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