mm/hugetlb.c

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