstate_attr_group);
1300         if (retval)
1301                 kobject_put(hstate_kobjs[h - hstates]);
1302
1303         return retval;
1304 }
1305
1306 static void __init hugetlb_sysfs_init(void)
1307 {
1308         struct hstate *h;
1309         int err;
1310
1311         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1312         if (!hugepages_kobj)
1313                 return;
1314
1315         for_each_hstate(h) {
1316                 err = hugetlb_sysfs_add_hstate(h);
1317                 if (err)
1318                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1319                                                                 h->name);
1320         }
1321 }
1322
1323 static void __exit hugetlb_exit(void)
1324 {
1325         struct hstate *h;
1326
1327         for_each_hstate(h) {
1328                 kobject_put(hstate_kobjs[h - hstates]);
1329         }
1330
1331         kobject_put(hugepages_kobj);
1332 }
1333 module_exit(hugetlb_exit);
1334
1335 static int __init hugetlb_init(void)
1336 {
1337         /* Some platform decide whether they support huge pages at boot
1338          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1339          * there is no such support
1340          */
1341         if (HPAGE_SHIFT == 0)
1342                 return 0;
1343
1344         if (!size_to_hstate(default_hstate_size)) {
1345                 default_hstate_size = HPAGE_SIZE;
1346                 if (!size_to_hstate(default_hstate_size))
1347                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1348         }
1349         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1350         if (default_hstate_max_huge_pages)
1351                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1352
1353         hugetlb_init_hstates();
1354
1355         gather_bootmem_prealloc();
1356
1357         report_hugepages();
1358
1359         hugetlb_sysfs_init();
1360
1361         return 0;
1362 }
1363 module_init(hugetlb_init);
1364
1365 /* Should be called on processing a hugepagesz=... option */
1366 void __init hugetlb_add_hstate(unsigned order)
1367 {
1368         struct hstate *h;
1369         unsigned long i;
1370
1371         if (size_to_hstate(PAGE_SIZE << order)) {
1372                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1373                 return;
1374         }
1375         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1376         BUG_ON(order == 0);
1377         h = &hstates[max_hstate++];
1378         h->order = order;
1379         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1380         h->nr_huge_pages = 0;
1381         h->free_huge_pages = 0;
1382         for (i = 0; i < MAX_NUMNODES; ++i)
1383                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1384         h->hugetlb_next_nid = first_node(node_online_map);
1385         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1386                                         huge_page_size(h)/1024);
1387
1388         parsed_hstate = h;
1389 }
1390
1391 static int __init hugetlb_nrpages_setup(char *s)
1392 {
1393         unsigned long *mhp;
1394         static unsigned long *last_mhp;
1395
1396         /*
1397          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1398          * so this hugepages= parameter goes to the "default hstate".
1399          */
1400         if (!max_hstate)
1401                 mhp = &default_hstate_max_huge_pages;
1402         else
1403                 mhp = &parsed_hstate->max_huge_pages;
1404
1405         if (mhp == last_mhp) {
1406                 printk(KERN_WARNING "hugepages= specified twice without "
1407                         "interleaving hugepagesz=, ignoring\n");
1408                 return 1;
1409         }
1410
1411         if (sscanf(s, "%lu", mhp) <= 0)
1412                 *mhp = 0;
1413
1414         /*
1415          * Global state is always initialized later in hugetlb_init.
1416          * But we need to allocate >= MAX_ORDER hstates here early to still
1417          * use the bootmem allocator.
1418          */
1419         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1420                 hugetlb_hstate_alloc_pages(parsed_hstate);
1421
1422         last_mhp = mhp;
1423
1424         return 1;
1425 }
1426 __setup("hugepages=", hugetlb_nrpages_setup);
1427
1428 static int __init hugetlb_default_setup(char *s)
1429 {
1430         default_hstate_size = memparse(s, &s);
1431         return 1;
1432 }
1433 __setup("default_hugepagesz=", hugetlb_default_setup);
1434
1435 static unsigned int cpuset_mems_nr(unsigned int *array)
1436 {
1437         int node;
1438         unsigned int nr = 0;
1439
1440         for_each_node_mask(node, cpuset_current_mems_allowed)
1441                 nr += array[node];
1442
1443         return nr;
1444 }
1445
1446 #ifdef CONFIG_SYSCTL
1447 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1448                            struct file *file, void __user *buffer,
1449                            size_t *length, loff_t *ppos)
1450 {
1451         struct hstate *h = &default_hstate;
1452         unsigned long tmp;
1453
1454         if (!write)
1455                 tmp = h->max_huge_pages;
1456
1457         table->data = &tmp;
1458         table->maxlen = sizeof(unsigned long);
1459         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1460
1461         if (write)
1462                 h->max_huge_pages = set_max_huge_pages(h, tmp);
1463
1464         return 0;
1465 }
1466
1467 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1468                         struct file *file, void __user *buffer,
1469                         size_t *length, loff_t *ppos)
1470 {
1471         proc_dointvec(table, write, file, buffer, length, ppos);
1472         if (hugepages_treat_as_movable)
1473                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1474         else
1475                 htlb_alloc_mask = GFP_HIGHUSER;
1476         return 0;
1477 }
1478
1479 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1480                         struct file *file, void __user *buffer,
1481                         size_t *length, loff_t *ppos)
1482 {
1483         struct hstate *h = &default_hstate;
1484         unsigned long tmp;
1485
1486         if (!write)
1487                 tmp = h->nr_overcommit_huge_pages;
1488
1489         table->data = &tmp;
1490         table->maxlen = sizeof(unsigned long);
1491         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1492
1493         if (write) {
1494                 spin_lock(&hugetlb_lock);
1495                 h->nr_overcommit_huge_pages = tmp;
1496                 spin_unlock(&hugetlb_lock);
1497         }
1498
1499         return 0;
1500 }
1501
1502 #endif /* CONFIG_SYSCTL */
1503
1504 void hugetlb_report_meminfo(struct seq_file *m)
1505 {
1506         struct hstate *h = &default_hstate;
1507         seq_printf(m,
1508                         "HugePages_Total:   %5lu\n"
1509                         "HugePages_Free:    %5lu\n"
1510                         "HugePages_Rsvd:    %5lu\n"
1511                         "HugePages_Surp:    %5lu\n"
1512                         "Hugepagesize:   %8lu kB\n",
1513                         h->nr_huge_pages,
1514                         h->free_huge_pages,
1515                         h->resv_huge_pages,
1516                         h->surplus_huge_pages,
1517                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1518 }
1519
1520 int hugetlb_report_node_meminfo(int nid, char *buf)
1521 {
1522         struct hstate *h = &default_hstate;
1523         return sprintf(buf,
1524                 "Node %d HugePages_Total: %5u\n"
1525                 "Node %d HugePages_Free:  %5u\n"
1526                 "Node %d HugePages_Surp:  %5u\n",
1527                 nid, h->nr_huge_pages_node[nid],
1528                 nid, h->free_huge_pages_node[nid],
1529                 nid, h->surplus_huge_pages_node[nid]);
1530 }
1531
1532 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1533 unsigned long hugetlb_total_pages(void)
1534 {
1535         struct hstate *h = &default_hstate;
1536         return h->nr_huge_pages * pages_per_huge_page(h);
1537 }
1538
1539 static int hugetlb_acct_memory(struct hstate *h, long delta)
1540 {
1541         int ret = -ENOMEM;
1542
1543         spin_lock(&hugetlb_lock);
1544         /*
1545          * When cpuset is configured, it breaks the strict hugetlb page
1546          * reservation as the accounting is done on a global variable. Such
1547          * reservation is completely rubbish in the presence of cpuset because
1548          * the reservation is not checked against page availability for the
1549          * current cpuset. Application can still potentially OOM'ed by kernel
1550          * with lack of free htlb page in cpuset that the task is in.
1551          * Attempt to enforce strict accounting with cpuset is almost
1552          * impossible (or too ugly) because cpuset is too fluid that
1553          * task or memory node can be dynamically moved between cpusets.
1554          *
1555          * The change of semantics for shared hugetlb mapping with cpuset is
1556          * undesirable. However, in order to preserve some of the semantics,
1557          * we fall back to check against current free page availability as
1558          * a best attempt and hopefully to minimize the impact of changing
1559          * semantics that cpuset has.
1560          */
1561         if (delta > 0) {
1562                 if (gather_surplus_pages(h, delta) < 0)
1563                         goto out;
1564
1565                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1566                         return_unused_surplus_pages(h, delta);
1567                         goto out;
1568                 }
1569         }
1570
1571         ret = 0;
1572         if (delta < 0)
1573                 return_unused_surplus_pages(h, (unsigned long) -delta);
1574
1575 out:
1576         spin_unlock(&hugetlb_lock);
1577         return ret;
1578 }
1579
1580 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1581 {
1582         struct resv_map *reservations = vma_resv_map(vma);
1583
1584         /*
1585          * This new VMA should share its siblings reservation map if present.
1586          * The VMA will only ever have a valid reservation map pointer where
1587          * it is being copied for another still existing VMA.  As that VMA
1588          * has a reference to the reservation map it cannot dissappear until
1589          * after this open call completes.  It is therefore safe to take a
1590          * new reference here without additional locking.
1591          */
1592         if (reservations)
1593                 kref_get(&reservations->refs);
1594 }
1595
1596 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1597 {
1598         struct hstate *h = hstate_vma(vma);
1599         struct resv_map *reservations = vma_resv_map(vma);
1600         unsigned long reserve;
1601         unsigned long start;
1602         unsigned long end;
1603
1604         if (reservations) {
1605                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1606                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1607
1608                 reserve = (end - start) -
1609                         region_count(&reservations->regions, start, end);
1610
1611                 kref_put(&reservations->refs, resv_map_release);
1612
1613                 if (reserve) {
1614                         hugetlb_acct_memory(h, -reserve);
1615                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1616                 }
1617         }
1618 }
1619
1620 /*
1621  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1622  * handle_mm_fault() to try to instantiate regular-sized pages in the
1623  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1624  * this far.
1625  */
1626 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1627 {
1628         BUG();
1629         return 0;
1630 }
1631
1632 struct vm_operations_struct hugetlb_vm_ops = {
1633         .fault = hugetlb_vm_op_fault,
1634         .open = hugetlb_vm_op_open,
1635         .close = hugetlb_vm_op_close,
1636 };
1637
1638 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1639                                 int writable)
1640 {
1641         pte_t entry;
1642
1643         if (writable) {
1644                 entry =
1645                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1646         } else {
1647                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1648         }
1649         entry = pte_mkyoung(entry);
1650         entry = pte_mkhuge(entry);
1651
1652         return entry;
1653 }
1654
1655 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1656                                    unsigned long address, pte_t *ptep)
1657 {
1658         pte_t entry;
1659
1660         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1661         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1662                 update_mmu_cache(vma, address, entry);
1663         }
1664 }
1665
1666
1667 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1668                             struct vm_area_struct *vma)
1669 {
1670         pte_t *src_pte, *dst_pte, entry;
1671         struct page *ptepage;
1672         unsigned long addr;
1673         int cow;
1674         struct hstate *h = hstate_vma(vma);
1675         unsigned long sz = huge_page_size(h);
1676
1677         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1678
1679         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1680                 src_pte = huge_pte_offset(src, addr);
1681                 if (!src_pte)
1682                         continue;
1683                 dst_pte = huge_pte_alloc(dst, addr, sz);
1684                 if (!dst_pte)
1685                         goto nomem;
1686
1687                 /* If the pagetables are shared don't copy or take references */
1688                 if (dst_pte == src_pte)
1689                         continue;
1690
1691                 spin_lock(&dst->page_table_lock);
1692                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1693                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1694                         if (cow)
1695                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1696                         entry = huge_ptep_get(src_pte);
1697                         ptepage = pte_page(entry);
1698                         get_page(ptepage);
1699                         set_huge_pte_at(dst, addr, dst_pte, entry);
1700                 }
1701                 spin_unlock(&src->page_table_lock);
1702                 spin_unlock(&dst->page_table_lock);
1703         }
1704         return 0;
1705
1706 nomem:
1707         return -ENOMEM;
1708 }
1709
1710 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1711                             unsigned long end, struct page *ref_page)
1712 {
1713         struct mm_struct *mm = vma->vm_mm;
1714         unsigned long address;
1715         pte_t *ptep;
1716         pte_t pte;
1717         struct page *page;
1718         struct page *tmp;
1719         struct hstate *h = hstate_vma(vma);
1720         unsigned long sz = huge_page_size(h);
1721
1722         /*
1723          * A page gathering list, protected by per file i_mmap_lock. The
1724          * lock is used to avoid list corruption from multiple unmapping
1725          * of the same page since we are using page->lru.
1726          */
1727         LIST_HEAD(page_list);
1728
1729         WARN_ON(!is_vm_hugetlb_page(vma));
1730         BUG_ON(start & ~huge_page_mask(h));
1731         BUG_ON(end & ~huge_page_mask(h));
1732
1733         mmu_notifier_invalidate_range_start(mm, start, end);
1734         spin_lock(&mm->page_table_lock);
1735         for (address = start; address < end; address += sz) {
1736                 ptep = huge_pte_offset(mm, address);
1737                 if (!ptep)
1738                         continue;
1739
1740                 if (huge_pmd_unshare(mm, &address, ptep))
1741                         continue;
1742
1743                 /*
1744                  * If a reference page is supplied, it is because a specific
1745                  * page is being unmapped, not a range. Ensure the page we
1746                  * are about to unmap is the actual page of interest.
1747                  */
1748                 if (ref_page) {
1749                         pte = huge_ptep_get(ptep);
1750                         if (huge_pte_none(pte))
1751                                 continue;
1752                         page = pte_page(pte);
1753                         if (page != ref_page)
1754                                 continue;
1755
1756                         /*
1757                          * Mark the VMA as having unmapped its page so that
1758                          * future faults in this VMA will fail rather than
1759                          * looking like data was lost
1760                          */
1761                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1762                 }
1763
1764                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1765                 if (huge_pte_none(pte))
1766                         continue;
1767
1768                 page = pte_page(pte);
1769                 if (pte_dirty(pte))
1770                         set_page_dirty(page);
1771                 list_add(&page->lru, &page_list);
1772         }
1773         spin_unlock(&mm->page_table_lock);
1774         flush_tlb_range(vma, start, end);
1775         mmu_notifier_invalidate_range_end(mm, start, end);
1776         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1777                 list_del(&page->lru);
1778                 put_page(page);
1779         }
1780 }
1781
1782 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1783                           unsigned long end, struct page *ref_page)
1784 {
1785         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1786         __unmap_hugepage_range(vma, start, end, ref_page);
1787         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1788 }
1789
1790 /*
1791  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1792  * mappping it owns the reserve page for. The intention is to unmap the page
1793  * from other VMAs and let the children be SIGKILLed if they are faulting the
1794  * same region.
1795  */
1796 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1797                                 struct page *page, unsigned long address)
1798 {
1799         struct vm_area_struct *iter_vma;
1800         struct address_space *mapping;
1801         struct prio_tree_iter iter;
1802         pgoff_t pgoff;
1803
1804         /*
1805          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1806          * from page cache lookup which is in HPAGE_SIZE units.
1807          */
1808         address = address & huge_page_mask(hstate_vma(vma));
1809         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1810                 + (vma->vm_pgoff >> PAGE_SHIFT);
1811         mapping = (struct address_space *)page_private(page);
1812
1813         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1814                 /* Do not unmap the current VMA */
1815                 if (iter_vma == vma)
1816                         continue;
1817
1818                 /*
1819                  * Unmap the page from other VMAs without their own reserves.
1820                  * They get marked to be SIGKILLed if they fault in these
1821                  * areas. This is because a future no-page fault on this VMA
1822                  * could insert a zeroed page instead of the data existing
1823                  * from the time of fork. This would look like data corruption
1824                  */
1825                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1826                         unmap_hugepage_range(iter_vma,
1827                                 address, address + HPAGE_SIZE,
1828                                 page);
1829         }
1830
1831         return 1;
1832 }
1833
1834 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1835                         unsigned long address, pte_t *ptep, pte_t pte,
1836                         struct page *pagecache_page)
1837 {
1838         struct hstate *h = hstate_vma(vma);
1839         struct page *old_page, *new_page;
1840         int avoidcopy;
1841         int outside_reserve = 0;
1842
1843         old_page = pte_page(pte);
1844
1845 retry_avoidcopy:
1846         /* If no-one else is actually using this page, avoid the copy
1847          * and just make the page writable */
1848         avoidcopy = (page_count(old_page) == 1);
1849         if (avoidcopy) {
1850                 set_huge_ptep_writable(vma, address, ptep);
1851                 return 0;
1852         }
1853
1854         /*
1855          * If the process that created a MAP_PRIVATE mapping is about to
1856          * perform a COW due to a shared page count, attempt to satisfy
1857          * the allocation without using the existing reserves. The pagecache
1858          * page is used to determine if the reserve at this address was
1859          * consumed or not. If reserves were used, a partial faulted mapping
1860          * at the time of fork() could consume its reserves on COW instead
1861          * of the full address range.
1862          */
1863         if (!(vma->vm_flags & VM_SHARED) &&
1864                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1865                         old_page != pagecache_page)
1866                 outside_reserve = 1;
1867
1868         page_cache_get(old_page);
1869         new_page = alloc_huge_page(vma, address, outside_reserve);
1870
1871         if (IS_ERR(new_page)) {
1872                 page_cache_release(old_page);
1873
1874                 /*
1875                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1876                  * it is due to references held by a child and an insufficient
1877                  * huge page pool. To guarantee the original mappers
1878                  * reliability, unmap the page from child processes. The child
1879                  * may get SIGKILLed if it later faults.
1880                  */
1881                 if (outside_reserve) {
1882                         BUG_ON(huge_pte_none(pte));
1883                         if (unmap_ref_private(mm, vma, old_page, address)) {
1884                                 BUG_ON(page_count(old_page) != 1);
1885                                 BUG_ON(huge_pte_none(pte));
1886                                 goto retry_avoidcopy;
1887                         }
1888                         WARN_ON_ONCE(1);
1889                 }
1890
1891                 return -PTR_ERR(new_page);
1892         }
1893
1894         spin_unlock(&mm->page_table_lock);
1895         copy_huge_page(new_page, old_page, address, vma);
1896         __SetPageUptodate(new_page);
1897         spin_lock(&mm->page_table_lock);
1898
1899         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1900         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1901                 /* Break COW */
1902                 huge_ptep_clear_flush(vma, address, ptep);
1903                 set_huge_pte_at(mm, address, ptep,
1904                                 make_huge_pte(vma, new_page, 1));
1905                 /* Make the old page be freed below */
1906                 new_page = old_page;
1907         }
1908         page_cache_release(new_page);
1909         page_cache_release(old_page);
1910         return 0;
1911 }
1912
1913 /* Return the pagecache page at a given address within a VMA */
1914 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1915                         struct vm_area_struct *vma, unsigned long address)
1916 {
1917         struct address_space *mapping;
1918         pgoff_t idx;
1919
1920         mapping = vma->vm_file->f_mapping;
1921         idx = vma_hugecache_offset(h, vma, address);
1922
1923         return find_lock_page(mapping, idx);
1924 }
1925
1926 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1927                         unsigned long address, pte_t *ptep, int write_access)
1928 {
1929         struct hstate *h = hstate_vma(vma);
1930         int ret = VM_FAULT_SIGBUS;
1931         pgoff_t idx;
1932         unsigned long size;
1933         struct page *page;
1934         struct address_space *mapping;
1935         pte_t new_pte;
1936
1937         /*
1938          * Currently, we are forced to kill the process in the event the
1939          * original mapper has unmapped pages from the child due to a failed
1940          * COW. Warn that such a situation has occured as it may not be obvious
1941          */
1942         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1943                 printk(KERN_WARNING
1944                         "PID %d killed due to inadequate hugepage pool\n",
1945                         current->pid);
1946                 return ret;
1947         }
1948
1949         mapping = vma->vm_file->f_mapping;
1950         idx = vma_hugecache_offset(h, vma, address);
1951
1952         /*
1953          * Use page lock to guard against racing truncation
1954          * before we get page_table_lock.
1955          */
1956 retry:
1957         page = find_lock_page(mapping, idx);
1958         if (!page) {
1959                 size = i_size_read(mapping->host) >> huge_page_shift(h);
1960                 if (idx >= size)
1961                         goto out;
1962                 page = alloc_huge_page(vma, address, 0);
1963                 if (IS_ERR(page)) {
1964                         ret = -PTR_ERR(page);
1965                         goto out;
1966                 }
1967                 clear_huge_page(page, address, huge_page_size(h));
1968                 __SetPageUptodate(page);
1969
1970                 if (vma->vm_flags & VM_SHARED) {
1971                         int err;
1972                         struct inode *inode = mapping->host;
1973
1974                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1975                         if (err) {
1976                                 put_page(page);
1977                                 if (err == -EEXIST)
1978                                         goto retry;
1979                                 goto out;
1980                         }
1981
1982                         spin_lock(&inode->i_lock);
1983                         inode->i_blocks += blocks_per_huge_page(h);
1984                         spin_unlock(&inode->i_lock);
1985                 } else
1986                         lock_page(page);
1987         }
1988
1989         /*
1990          * If we are going to COW a private mapping later, we examine the
1991          * pending reservations for this page now. This will ensure that
1992          * any allocations necessary to record that reservation occur outside
1993          * the spinlock.
1994          */
1995         if (write_access && !(vma->vm_flags & VM_SHARED))
1996                 if (vma_needs_reservation(h, vma, address) < 0) {
1997                         ret = VM_FAULT_OOM;
1998                         goto backout_unlocked;
1999                 }
2000
2001         spin_lock(&mm->page_table_lock);
2002         size = i_size_read(mapping->host) >> huge_page_shift(h);
2003         if (idx >= size)
2004                 goto backout;
2005
2006         ret = 0;
2007         if (!huge_pte_none(huge_ptep_get(ptep)))
2008                 goto backout;
2009
2010         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2011                                 && (vma->vm_flags & VM_SHARED)));
2012         set_huge_pte_at(mm, address, ptep, new_pte);
2013
2014         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2015                 /* Optimization, do the COW without a second fault */
2016                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2017         }
2018
2019         spin_unlock(&mm->page_table_lock);
2020         unlock_page(page);
2021 out:
2022         return ret;
2023
2024 backout:
2025         spin_unlock(&mm->page_table_lock);
2026 backout_unlocked:
2027         unlock_page(page);
2028         put_page(page);
2029         goto out;
2030 }
2031
2032 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2033                         unsigned long address, int write_access)
2034 {
2035         pte_t *ptep;
2036         pte_t entry;
2037         int ret;
2038         struct page *pagecache_page = NULL;
2039         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2040         struct hstate *h = hstate_vma(vma);
2041
2042         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2043         if (!ptep)
2044                 return VM_FAULT_OOM;
2045
2046         /*
2047          * Serialize hugepage allocation and instantiation, so that we don't
2048          * get spurious allocation failures if two CPUs race to instantiate
2049          * the same page in the page cache.
2050          */
2051         mutex_lock(&hugetlb_instantiation_mutex);
2052         entry = huge_ptep_get(ptep);
2053         if (huge_pte_none(entry)) {
2054                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2055                 goto out_mutex;
2056         }
2057
2058         ret = 0;
2059
2060         /*
2061          * If we are going to COW the mapping later, we examine the pending
2062          * reservations for this page now. This will ensure that any
2063          * allocations necessary to record that reservation occur outside the
2064          * spinlock. For private mappings, we also lookup the pagecache
2065          * page now as it is used to determine if a reservation has been
2066          * consumed.
2067          */
2068         if (write_access && !pte_write(entry)) {
2069                 if (vma_needs_reservation(h, vma, address) < 0) {
2070                         ret = VM_FAULT_OOM;
2071                         goto out_mutex;
2072                 }
2073
2074                 if (!(vma->vm_flags & VM_SHARED))
2075                         pagecache_page = hugetlbfs_pagecache_page(h,
2076                                                                 vma, address);
2077         }
2078
2079         spin_lock(&mm->page_table_lock);
2080         /* Check for a racing update before calling hugetlb_cow */
2081         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2082                 goto out_page_table_lock;
2083
2084
2085         if (write_access) {
2086                 if (!pte_write(entry)) {
2087                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2088                                                         pagecache_page);
2089                         goto out_page_table_lock;
2090                 }
2091                 entry = pte_mkdirty(entry);
2092         }
2093         entry = pte_mkyoung(entry);
2094         if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2095                 update_mmu_cache(vma, address, entry);
2096
2097 out_page_table_lock:
2098         spin_unlock(&mm->page_table_lock);
2099
2100         if (pagecache_page) {
2101                 unlock_page(pagecache_page);
2102                 put_page(pagecache_page);
2103         }
2104
2105 out_mutex:
2106         mutex_unlock(&hugetlb_instantiation_mutex);
2107
2108         return ret;
2109 }
2110
2111 /* Can be overriden by architectures */
2112 __attribute__((weak)) struct page *
2113 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2114                pud_t *pud, int write)
2115 {
2116         BUG();
2117         return NULL;
2118 }
2119
2120 static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2121 {
2122         if (!ptep || write || shared)
2123                 return 0;
2124         else
2125                 return huge_pte_none(huge_ptep_get(ptep));
2126 }
2127
2128 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2129                         struct page **pages, struct vm_area_struct **vmas,
2130                         unsigned long *position, int *length, int i,
2131                         int write)
2132 {
2133         unsigned long pfn_offset;
2134         unsigned long vaddr = *position;
2135         int remainder = *length;
2136         struct hstate *h = hstate_vma(vma);
2137         int zeropage_ok = 0;
2138         int shared = vma->vm_flags & VM_SHARED;
2139
2140         spin_lock(&mm->page_table_lock);
2141         while (vaddr < vma->vm_end && remainder) {
2142                 pte_t *pte;
2143                 struct page *page;
2144
2145                 /*
2146                  * Some archs (sparc64, sh*) have multiple pte_ts to
2147                  * each hugepage.  We have to make * sure we get the
2148                  * first, for the page indexing below to work.
2149                  */
2150                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2151                 if (huge_zeropage_ok(pte, write, shared))
2152                         zeropage_ok = 1;
2153
2154                 if (!pte ||
2155                     (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
2156                     (write && !pte_write(huge_ptep_get(pte)))) {
2157                         int ret;
2158
2159                         spin_unlock(&mm->page_table_lock);
2160                         ret = hugetlb_fault(mm, vma, vaddr, write);
2161                         spin_lock(&mm->page_table_lock);
2162                         if (!(ret & VM_FAULT_ERROR))
2163                                 continue;
2164
2165                         remainder = 0;
2166                         if (!i)
2167                                 i = -EFAULT;
2168                         break;
2169                 }
2170
2171                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2172                 page = pte_page(huge_ptep_get(pte));
2173 same_page:
2174                 if (pages) {
2175                         if (zeropage_ok)
2176                                 pages[i] = ZERO_PAGE(0);
2177                         else
2178                                 pages[i] = mem_map_offset(page, pfn_offset);
2179                         get_page(pages[i]);
2180                 }
2181
2182                 if (vmas)
2183                         vmas[i] = vma;
2184
2185                 vaddr += PAGE_SIZE;
2186                 ++pfn_offset;
2187                 --remainder;
2188                 ++i;
2189                 if (vaddr < vma->vm_end && remainder &&
2190                                 pfn_offset < pages_per_huge_page(h)) {
2191                         /*
2192                          * We use pfn_offset to avoid touching the pageframes
2193                          * of this compound page.
2194                          */
2195                         goto same_page;
2196                 }
2197         }
2198         spin_unlock(&mm->page_table_lock);
2199         *length = remainder;
2200         *position = vaddr;
2201
2202         return i;
2203 }
2204
2205 void hugetlb_change_protection(struct vm_area_struct *vma,
2206                 unsigned long address, unsigned long end, pgprot_t newprot)
2207 {
2208         struct mm_struct *mm = vma->vm_mm;
2209         unsigned long start = address;
2210         pte_t *ptep;
2211         pte_t pte;
2212         struct hstate *h = hstate_vma(vma);
2213
2214         BUG_ON(address >= end);
2215         flush_cache_range(vma, address, end);
2216
2217         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2218         spin_lock(&mm->page_table_lock);
2219         for (; address < end; address += huge_page_size(h)) {
2220                 ptep = huge_pte_offset(mm, address);
2221                 if (!ptep)
2222                         continue;
2223                 if (huge_pmd_unshare(mm, &address, ptep))
2224                         continue;
2225                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2226                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2227                         pte = pte_mkhuge(pte_modify(pte, newprot));
2228                         set_huge_pte_at(mm, address, ptep, pte);
2229                 }
2230         }
2231         spin_unlock(&mm->page_table_lock);
2232         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2233
2234         flush_tlb_range(vma, start, end);
2235 }
2236
2237 int hugetlb_reserve_pages(struct inode *inode,
2238                                         long from, long to,
2239                                         struct vm_area_struct *vma)
2240 {
2241         long ret, chg;
2242         struct hstate *h = hstate_inode(inode);
2243
2244         if (vma && vma->vm_flags & VM_NORESERVE)
2245                 return 0;
2246
2247         /*
2248          * Shared mappings base their reservation on the number of pages that
2249          * are already allocated on behalf of the file. Private mappings need
2250          * to reserve the full area even if read-only as mprotect() may be
2251          * called to make the mapping read-write. Assume !vma is a shm mapping
2252          */
2253         if (!vma || vma->vm_flags & VM_SHARED)
2254                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2255         else {
2256                 struct resv_map *resv_map = resv_map_alloc();
2257                 if (!resv_map)
2258                         return -ENOMEM;
2259
2260                 chg = to - from;
2261
2262                 set_vma_resv_map(vma, resv_map);
2263                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2264         }
2265
2266         if (chg < 0)
2267                 return chg;
2268
2269         if (hugetlb_get_quota(inode->i_mapping, chg))
2270                 return -ENOSPC;
2271         ret = hugetlb_acct_memory(h, chg);
2272         if (ret < 0) {
2273                 hugetlb_put_quota(inode->i_mapping, chg);
2274                 return ret;
2275         }
2276         if (!vma || vma->vm_flags & VM_SHARED)
2277                 region_add(&inode->i_mapping->private_list, from, to);
2278         return 0;
2279 }
2280
2281 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2282 {
2283         struct hstate *h = hstate_inode(inode);
2284         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2285
2286         spin_lock(&inode->i_lock);
2287         inode->i_blocks -= blocks_per_huge_page(h);
2288         spin_unlock(&inode->i_lock);
2289
2290         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2291         hugetlb_acct_memory(h, -(chg - freed));
2292 }