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  * Return the size of the pages allocated when backing a VMA. In the majority
 224  * cases this will be same size as used by the page table entries.
 225  */
 226 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 227 {
 228         struct hstate *hstate;
 229
 230         if (!is_vm_hugetlb_page(vma))
 231                 return PAGE_SIZE;
 232
 233         hstate = hstate_vma(vma);
 234
 235         return 1UL << (hstate->order + PAGE_SHIFT);
 236 }
 237
 238 /*
 239  * Return the page size being used by the MMU to back a VMA. In the majority
 240  * of cases, the page size used by the kernel matches the MMU size. On
 241  * architectures where it differs, an architecture-specific version of this
 242  * function is required.
 243  */
 244 #ifndef vma_mmu_pagesize
 245 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 246 {
 247         return vma_kernel_pagesize(vma);
 248 }
 249 #endif
 250
 251 /*
 252  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 253  * bits of the reservation map pointer, which are always clear due to
 254  * alignment.
 255  */
 256 #define HPAGE_RESV_OWNER    (1UL << 0)
 257 #define HPAGE_RESV_UNMAPPED (1UL << 1)
 258 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 259
 260 /*
 261  * These helpers are used to track how many pages are reserved for
 262  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 263  * is guaranteed to have their future faults succeed.
 264  *
 265  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 266  * the reserve counters are updated with the hugetlb_lock held. It is safe
 267  * to reset the VMA at fork() time as it is not in use yet and there is no
 268  * chance of the global counters getting corrupted as a result of the values.
 269  *
 270  * The private mapping reservation is represented in a subtly different
 271  * manner to a shared mapping.  A shared mapping has a region map associated
 272  * with the underlying file, this region map represents the backing file
 273  * pages which have ever had a reservation assigned which this persists even
 274  * after the page is instantiated.  A private mapping has a region map
 275  * associated with the original mmap which is attached to all VMAs which
 276  * reference it, this region map represents those offsets which have consumed
 277  * reservation ie. where pages have been instantiated.
 278  */
 279 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 280 {
 281         return (unsigned long)vma->vm_private_data;
 282 }
 283
 284 static void set_vma_private_data(struct vm_area_struct *vma,
 285                                                         unsigned long value)
 286 {
 287         vma->vm_private_data = (void *)value;
 288 }
 289
 290 struct resv_map {
 291         struct kref refs;
 292         struct list_head regions;
 293 };
 294
 295 static struct resv_map *resv_map_alloc(void)
 296 {
 297         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 298         if (!resv_map)
 299                 return NULL;
 300
 301         kref_init(&resv_map->refs);
 302         INIT_LIST_HEAD(&resv_map->regions);
 303
 304         return resv_map;
 305 }
 306
 307 static void resv_map_release(struct kref *ref)
 308 {
 309         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 310
 311         /* Clear out any active regions before we release the map. */
 312         region_truncate(&resv_map->regions, 0);
 313         kfree(resv_map);
 314 }
 315
 316 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 317 {
 318         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 319         if (!(vma->vm_flags & VM_SHARED))
 320                 return (struct resv_map *)(get_vma_private_data(vma) &
 321                                                         ~HPAGE_RESV_MASK);
 322         return NULL;
 323 }
 324
 325 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 326 {
 327         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 328         VM_BUG_ON(vma->vm_flags & VM_SHARED);
 329
 330         set_vma_private_data(vma, (get_vma_private_data(vma) &
 331                                 HPAGE_RESV_MASK) | (unsigned long)map);
 332 }
 333
 334 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 335 {
 336         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 337         VM_BUG_ON(vma->vm_flags & VM_SHARED);
 338
 339         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 340 }
 341
 342 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 343 {
 344         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 345
 346         return (get_vma_private_data(vma) & flag) != 0;
 347 }
 348
 349 /* Decrement the reserved pages in the hugepage pool by one */
 350 static void decrement_hugepage_resv_vma(struct hstate *h,
 351                         struct vm_area_struct *vma)
 352 {
 353         if (vma->vm_flags & VM_NORESERVE)
 354                 return;
 355
 356         if (vma->vm_flags & VM_SHARED) {
 357                 /* Shared mappings always use reserves */
 358                 h->resv_huge_pages--;
 359         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 360                 /*
 361                  * Only the process that called mmap() has reserves for
 362                  * private mappings.
 363                  */
 364                 h->resv_huge_pages--;
 365         }
 366 }
 367
 368 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 369 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 370 {
 371         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 372         if (!(vma->vm_flags & VM_SHARED))
 373                 vma->vm_private_data = (void *)0;
 374 }
 375
 376 /* Returns true if the VMA has associated reserve pages */
 377 static int vma_has_reserves(struct vm_area_struct *vma)
 378 {
 379         if (vma->vm_flags & VM_SHARED)
 380                 return 1;
 381         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 382                 return 1;
 383         return 0;
 384 }
 385
 386 static void clear_gigantic_page(struct page *page,
 387                         unsigned long addr, unsigned long sz)
 388 {
 389         int i;
 390         struct page *p = page;
 391
 392         might_sleep();
 393         for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
 394                 cond_resched();
 395                 clear_user_highpage(p, addr + i * PAGE_SIZE);
 396         }
 397 }
 398 static void clear_huge_page(struct page *page,
 399                         unsigned long addr, unsigned long sz)
 400 {
 401         int i;
 402
 403         if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
 404                 clear_gigantic_page(page, addr, sz);
 405                 return;
 406         }
 407
 408         might_sleep();
 409         for (i = 0; i < sz/PAGE_SIZE; i++) {
 410                 cond_resched();
 411                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
 412         }
 413 }
 414
 415 static void copy_gigantic_page(struct page *dst, struct page *src,
 416                            unsigned long addr, struct vm_area_struct *vma)
 417 {
 418         int i;
 419         struct hstate *h = hstate_vma(vma);
 420         struct page *dst_base = dst;
 421         struct page *src_base = src;
 422         might_sleep();
 423         for (i = 0; i < pages_per_huge_page(h); ) {
 424                 cond_resched();
 425                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
 426
 427                 i++;
 428                 dst = mem_map_next(dst, dst_base, i);
 429                 src = mem_map_next(src, src_base, i);
 430         }
 431 }
 432 static void copy_huge_page(struct page *dst, struct page *src,
 433                            unsigned long addr, struct vm_area_struct *vma)
 434 {
 435         int i;
 436         struct hstate *h = hstate_vma(vma);
 437
 438         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
 439                 copy_gigantic_page(dst, src, addr, vma);
 440                 return;
 441         }
 442
 443         might_sleep();
 444         for (i = 0; i < pages_per_huge_page(h); i++) {
 445                 cond_resched();
 446                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
 447         }
 448 }
 449
 450 static void enqueue_huge_page(struct hstate *h, struct page *page)
 451 {
 452         int nid = page_to_nid(page);
 453         list_add(&page->lru, &h->hugepage_freelists[nid]);
 454         h->free_huge_pages++;
 455         h->free_huge_pages_node[nid]++;
 456 }
 457
 458 static struct page *dequeue_huge_page(struct hstate *h)
 459 {
 460         int nid;
 461         struct page *page = NULL;
 462
 463         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
 464                 if (!list_empty(&h->hugepage_freelists[nid])) {
 465                         page = list_entry(h->hugepage_freelists[nid].next,
 466                                           struct page, lru);
 467                         list_del(&page->lru);
 468                         h->free_huge_pages--;
 469                         h->free_huge_pages_node[nid]--;
 470                         break;
 471                 }
 472         }
 473         return page;
 474 }
 475
 476 static struct page *dequeue_huge_page_vma(struct hstate *h,
 477                                 struct vm_area_struct *vma,
 478                                 unsigned long address, int avoid_reserve)
 479 {
 480         int nid;
 481         struct page *page = NULL;
 482         struct mempolicy *mpol;
 483         nodemask_t *nodemask;
 484         struct zonelist *zonelist = huge_zonelist(vma, address,
 485                                         htlb_alloc_mask, &mpol, &nodemask);
 486         struct zone *zone;
 487         struct zoneref *z;
 488
 489         /*
 490          * A child process with MAP_PRIVATE mappings created by their parent
 491          * have no page reserves. This check ensures that reservations are
 492          * not "stolen". The child may still get SIGKILLed
 493          */
 494         if (!vma_has_reserves(vma) &&
 495                         h->free_huge_pages - h->resv_huge_pages == 0)
 496                 return NULL;
 497
 498         /* If reserves cannot be used, ensure enough pages are in the pool */
 499         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 500                 return NULL;
 501
 502         for_each_zone_zonelist_nodemask(zone, z, zonelist,
 503                                                 MAX_NR_ZONES - 1, nodemask) {
 504                 nid = zone_to_nid(zone);
 505                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
 506                     !list_empty(&h->hugepage_freelists[nid])) {
 507                         page = list_entry(h->hugepage_freelists[nid].next,
 508                                           struct page, lru);
 509                         list_del(&page->lru);
 510                         h->free_huge_pages--;
 511                         h->free_huge_pages_node[nid]--;
 512
 513                         if (!avoid_reserve)
 514                                 decrement_hugepage_resv_vma(h, vma);
 515
 516                         break;
 517                 }
 518         }
 519         mpol_cond_put(mpol);
 520         return page;
 521 }
 522
 523 static void update_and_free_page(struct hstate *h, struct page *page)
 524 {
 525         int i;
 526
 527         VM_BUG_ON(h->order >= MAX_ORDER);
 528
 529         h->nr_huge_pages--;
 530         h->nr_huge_pages_node[page_to_nid(page)]--;
 531         for (i = 0; i < pages_per_huge_page(h); i++) {
 532                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 533                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 534                                 1 << PG_private | 1<< PG_writeback);
 535         }
 536         set_compound_page_dtor(page, NULL);
 537         set_page_refcounted(page);
 538         arch_release_hugepage(page);
 539         __free_pages(page, huge_page_order(h));
 540 }
 541
 542 struct hstate *size_to_hstate(unsigned long size)
 543 {
 544         struct hstate *h;
 545
 546         for_each_hstate(h) {
 547                 if (huge_page_size(h) == size)
 548                         return h;
 549         }
 550         return NULL;
 551 }
 552
 553 static void free_huge_page(struct page *page)
 554 {
 555         /*
 556          * Can't pass hstate in here because it is called from the
 557          * compound page destructor.
 558          */
 559         struct hstate *h = page_hstate(page);
 560         int nid = page_to_nid(page);
 561         struct address_space *mapping;
 562
 563         mapping = (struct address_space *) page_private(page);
 564         set_page_private(page, 0);
 565         BUG_ON(page_count(page));
 566         INIT_LIST_HEAD(&page->lru);
 567
 568         spin_lock(&hugetlb_lock);
 569         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 570                 update_and_free_page(h, page);
 571                 h->surplus_huge_pages--;
 572                 h->surplus_huge_pages_node[nid]--;
 573         } else {
 574                 enqueue_huge_page(h, page);
 575         }
 576         spin_unlock(&hugetlb_lock);
 577         if (mapping)
 578                 hugetlb_put_quota(mapping, 1);
 579 }
 580
 581 /*
 582  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 583  * balanced by operating on them in a round-robin fashion.
 584  * Returns 1 if an adjustment was made.
 585  */
 586 static int adjust_pool_surplus(struct hstate *h, int delta)
 587 {
 588         static int prev_nid;
 589         int nid = prev_nid;
 590         int ret = 0;
 591
 592         VM_BUG_ON(delta != -1 && delta != 1);
 593         do {
 594                 nid = next_node(nid, node_online_map);
 595                 if (nid == MAX_NUMNODES)
 596                         nid = first_node(node_online_map);
 597
 598                 /* To shrink on this node, there must be a surplus page */
 599                 if (delta < 0 && !h->surplus_huge_pages_node[nid])
 600                         continue;
 601                 /* Surplus cannot exceed the total number of pages */
 602                 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
 603                                                 h->nr_huge_pages_node[nid])
 604                         continue;
 605
 606                 h->surplus_huge_pages += delta;
 607                 h->surplus_huge_pages_node[nid] += delta;
 608                 ret = 1;
 609                 break;
 610         } while (nid != prev_nid);
 611
 612         prev_nid = nid;
 613         return ret;
 614 }
 615
 616 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 617 {
 618         set_compound_page_dtor(page, free_huge_page);
 619         spin_lock(&hugetlb_lock);
 620         h->nr_huge_pages++;
 621         h->nr_huge_pages_node[nid]++;
 622         spin_unlock(&hugetlb_lock);
 623         put_page(page); /* free it into the hugepage allocator */
 624 }
 625
 626 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 627 {
 628         struct page *page;
 629
 630         if (h->order >= MAX_ORDER)
 631                 return NULL;
 632
 633         page = alloc_pages_node(nid,
 634                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 635                                                 __GFP_REPEAT|__GFP_NOWARN,
 636                 huge_page_order(h));
 637         if (page) {
 638                 if (arch_prepare_hugepage(page)) {
 639                         __free_pages(page, huge_page_order(h));
 640                         return NULL;
 641                 }
 642                 prep_new_huge_page(h, page, nid);
 643         }
 644
 645         return page;
 646 }
 647
 648 /*
 649  * Use a helper variable to find the next node and then
 650  * copy it back to hugetlb_next_nid afterwards:
 651  * otherwise there's a window in which a racer might
 652  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 653  * But we don't need to use a spin_lock here: it really
 654  * doesn't matter if occasionally a racer chooses the
 655  * same nid as we do.  Move nid forward in the mask even
 656  * if we just successfully allocated a hugepage so that
 657  * the next caller gets hugepages on the next node.
 658  */
 659 static int hstate_next_node(struct hstate *h)
 660 {
 661         int next_nid;
 662         next_nid = next_node(h->hugetlb_next_nid, node_online_map);
 663         if (next_nid == MAX_NUMNODES)
 664                 next_nid = first_node(node_online_map);
 665         h->hugetlb_next_nid = next_nid;
 666         return next_nid;
 667 }
 668
 669 static int alloc_fresh_huge_page(struct hstate *h)
 670 {
 671         struct page *page;
 672         int start_nid;
 673         int next_nid;
 674         int ret = 0;
 675
 676         start_nid = h->hugetlb_next_nid;
 677
 678         do {
 679                 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
 680                 if (page)
 681                         ret = 1;
 682                 next_nid = hstate_next_node(h);
 683         } while (!page && h->hugetlb_next_nid != start_nid);
 684
 685         if (ret)
 686                 count_vm_event(HTLB_BUDDY_PGALLOC);
 687         else
 688                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 689
 690         return ret;
 691 }
 692
 693 static struct page *alloc_buddy_huge_page(struct hstate *h,
 694                         struct vm_area_struct *vma, unsigned long address)
 695 {
 696         struct page *page;
 697         unsigned int nid;
 698
 699         if (h->order >= MAX_ORDER)
 700                 return NULL;
 701
 702         /*
 703          * Assume we will successfully allocate the surplus page to
 704          * prevent racing processes from causing the surplus to exceed
 705          * overcommit
 706          *
 707          * This however introduces a different race, where a process B
 708          * tries to grow the static hugepage pool while alloc_pages() is
 709          * called by process A. B will only examine the per-node
 710          * counters in determining if surplus huge pages can be
 711          * converted to normal huge pages in adjust_pool_surplus(). A
 712          * won't be able to increment the per-node counter, until the
 713          * lock is dropped by B, but B doesn't drop hugetlb_lock until
 714          * no more huge pages can be converted from surplus to normal
 715          * state (and doesn't try to convert again). Thus, we have a
 716          * case where a surplus huge page exists, the pool is grown, and
 717          * the surplus huge page still exists after, even though it
 718          * should just have been converted to a normal huge page. This
 719          * does not leak memory, though, as the hugepage will be freed
 720          * once it is out of use. It also does not allow the counters to
 721          * go out of whack in adjust_pool_surplus() as we don't modify
 722          * the node values until we've gotten the hugepage and only the
 723          * per-node value is checked there.
 724          */
 725         spin_lock(&hugetlb_lock);
 726         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 727                 spin_unlock(&hugetlb_lock);
 728                 return NULL;
 729         } else {
 730                 h->nr_huge_pages++;
 731                 h->surplus_huge_pages++;
 732         }
 733         spin_unlock(&hugetlb_lock);
 734
 735         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 736                                         __GFP_REPEAT|__GFP_NOWARN,
 737                                         huge_page_order(h));
 738
 739         if (page && arch_prepare_hugepage(page)) {
 740                 __free_pages(page, huge_page_order(h));
 741                 return NULL;
 742         }
 743
 744         spin_lock(&hugetlb_lock);
 745         if (page) {
 746                 /*
 747                  * This page is now managed by the hugetlb allocator and has
 748                  * no users -- drop the buddy allocator's reference.
 749                  */
 750                 put_page_testzero(page);
 751                 VM_BUG_ON(page_count(page));
 752                 nid = page_to_nid(page);
 753                 set_compound_page_dtor(page, free_huge_page);
 754                 /*
 755                  * We incremented the global counters already
 756                  */
 757                 h->nr_huge_pages_node[nid]++;
 758                 h->surplus_huge_pages_node[nid]++;
 759                 __count_vm_event(HTLB_BUDDY_PGALLOC);
 760         } else {
 761                 h->nr_huge_pages--;
 762                 h->surplus_huge_pages--;
 763                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 764         }
 765         spin_unlock(&hugetlb_lock);
 766
 767         return page;
 768 }
 769
 770 /*
 771  * Increase the hugetlb pool such that it can accomodate a reservation
 772  * of size 'delta'.
 773  */
 774 static int gather_surplus_pages(struct hstate *h, int delta)
 775 {
 776         struct list_head surplus_list;
 777         struct page *page, *tmp;
 778         int ret, i;
 779         int needed, allocated;
 780
 781         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
 782         if (needed <= 0) {
 783                 h->resv_huge_pages += delta;
 784                 return 0;
 785         }
 786
 787         allocated = 0;
 788         INIT_LIST_HEAD(&surplus_list);
 789
 790         ret = -ENOMEM;
 791 retry:
 792         spin_unlock(&hugetlb_lock);
 793         for (i = 0; i < needed; i++) {
 794                 page = alloc_buddy_huge_page(h, NULL, 0);
 795                 if (!page) {
 796                         /*
 797                          * We were not able to allocate enough pages to
 798                          * satisfy the entire reservation so we free what
 799                          * we've allocated so far.
 800                          */
 801                         spin_lock(&hugetlb_lock);
 802                         needed = 0;
 803                         goto free;
 804                 }
 805
 806                 list_add(&page->lru, &surplus_list);
 807         }
 808         allocated += needed;
 809
 810         /*
 811          * After retaking hugetlb_lock, we need to recalculate 'needed'
 812          * because either resv_huge_pages or free_huge_pages may have changed.
 813          */
 814         spin_lock(&hugetlb_lock);
 815         needed = (h->resv_huge_pages + delta) -
 816                         (h->free_huge_pages + allocated);
 817         if (needed > 0)
 818                 goto retry;
 819
 820         /*
 821          * The surplus_list now contains _at_least_ the number of extra pages
 822          * needed to accomodate the reservation.  Add the appropriate number
 823          * of pages to the hugetlb pool and free the extras back to the buddy
 824          * allocator.  Commit the entire reservation here to prevent another
 825          * process from stealing the pages as they are added to the pool but
 826          * before they are reserved.
 827          */
 828         needed += allocated;
 829         h->resv_huge_pages += delta;
 830         ret = 0;
 831 free:
 832         /* Free the needed pages to the hugetlb pool */
 833         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 834                 if ((--needed) < 0)
 835                         break;
 836                 list_del(&page->lru);
 837                 enqueue_huge_page(h, page);
 838         }
 839
 840         /* Free unnecessary surplus pages to the buddy allocator */
 841         if (!list_empty(&surplus_list)) {
 842                 spin_unlock(&hugetlb_lock);
 843                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 844                         list_del(&page->lru);
 845                         /*
 846                          * The page has a reference count of zero already, so
 847                          * call free_huge_page directly instead of using
 848                          * put_page.  This must be done with hugetlb_lock
 849                          * unlocked which is safe because free_huge_page takes
 850                          * hugetlb_lock before deciding how to free the page.
 851                          */
 852                         free_huge_page(page);
 853                 }
 854                 spin_lock(&hugetlb_lock);
 855         }
 856
 857         return ret;
 858 }
 859
 860 /*
 861  * When releasing a hugetlb pool reservation, any surplus pages that were
 862  * allocated to satisfy the reservation must be explicitly freed if they were
 863  * never used.
 864  */
 865 static void return_unused_surplus_pages(struct hstate *h,
 866                                         unsigned long unused_resv_pages)
 867 {
 868         static int nid = -1;
 869         struct page *page;
 870         unsigned long nr_pages;
 871
 872         /*
 873          * We want to release as many surplus pages as possible, spread
 874          * evenly across all nodes. Iterate across all nodes until we
 875          * can no longer free unreserved surplus pages. This occurs when
 876          * the nodes with surplus pages have no free pages.
 877          */
 878         unsigned long remaining_iterations = num_online_nodes();
 879
 880         /* Uncommit the reservation */
 881         h->resv_huge_pages -= unused_resv_pages;
 882
 883         /* Cannot return gigantic pages currently */
 884         if (h->order >= MAX_ORDER)
 885                 return;
 886
 887         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
 888
 889         while (remaining_iterations-- && nr_pages) {
 890                 nid = next_node(nid, node_online_map);
 891                 if (nid == MAX_NUMNODES)
 892                         nid = first_node(node_online_map);
 893
 894                 if (!h->surplus_huge_pages_node[nid])
 895                         continue;
 896
 897                 if (!list_empty(&h->hugepage_freelists[nid])) {
 898                         page = list_entry(h->hugepage_freelists[nid].next,
 899                                           struct page, lru);
 900                         list_del(&page->lru);
 901                         update_and_free_page(h, page);
 902                         h->free_huge_pages--;
 903                         h->free_huge_pages_node[nid]--;
 904                         h->surplus_huge_pages--;
 905                         h->surplus_huge_pages_node[nid]--;
 906                         nr_pages--;
 907                         remaining_iterations = num_online_nodes();
 908                 }
 909         }
 910 }
 911
 912 /*
 913  * Determine if the huge page at addr within the vma has an associated
 914  * reservation.  Where it does not we will need to logically increase
 915  * reservation and actually increase quota before an allocation can occur.
 916  * Where any new reservation would be required the reservation change is
 917  * prepared, but not committed.  Once the page has been quota'd allocated
 918  * an instantiated the change should be committed via vma_commit_reservation.
 919  * No action is required on failure.
 920  */
 921 static long vma_needs_reservation(struct hstate *h,
 922                         struct vm_area_struct *vma, unsigned long addr)
 923 {
 924         struct address_space *mapping = vma->vm_file->f_mapping;
 925         struct inode *inode = mapping->host;
 926
 927         if (vma->vm_flags & VM_SHARED) {
 928                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 929                 return region_chg(&inode->i_mapping->private_list,
 930                                                         idx, idx + 1);
 931
 932         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 933                 return 1;
 934
 935         } else  {
 936                 long err;
 937                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 938                 struct resv_map *reservations = vma_resv_map(vma);
 939
 940                 err = region_chg(&reservations->regions, idx, idx + 1);
 941                 if (err < 0)
 942                         return err;
 943                 return 0;
 944         }
 945 }
 946 static void vma_commit_reservation(struct hstate *h,
 947                         struct vm_area_struct *vma, unsigned long addr)
 948 {
 949         struct address_space *mapping = vma->vm_file->f_mapping;
 950         struct inode *inode = mapping->host;
 951
 952         if (vma->vm_flags & VM_SHARED) {
 953                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 954                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
 955
 956         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 957                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
 958                 struct resv_map *reservations = vma_resv_map(vma);
 959
 960                 /* Mark this page used in the map. */
 961                 region_add(&reservations->regions, idx, idx + 1);
 962         }
 963 }
 964
 965 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 966                                     unsigned long addr, int avoid_reserve)
 967 {
 968         struct hstate *h = hstate_vma(vma);
 969         struct page *page;
 970         struct address_space *mapping = vma->vm_file->f_mapping;
 971         struct inode *inode = mapping->host;
 972         long chg;
 973
 974         /*
 975          * Processes that did not create the mapping will have no reserves and
 976          * will not have accounted against quota. Check that the quota can be
 977          * made before satisfying the allocation
 978          * MAP_NORESERVE mappings may also need pages and quota allocated
 979          * if no reserve mapping overlaps.
 980          */
 981         chg = vma_needs_reservation(h, vma, addr);
 982         if (chg < 0)
 983                 return ERR_PTR(chg);
 984         if (chg)
 985                 if (hugetlb_get_quota(inode->i_mapping, chg))
 986                         return ERR_PTR(-ENOSPC);
 987
 988         spin_lock(&hugetlb_lock);
 989         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
 990         spin_unlock(&hugetlb_lock);
 991
 992         if (!page) {
 993                 page = alloc_buddy_huge_page(h, vma, addr);
 994                 if (!page) {
 995                         hugetlb_put_quota(inode->i_mapping, chg);
 996                         return ERR_PTR(-VM_FAULT_OOM);
 997                 }
 998         }
 999
1000         set_page_refcounted(page);
1001         set_page_private(page, (unsigned long) mapping);
1002
1003         vma_commit_reservation(h, vma, addr);
1004
1005         return page;
1006 }
1007
1008 int __weak alloc_bootmem_huge_page(struct hstate *h)
1009 {
1010         struct huge_bootmem_page *m;
1011         int nr_nodes = nodes_weight(node_online_map);
1012
1013         while (nr_nodes) {
1014                 void *addr;
1015
1016                 addr = __alloc_bootmem_node_nopanic(
1017                                 NODE_DATA(h->hugetlb_next_nid),
1018                                 huge_page_size(h), huge_page_size(h), 0);
1019
1020                 if (addr) {
1021                         /*
1022                          * Use the beginning of the huge page to store the
1023                          * huge_bootmem_page struct (until gather_bootmem
1024                          * puts them into the mem_map).
1025                          */
1026                         m = addr;
1027                         goto found;
1028                 }
1029                 hstate_next_node(h);
1030                 nr_nodes--;
1031         }
1032         return 0;
1033
1034 found:
1035         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1036         /* Put them into a private list first because mem_map is not up yet */
1037         list_add(&m->list, &huge_boot_pages);
1038         m->hstate = h;
1039         return 1;
1040 }
1041
1042 static void prep_compound_huge_page(struct page *page, int order)
1043 {
1044         if (unlikely(order > (MAX_ORDER - 1)))
1045                 prep_compound_gigantic_page(page, order);
1046         else
1047                 prep_compound_page(page, order);
1048 }
1049
1050 /* Put bootmem huge pages into the standard lists after mem_map is up */
1051 static void __init gather_bootmem_prealloc(void)
1052 {
1053         struct huge_bootmem_page *m;
1054
1055         list_for_each_entry(m, &huge_boot_pages, list) {
1056                 struct page *page = virt_to_page(m);
1057                 struct hstate *h = m->hstate;
1058                 __ClearPageReserved(page);
1059                 WARN_ON(page_count(page) != 1);
1060                 prep_compound_huge_page(page, h->order);
1061                 prep_new_huge_page(h, page, page_to_nid(page));
1062         }
1063 }
1064
1065 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1066 {
1067         unsigned long i;
1068
1069         for (i = 0; i < h->max_huge_pages; ++i) {
1070                 if (h->order >= MAX_ORDER) {
1071                         if (!alloc_bootmem_huge_page(h))
1072                                 break;
1073                 } else if (!alloc_fresh_huge_page(h))
1074                         break;
1075         }
1076         h->max_huge_pages = i;
1077 }
1078
1079 static void __init hugetlb_init_hstates(void)
1080 {
1081         struct hstate *h;
1082
1083         for_each_hstate(h) {
1084                 /* oversize hugepages were init'ed in early boot */
1085                 if (h->order < MAX_ORDER)
1086                         hugetlb_hstate_alloc_pages(h);
1087         }
1088 }
1089
1090 static char * __init memfmt(char *buf, unsigned long n)
1091 {
1092         if (n >= (1UL << 30))
1093                 sprintf(buf, "%lu GB", n >> 30);
1094         else if (n >= (1UL << 20))
1095                 sprintf(buf, "%lu MB", n >> 20);
1096         else
1097                 sprintf(buf, "%lu KB", n >> 10);
1098         return buf;
1099 }
1100
1101 static void __init report_hugepages(void)
1102 {
1103         struct hstate *h;
1104
1105         for_each_hstate(h) {
1106                 char buf[32];
1107                 printk(KERN_INFO "HugeTLB registered %s page size, "
1108                                  "pre-allocated %ld pages\n",
1109                         memfmt(buf, huge_page_size(h)),
1110                         h->free_huge_pages);
1111         }
1112 }
1113
1114 #ifdef CONFIG_HIGHMEM
1115 static void try_to_free_low(struct hstate *h, unsigned long count)
1116 {
1117         int i;
1118
1119         if (h->order >= MAX_ORDER)
1120                 return;
1121
1122         for (i = 0; i < MAX_NUMNODES; ++i) {
1123                 struct page *page, *next;
1124                 struct list_head *freel = &h->hugepage_freelists[i];
1125                 list_for_each_entry_safe(page, next, freel, lru) {
1126                         if (count >= h->nr_huge_pages)
1127                                 return;
1128                         if (PageHighMem(page))
1129                                 continue;
1130                         list_del(&page->lru);
1131                         update_and_free_page(h, page);
1132                         h->free_huge_pages--;
1133                         h->free_huge_pages_node[page_to_nid(page)]--;
1134                 }
1135         }
1136 }
1137 #else
1138 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1139 {
1140 }
1141 #endif
1142
1143 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1144 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1145 {
1146         unsigned long min_count, ret;
1147
1148         if (h->order >= MAX_ORDER)
1149                 return h->max_huge_pages;
1150
1151         /*
1152          * Increase the pool size
1153          * First take pages out of surplus state.  Then make up the
1154          * remaining difference by allocating fresh huge pages.
1155          *
1156          * We might race with alloc_buddy_huge_page() here and be unable
1157          * to convert a surplus huge page to a normal huge page. That is
1158          * not critical, though, it just means the overall size of the
1159          * pool might be one hugepage larger than it needs to be, but
1160          * within all the constraints specified by the sysctls.
1161          */
1162         spin_lock(&hugetlb_lock);
1163         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1164                 if (!adjust_pool_surplus(h, -1))
1165                         break;
1166         }
1167
1168         while (count > persistent_huge_pages(h)) {
1169                 /*
1170                  * If this allocation races such that we no longer need the
1171                  * page, free_huge_page will handle it by freeing the page
1172                  * and reducing the surplus.
1173                  */
1174                 spin_unlock(&hugetlb_lock);
1175                 ret = alloc_fresh_huge_page(h);
1176                 spin_lock(&hugetlb_lock);
1177                 if (!ret)
1178                         goto out;
1179
1180         }
1181
1182         /*
1183          * Decrease the pool size
1184          * First return free pages to the buddy allocator (being careful
1185          * to keep enough around to satisfy reservations).  Then place
1186          * pages into surplus state as needed so the pool will shrink
1187          * to the desired size as pages become free.
1188          *
1189          * By placing pages into the surplus state independent of the
1190          * overcommit value, we are allowing the surplus pool size to
1191          * exceed overcommit. There are few sane options here. Since
1192          * alloc_buddy_huge_page() is checking the global counter,
1193          * though, we'll note that we're not allowed to exceed surplus
1194          * and won't grow the pool anywhere else. Not until one of the
1195          * sysctls are changed, or the surplus pages go out of use.
1196          */
1197         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1198         min_count = max(count, min_count);
1199         try_to_free_low(h, min_count);
1200         while (min_count < persistent_huge_pages(h)) {
1201                 struct page *page = dequeue_huge_page(h);
1202                 if (!page)
1203                         break;
1204                 update_and_free_page(h, page);
1205         }
1206         while (count < persistent_huge_pages(h)) {
1207                 if (!adjust_pool_surplus(h, 1))
1208                         break;
1209         }
1210 out:
1211         ret = persistent_huge_pages(h);
1212         spin_unlock(&hugetlb_lock);
1213         return ret;
1214 }
1215
1216 #define HSTATE_ATTR_RO(_name) \
1217         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1218
1219 #define HSTATE_ATTR(_name) \
1220         static struct kobj_attribute _name##_attr = \
1221                 __ATTR(_name, 0644, _name##_show, _name##_store)
1222
1223 static struct kobject *hugepages_kobj;
1224 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1225
1226 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1227 {
1228         int i;
1229         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1230                 if (hstate_kobjs[i] == kobj)
1231                         return &hstates[i];
1232         BUG();
1233         return NULL;
1234 }
1235
1236 static ssize_t nr_hugepages_show(struct kobject *kobj,
1237                                         struct kobj_attribute *attr, char *buf)
1238 {
1239         struct hstate *h = kobj_to_hstate(kobj);
1240         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1241 }
1242 static ssize_t nr_hugepages_store(struct kobject *kobj,
1243                 struct kobj_attribute *attr, const char *buf, size_t count)
1244 {
1245         int err;
1246         unsigned long input;
1247         struct hstate *h = kobj_to_hstate(kobj);
1248
1249         err = strict_strtoul(buf, 10, &input);
1250         if (err)
1251                 return 0;
1252
1253         h->max_huge_pages = set_max_huge_pages(h, input);
1254
1255         return count;
1256 }
1257 HSTATE_ATTR(nr_hugepages);
1258
1259 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1260                                         struct kobj_attribute *attr, char *buf)
1261 {
1262         struct hstate *h = kobj_to_hstate(kobj);
1263         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1264 }
1265 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1266                 struct kobj_attribute *attr, const char *buf, size_t count)
1267 {
1268         int err;
1269         unsigned long input;
1270         struct hstate *h = kobj_to_hstate(kobj);
1271
1272         err = strict_strtoul(buf, 10, &input);
1273         if (err)
1274                 return 0;
1275
1276         spin_lock(&hugetlb_lock);
1277         h->nr_overcommit_huge_pages = input;
1278         spin_unlock(&hugetlb_lock);
1279
1280         return count;
1281 }
1282 HSTATE_ATTR(nr_overcommit_hugepages);
1283
1284 static ssize_t free_hugepages_show(struct kobject *kobj,
1285                                         struct kobj_attribute *attr, char *buf)
1286 {
1287         struct hstate *h = kobj_to_hstate(kobj);
1288         return sprintf(buf, "%lu\n", h->free_huge_pages);
1289 }
1290 HSTATE_ATTR_RO(free_hugepages);
1291
1292 static ssize_t resv_hugepages_show(struct kobject *kobj,
1293                                         struct kobj_attribute *attr, char *buf)
1294 {
1295         struct hstate *h = kobj_to_hstate(kobj);
1296         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1297 }
1298 HSTATE_ATTR_RO(resv_hugepages);
1299
1300 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1301                                         struct kobj_attribute *attr, char *buf)
1302 {
1303         struct hstate *h = kobj_to_hstate(kobj);
1304         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1305 }
1306 HSTATE_ATTR_RO(surplus_hugepages);
1307
1308 static struct attribute *hstate_attrs[] = {
1309         &nr_hugepages_attr.attr,
1310         &nr_overcommit_hugepages_attr.attr,
1311         &free_hugepages_attr.attr,
1312         &resv_hugepages_attr.attr,
1313         &surplus_hugepages_attr.attr,
1314         NULL,
1315 };
1316
1317 static struct attribute_group hstate_attr_group = {
1318         .attrs = hstate_attrs,
1319 };
1320
1321 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1322 {
1323         int retval;
1324
1325         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1326                                                         hugepages_kobj);
1327         if (!hstate_kobjs[h - hstates])
1328                 return -ENOMEM;
1329
1330         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1331                                                         &hstate_attr_group);
1332         if (retval)
1333                 kobject_put(hstate_kobjs[h - hstates]);
1334
1335         return retval;
1336 }
1337
1338 static void __init hugetlb_sysfs_init(void)
1339 {
1340         struct hstate *h;
1341         int err;
1342
1343         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1344         if (!hugepages_kobj)
1345                 return;
1346
1347         for_each_hstate(h) {
1348                 err = hugetlb_sysfs_add_hstate(h);
1349                 if (err)
1350                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1351                                                                 h->name);
1352         }
1353 }
1354
1355 static void __exit hugetlb_exit(void)
1356 {
1357         struct hstate *h;
1358
1359         for_each_hstate(h) {
1360                 kobject_put(hstate_kobjs[h - hstates]);
1361         }
1362
1363         kobject_put(hugepages_kobj);
1364 }
1365 module_exit(hugetlb_exit);
1366
1367 static int __init hugetlb_init(void)
1368 {
1369         /* Some platform decide whether they support huge pages at boot
1370          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1371          * there is no such support
1372          */
1373         if (HPAGE_SHIFT == 0)
1374                 return 0;
1375
1376         if (!size_to_hstate(default_hstate_size)) {
1377                 default_hstate_size = HPAGE_SIZE;
1378                 if (!size_to_hstate(default_hstate_size))
1379                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1380         }
1381         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1382         if (default_hstate_max_huge_pages)
1383                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1384
1385         hugetlb_init_hstates();
1386
1387         gather_bootmem_prealloc();
1388
1389         report_hugepages();
1390
1391         hugetlb_sysfs_init();
1392
1393         return 0;
1394 }
1395 module_init(hugetlb_init);
1396
1397 /* Should be called on processing a hugepagesz=... option */
1398 void __init hugetlb_add_hstate(unsigned order)
1399 {
1400         struct hstate *h;
1401         unsigned long i;
1402
1403         if (size_to_hstate(PAGE_SIZE << order)) {
1404                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1405                 return;
1406         }
1407         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1408         BUG_ON(order == 0);
1409         h = &hstates[max_hstate++];
1410         h->order = order;
1411         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1412         h->nr_huge_pages = 0;
1413         h->free_huge_pages = 0;
1414         for (i = 0; i < MAX_NUMNODES; ++i)
1415                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1416         h->hugetlb_next_nid = first_node(node_online_map);
1417         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1418                                         huge_page_size(h)/1024);
1419
1420         parsed_hstate = h;
1421 }
1422
1423 static int __init hugetlb_nrpages_setup(char *s)
1424 {
1425         unsigned long *mhp;
1426         static unsigned long *last_mhp;
1427
1428         /*
1429          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1430          * so this hugepages= parameter goes to the "default hstate".
1431          */
1432         if (!max_hstate)
1433                 mhp = &default_hstate_max_huge_pages;
1434         else
1435                 mhp = &parsed_hstate->max_huge_pages;
1436
1437         if (mhp == last_mhp) {
1438                 printk(KERN_WARNING "hugepages= specified twice without "
1439                         "interleaving hugepagesz=, ignoring\n");
1440                 return 1;
1441         }
1442
1443         if (sscanf(s, "%lu", mhp) <= 0)
1444                 *mhp = 0;
1445
1446         /*
1447          * Global state is always initialized later in hugetlb_init.
1448          * But we need to allocate >= MAX_ORDER hstates here early to still
1449          * use the bootmem allocator.
1450          */
1451         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1452                 hugetlb_hstate_alloc_pages(parsed_hstate);
1453
1454         last_mhp = mhp;
1455
1456         return 1;
1457 }
1458 __setup("hugepages=", hugetlb_nrpages_setup);
1459
1460 static int __init hugetlb_default_setup(char *s)
1461 {
1462         default_hstate_size = memparse(s, &s);
1463         return 1;
1464 }
1465 __setup("default_hugepagesz=", hugetlb_default_setup);
1466
1467 static unsigned int cpuset_mems_nr(unsigned int *array)
1468 {
1469         int node;
1470         unsigned int nr = 0;
1471
1472         for_each_node_mask(node, cpuset_current_mems_allowed)
1473                 nr += array[node];
1474
1475         return nr;
1476 }
1477
1478 #ifdef CONFIG_SYSCTL
1479 int hugetlb_sysctl_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->max_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                 h->max_huge_pages = set_max_huge_pages(h, tmp);
1495
1496         return 0;
1497 }
1498
1499 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1500                         struct file *file, void __user *buffer,
1501                         size_t *length, loff_t *ppos)
1502 {
1503         proc_dointvec(table, write, file, buffer, length, ppos);
1504         if (hugepages_treat_as_movable)
1505                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1506         else
1507                 htlb_alloc_mask = GFP_HIGHUSER;
1508         return 0;
1509 }
1510
1511 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1512                         struct file *file, void __user *buffer,
1513                         size_t *length, loff_t *ppos)
1514 {
1515         struct hstate *h = &default_hstate;
1516         unsigned long tmp;
1517
1518         if (!write)
1519                 tmp = h->nr_overcommit_huge_pages;
1520
1521         table->data = &tmp;
1522         table->maxlen = sizeof(unsigned long);
1523         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1524
1525         if (write) {
1526                 spin_lock(&hugetlb_lock);
1527                 h->nr_overcommit_huge_pages = tmp;
1528                 spin_unlock(&hugetlb_lock);
1529         }
1530
1531         return 0;
1532 }
1533
1534 #endif /* CONFIG_SYSCTL */
1535
1536 void hugetlb_report_meminfo(struct seq_file *m)
1537 {
1538         struct hstate *h = &default_hstate;
1539         seq_printf(m,
1540                         "HugePages_Total:   %5lu\n"
1541                         "HugePages_Free:    %5lu\n"
1542                         "HugePages_Rsvd:    %5lu\n"
1543                         "HugePages_Surp:    %5lu\n"
1544                         "Hugepagesize:   %8lu kB\n",
1545                         h->nr_huge_pages,
1546                         h->free_huge_pages,
1547                         h->resv_huge_pages,
1548                         h->surplus_huge_pages,
1549                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1550 }
1551
1552 int hugetlb_report_node_meminfo(int nid, char *buf)
1553 {
1554         struct hstate *h = &default_hstate;
1555         return sprintf(buf,
1556                 "Node %d HugePages_Total: %5u\n"
1557                 "Node %d HugePages_Free:  %5u\n"
1558                 "Node %d HugePages_Surp:  %5u\n",
1559                 nid, h->nr_huge_pages_node[nid],
1560                 nid, h->free_huge_pages_node[nid],
1561                 nid, h->surplus_huge_pages_node[nid]);
1562 }
1563
1564 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1565 unsigned long hugetlb_total_pages(void)
1566 {
1567         struct hstate *h = &default_hstate;
1568         return h->nr_huge_pages * pages_per_huge_page(h);
1569 }
1570
1571 static int hugetlb_acct_memory(struct hstate *h, long delta)
1572 {
1573         int ret = -ENOMEM;
1574
1575         spin_lock(&hugetlb_lock);
1576         /*
1577          * When cpuset is configured, it breaks the strict hugetlb page
1578          * reservation as the accounting is done on a global variable. Such
1579          * reservation is completely rubbish in the presence of cpuset because
1580          * the reservation is not checked against page availability for the
1581          * current cpuset. Application can still potentially OOM'ed by kernel
1582          * with lack of free htlb page in cpuset that the task is in.
1583          * Attempt to enforce strict accounting with cpuset is almost
1584          * impossible (or too ugly) because cpuset is too fluid that
1585          * task or memory node can be dynamically moved between cpusets.
1586          *
1587          * The change of semantics for shared hugetlb mapping with cpuset is
1588          * undesirable. However, in order to preserve some of the semantics,
1589          * we fall back to check against current free page availability as
1590          * a best attempt and hopefully to minimize the impact of changing
1591          * semantics that cpuset has.
1592          */
1593         if (delta > 0) {
1594                 if (gather_surplus_pages(h, delta) < 0)
1595                         goto out;
1596
1597                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1598                         return_unused_surplus_pages(h, delta);
1599                         goto out;
1600                 }
1601         }
1602
1603         ret = 0;
1604         if (delta < 0)
1605                 return_unused_surplus_pages(h, (unsigned long) -delta);
1606
1607 out:
1608         spin_unlock(&hugetlb_lock);
1609         return ret;
1610 }
1611
1612 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1613 {
1614         struct resv_map *reservations = vma_resv_map(vma);
1615
1616         /*
1617          * This new VMA should share its siblings reservation map if present.
1618          * The VMA will only ever have a valid reservation map pointer where
1619          * it is being copied for another still existing VMA.  As that VMA
1620          * has a reference to the reservation map it cannot dissappear until
1621          * after this open call completes.  It is therefore safe to take a
1622          * new reference here without additional locking.
1623          */
1624         if (reservations)
1625                 kref_get(&reservations->refs);
1626 }
1627
1628 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1629 {
1630         struct hstate *h = hstate_vma(vma);
1631         struct resv_map *reservations = vma_resv_map(vma);
1632         unsigned long reserve;
1633         unsigned long start;
1634         unsigned long end;
1635
1636         if (reservations) {
1637                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1638                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1639
1640                 reserve = (end - start) -
1641                         region_count(&reservations->regions, start, end);
1642
1643                 kref_put(&reservations->refs, resv_map_release);
1644
1645                 if (reserve) {
1646                         hugetlb_acct_memory(h, -reserve);
1647                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1648                 }
1649         }
1650 }
1651
1652 /*
1653  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1654  * handle_mm_fault() to try to instantiate regular-sized pages in the
1655  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1656  * this far.
1657  */
1658 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1659 {
1660         BUG();
1661         return 0;
1662 }
1663
1664 struct vm_operations_struct hugetlb_vm_ops = {
1665         .fault = hugetlb_vm_op_fault,
1666         .open = hugetlb_vm_op_open,
1667         .close = hugetlb_vm_op_close,
1668 };
1669
1670 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1671                                 int writable)
1672 {
1673         pte_t entry;
1674
1675         if (writable) {
1676                 entry =
1677                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1678         } else {
1679                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1680         }
1681         entry = pte_mkyoung(entry);
1682         entry = pte_mkhuge(entry);
1683
1684         return entry;
1685 }
1686
1687 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1688                                    unsigned long address, pte_t *ptep)
1689 {
1690         pte_t entry;
1691
1692         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1693         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1694                 update_mmu_cache(vma, address, entry);
1695         }
1696 }
1697
1698
1699 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1700                             struct vm_area_struct *vma)
1701 {
1702         pte_t *src_pte, *dst_pte, entry;
1703         struct page *ptepage;
1704         unsigned long addr;
1705         int cow;
1706         struct hstate *h = hstate_vma(vma);
1707         unsigned long sz = huge_page_size(h);
1708
1709         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1710
1711         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1712                 src_pte = huge_pte_offset(src, addr);
1713                 if (!src_pte)
1714                         continue;
1715                 dst_pte = huge_pte_alloc(dst, addr, sz);
1716                 if (!dst_pte)
1717                         goto nomem;
1718
1719                 /* If the pagetables are shared don't copy or take references */
1720                 if (dst_pte == src_pte)
1721                         continue;
1722
1723                 spin_lock(&dst->page_table_lock);
1724                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1725                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1726                         if (cow)
1727                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1728                         entry = huge_ptep_get(src_pte);
1729                         ptepage = pte_page(entry);
1730                         get_page(ptepage);
1731                         set_huge_pte_at(dst, addr, dst_pte, entry);
1732                 }
1733                 spin_unlock(&src->page_table_lock);
1734                 spin_unlock(&dst->page_table_lock);
1735         }
1736         return 0;
1737
1738 nomem:
1739         return -ENOMEM;
1740 }
1741
1742 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1743                             unsigned long end, struct page *ref_page)
1744 {
1745         struct mm_struct *mm = vma->vm_mm;
1746         unsigned long address;
1747         pte_t *ptep;
1748         pte_t pte;
1749         struct page *page;
1750         struct page *tmp;
1751         struct hstate *h = hstate_vma(vma);
1752         unsigned long sz = huge_page_size(h);
1753
1754         /*
1755          * A page gathering list, protected by per file i_mmap_lock. The
1756          * lock is used to avoid list corruption from multiple unmapping
1757          * of the same page since we are using page->lru.
1758          */
1759         LIST_HEAD(page_list);
1760
1761         WARN_ON(!is_vm_hugetlb_page(vma));
1762         BUG_ON(start & ~huge_page_mask(h));
1763         BUG_ON(end & ~huge_page_mask(h));
1764
1765         mmu_notifier_invalidate_range_start(mm, start, end);
1766         spin_lock(&mm->page_table_lock);
1767         for (address = start; address < end; address += sz) {
1768                 ptep = huge_pte_offset(mm, address);
1769                 if (!ptep)
1770                         continue;
1771
1772                 if (huge_pmd_unshare(mm, &address, ptep))
1773                         continue;
1774
1775                 /*
1776                  * If a reference page is supplied, it is because a specific
1777                  * page is being unmapped, not a range. Ensure the page we
1778                  * are about to unmap is the actual page of interest.
1779                  */
1780                 if (ref_page) {
1781                         pte = huge_ptep_get(ptep);
1782                         if (huge_pte_none(pte))
1783                                 continue;
1784                         page = pte_page(pte);
1785                         if (page != ref_page)
1786                                 continue;
1787
1788                         /*
1789                          * Mark the VMA as having unmapped its page so that
1790                          * future faults in this VMA will fail rather than
1791                          * looking like data was lost
1792                          */
1793                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1794                 }
1795
1796                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1797                 if (huge_pte_none(pte))
1798                         continue;
1799
1800                 page = pte_page(pte);
1801                 if (pte_dirty(pte))
1802                         set_page_dirty(page);
1803                 list_add(&page->lru, &page_list);
1804         }
1805         spin_unlock(&mm->page_table_lock);
1806         flush_tlb_range(vma, start, end);
1807         mmu_notifier_invalidate_range_end(mm, start, end);
1808         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1809                 list_del(&page->lru);
1810                 put_page(page);
1811         }
1812 }
1813
1814 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1815                           unsigned long end, struct page *ref_page)
1816 {
1817         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1818         __unmap_hugepage_range(vma, start, end, ref_page);
1819         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1820 }
1821
1822 /*
1823  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1824  * mappping it owns the reserve page for. The intention is to unmap the page
1825  * from other VMAs and let the children be SIGKILLed if they are faulting the
1826  * same region.
1827  */
1828 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1829                                 struct page *page, unsigned long address)
1830 {
1831         struct hstate *h = hstate_vma(vma);
1832         struct vm_area_struct *iter_vma;
1833         struct address_space *mapping;
1834         struct prio_tree_iter iter;
1835         pgoff_t pgoff;
1836
1837         /*
1838          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1839          * from page cache lookup which is in HPAGE_SIZE units.
1840          */
1841         address = address & huge_page_mask(h);
1842         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1843                 + (vma->vm_pgoff >> PAGE_SHIFT);
1844         mapping = (struct address_space *)page_private(page);
1845
1846         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1847                 /* Do not unmap the current VMA */
1848                 if (iter_vma == vma)
1849                         continue;
1850
1851                 /*
1852                  * Unmap the page from other VMAs without their own reserves.
1853                  * They get marked to be SIGKILLed if they fault in these
1854                  * areas. This is because a future no-page fault on this VMA
1855                  * could insert a zeroed page instead of the data existing
1856                  * from the time of fork. This would look like data corruption
1857                  */
1858                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1859                         unmap_hugepage_range(iter_vma,
1860                                 address, address + huge_page_size(h),
1861                                 page);
1862         }
1863
1864         return 1;
1865 }
1866
1867 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1868                         unsigned long address, pte_t *ptep, pte_t pte,
1869                         struct page *pagecache_page)
1870 {
1871         struct hstate *h = hstate_vma(vma);
1872         struct page *old_page, *new_page;
1873         int avoidcopy;
1874         int outside_reserve = 0;
1875
1876         old_page = pte_page(pte);
1877
1878 retry_avoidcopy:
1879         /* If no-one else is actually using this page, avoid the copy
1880          * and just make the page writable */
1881         avoidcopy = (page_count(old_page) == 1);
1882         if (avoidcopy) {
1883                 set_huge_ptep_writable(vma, address, ptep);
1884                 return 0;
1885         }
1886
1887         /*
1888          * If the process that created a MAP_PRIVATE mapping is about to
1889          * perform a COW due to a shared page count, attempt to satisfy
1890          * the allocation without using the existing reserves. The pagecache
1891          * page is used to determine if the reserve at this address was
1892          * consumed or not. If reserves were used, a partial faulted mapping
1893          * at the time of fork() could consume its reserves on COW instead
1894          * of the full address range.
1895          */
1896         if (!(vma->vm_flags & VM_SHARED) &&
1897                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1898                         old_page != pagecache_page)
1899                 outside_reserve = 1;
1900
1901         page_cache_get(old_page);
1902         new_page = alloc_huge_page(vma, address, outside_reserve);
1903
1904         if (IS_ERR(new_page)) {
1905                 page_cache_release(old_page);
1906
1907                 /*
1908                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1909                  * it is due to references held by a child and an insufficient
1910                  * huge page pool. To guarantee the original mappers
1911                  * reliability, unmap the page from child processes. The child
1912                  * may get SIGKILLed if it later faults.
1913                  */
1914                 if (outside_reserve) {
1915                         BUG_ON(huge_pte_none(pte));
1916                         if (unmap_ref_private(mm, vma, old_page, address)) {
1917                                 BUG_ON(page_count(old_page) != 1);
1918                                 BUG_ON(huge_pte_none(pte));
1919                                 goto retry_avoidcopy;
1920                         }
1921                         WARN_ON_ONCE(1);
1922                 }
1923
1924                 return -PTR_ERR(new_page);
1925         }
1926
1927         spin_unlock(&mm->page_table_lock);
1928         copy_huge_page(new_page, old_page, address, vma);
1929         __SetPageUptodate(new_page);
1930         spin_lock(&mm->page_table_lock);
1931
1932         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1933         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1934                 /* Break COW */
1935                 huge_ptep_clear_flush(vma, address, ptep);
1936                 set_huge_pte_at(mm, address, ptep,
1937                                 make_huge_pte(vma, new_page, 1));
1938                 /* Make the old page be freed below */
1939                 new_page = old_page;
1940         }
1941         page_cache_release(new_page);
1942         page_cache_release(old_page);
1943         return 0;
1944 }
1945
1946 /* Return the pagecache page at a given address within a VMA */
1947 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1948                         struct vm_area_struct *vma, unsigned long address)
1949 {
1950         struct address_space *mapping;
1951         pgoff_t idx;
1952
1953         mapping = vma->vm_file->f_mapping;
1954         idx = vma_hugecache_offset(h, vma, address);
1955
1956         return find_lock_page(mapping, idx);
1957 }
1958
1959 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1960                         unsigned long address, pte_t *ptep, int write_access)
1961 {
1962         struct hstate *h = hstate_vma(vma);
1963         int ret = VM_FAULT_SIGBUS;
1964         pgoff_t idx;
1965         unsigned long size;
1966         struct page *page;
1967         struct address_space *mapping;
1968         pte_t new_pte;
1969
1970         /*
1971          * Currently, we are forced to kill the process in the event the
1972          * original mapper has unmapped pages from the child due to a failed
1973          * COW. Warn that such a situation has occured as it may not be obvious
1974          */
1975         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1976                 printk(KERN_WARNING
1977                         "PID %d killed due to inadequate hugepage pool\n",
1978                         current->pid);
1979                 return ret;
1980         }
1981
1982         mapping = vma->vm_file->f_mapping;
1983         idx = vma_hugecache_offset(h, vma, address);
1984
1985         /*
1986          * Use page lock to guard against racing truncation
1987          * before we get page_table_lock.
1988          */
1989 retry:
1990         page = find_lock_page(mapping, idx);
1991         if (!page) {
1992                 size = i_size_read(mapping->host) >> huge_page_shift(h);
1993                 if (idx >= size)
1994                         goto out;
1995                 page = alloc_huge_page(vma, address, 0);
1996                 if (IS_ERR(page)) {
1997                         ret = -PTR_ERR(page);
1998                         goto out;
1999                 }
2000                 clear_huge_page(page, address, huge_page_size(h));
2001                 __SetPageUptodate(page);
2002
2003                 if (vma->vm_flags & VM_SHARED) {
2004                         int err;
2005                         struct inode *inode = mapping->host;
2006
2007                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2008                         if (err) {
2009                                 put_page(page);
2010                                 if (err == -EEXIST)
2011                                         goto retry;
2012                                 goto out;
2013                         }
2014
2015                         spin_lock(&inode->i_lock);
2016                         inode->i_blocks += blocks_per_huge_page(h);
2017                         spin_unlock(&inode->i_lock);
2018                 } else
2019                         lock_page(page);
2020         }
2021
2022         /*
2023          * If we are going to COW a private mapping later, we examine the
2024          * pending reservations for this page now. This will ensure that
2025          * any allocations necessary to record that reservation occur outside
2026          * the spinlock.
2027          */
2028         if (write_access && !(vma->vm_flags & VM_SHARED))
2029                 if (vma_needs_reservation(h, vma, address) < 0) {
2030                         ret = VM_FAULT_OOM;
2031                         goto backout_unlocked;
2032                 }
2033
2034         spin_lock(&mm->page_table_lock);
2035         size = i_size_read(mapping->host) >> huge_page_shift(h);
2036         if (idx >= size)
2037                 goto backout;
2038
2039         ret = 0;
2040         if (!huge_pte_none(huge_ptep_get(ptep)))
2041                 goto backout;
2042
2043         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2044                                 && (vma->vm_flags & VM_SHARED)));
2045         set_huge_pte_at(mm, address, ptep, new_pte);
2046
2047         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2048                 /* Optimization, do the COW without a second fault */
2049                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2050         }
2051
2052         spin_unlock(&mm->page_table_lock);
2053         unlock_page(page);
2054 out:
2055         return ret;
2056
2057 backout:
2058         spin_unlock(&mm->page_table_lock);
2059 backout_unlocked:
2060         unlock_page(page);
2061         put_page(page);
2062         goto out;
2063 }
2064
2065 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2066                         unsigned long address, int write_access)
2067 {
2068         pte_t *ptep;
2069         pte_t entry;
2070         int ret;
2071         struct page *pagecache_page = NULL;
2072         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2073         struct hstate *h = hstate_vma(vma);
2074
2075         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2076         if (!ptep)
2077                 return VM_FAULT_OOM;
2078
2079         /*
2080          * Serialize hugepage allocation and instantiation, so that we don't
2081          * get spurious allocation failures if two CPUs race to instantiate
2082          * the same page in the page cache.
2083          */
2084         mutex_lock(&hugetlb_instantiation_mutex);
2085         entry = huge_ptep_get(ptep);
2086         if (huge_pte_none(entry)) {
2087                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2088                 goto out_mutex;
2089         }
2090
2091         ret = 0;
2092
2093         /*
2094          * If we are going to COW the mapping later, we examine the pending
2095          * reservations for this page now. This will ensure that any
2096          * allocations necessary to record that reservation occur outside the
2097          * spinlock. For private mappings, we also lookup the pagecache
2098          * page now as it is used to determine if a reservation has been
2099          * consumed.
2100          */
2101         if (write_access && !pte_write(entry)) {
2102                 if (vma_needs_reservation(h, vma, address) < 0) {
2103                         ret = VM_FAULT_OOM;
2104                         goto out_mutex;
2105                 }
2106
2107                 if (!(vma->vm_flags & VM_SHARED))
2108                         pagecache_page = hugetlbfs_pagecache_page(h,
2109                                                                 vma, address);
2110         }
2111
2112         spin_lock(&mm->page_table_lock);
2113         /* Check for a racing update before calling hugetlb_cow */
2114         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2115                 goto out_page_table_lock;
2116
2117
2118         if (write_access) {
2119                 if (!pte_write(entry)) {
2120                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2121                                                         pagecache_page);
2122                         goto out_page_table_lock;
2123                 }
2124                 entry = pte_mkdirty(entry);
2125         }
2126         entry = pte_mkyoung(entry);
2127         if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2128                 update_mmu_cache(vma, address, entry);
2129
2130 out_page_table_lock:
2131         spin_unlock(&mm->page_table_lock);
2132
2133         if (pagecache_page) {
2134                 unlock_page(pagecache_page);
2135                 put_page(pagecache_page);
2136         }
2137
2138 out_mutex:
2139         mutex_unlock(&hugetlb_instantiation_mutex);
2140
2141         return ret;
2142 }
2143
2144 /* Can be overriden by architectures */
2145 __attribute__((weak)) struct page *
2146 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2147                pud_t *pud, int write)
2148 {
2149         BUG();
2150         return NULL;
2151 }
2152
2153 static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2154 {
2155         if (!ptep || write || shared)
2156                 return 0;
2157         else
2158                 return huge_pte_none(huge_ptep_get(ptep));
2159 }
2160
2161 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2162                         struct page **pages, struct vm_area_struct **vmas,
2163                         unsigned long *position, int *length, int i,
2164                         int write)
2165 {
2166         unsigned long pfn_offset;
2167         unsigned long vaddr = *position;
2168         int remainder = *length;
2169         struct hstate *h = hstate_vma(vma);
2170         int zeropage_ok = 0;
2171         int shared = vma->vm_flags & VM_SHARED;
2172
2173         spin_lock(&mm->page_table_lock);
2174         while (vaddr < vma->vm_end && remainder) {
2175                 pte_t *pte;
2176                 struct page *page;
2177
2178                 /*
2179                  * Some archs (sparc64, sh*) have multiple pte_ts to
2180                  * each hugepage.  We have to make * sure we get the
2181                  * first, for the page indexing below to work.
2182                  */
2183                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2184                 if (huge_zeropage_ok(pte, write, shared))
2185                         zeropage_ok = 1;
2186
2187                 if (!pte ||
2188                     (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
2189                     (write && !pte_write(huge_ptep_get(pte)))) {
2190                         int ret;
2191
2192                         spin_unlock(&mm->page_table_lock);
2193                         ret = hugetlb_fault(mm, vma, vaddr, write);
2194                         spin_lock(&mm->page_table_lock);
2195                         if (!(ret & VM_FAULT_ERROR))
2196                                 continue;
2197
2198                         remainder = 0;
2199                         if (!i)
2200                                 i = -EFAULT;
2201                         break;
2202                 }
2203
2204                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2205                 page = pte_page(huge_ptep_get(pte));
2206 same_page:
2207                 if (pages) {
2208                         if (zeropage_ok)
2209                                 pages[i] = ZERO_PAGE(0);
2210                         else
2211                                 pages[i] = mem_map_offset(page, pfn_offset);
2212                         get_page(pages[i]);
2213                 }
2214
2215                 if (vmas)
2216                         vmas[i] = vma;
2217
2218                 vaddr += PAGE_SIZE;
2219                 ++pfn_offset;
2220                 --remainder;
2221                 ++i;
2222                 if (vaddr < vma->vm_end && remainder &&
2223                                 pfn_offset < pages_per_huge_page(h)) {
2224                         /*
2225                          * We use pfn_offset to avoid touching the pageframes
2226                          * of this compound page.
2227                          */
2228                         goto same_page;
2229                 }
2230         }
2231         spin_unlock(&mm->page_table_lock);
2232         *length = remainder;
2233         *position = vaddr;
2234
2235         return i;
2236 }
2237
2238 void hugetlb_change_protection(struct vm_area_struct *vma,
2239                 unsigned long address, unsigned long end, pgprot_t newprot)
2240 {
2241         struct mm_struct *mm = vma->vm_mm;
2242         unsigned long start = address;
2243         pte_t *ptep;
2244         pte_t pte;
2245         struct hstate *h = hstate_vma(vma);
2246
2247         BUG_ON(address >= end);
2248         flush_cache_range(vma, address, end);
2249
2250         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2251         spin_lock(&mm->page_table_lock);
2252         for (; address < end; address += huge_page_size(h)) {
2253                 ptep = huge_pte_offset(mm, address);
2254                 if (!ptep)
2255                         continue;
2256                 if (huge_pmd_unshare(mm, &address, ptep))
2257                         continue;
2258                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2259                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2260                         pte = pte_mkhuge(pte_modify(pte, newprot));
2261                         set_huge_pte_at(mm, address, ptep, pte);
2262                 }
2263         }
2264         spin_unlock(&mm->page_table_lock);
2265         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2266
2267         flush_tlb_range(vma, start, end);
2268 }
2269
2270 int hugetlb_reserve_pages(struct inode *inode,
2271                                         long from, long to,
2272                                         struct vm_area_struct *vma,
2273                                         int acctflag)
2274 {
2275         long ret, chg;
2276         struct hstate *h = hstate_inode(inode);
2277
2278         /*
2279          * Only apply hugepage reservation if asked. At fault time, an
2280          * attempt will be made for VM_NORESERVE to allocate a page
2281          * and filesystem quota without using reserves
2282          */
2283         if (acctflag & VM_NORESERVE)
2284                 return 0;
2285
2286         /*
2287          * Shared mappings base their reservation on the number of pages that
2288          * are already allocated on behalf of the file. Private mappings need
2289          * to reserve the full area even if read-only as mprotect() may be
2290          * called to make the mapping read-write. Assume !vma is a shm mapping
2291          */
2292         if (!vma || vma->vm_flags & VM_SHARED)
2293                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2294         else {
2295                 struct resv_map *resv_map = resv_map_alloc();
2296                 if (!resv_map)
2297                         return -ENOMEM;
2298
2299                 chg = to - from;
2300
2301                 set_vma_resv_map(vma, resv_map);
2302                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2303         }
2304
2305         if (chg < 0)
2306                 return chg;
2307
2308         /* There must be enough filesystem quota for the mapping */
2309         if (hugetlb_get_quota(inode->i_mapping, chg))
2310                 return -ENOSPC;
2311
2312         /*
2313          * Check enough hugepages are available for the reservation.
2314          * Hand back the quota if there are not
2315          */
2316         ret = hugetlb_acct_memory(h, chg);
2317         if (ret < 0) {
2318                 hugetlb_put_quota(inode->i_mapping, chg);
2319                 return ret;
2320         }
2321
2322         /*
2323          * Account for the reservations made. Shared mappings record regions
2324          * that have reservations as they are shared by multiple VMAs.
2325          * When the last VMA disappears, the region map says how much
2326          * the reservation was and the page cache tells how much of
2327          * the reservation was consumed. Private mappings are per-VMA and
2328          * only the consumed reservations are tracked. When the VMA
2329          * disappears, the original reservation is the VMA size and the
2330          * consumed reservations are stored in the map. Hence, nothing
2331          * else has to be done for private mappings here
2332          */
2333         if (!vma || vma->vm_flags & VM_SHARED)
2334                 region_add(&inode->i_mapping->private_list, from, to);
2335         return 0;
2336 }
2337
2338 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2339 {
2340         struct hstate *h = hstate_inode(inode);
2341         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2342
2343         spin_lock(&inode->i_lock);
2344         inode->i_blocks -= blocks_per_huge_page(h);
2345         spin_unlock(&inode->i_lock);
2346
2347         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2348         hugetlb_acct_memory(h, -(chg - freed));
2349 }