mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) Nadia Yvette Chambers, April 2004
   4  */
   5 #include <linux/list.h>
   6 #include <linux/init.h>
   7 #include <linux/module.h>
   8 #include <linux/mm.h>
   9 #include <linux/seq_file.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/highmem.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/nodemask.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/mempolicy.h>
  16 #include <linux/cpuset.h>
  17 #include <linux/mutex.h>
  18 #include <linux/bootmem.h>
  19 #include <linux/sysfs.h>
  20 #include <linux/slab.h>
  21 #include <linux/rmap.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24
  25 #include <asm/page.h>
  26 #include <asm/pgtable.h>
  27 #include <asm/tlb.h>
  28
  29 #include <linux/io.h>
  30 #include <linux/hugetlb.h>
  31 #include <linux/hugetlb_cgroup.h>
  32 #include <linux/node.h>
  33 #include "internal.h"
  34
  35 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  36 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  37 unsigned long hugepages_treat_as_movable;
  38
  39 int hugetlb_max_hstate __read_mostly;
  40 unsigned int default_hstate_idx;
  41 struct hstate hstates[HUGE_MAX_HSTATE];
  42
  43 __initdata LIST_HEAD(huge_boot_pages);
  44
  45 /* for command line parsing */
  46 static struct hstate * __initdata parsed_hstate;
  47 static unsigned long __initdata default_hstate_max_huge_pages;
  48 static unsigned long __initdata default_hstate_size;
  49
  50 /*
  51  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  52  */
  53 DEFINE_SPINLOCK(hugetlb_lock);
  54
  55 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
  56 {
  57         bool free = (spool->count == 0) && (spool->used_hpages == 0);
  58
  59         spin_unlock(&spool->lock);
  60
  61         /* If no pages are used, and no other handles to the subpool
  62          * remain, free the subpool the subpool remain */
  63         if (free)
  64                 kfree(spool);
  65 }
  66
  67 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
  68 {
  69         struct hugepage_subpool *spool;
  70
  71         spool = kmalloc(sizeof(*spool), GFP_KERNEL);
  72         if (!spool)
  73                 return NULL;
  74
  75         spin_lock_init(&spool->lock);
  76         spool->count = 1;
  77         spool->max_hpages = nr_blocks;
  78         spool->used_hpages = 0;
  79
  80         return spool;
  81 }
  82
  83 void hugepage_put_subpool(struct hugepage_subpool *spool)
  84 {
  85         spin_lock(&spool->lock);
  86         BUG_ON(!spool->count);
  87         spool->count--;
  88         unlock_or_release_subpool(spool);
  89 }
  90
  91 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
  92                                       long delta)
  93 {
  94         int ret = 0;
  95
  96         if (!spool)
  97                 return 0;
  98
  99         spin_lock(&spool->lock);
 100         if ((spool->used_hpages + delta) <= spool->max_hpages) {
 101                 spool->used_hpages += delta;
 102         } else {
 103                 ret = -ENOMEM;
 104         }
 105         spin_unlock(&spool->lock);
 106
 107         return ret;
 108 }
 109
 110 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
 111                                        long delta)
 112 {
 113         if (!spool)
 114                 return;
 115
 116         spin_lock(&spool->lock);
 117         spool->used_hpages -= delta;
 118         /* If hugetlbfs_put_super couldn't free spool due to
 119         * an outstanding quota reference, free it now. */
 120         unlock_or_release_subpool(spool);
 121 }
 122
 123 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
 124 {
 125         return HUGETLBFS_SB(inode->i_sb)->spool;
 126 }
 127
 128 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
 129 {
 130         return subpool_inode(file_inode(vma->vm_file));
 131 }
 132
 133 /*
 134  * Region tracking -- allows tracking of reservations and instantiated pages
 135  *                    across the pages in a mapping.
 136  *
 137  * The region data structures are protected by a combination of the mmap_sem
 138  * and the hugetlb_instantiation_mutex.  To access or modify a region the caller
 139  * must either hold the mmap_sem for write, or the mmap_sem for read and
 140  * the hugetlb_instantiation_mutex:
 141  *
 142  *      down_write(&mm->mmap_sem);
 143  * or
 144  *      down_read(&mm->mmap_sem);
 145  *      mutex_lock(&hugetlb_instantiation_mutex);
 146  */
 147 struct file_region {
 148         struct list_head link;
 149         long from;
 150         long to;
 151 };
 152
 153 static long region_add(struct list_head *head, long f, long t)
 154 {
 155         struct file_region *rg, *nrg, *trg;
 156
 157         /* Locate the region we are either in or before. */
 158         list_for_each_entry(rg, head, link)
 159                 if (f <= rg->to)
 160                         break;
 161
 162         /* Round our left edge to the current segment if it encloses us. */
 163         if (f > rg->from)
 164                 f = rg->from;
 165
 166         /* Check for and consume any regions we now overlap with. */
 167         nrg = rg;
 168         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 169                 if (&rg->link == head)
 170                         break;
 171                 if (rg->from > t)
 172                         break;
 173
 174                 /* If this area reaches higher then extend our area to
 175                  * include it completely.  If this is not the first area
 176                  * which we intend to reuse, free it. */
 177                 if (rg->to > t)
 178                         t = rg->to;
 179                 if (rg != nrg) {
 180                         list_del(&rg->link);
 181                         kfree(rg);
 182                 }
 183         }
 184         nrg->from = f;
 185         nrg->to = t;
 186         return 0;
 187 }
 188
 189 static long region_chg(struct list_head *head, long f, long t)
 190 {
 191         struct file_region *rg, *nrg;
 192         long chg = 0;
 193
 194         /* Locate the region we are before or in. */
 195         list_for_each_entry(rg, head, link)
 196                 if (f <= rg->to)
 197                         break;
 198
 199         /* If we are below the current region then a new region is required.
 200          * Subtle, allocate a new region at the position but make it zero
 201          * size such that we can guarantee to record the reservation. */
 202         if (&rg->link == head || t < rg->from) {
 203                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 204                 if (!nrg)
 205                         return -ENOMEM;
 206                 nrg->from = f;
 207                 nrg->to   = f;
 208                 INIT_LIST_HEAD(&nrg->link);
 209                 list_add(&nrg->link, rg->link.prev);
 210
 211                 return t - f;
 212         }
 213
 214         /* Round our left edge to the current segment if it encloses us. */
 215         if (f > rg->from)
 216                 f = rg->from;
 217         chg = t - f;
 218
 219         /* Check for and consume any regions we now overlap with. */
 220         list_for_each_entry(rg, rg->link.prev, link) {
 221                 if (&rg->link == head)
 222                         break;
 223                 if (rg->from > t)
 224                         return chg;
 225
 226                 /* We overlap with this area, if it extends further than
 227                  * us then we must extend ourselves.  Account for its
 228                  * existing reservation. */
 229                 if (rg->to > t) {
 230                         chg += rg->to - t;
 231                         t = rg->to;
 232                 }
 233                 chg -= rg->to - rg->from;
 234         }
 235         return chg;
 236 }
 237
 238 static long region_truncate(struct list_head *head, long end)
 239 {
 240         struct file_region *rg, *trg;
 241         long chg = 0;
 242
 243         /* Locate the region we are either in or before. */
 244         list_for_each_entry(rg, head, link)
 245                 if (end <= rg->to)
 246                         break;
 247         if (&rg->link == head)
 248                 return 0;
 249
 250         /* If we are in the middle of a region then adjust it. */
 251         if (end > rg->from) {
 252                 chg = rg->to - end;
 253                 rg->to = end;
 254                 rg = list_entry(rg->link.next, typeof(*rg), link);
 255         }
 256
 257         /* Drop any remaining regions. */
 258         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 259                 if (&rg->link == head)
 260                         break;
 261                 chg += rg->to - rg->from;
 262                 list_del(&rg->link);
 263                 kfree(rg);
 264         }
 265         return chg;
 266 }
 267
 268 static long region_count(struct list_head *head, long f, long t)
 269 {
 270         struct file_region *rg;
 271         long chg = 0;
 272
 273         /* Locate each segment we overlap with, and count that overlap. */
 274         list_for_each_entry(rg, head, link) {
 275                 long seg_from;
 276                 long seg_to;
 277
 278                 if (rg->to <= f)
 279                         continue;
 280                 if (rg->from >= t)
 281                         break;
 282
 283                 seg_from = max(rg->from, f);
 284                 seg_to = min(rg->to, t);
 285
 286                 chg += seg_to - seg_from;
 287         }
 288
 289         return chg;
 290 }
 291
 292 /*
 293  * Convert the address within this vma to the page offset within
 294  * the mapping, in pagecache page units; huge pages here.
 295  */
 296 static pgoff_t vma_hugecache_offset(struct hstate *h,
 297                         struct vm_area_struct *vma, unsigned long address)
 298 {
 299         return ((address - vma->vm_start) >> huge_page_shift(h)) +
 300                         (vma->vm_pgoff >> huge_page_order(h));
 301 }
 302
 303 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
 304                                      unsigned long address)
 305 {
 306         return vma_hugecache_offset(hstate_vma(vma), vma, address);
 307 }
 308
 309 /*
 310  * Return the size of the pages allocated when backing a VMA. In the majority
 311  * cases this will be same size as used by the page table entries.
 312  */
 313 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 314 {
 315         struct hstate *hstate;
 316
 317         if (!is_vm_hugetlb_page(vma))
 318                 return PAGE_SIZE;
 319
 320         hstate = hstate_vma(vma);
 321
 322         return 1UL << huge_page_shift(hstate);
 323 }
 324 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
 325
 326 /*
 327  * Return the page size being used by the MMU to back a VMA. In the majority
 328  * of cases, the page size used by the kernel matches the MMU size. On
 329  * architectures where it differs, an architecture-specific version of this
 330  * function is required.
 331  */
 332 #ifndef vma_mmu_pagesize
 333 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 334 {
 335         return vma_kernel_pagesize(vma);
 336 }
 337 #endif
 338
 339 /*
 340  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 341  * bits of the reservation map pointer, which are always clear due to
 342  * alignment.
 343  */
 344 #define HPAGE_RESV_OWNER    (1UL << 0)
 345 #define HPAGE_RESV_UNMAPPED (1UL << 1)
 346 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 347
 348 /*
 349  * These helpers are used to track how many pages are reserved for
 350  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 351  * is guaranteed to have their future faults succeed.
 352  *
 353  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 354  * the reserve counters are updated with the hugetlb_lock held. It is safe
 355  * to reset the VMA at fork() time as it is not in use yet and there is no
 356  * chance of the global counters getting corrupted as a result of the values.
 357  *
 358  * The private mapping reservation is represented in a subtly different
 359  * manner to a shared mapping.  A shared mapping has a region map associated
 360  * with the underlying file, this region map represents the backing file
 361  * pages which have ever had a reservation assigned which this persists even
 362  * after the page is instantiated.  A private mapping has a region map
 363  * associated with the original mmap which is attached to all VMAs which
 364  * reference it, this region map represents those offsets which have consumed
 365  * reservation ie. where pages have been instantiated.
 366  */
 367 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 368 {
 369         return (unsigned long)vma->vm_private_data;
 370 }
 371
 372 static void set_vma_private_data(struct vm_area_struct *vma,
 373                                                         unsigned long value)
 374 {
 375         vma->vm_private_data = (void *)value;
 376 }
 377
 378 struct resv_map {
 379         struct kref refs;
 380         struct list_head regions;
 381 };
 382
 383 static struct resv_map *resv_map_alloc(void)
 384 {
 385         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 386         if (!resv_map)
 387                 return NULL;
 388
 389         kref_init(&resv_map->refs);
 390         INIT_LIST_HEAD(&resv_map->regions);
 391
 392         return resv_map;
 393 }
 394
 395 static void resv_map_release(struct kref *ref)
 396 {
 397         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 398
 399         /* Clear out any active regions before we release the map. */
 400         region_truncate(&resv_map->regions, 0);
 401         kfree(resv_map);
 402 }
 403
 404 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 405 {
 406         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 407         if (!(vma->vm_flags & VM_MAYSHARE))
 408                 return (struct resv_map *)(get_vma_private_data(vma) &
 409                                                         ~HPAGE_RESV_MASK);
 410         return NULL;
 411 }
 412
 413 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 414 {
 415         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 416         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 417
 418         set_vma_private_data(vma, (get_vma_private_data(vma) &
 419                                 HPAGE_RESV_MASK) | (unsigned long)map);
 420 }
 421
 422 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 423 {
 424         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 425         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 426
 427         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 428 }
 429
 430 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 431 {
 432         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 433
 434         return (get_vma_private_data(vma) & flag) != 0;
 435 }
 436
 437 /* Decrement the reserved pages in the hugepage pool by one */
 438 static void decrement_hugepage_resv_vma(struct hstate *h,
 439                         struct vm_area_struct *vma)
 440 {
 441         if (vma->vm_flags & VM_NORESERVE)
 442                 return;
 443
 444         if (vma->vm_flags & VM_MAYSHARE) {
 445                 /* Shared mappings always use reserves */
 446                 h->resv_huge_pages--;
 447         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 448                 /*
 449                  * Only the process that called mmap() has reserves for
 450                  * private mappings.
 451                  */
 452                 h->resv_huge_pages--;
 453         }
 454 }
 455
 456 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 457 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 458 {
 459         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 460         if (!(vma->vm_flags & VM_MAYSHARE))
 461                 vma->vm_private_data = (void *)0;
 462 }
 463
 464 /* Returns true if the VMA has associated reserve pages */
 465 static int vma_has_reserves(struct vm_area_struct *vma)
 466 {
 467         if (vma->vm_flags & VM_MAYSHARE)
 468                 return 1;
 469         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 470                 return 1;
 471         return 0;
 472 }
 473
 474 static void copy_gigantic_page(struct page *dst, struct page *src)
 475 {
 476         int i;
 477         struct hstate *h = page_hstate(src);
 478         struct page *dst_base = dst;
 479         struct page *src_base = src;
 480
 481         for (i = 0; i < pages_per_huge_page(h); ) {
 482                 cond_resched();
 483                 copy_highpage(dst, src);
 484
 485                 i++;
 486                 dst = mem_map_next(dst, dst_base, i);
 487                 src = mem_map_next(src, src_base, i);
 488         }
 489 }
 490
 491 void copy_huge_page(struct page *dst, struct page *src)
 492 {
 493         int i;
 494         struct hstate *h = page_hstate(src);
 495
 496         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
 497                 copy_gigantic_page(dst, src);
 498                 return;
 499         }
 500
 501         might_sleep();
 502         for (i = 0; i < pages_per_huge_page(h); i++) {
 503                 cond_resched();
 504                 copy_highpage(dst + i, src + i);
 505         }
 506 }
 507
 508 static void enqueue_huge_page(struct hstate *h, struct page *page)
 509 {
 510         int nid = page_to_nid(page);
 511         list_move(&page->lru, &h->hugepage_freelists[nid]);
 512         h->free_huge_pages++;
 513         h->free_huge_pages_node[nid]++;
 514 }
 515
 516 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 517 {
 518         struct page *page;
 519
 520         if (list_empty(&h->hugepage_freelists[nid]))
 521                 return NULL;
 522         page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
 523         list_move(&page->lru, &h->hugepage_activelist);
 524         set_page_refcounted(page);
 525         h->free_huge_pages--;
 526         h->free_huge_pages_node[nid]--;
 527         return page;
 528 }
 529
 530 static struct page *dequeue_huge_page_vma(struct hstate *h,
 531                                 struct vm_area_struct *vma,
 532                                 unsigned long address, int avoid_reserve)
 533 {
 534         struct page *page = NULL;
 535         struct mempolicy *mpol;
 536         nodemask_t *nodemask;
 537         struct zonelist *zonelist;
 538         struct zone *zone;
 539         struct zoneref *z;
 540         unsigned int cpuset_mems_cookie;
 541
 542         /*
 543          * A child process with MAP_PRIVATE mappings created by their parent
 544          * have no page reserves. This check ensures that reservations are
 545          * not "stolen". The child may still get SIGKILLed
 546          */
 547         if (!vma_has_reserves(vma) &&
 548                         h->free_huge_pages - h->resv_huge_pages == 0)
 549                 goto err;
 550
 551         /* If reserves cannot be used, ensure enough pages are in the pool */
 552         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 553                 goto err;
 554
 555 retry_cpuset:
 556         cpuset_mems_cookie = get_mems_allowed();
 557         zonelist = huge_zonelist(vma, address,
 558                                         htlb_alloc_mask, &mpol, &nodemask);
 559
 560         for_each_zone_zonelist_nodemask(zone, z, zonelist,
 561                                                 MAX_NR_ZONES - 1, nodemask) {
 562                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
 563                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
 564                         if (page) {
 565                                 if (!avoid_reserve)
 566                                         decrement_hugepage_resv_vma(h, vma);
 567                                 break;
 568                         }
 569                 }
 570         }
 571
 572         mpol_cond_put(mpol);
 573         if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
 574                 goto retry_cpuset;
 575         return page;
 576
 577 err:
 578         return NULL;
 579 }
 580
 581 static void update_and_free_page(struct hstate *h, struct page *page)
 582 {
 583         int i;
 584
 585         VM_BUG_ON(h->order >= MAX_ORDER);
 586
 587         h->nr_huge_pages--;
 588         h->nr_huge_pages_node[page_to_nid(page)]--;
 589         for (i = 0; i < pages_per_huge_page(h); i++) {
 590                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
 591                                 1 << PG_referenced | 1 << PG_dirty |
 592                                 1 << PG_active | 1 << PG_reserved |
 593                                 1 << PG_private | 1 << PG_writeback);
 594         }
 595         VM_BUG_ON(hugetlb_cgroup_from_page(page));
 596         set_compound_page_dtor(page, NULL);
 597         set_page_refcounted(page);
 598         arch_release_hugepage(page);
 599         __free_pages(page, huge_page_order(h));
 600 }
 601
 602 struct hstate *size_to_hstate(unsigned long size)
 603 {
 604         struct hstate *h;
 605
 606         for_each_hstate(h) {
 607                 if (huge_page_size(h) == size)
 608                         return h;
 609         }
 610         return NULL;
 611 }
 612
 613 static void free_huge_page(struct page *page)
 614 {
 615         /*
 616          * Can't pass hstate in here because it is called from the
 617          * compound page destructor.
 618          */
 619         struct hstate *h = page_hstate(page);
 620         int nid = page_to_nid(page);
 621         struct hugepage_subpool *spool =
 622                 (struct hugepage_subpool *)page_private(page);
 623
 624         set_page_private(page, 0);
 625         page->mapping = NULL;
 626         BUG_ON(page_count(page));
 627         BUG_ON(page_mapcount(page));
 628
 629         spin_lock(&hugetlb_lock);
 630         hugetlb_cgroup_uncharge_page(hstate_index(h),
 631                                      pages_per_huge_page(h), page);
 632         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 633                 /* remove the page from active list */
 634                 list_del(&page->lru);
 635                 update_and_free_page(h, page);
 636                 h->surplus_huge_pages--;
 637                 h->surplus_huge_pages_node[nid]--;
 638         } else {
 639                 arch_clear_hugepage_flags(page);
 640                 enqueue_huge_page(h, page);
 641         }
 642         spin_unlock(&hugetlb_lock);
 643         hugepage_subpool_put_pages(spool, 1);
 644 }
 645
 646 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 647 {
 648         INIT_LIST_HEAD(&page->lru);
 649         set_compound_page_dtor(page, free_huge_page);
 650         spin_lock(&hugetlb_lock);
 651         set_hugetlb_cgroup(page, NULL);
 652         h->nr_huge_pages++;
 653         h->nr_huge_pages_node[nid]++;
 654         spin_unlock(&hugetlb_lock);
 655         put_page(page); /* free it into the hugepage allocator */
 656 }
 657
 658 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
 659 {
 660         int i;
 661         int nr_pages = 1 << order;
 662         struct page *p = page + 1;
 663
 664         /* we rely on prep_new_huge_page to set the destructor */
 665         set_compound_order(page, order);
 666         __SetPageHead(page);
 667         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
 668                 __SetPageTail(p);
 669                 set_page_count(p, 0);
 670                 p->first_page = page;
 671         }
 672 }
 673
 674 /*
 675  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
 676  * transparent huge pages.  See the PageTransHuge() documentation for more
 677  * details.
 678  */
 679 int PageHuge(struct page *page)
 680 {
 681         compound_page_dtor *dtor;
 682
 683         if (!PageCompound(page))
 684                 return 0;
 685
 686         page = compound_head(page);
 687         dtor = get_compound_page_dtor(page);
 688
 689         return dtor == free_huge_page;
 690 }
 691 EXPORT_SYMBOL_GPL(PageHuge);
 692
 693 pgoff_t __basepage_index(struct page *page)
 694 {
 695         struct page *page_head = compound_head(page);
 696         pgoff_t index = page_index(page_head);
 697         unsigned long compound_idx;
 698
 699         if (!PageHuge(page_head))
 700                 return page_index(page);
 701
 702         if (compound_order(page_head) >= MAX_ORDER)
 703                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 704         else
 705                 compound_idx = page - page_head;
 706
 707         return (index << compound_order(page_head)) + compound_idx;
 708 }
 709
 710 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 711 {
 712         struct page *page;
 713
 714         if (h->order >= MAX_ORDER)
 715                 return NULL;
 716
 717         page = alloc_pages_exact_node(nid,
 718                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 719                                                 __GFP_REPEAT|__GFP_NOWARN,
 720                 huge_page_order(h));
 721         if (page) {
 722                 if (arch_prepare_hugepage(page)) {
 723                         __free_pages(page, huge_page_order(h));
 724                         return NULL;
 725                 }
 726                 prep_new_huge_page(h, page, nid);
 727         }
 728
 729         return page;
 730 }
 731
 732 /*
 733  * common helper functions for hstate_next_node_to_{alloc|free}.
 734  * We may have allocated or freed a huge page based on a different
 735  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
 736  * be outside of *nodes_allowed.  Ensure that we use an allowed
 737  * node for alloc or free.
 738  */
 739 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 740 {
 741         nid = next_node(nid, *nodes_allowed);
 742         if (nid == MAX_NUMNODES)
 743                 nid = first_node(*nodes_allowed);
 744         VM_BUG_ON(nid >= MAX_NUMNODES);
 745
 746         return nid;
 747 }
 748
 749 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
 750 {
 751         if (!node_isset(nid, *nodes_allowed))
 752                 nid = next_node_allowed(nid, nodes_allowed);
 753         return nid;
 754 }
 755
 756 /*
 757  * returns the previously saved node ["this node"] from which to
 758  * allocate a persistent huge page for the pool and advance the
 759  * next node from which to allocate, handling wrap at end of node
 760  * mask.
 761  */
 762 static int hstate_next_node_to_alloc(struct hstate *h,
 763                                         nodemask_t *nodes_allowed)
 764 {
 765         int nid;
 766
 767         VM_BUG_ON(!nodes_allowed);
 768
 769         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
 770         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
 771
 772         return nid;
 773 }
 774
 775 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
 776 {
 777         struct page *page;
 778         int start_nid;
 779         int next_nid;
 780         int ret = 0;
 781
 782         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 783         next_nid = start_nid;
 784
 785         do {
 786                 page = alloc_fresh_huge_page_node(h, next_nid);
 787                 if (page) {
 788                         ret = 1;
 789                         break;
 790                 }
 791                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 792         } while (next_nid != start_nid);
 793
 794         if (ret)
 795                 count_vm_event(HTLB_BUDDY_PGALLOC);
 796         else
 797                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 798
 799         return ret;
 800 }
 801
 802 /*
 803  * helper for free_pool_huge_page() - return the previously saved
 804  * node ["this node"] from which to free a huge page.  Advance the
 805  * next node id whether or not we find a free huge page to free so
 806  * that the next attempt to free addresses the next node.
 807  */
 808 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
 809 {
 810         int nid;
 811
 812         VM_BUG_ON(!nodes_allowed);
 813
 814         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
 815         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
 816
 817         return nid;
 818 }
 819
 820 /*
 821  * Free huge page from pool from next node to free.
 822  * Attempt to keep persistent huge pages more or less
 823  * balanced over allowed nodes.
 824  * Called with hugetlb_lock locked.
 825  */
 826 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
 827                                                          bool acct_surplus)
 828 {
 829         int start_nid;
 830         int next_nid;
 831         int ret = 0;
 832
 833         start_nid = hstate_next_node_to_free(h, nodes_allowed);
 834         next_nid = start_nid;
 835
 836         do {
 837                 /*
 838                  * If we're returning unused surplus pages, only examine
 839                  * nodes with surplus pages.
 840                  */
 841                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
 842                     !list_empty(&h->hugepage_freelists[next_nid])) {
 843                         struct page *page =
 844                                 list_entry(h->hugepage_freelists[next_nid].next,
 845                                           struct page, lru);
 846                         list_del(&page->lru);
 847                         h->free_huge_pages--;
 848                         h->free_huge_pages_node[next_nid]--;
 849                         if (acct_surplus) {
 850                                 h->surplus_huge_pages--;
 851                                 h->surplus_huge_pages_node[next_nid]--;
 852                         }
 853                         update_and_free_page(h, page);
 854                         ret = 1;
 855                         break;
 856                 }
 857                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
 858         } while (next_nid != start_nid);
 859
 860         return ret;
 861 }
 862
 863 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 864 {
 865         struct page *page;
 866         unsigned int r_nid;
 867
 868         if (h->order >= MAX_ORDER)
 869                 return NULL;
 870
 871         /*
 872          * Assume we will successfully allocate the surplus page to
 873          * prevent racing processes from causing the surplus to exceed
 874          * overcommit
 875          *
 876          * This however introduces a different race, where a process B
 877          * tries to grow the static hugepage pool while alloc_pages() is
 878          * called by process A. B will only examine the per-node
 879          * counters in determining if surplus huge pages can be
 880          * converted to normal huge pages in adjust_pool_surplus(). A
 881          * won't be able to increment the per-node counter, until the
 882          * lock is dropped by B, but B doesn't drop hugetlb_lock until
 883          * no more huge pages can be converted from surplus to normal
 884          * state (and doesn't try to convert again). Thus, we have a
 885          * case where a surplus huge page exists, the pool is grown, and
 886          * the surplus huge page still exists after, even though it
 887          * should just have been converted to a normal huge page. This
 888          * does not leak memory, though, as the hugepage will be freed
 889          * once it is out of use. It also does not allow the counters to
 890          * go out of whack in adjust_pool_surplus() as we don't modify
 891          * the node values until we've gotten the hugepage and only the
 892          * per-node value is checked there.
 893          */
 894         spin_lock(&hugetlb_lock);
 895         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 896                 spin_unlock(&hugetlb_lock);
 897                 return NULL;
 898         } else {
 899                 h->nr_huge_pages++;
 900                 h->surplus_huge_pages++;
 901         }
 902         spin_unlock(&hugetlb_lock);
 903
 904         if (nid == NUMA_NO_NODE)
 905                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 906                                    __GFP_REPEAT|__GFP_NOWARN,
 907                                    huge_page_order(h));
 908         else
 909                 page = alloc_pages_exact_node(nid,
 910                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 911                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
 912
 913         if (page && arch_prepare_hugepage(page)) {
 914                 __free_pages(page, huge_page_order(h));
 915                 page = NULL;
 916         }
 917
 918         spin_lock(&hugetlb_lock);
 919         if (page) {
 920                 INIT_LIST_HEAD(&page->lru);
 921                 r_nid = page_to_nid(page);
 922                 set_compound_page_dtor(page, free_huge_page);
 923                 set_hugetlb_cgroup(page, NULL);
 924                 /*
 925                  * We incremented the global counters already
 926                  */
 927                 h->nr_huge_pages_node[r_nid]++;
 928                 h->surplus_huge_pages_node[r_nid]++;
 929                 __count_vm_event(HTLB_BUDDY_PGALLOC);
 930         } else {
 931                 h->nr_huge_pages--;
 932                 h->surplus_huge_pages--;
 933                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 934         }
 935         spin_unlock(&hugetlb_lock);
 936
 937         return page;
 938 }
 939
 940 /*
 941  * This allocation function is useful in the context where vma is irrelevant.
 942  * E.g. soft-offlining uses this function because it only cares physical
 943  * address of error page.
 944  */
 945 struct page *alloc_huge_page_node(struct hstate *h, int nid)
 946 {
 947         struct page *page;
 948
 949         spin_lock(&hugetlb_lock);
 950         page = dequeue_huge_page_node(h, nid);
 951         spin_unlock(&hugetlb_lock);
 952
 953         if (!page)
 954                 page = alloc_buddy_huge_page(h, nid);
 955
 956         return page;
 957 }
 958
 959 /*
 960  * Increase the hugetlb pool such that it can accommodate a reservation
 961  * of size 'delta'.
 962  */
 963 static int gather_surplus_pages(struct hstate *h, int delta)
 964 {
 965         struct list_head surplus_list;
 966         struct page *page, *tmp;
 967         int ret, i;
 968         int needed, allocated;
 969         bool alloc_ok = true;
 970
 971         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
 972         if (needed <= 0) {
 973                 h->resv_huge_pages += delta;
 974                 return 0;
 975         }
 976
 977         allocated = 0;
 978         INIT_LIST_HEAD(&surplus_list);
 979
 980         ret = -ENOMEM;
 981 retry:
 982         spin_unlock(&hugetlb_lock);
 983         for (i = 0; i < needed; i++) {
 984                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
 985                 if (!page) {
 986                         alloc_ok = false;
 987                         break;
 988                 }
 989                 list_add(&page->lru, &surplus_list);
 990         }
 991         allocated += i;
 992
 993         /*
 994          * After retaking hugetlb_lock, we need to recalculate 'needed'
 995          * because either resv_huge_pages or free_huge_pages may have changed.
 996          */
 997         spin_lock(&hugetlb_lock);
 998         needed = (h->resv_huge_pages + delta) -
 999                         (h->free_huge_pages + allocated);
1000         if (needed > 0) {
1001                 if (alloc_ok)
1002                         goto retry;
1003                 /*
1004                  * We were not able to allocate enough pages to
1005                  * satisfy the entire reservation so we free what
1006                  * we've allocated so far.
1007                  */
1008                 goto free;
1009         }
1010         /*
1011          * The surplus_list now contains _at_least_ the number of extra pages
1012          * needed to accommodate the reservation.  Add the appropriate number
1013          * of pages to the hugetlb pool and free the extras back to the buddy
1014          * allocator.  Commit the entire reservation here to prevent another
1015          * process from stealing the pages as they are added to the pool but
1016          * before they are reserved.
1017          */
1018         needed += allocated;
1019         h->resv_huge_pages += delta;
1020         ret = 0;
1021
1022         /* Free the needed pages to the hugetlb pool */
1023         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1024                 if ((--needed) < 0)
1025                         break;
1026                 /*
1027                  * This page is now managed by the hugetlb allocator and has
1028                  * no users -- drop the buddy allocator's reference.
1029                  */
1030                 put_page_testzero(page);
1031                 VM_BUG_ON(page_count(page));
1032                 enqueue_huge_page(h, page);
1033         }
1034 free:
1035         spin_unlock(&hugetlb_lock);
1036
1037         /* Free unnecessary surplus pages to the buddy allocator */
1038         if (!list_empty(&surplus_list)) {
1039                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1040                         put_page(page);
1041                 }
1042         }
1043         spin_lock(&hugetlb_lock);
1044
1045         return ret;
1046 }
1047
1048 /*
1049  * When releasing a hugetlb pool reservation, any surplus pages that were
1050  * allocated to satisfy the reservation must be explicitly freed if they were
1051  * never used.
1052  * Called with hugetlb_lock held.
1053  */
1054 static void return_unused_surplus_pages(struct hstate *h,
1055                                         unsigned long unused_resv_pages)
1056 {
1057         unsigned long nr_pages;
1058
1059         /* Uncommit the reservation */
1060         h->resv_huge_pages -= unused_resv_pages;
1061
1062         /* Cannot return gigantic pages currently */
1063         if (h->order >= MAX_ORDER)
1064                 return;
1065
1066         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1067
1068         /*
1069          * We want to release as many surplus pages as possible, spread
1070          * evenly across all nodes with memory. Iterate across these nodes
1071          * until we can no longer free unreserved surplus pages. This occurs
1072          * when the nodes with surplus pages have no free pages.
1073          * free_pool_huge_page() will balance the the freed pages across the
1074          * on-line nodes with memory and will handle the hstate accounting.
1075          */
1076         while (nr_pages--) {
1077                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1078                         break;
1079         }
1080 }
1081
1082 /*
1083  * Determine if the huge page at addr within the vma has an associated
1084  * reservation.  Where it does not we will need to logically increase
1085  * reservation and actually increase subpool usage before an allocation
1086  * can occur.  Where any new reservation would be required the
1087  * reservation change is prepared, but not committed.  Once the page
1088  * has been allocated from the subpool and instantiated the change should
1089  * be committed via vma_commit_reservation.  No action is required on
1090  * failure.
1091  */
1092 static long vma_needs_reservation(struct hstate *h,
1093                         struct vm_area_struct *vma, unsigned long addr)
1094 {
1095         struct address_space *mapping = vma->vm_file->f_mapping;
1096         struct inode *inode = mapping->host;
1097
1098         if (vma->vm_flags & VM_MAYSHARE) {
1099                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1100                 return region_chg(&inode->i_mapping->private_list,
1101                                                         idx, idx + 1);
1102
1103         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1104                 return 1;
1105
1106         } else  {
1107                 long err;
1108                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1109                 struct resv_map *reservations = vma_resv_map(vma);
1110
1111                 err = region_chg(&reservations->regions, idx, idx + 1);
1112                 if (err < 0)
1113                         return err;
1114                 return 0;
1115         }
1116 }
1117 static void vma_commit_reservation(struct hstate *h,
1118                         struct vm_area_struct *vma, unsigned long addr)
1119 {
1120         struct address_space *mapping = vma->vm_file->f_mapping;
1121         struct inode *inode = mapping->host;
1122
1123         if (vma->vm_flags & VM_MAYSHARE) {
1124                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1125                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1126
1127         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1128                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1129                 struct resv_map *reservations = vma_resv_map(vma);
1130
1131                 /* Mark this page used in the map. */
1132                 region_add(&reservations->regions, idx, idx + 1);
1133         }
1134 }
1135
1136 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1137                                     unsigned long addr, int avoid_reserve)
1138 {
1139         struct hugepage_subpool *spool = subpool_vma(vma);
1140         struct hstate *h = hstate_vma(vma);
1141         struct page *page;
1142         long chg;
1143         int ret, idx;
1144         struct hugetlb_cgroup *h_cg;
1145
1146         idx = hstate_index(h);
1147         /*
1148          * Processes that did not create the mapping will have no
1149          * reserves and will not have accounted against subpool
1150          * limit. Check that the subpool limit can be made before
1151          * satisfying the allocation MAP_NORESERVE mappings may also
1152          * need pages and subpool limit allocated allocated if no reserve
1153          * mapping overlaps.
1154          */
1155         chg = vma_needs_reservation(h, vma, addr);
1156         if (chg < 0)
1157                 return ERR_PTR(-ENOMEM);
1158         if (chg)
1159                 if (hugepage_subpool_get_pages(spool, chg))
1160                         return ERR_PTR(-ENOSPC);
1161
1162         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1163         if (ret) {
1164                 hugepage_subpool_put_pages(spool, chg);
1165                 return ERR_PTR(-ENOSPC);
1166         }
1167         spin_lock(&hugetlb_lock);
1168         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1169         if (!page) {
1170                 spin_unlock(&hugetlb_lock);
1171                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1172                 if (!page) {
1173                         hugetlb_cgroup_uncharge_cgroup(idx,
1174                                                        pages_per_huge_page(h),
1175                                                        h_cg);
1176                         hugepage_subpool_put_pages(spool, chg);
1177                         return ERR_PTR(-ENOSPC);
1178                 }
1179                 spin_lock(&hugetlb_lock);
1180                 list_move(&page->lru, &h->hugepage_activelist);
1181                 /* Fall through */
1182         }
1183         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1184         spin_unlock(&hugetlb_lock);
1185
1186         set_page_private(page, (unsigned long)spool);
1187
1188         vma_commit_reservation(h, vma, addr);
1189         return page;
1190 }
1191
1192 int __weak alloc_bootmem_huge_page(struct hstate *h)
1193 {
1194         struct huge_bootmem_page *m;
1195         int nr_nodes = nodes_weight(node_states[N_MEMORY]);
1196
1197         while (nr_nodes) {
1198                 void *addr;
1199
1200                 addr = __alloc_bootmem_node_nopanic(
1201                                 NODE_DATA(hstate_next_node_to_alloc(h,
1202                                                 &node_states[N_MEMORY])),
1203                                 huge_page_size(h), huge_page_size(h), 0);
1204
1205                 if (addr) {
1206                         /*
1207                          * Use the beginning of the huge page to store the
1208                          * huge_bootmem_page struct (until gather_bootmem
1209                          * puts them into the mem_map).
1210                          */
1211                         m = addr;
1212                         goto found;
1213                 }
1214                 nr_nodes--;
1215         }
1216         return 0;
1217
1218 found:
1219         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1220         /* Put them into a private list first because mem_map is not up yet */
1221         list_add(&m->list, &huge_boot_pages);
1222         m->hstate = h;
1223         return 1;
1224 }
1225
1226 static void prep_compound_huge_page(struct page *page, int order)
1227 {
1228         if (unlikely(order > (MAX_ORDER - 1)))
1229                 prep_compound_gigantic_page(page, order);
1230         else
1231                 prep_compound_page(page, order);
1232 }
1233
1234 /* Put bootmem huge pages into the standard lists after mem_map is up */
1235 static void __init gather_bootmem_prealloc(void)
1236 {
1237         struct huge_bootmem_page *m;
1238
1239         list_for_each_entry(m, &huge_boot_pages, list) {
1240                 struct hstate *h = m->hstate;
1241                 struct page *page;
1242
1243 #ifdef CONFIG_HIGHMEM
1244                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1245                 free_bootmem_late((unsigned long)m,
1246                                   sizeof(struct huge_bootmem_page));
1247 #else
1248                 page = virt_to_page(m);
1249 #endif
1250                 __ClearPageReserved(page);
1251                 WARN_ON(page_count(page) != 1);
1252                 prep_compound_huge_page(page, h->order);
1253                 prep_new_huge_page(h, page, page_to_nid(page));
1254                 /*
1255                  * If we had gigantic hugepages allocated at boot time, we need
1256                  * to restore the 'stolen' pages to totalram_pages in order to
1257                  * fix confusing memory reports from free(1) and another
1258                  * side-effects, like CommitLimit going negative.
1259                  */
1260                 if (h->order > (MAX_ORDER - 1))
1261                         adjust_managed_page_count(page, 1 << h->order);
1262         }
1263 }
1264
1265 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1266 {
1267         unsigned long i;
1268
1269         for (i = 0; i < h->max_huge_pages; ++i) {
1270                 if (h->order >= MAX_ORDER) {
1271                         if (!alloc_bootmem_huge_page(h))
1272                                 break;
1273                 } else if (!alloc_fresh_huge_page(h,
1274                                          &node_states[N_MEMORY]))
1275                         break;
1276         }
1277         h->max_huge_pages = i;
1278 }
1279
1280 static void __init hugetlb_init_hstates(void)
1281 {
1282         struct hstate *h;
1283
1284         for_each_hstate(h) {
1285                 /* oversize hugepages were init'ed in early boot */
1286                 if (h->order < MAX_ORDER)
1287                         hugetlb_hstate_alloc_pages(h);
1288         }
1289 }
1290
1291 static char * __init memfmt(char *buf, unsigned long n)
1292 {
1293         if (n >= (1UL << 30))
1294                 sprintf(buf, "%lu GB", n >> 30);
1295         else if (n >= (1UL << 20))
1296                 sprintf(buf, "%lu MB", n >> 20);
1297         else
1298                 sprintf(buf, "%lu KB", n >> 10);
1299         return buf;
1300 }
1301
1302 static void __init report_hugepages(void)
1303 {
1304         struct hstate *h;
1305
1306         for_each_hstate(h) {
1307                 char buf[32];
1308                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1309                         memfmt(buf, huge_page_size(h)),
1310                         h->free_huge_pages);
1311         }
1312 }
1313
1314 #ifdef CONFIG_HIGHMEM
1315 static void try_to_free_low(struct hstate *h, unsigned long count,
1316                                                 nodemask_t *nodes_allowed)
1317 {
1318         int i;
1319
1320         if (h->order >= MAX_ORDER)
1321                 return;
1322
1323         for_each_node_mask(i, *nodes_allowed) {
1324                 struct page *page, *next;
1325                 struct list_head *freel = &h->hugepage_freelists[i];
1326                 list_for_each_entry_safe(page, next, freel, lru) {
1327                         if (count >= h->nr_huge_pages)
1328                                 return;
1329                         if (PageHighMem(page))
1330                                 continue;
1331                         list_del(&page->lru);
1332                         update_and_free_page(h, page);
1333                         h->free_huge_pages--;
1334                         h->free_huge_pages_node[page_to_nid(page)]--;
1335                 }
1336         }
1337 }
1338 #else
1339 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1340                                                 nodemask_t *nodes_allowed)
1341 {
1342 }
1343 #endif
1344
1345 /*
1346  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1347  * balanced by operating on them in a round-robin fashion.
1348  * Returns 1 if an adjustment was made.
1349  */
1350 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1351                                 int delta)
1352 {
1353         int start_nid, next_nid;
1354         int ret = 0;
1355
1356         VM_BUG_ON(delta != -1 && delta != 1);
1357
1358         if (delta < 0)
1359                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1360         else
1361                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1362         next_nid = start_nid;
1363
1364         do {
1365                 int nid = next_nid;
1366                 if (delta < 0)  {
1367                         /*
1368                          * To shrink on this node, there must be a surplus page
1369                          */
1370                         if (!h->surplus_huge_pages_node[nid]) {
1371                                 next_nid = hstate_next_node_to_alloc(h,
1372                                                                 nodes_allowed);
1373                                 continue;
1374                         }
1375                 }
1376                 if (delta > 0) {
1377                         /*
1378                          * Surplus cannot exceed the total number of pages
1379                          */
1380                         if (h->surplus_huge_pages_node[nid] >=
1381                                                 h->nr_huge_pages_node[nid]) {
1382                                 next_nid = hstate_next_node_to_free(h,
1383                                                                 nodes_allowed);
1384                                 continue;
1385                         }
1386                 }
1387
1388                 h->surplus_huge_pages += delta;
1389                 h->surplus_huge_pages_node[nid] += delta;
1390                 ret = 1;
1391                 break;
1392         } while (next_nid != start_nid);
1393
1394         return ret;
1395 }
1396
1397 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1398 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1399                                                 nodemask_t *nodes_allowed)
1400 {
1401         unsigned long min_count, ret;
1402
1403         if (h->order >= MAX_ORDER)
1404                 return h->max_huge_pages;
1405
1406         /*
1407          * Increase the pool size
1408          * First take pages out of surplus state.  Then make up the
1409          * remaining difference by allocating fresh huge pages.
1410          *
1411          * We might race with alloc_buddy_huge_page() here and be unable
1412          * to convert a surplus huge page to a normal huge page. That is
1413          * not critical, though, it just means the overall size of the
1414          * pool might be one hugepage larger than it needs to be, but
1415          * within all the constraints specified by the sysctls.
1416          */
1417         spin_lock(&hugetlb_lock);
1418         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1419                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1420                         break;
1421         }
1422
1423         while (count > persistent_huge_pages(h)) {
1424                 /*
1425                  * If this allocation races such that we no longer need the
1426                  * page, free_huge_page will handle it by freeing the page
1427                  * and reducing the surplus.
1428                  */
1429                 spin_unlock(&hugetlb_lock);
1430                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1431                 spin_lock(&hugetlb_lock);
1432                 if (!ret)
1433                         goto out;
1434
1435                 /* Bail for signals. Probably ctrl-c from user */
1436                 if (signal_pending(current))
1437                         goto out;
1438         }
1439
1440         /*
1441          * Decrease the pool size
1442          * First return free pages to the buddy allocator (being careful
1443          * to keep enough around to satisfy reservations).  Then place
1444          * pages into surplus state as needed so the pool will shrink
1445          * to the desired size as pages become free.
1446          *
1447          * By placing pages into the surplus state independent of the
1448          * overcommit value, we are allowing the surplus pool size to
1449          * exceed overcommit. There are few sane options here. Since
1450          * alloc_buddy_huge_page() is checking the global counter,
1451          * though, we'll note that we're not allowed to exceed surplus
1452          * and won't grow the pool anywhere else. Not until one of the
1453          * sysctls are changed, or the surplus pages go out of use.
1454          */
1455         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1456         min_count = max(count, min_count);
1457         try_to_free_low(h, min_count, nodes_allowed);
1458         while (min_count < persistent_huge_pages(h)) {
1459                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1460                         break;
1461         }
1462         while (count < persistent_huge_pages(h)) {
1463                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1464                         break;
1465         }
1466 out:
1467         ret = persistent_huge_pages(h);
1468         spin_unlock(&hugetlb_lock);
1469         return ret;
1470 }
1471
1472 #define HSTATE_ATTR_RO(_name) \
1473         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1474
1475 #define HSTATE_ATTR(_name) \
1476         static struct kobj_attribute _name##_attr = \
1477                 __ATTR(_name, 0644, _name##_show, _name##_store)
1478
1479 static struct kobject *hugepages_kobj;
1480 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1481
1482 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1483
1484 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1485 {
1486         int i;
1487
1488         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1489                 if (hstate_kobjs[i] == kobj) {
1490                         if (nidp)
1491                                 *nidp = NUMA_NO_NODE;
1492                         return &hstates[i];
1493                 }
1494
1495         return kobj_to_node_hstate(kobj, nidp);
1496 }
1497
1498 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1499                                         struct kobj_attribute *attr, char *buf)
1500 {
1501         struct hstate *h;
1502         unsigned long nr_huge_pages;
1503         int nid;
1504
1505         h = kobj_to_hstate(kobj, &nid);
1506         if (nid == NUMA_NO_NODE)
1507                 nr_huge_pages = h->nr_huge_pages;
1508         else
1509                 nr_huge_pages = h->nr_huge_pages_node[nid];
1510
1511         return sprintf(buf, "%lu\n", nr_huge_pages);
1512 }
1513
1514 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1515                         struct kobject *kobj, struct kobj_attribute *attr,
1516                         const char *buf, size_t len)
1517 {
1518         int err;
1519         int nid;
1520         unsigned long count;
1521         struct hstate *h;
1522         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1523
1524         err = kstrtoul(buf, 10, &count);
1525         if (err)
1526                 goto out;
1527
1528         h = kobj_to_hstate(kobj, &nid);
1529         if (h->order >= MAX_ORDER) {
1530                 err = -EINVAL;
1531                 goto out;
1532         }
1533
1534         if (nid == NUMA_NO_NODE) {
1535                 /*
1536                  * global hstate attribute
1537                  */
1538                 if (!(obey_mempolicy &&
1539                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1540                         NODEMASK_FREE(nodes_allowed);
1541                         nodes_allowed = &node_states[N_MEMORY];
1542                 }
1543         } else if (nodes_allowed) {
1544                 /*
1545                  * per node hstate attribute: adjust count to global,
1546                  * but restrict alloc/free to the specified node.
1547                  */
1548                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1549                 init_nodemask_of_node(nodes_allowed, nid);
1550         } else
1551                 nodes_allowed = &node_states[N_MEMORY];
1552
1553         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1554
1555         if (nodes_allowed != &node_states[N_MEMORY])
1556                 NODEMASK_FREE(nodes_allowed);
1557
1558         return len;
1559 out:
1560         NODEMASK_FREE(nodes_allowed);
1561         return err;
1562 }
1563
1564 static ssize_t nr_hugepages_show(struct kobject *kobj,
1565                                        struct kobj_attribute *attr, char *buf)
1566 {
1567         return nr_hugepages_show_common(kobj, attr, buf);
1568 }
1569
1570 static ssize_t nr_hugepages_store(struct kobject *kobj,
1571                struct kobj_attribute *attr, const char *buf, size_t len)
1572 {
1573         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1574 }
1575 HSTATE_ATTR(nr_hugepages);
1576
1577 #ifdef CONFIG_NUMA
1578
1579 /*
1580  * hstate attribute for optionally mempolicy-based constraint on persistent
1581  * huge page alloc/free.
1582  */
1583 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1584                                        struct kobj_attribute *attr, char *buf)
1585 {
1586         return nr_hugepages_show_common(kobj, attr, buf);
1587 }
1588
1589 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1590                struct kobj_attribute *attr, const char *buf, size_t len)
1591 {
1592         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1593 }
1594 HSTATE_ATTR(nr_hugepages_mempolicy);
1595 #endif
1596
1597
1598 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1599                                         struct kobj_attribute *attr, char *buf)
1600 {
1601         struct hstate *h = kobj_to_hstate(kobj, NULL);
1602         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1603 }
1604
1605 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1606                 struct kobj_attribute *attr, const char *buf, size_t count)
1607 {
1608         int err;
1609         unsigned long input;
1610         struct hstate *h = kobj_to_hstate(kobj, NULL);
1611
1612         if (h->order >= MAX_ORDER)
1613                 return -EINVAL;
1614
1615         err = kstrtoul(buf, 10, &input);
1616         if (err)
1617                 return err;
1618
1619         spin_lock(&hugetlb_lock);
1620         h->nr_overcommit_huge_pages = input;
1621         spin_unlock(&hugetlb_lock);
1622
1623         return count;
1624 }
1625 HSTATE_ATTR(nr_overcommit_hugepages);
1626
1627 static ssize_t free_hugepages_show(struct kobject *kobj,
1628                                         struct kobj_attribute *attr, char *buf)
1629 {
1630         struct hstate *h;
1631         unsigned long free_huge_pages;
1632         int nid;
1633
1634         h = kobj_to_hstate(kobj, &nid);
1635         if (nid == NUMA_NO_NODE)
1636                 free_huge_pages = h->free_huge_pages;
1637         else
1638                 free_huge_pages = h->free_huge_pages_node[nid];
1639
1640         return sprintf(buf, "%lu\n", free_huge_pages);
1641 }
1642 HSTATE_ATTR_RO(free_hugepages);
1643
1644 static ssize_t resv_hugepages_show(struct kobject *kobj,
1645                                         struct kobj_attribute *attr, char *buf)
1646 {
1647         struct hstate *h = kobj_to_hstate(kobj, NULL);
1648         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1649 }
1650 HSTATE_ATTR_RO(resv_hugepages);
1651
1652 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1653                                         struct kobj_attribute *attr, char *buf)
1654 {
1655         struct hstate *h;
1656         unsigned long surplus_huge_pages;
1657         int nid;
1658
1659         h = kobj_to_hstate(kobj, &nid);
1660         if (nid == NUMA_NO_NODE)
1661                 surplus_huge_pages = h->surplus_huge_pages;
1662         else
1663                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1664
1665         return sprintf(buf, "%lu\n", surplus_huge_pages);
1666 }
1667 HSTATE_ATTR_RO(surplus_hugepages);
1668
1669 static struct attribute *hstate_attrs[] = {
1670         &nr_hugepages_attr.attr,
1671         &nr_overcommit_hugepages_attr.attr,
1672         &free_hugepages_attr.attr,
1673         &resv_hugepages_attr.attr,
1674         &surplus_hugepages_attr.attr,
1675 #ifdef CONFIG_NUMA
1676         &nr_hugepages_mempolicy_attr.attr,
1677 #endif
1678         NULL,
1679 };
1680
1681 static struct attribute_group hstate_attr_group = {
1682         .attrs = hstate_attrs,
1683 };
1684
1685 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1686                                     struct kobject **hstate_kobjs,
1687                                     struct attribute_group *hstate_attr_group)
1688 {
1689         int retval;
1690         int hi = hstate_index(h);
1691
1692         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1693         if (!hstate_kobjs[hi])
1694                 return -ENOMEM;
1695
1696         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1697         if (retval)
1698                 kobject_put(hstate_kobjs[hi]);
1699
1700         return retval;
1701 }
1702
1703 static void __init hugetlb_sysfs_init(void)
1704 {
1705         struct hstate *h;
1706         int err;
1707
1708         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1709         if (!hugepages_kobj)
1710                 return;
1711
1712         for_each_hstate(h) {
1713                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1714                                          hstate_kobjs, &hstate_attr_group);
1715                 if (err)
1716                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
1717         }
1718 }
1719
1720 #ifdef CONFIG_NUMA
1721
1722 /*
1723  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1724  * with node devices in node_devices[] using a parallel array.  The array
1725  * index of a node device or _hstate == node id.
1726  * This is here to avoid any static dependency of the node device driver, in
1727  * the base kernel, on the hugetlb module.
1728  */
1729 struct node_hstate {
1730         struct kobject          *hugepages_kobj;
1731         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1732 };
1733 struct node_hstate node_hstates[MAX_NUMNODES];
1734
1735 /*
1736  * A subset of global hstate attributes for node devices
1737  */
1738 static struct attribute *per_node_hstate_attrs[] = {
1739         &nr_hugepages_attr.attr,
1740         &free_hugepages_attr.attr,
1741         &surplus_hugepages_attr.attr,
1742         NULL,
1743 };
1744
1745 static struct attribute_group per_node_hstate_attr_group = {
1746         .attrs = per_node_hstate_attrs,
1747 };
1748
1749 /*
1750  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1751  * Returns node id via non-NULL nidp.
1752  */
1753 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1754 {
1755         int nid;
1756
1757         for (nid = 0; nid < nr_node_ids; nid++) {
1758                 struct node_hstate *nhs = &node_hstates[nid];
1759                 int i;
1760                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1761                         if (nhs->hstate_kobjs[i] == kobj) {
1762                                 if (nidp)
1763                                         *nidp = nid;
1764                                 return &hstates[i];
1765                         }
1766         }
1767
1768         BUG();
1769         return NULL;
1770 }
1771
1772 /*
1773  * Unregister hstate attributes from a single node device.
1774  * No-op if no hstate attributes attached.
1775  */
1776 static void hugetlb_unregister_node(struct node *node)
1777 {
1778         struct hstate *h;
1779         struct node_hstate *nhs = &node_hstates[node->dev.id];
1780
1781         if (!nhs->hugepages_kobj)
1782                 return;         /* no hstate attributes */
1783
1784         for_each_hstate(h) {
1785                 int idx = hstate_index(h);
1786                 if (nhs->hstate_kobjs[idx]) {
1787                         kobject_put(nhs->hstate_kobjs[idx]);
1788                         nhs->hstate_kobjs[idx] = NULL;
1789                 }
1790         }
1791
1792         kobject_put(nhs->hugepages_kobj);
1793         nhs->hugepages_kobj = NULL;
1794 }
1795
1796 /*
1797  * hugetlb module exit:  unregister hstate attributes from node devices
1798  * that have them.
1799  */
1800 static void hugetlb_unregister_all_nodes(void)
1801 {
1802         int nid;
1803
1804         /*
1805          * disable node device registrations.
1806          */
1807         register_hugetlbfs_with_node(NULL, NULL);
1808
1809         /*
1810          * remove hstate attributes from any nodes that have them.
1811          */
1812         for (nid = 0; nid < nr_node_ids; nid++)
1813                 hugetlb_unregister_node(node_devices[nid]);
1814 }
1815
1816 /*
1817  * Register hstate attributes for a single node device.
1818  * No-op if attributes already registered.
1819  */
1820 static void hugetlb_register_node(struct node *node)
1821 {
1822         struct hstate *h;
1823         struct node_hstate *nhs = &node_hstates[node->dev.id];
1824         int err;
1825
1826         if (nhs->hugepages_kobj)
1827                 return;         /* already allocated */
1828
1829         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1830                                                         &node->dev.kobj);
1831         if (!nhs->hugepages_kobj)
1832                 return;
1833
1834         for_each_hstate(h) {
1835                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1836                                                 nhs->hstate_kobjs,
1837                                                 &per_node_hstate_attr_group);
1838                 if (err) {
1839                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1840                                 h->name, node->dev.id);
1841                         hugetlb_unregister_node(node);
1842                         break;
1843                 }
1844         }
1845 }
1846
1847 /*
1848  * hugetlb init time:  register hstate attributes for all registered node
1849  * devices of nodes that have memory.  All on-line nodes should have
1850  * registered their associated device by this time.
1851  */
1852 static void hugetlb_register_all_nodes(void)
1853 {
1854         int nid;
1855
1856         for_each_node_state(nid, N_MEMORY) {
1857                 struct node *node = node_devices[nid];
1858                 if (node->dev.id == nid)
1859                         hugetlb_register_node(node);
1860         }
1861
1862         /*
1863          * Let the node device driver know we're here so it can
1864          * [un]register hstate attributes on node hotplug.
1865          */
1866         register_hugetlbfs_with_node(hugetlb_register_node,
1867                                      hugetlb_unregister_node);
1868 }
1869 #else   /* !CONFIG_NUMA */
1870
1871 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1872 {
1873         BUG();
1874         if (nidp)
1875                 *nidp = -1;
1876         return NULL;
1877 }
1878
1879 static void hugetlb_unregister_all_nodes(void) { }
1880
1881 static void hugetlb_register_all_nodes(void) { }
1882
1883 #endif
1884
1885 static void __exit hugetlb_exit(void)
1886 {
1887         struct hstate *h;
1888
1889         hugetlb_unregister_all_nodes();
1890
1891         for_each_hstate(h) {
1892                 kobject_put(hstate_kobjs[hstate_index(h)]);
1893         }
1894
1895         kobject_put(hugepages_kobj);
1896 }
1897 module_exit(hugetlb_exit);
1898
1899 static int __init hugetlb_init(void)
1900 {
1901         /* Some platform decide whether they support huge pages at boot
1902          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1903          * there is no such support
1904          */
1905         if (HPAGE_SHIFT == 0)
1906                 return 0;
1907
1908         if (!size_to_hstate(default_hstate_size)) {
1909                 default_hstate_size = HPAGE_SIZE;
1910                 if (!size_to_hstate(default_hstate_size))
1911                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1912         }
1913         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1914         if (default_hstate_max_huge_pages)
1915                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1916
1917         hugetlb_init_hstates();
1918         gather_bootmem_prealloc();
1919         report_hugepages();
1920
1921         hugetlb_sysfs_init();
1922         hugetlb_register_all_nodes();
1923         hugetlb_cgroup_file_init();
1924
1925         return 0;
1926 }
1927 module_init(hugetlb_init);
1928
1929 /* Should be called on processing a hugepagesz=... option */
1930 void __init hugetlb_add_hstate(unsigned order)
1931 {
1932         struct hstate *h;
1933         unsigned long i;
1934
1935         if (size_to_hstate(PAGE_SIZE << order)) {
1936                 pr_warning("hugepagesz= specified twice, ignoring\n");
1937                 return;
1938         }
1939         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1940         BUG_ON(order == 0);
1941         h = &hstates[hugetlb_max_hstate++];
1942         h->order = order;
1943         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1944         h->nr_huge_pages = 0;
1945         h->free_huge_pages = 0;
1946         for (i = 0; i < MAX_NUMNODES; ++i)
1947                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1948         INIT_LIST_HEAD(&h->hugepage_activelist);
1949         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1950         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1951         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1952                                         huge_page_size(h)/1024);
1953
1954         parsed_hstate = h;
1955 }
1956
1957 static int __init hugetlb_nrpages_setup(char *s)
1958 {
1959         unsigned long *mhp;
1960         static unsigned long *last_mhp;
1961
1962         /*
1963          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1964          * so this hugepages= parameter goes to the "default hstate".
1965          */
1966         if (!hugetlb_max_hstate)
1967                 mhp = &default_hstate_max_huge_pages;
1968         else
1969                 mhp = &parsed_hstate->max_huge_pages;
1970
1971         if (mhp == last_mhp) {
1972                 pr_warning("hugepages= specified twice without "
1973                            "interleaving hugepagesz=, ignoring\n");
1974                 return 1;
1975         }
1976
1977         if (sscanf(s, "%lu", mhp) <= 0)
1978                 *mhp = 0;
1979
1980         /*
1981          * Global state is always initialized later in hugetlb_init.
1982          * But we need to allocate >= MAX_ORDER hstates here early to still
1983          * use the bootmem allocator.
1984          */
1985         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1986                 hugetlb_hstate_alloc_pages(parsed_hstate);
1987
1988         last_mhp = mhp;
1989
1990         return 1;
1991 }
1992 __setup("hugepages=", hugetlb_nrpages_setup);
1993
1994 static int __init hugetlb_default_setup(char *s)
1995 {
1996         default_hstate_size = memparse(s, &s);
1997         return 1;
1998 }
1999 __setup("default_hugepagesz=", hugetlb_default_setup);
2000
2001 static unsigned int cpuset_mems_nr(unsigned int *array)
2002 {
2003         int node;
2004         unsigned int nr = 0;
2005
2006         for_each_node_mask(node, cpuset_current_mems_allowed)
2007                 nr += array[node];
2008
2009         return nr;
2010 }
2011
2012 #ifdef CONFIG_SYSCTL
2013 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2014                          struct ctl_table *table, int write,
2015                          void __user *buffer, size_t *length, loff_t *ppos)
2016 {
2017         struct hstate *h = &default_hstate;
2018         unsigned long tmp;
2019         int ret;
2020
2021         tmp = h->max_huge_pages;
2022
2023         if (write && h->order >= MAX_ORDER)
2024                 return -EINVAL;
2025
2026         table->data = &tmp;
2027         table->maxlen = sizeof(unsigned long);
2028         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2029         if (ret)
2030                 goto out;
2031
2032         if (write) {
2033                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2034                                                 GFP_KERNEL | __GFP_NORETRY);
2035                 if (!(obey_mempolicy &&
2036                                init_nodemask_of_mempolicy(nodes_allowed))) {
2037                         NODEMASK_FREE(nodes_allowed);
2038                         nodes_allowed = &node_states[N_MEMORY];
2039                 }
2040                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2041
2042                 if (nodes_allowed != &node_states[N_MEMORY])
2043                         NODEMASK_FREE(nodes_allowed);
2044         }
2045 out:
2046         return ret;
2047 }
2048
2049 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2050                           void __user *buffer, size_t *length, loff_t *ppos)
2051 {
2052
2053         return hugetlb_sysctl_handler_common(false, table, write,
2054                                                         buffer, length, ppos);
2055 }
2056
2057 #ifdef CONFIG_NUMA
2058 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2059                           void __user *buffer, size_t *length, loff_t *ppos)
2060 {
2061         return hugetlb_sysctl_handler_common(true, table, write,
2062                                                         buffer, length, ppos);
2063 }
2064 #endif /* CONFIG_NUMA */
2065
2066 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2067                         void __user *buffer,
2068                         size_t *length, loff_t *ppos)
2069 {
2070         proc_dointvec(table, write, buffer, length, ppos);
2071         if (hugepages_treat_as_movable)
2072                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2073         else
2074                 htlb_alloc_mask = GFP_HIGHUSER;
2075         return 0;
2076 }
2077
2078 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2079                         void __user *buffer,
2080                         size_t *length, loff_t *ppos)
2081 {
2082         struct hstate *h = &default_hstate;
2083         unsigned long tmp;
2084         int ret;
2085
2086         tmp = h->nr_overcommit_huge_pages;
2087
2088         if (write && h->order >= MAX_ORDER)
2089                 return -EINVAL;
2090
2091         table->data = &tmp;
2092         table->maxlen = sizeof(unsigned long);
2093         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2094         if (ret)
2095                 goto out;
2096
2097         if (write) {
2098                 spin_lock(&hugetlb_lock);
2099                 h->nr_overcommit_huge_pages = tmp;
2100                 spin_unlock(&hugetlb_lock);
2101         }
2102 out:
2103         return ret;
2104 }
2105
2106 #endif /* CONFIG_SYSCTL */
2107
2108 void hugetlb_report_meminfo(struct seq_file *m)
2109 {
2110         struct hstate *h = &default_hstate;
2111         seq_printf(m,
2112                         "HugePages_Total:   %5lu\n"
2113                         "HugePages_Free:    %5lu\n"
2114                         "HugePages_Rsvd:    %5lu\n"
2115                         "HugePages_Surp:    %5lu\n"
2116                         "Hugepagesize:   %8lu kB\n",
2117                         h->nr_huge_pages,
2118                         h->free_huge_pages,
2119                         h->resv_huge_pages,
2120                         h->surplus_huge_pages,
2121                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2122 }
2123
2124 int hugetlb_report_node_meminfo(int nid, char *buf)
2125 {
2126         struct hstate *h = &default_hstate;
2127         return sprintf(buf,
2128                 "Node %d HugePages_Total: %5u\n"
2129                 "Node %d HugePages_Free:  %5u\n"
2130                 "Node %d HugePages_Surp:  %5u\n",
2131                 nid, h->nr_huge_pages_node[nid],
2132                 nid, h->free_huge_pages_node[nid],
2133                 nid, h->surplus_huge_pages_node[nid]);
2134 }
2135
2136 void hugetlb_show_meminfo(void)
2137 {
2138         struct hstate *h;
2139         int nid;
2140
2141         for_each_node_state(nid, N_MEMORY)
2142                 for_each_hstate(h)
2143                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2144                                 nid,
2145                                 h->nr_huge_pages_node[nid],
2146                                 h->free_huge_pages_node[nid],
2147                                 h->surplus_huge_pages_node[nid],
2148                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2149 }
2150
2151 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2152 unsigned long hugetlb_total_pages(void)
2153 {
2154         struct hstate *h;
2155         unsigned long nr_total_pages = 0;
2156
2157         for_each_hstate(h)
2158                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2159         return nr_total_pages;
2160 }
2161
2162 static int hugetlb_acct_memory(struct hstate *h, long delta)
2163 {
2164         int ret = -ENOMEM;
2165
2166         spin_lock(&hugetlb_lock);
2167         /*
2168          * When cpuset is configured, it breaks the strict hugetlb page
2169          * reservation as the accounting is done on a global variable. Such
2170          * reservation is completely rubbish in the presence of cpuset because
2171          * the reservation is not checked against page availability for the
2172          * current cpuset. Application can still potentially OOM'ed by kernel
2173          * with lack of free htlb page in cpuset that the task is in.
2174          * Attempt to enforce strict accounting with cpuset is almost
2175          * impossible (or too ugly) because cpuset is too fluid that
2176          * task or memory node can be dynamically moved between cpusets.
2177          *
2178          * The change of semantics for shared hugetlb mapping with cpuset is
2179          * undesirable. However, in order to preserve some of the semantics,
2180          * we fall back to check against current free page availability as
2181          * a best attempt and hopefully to minimize the impact of changing
2182          * semantics that cpuset has.
2183          */
2184         if (delta > 0) {
2185                 if (gather_surplus_pages(h, delta) < 0)
2186                         goto out;
2187
2188                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2189                         return_unused_surplus_pages(h, delta);
2190                         goto out;
2191                 }
2192         }
2193
2194         ret = 0;
2195         if (delta < 0)
2196                 return_unused_surplus_pages(h, (unsigned long) -delta);
2197
2198 out:
2199         spin_unlock(&hugetlb_lock);
2200         return ret;
2201 }
2202
2203 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2204 {
2205         struct resv_map *reservations = vma_resv_map(vma);
2206
2207         /*
2208          * This new VMA should share its siblings reservation map if present.
2209          * The VMA will only ever have a valid reservation map pointer where
2210          * it is being copied for another still existing VMA.  As that VMA
2211          * has a reference to the reservation map it cannot disappear until
2212          * after this open call completes.  It is therefore safe to take a
2213          * new reference here without additional locking.
2214          */
2215         if (reservations)
2216                 kref_get(&reservations->refs);
2217 }
2218
2219 static void resv_map_put(struct vm_area_struct *vma)
2220 {
2221         struct resv_map *reservations = vma_resv_map(vma);
2222
2223         if (!reservations)
2224                 return;
2225         kref_put(&reservations->refs, resv_map_release);
2226 }
2227
2228 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2229 {
2230         struct hstate *h = hstate_vma(vma);
2231         struct resv_map *reservations = vma_resv_map(vma);
2232         struct hugepage_subpool *spool = subpool_vma(vma);
2233         unsigned long reserve;
2234         unsigned long start;
2235         unsigned long end;
2236
2237         if (reservations) {
2238                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2239                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2240
2241                 reserve = (end - start) -
2242                         region_count(&reservations->regions, start, end);
2243
2244                 resv_map_put(vma);
2245
2246                 if (reserve) {
2247                         hugetlb_acct_memory(h, -reserve);
2248                         hugepage_subpool_put_pages(spool, reserve);
2249                 }
2250         }
2251 }
2252
2253 /*
2254  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2255  * handle_mm_fault() to try to instantiate regular-sized pages in the
2256  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2257  * this far.
2258  */
2259 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2260 {
2261         BUG();
2262         return 0;
2263 }
2264
2265 const struct vm_operations_struct hugetlb_vm_ops = {
2266         .fault = hugetlb_vm_op_fault,
2267         .open = hugetlb_vm_op_open,
2268         .close = hugetlb_vm_op_close,
2269 };
2270
2271 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2272                                 int writable)
2273 {
2274         pte_t entry;
2275
2276         if (writable) {
2277                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2278                                          vma->vm_page_prot)));
2279         } else {
2280                 entry = huge_pte_wrprotect(mk_huge_pte(page,
2281                                            vma->vm_page_prot));
2282         }
2283         entry = pte_mkyoung(entry);
2284         entry = pte_mkhuge(entry);
2285         entry = arch_make_huge_pte(entry, vma, page, writable);
2286
2287         return entry;
2288 }
2289
2290 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2291                                    unsigned long address, pte_t *ptep)
2292 {
2293         pte_t entry;
2294
2295         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2296         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2297                 update_mmu_cache(vma, address, ptep);
2298 }
2299
2300
2301 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2302                             struct vm_area_struct *vma)
2303 {
2304         pte_t *src_pte, *dst_pte, entry;
2305         struct page *ptepage;
2306         unsigned long addr;
2307         int cow;
2308         struct hstate *h = hstate_vma(vma);
2309         unsigned long sz = huge_page_size(h);
2310
2311         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2312
2313         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2314                 src_pte = huge_pte_offset(src, addr);
2315                 if (!src_pte)
2316                         continue;
2317                 dst_pte = huge_pte_alloc(dst, addr, sz);
2318                 if (!dst_pte)
2319                         goto nomem;
2320
2321                 /* If the pagetables are shared don't copy or take references */
2322                 if (dst_pte == src_pte)
2323                         continue;
2324
2325                 spin_lock(&dst->page_table_lock);
2326                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2327                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2328                         if (cow)
2329                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2330                         entry = huge_ptep_get(src_pte);
2331                         ptepage = pte_page(entry);
2332                         get_page(ptepage);
2333                         page_dup_rmap(ptepage);
2334                         set_huge_pte_at(dst, addr, dst_pte, entry);
2335                 }
2336                 spin_unlock(&src->page_table_lock);
2337                 spin_unlock(&dst->page_table_lock);
2338         }
2339         return 0;
2340
2341 nomem:
2342         return -ENOMEM;
2343 }
2344
2345 static int is_hugetlb_entry_migration(pte_t pte)
2346 {
2347         swp_entry_t swp;
2348
2349         if (huge_pte_none(pte) || pte_present(pte))
2350                 return 0;
2351         swp = pte_to_swp_entry(pte);
2352         if (non_swap_entry(swp) && is_migration_entry(swp))
2353                 return 1;
2354         else
2355                 return 0;
2356 }
2357
2358 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2359 {
2360         swp_entry_t swp;
2361
2362         if (huge_pte_none(pte) || pte_present(pte))
2363                 return 0;
2364         swp = pte_to_swp_entry(pte);
2365         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2366                 return 1;
2367         else
2368                 return 0;
2369 }
2370
2371 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2372                             unsigned long start, unsigned long end,
2373                             struct page *ref_page)
2374 {
2375         int force_flush = 0;
2376         struct mm_struct *mm = vma->vm_mm;
2377         unsigned long address;
2378         pte_t *ptep;
2379         pte_t pte;
2380         struct page *page;
2381         struct hstate *h = hstate_vma(vma);
2382         unsigned long sz = huge_page_size(h);
2383         const unsigned long mmun_start = start; /* For mmu_notifiers */
2384         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2385
2386         WARN_ON(!is_vm_hugetlb_page(vma));
2387         BUG_ON(start & ~huge_page_mask(h));
2388         BUG_ON(end & ~huge_page_mask(h));
2389
2390         tlb_start_vma(tlb, vma);
2391         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2392 again:
2393         spin_lock(&mm->page_table_lock);
2394         for (address = start; address < end; address += sz) {
2395                 ptep = huge_pte_offset(mm, address);
2396                 if (!ptep)
2397                         continue;
2398
2399                 if (huge_pmd_unshare(mm, &address, ptep))
2400                         continue;
2401
2402                 pte = huge_ptep_get(ptep);
2403                 if (huge_pte_none(pte))
2404                         continue;
2405
2406                 /*
2407                  * HWPoisoned hugepage is already unmapped and dropped reference
2408                  */
2409                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2410                         huge_pte_clear(mm, address, ptep);
2411                         continue;
2412                 }
2413
2414                 page = pte_page(pte);
2415                 /*
2416                  * If a reference page is supplied, it is because a specific
2417                  * page is being unmapped, not a range. Ensure the page we
2418                  * are about to unmap is the actual page of interest.
2419                  */
2420                 if (ref_page) {
2421                         if (page != ref_page)
2422                                 continue;
2423
2424                         /*
2425                          * Mark the VMA as having unmapped its page so that
2426                          * future faults in this VMA will fail rather than
2427                          * looking like data was lost
2428                          */
2429                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2430                 }
2431
2432                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2433                 tlb_remove_tlb_entry(tlb, ptep, address);
2434                 if (huge_pte_dirty(pte))
2435                         set_page_dirty(page);
2436
2437                 page_remove_rmap(page);
2438                 force_flush = !__tlb_remove_page(tlb, page);
2439                 if (force_flush)
2440                         break;
2441                 /* Bail out after unmapping reference page if supplied */
2442                 if (ref_page)
2443                         break;
2444         }
2445         spin_unlock(&mm->page_table_lock);
2446         /*
2447          * mmu_gather ran out of room to batch pages, we break out of
2448          * the PTE lock to avoid doing the potential expensive TLB invalidate
2449          * and page-free while holding it.
2450          */
2451         if (force_flush) {
2452                 force_flush = 0;
2453                 tlb_flush_mmu(tlb);
2454                 if (address < end && !ref_page)
2455                         goto again;
2456         }
2457         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2458         tlb_end_vma(tlb, vma);
2459 }
2460
2461 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2462                           struct vm_area_struct *vma, unsigned long start,
2463                           unsigned long end, struct page *ref_page)
2464 {
2465         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2466
2467         /*
2468          * Clear this flag so that x86's huge_pmd_share page_table_shareable
2469          * test will fail on a vma being torn down, and not grab a page table
2470          * on its way out.  We're lucky that the flag has such an appropriate
2471          * name, and can in fact be safely cleared here. We could clear it
2472          * before the __unmap_hugepage_range above, but all that's necessary
2473          * is to clear it before releasing the i_mmap_mutex. This works
2474          * because in the context this is called, the VMA is about to be
2475          * destroyed and the i_mmap_mutex is held.
2476          */
2477         vma->vm_flags &= ~VM_MAYSHARE;
2478 }
2479
2480 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2481                           unsigned long end, struct page *ref_page)
2482 {
2483         struct mm_struct *mm;
2484         struct mmu_gather tlb;
2485
2486         mm = vma->vm_mm;
2487
2488         tlb_gather_mmu(&tlb, mm, start, end);
2489         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2490         tlb_finish_mmu(&tlb, start, end);
2491 }
2492
2493 /*
2494  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2495  * mappping it owns the reserve page for. The intention is to unmap the page
2496  * from other VMAs and let the children be SIGKILLed if they are faulting the
2497  * same region.
2498  */
2499 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2500                                 struct page *page, unsigned long address)
2501 {
2502         struct hstate *h = hstate_vma(vma);
2503         struct vm_area_struct *iter_vma;
2504         struct address_space *mapping;
2505         pgoff_t pgoff;
2506
2507         /*
2508          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2509          * from page cache lookup which is in HPAGE_SIZE units.
2510          */
2511         address = address & huge_page_mask(h);
2512         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2513                         vma->vm_pgoff;
2514         mapping = file_inode(vma->vm_file)->i_mapping;
2515
2516         /*
2517          * Take the mapping lock for the duration of the table walk. As
2518          * this mapping should be shared between all the VMAs,
2519          * __unmap_hugepage_range() is called as the lock is already held
2520          */
2521         mutex_lock(&mapping->i_mmap_mutex);
2522         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2523                 /* Do not unmap the current VMA */
2524                 if (iter_vma == vma)
2525                         continue;
2526
2527                 /*
2528                  * Unmap the page from other VMAs without their own reserves.
2529                  * They get marked to be SIGKILLed if they fault in these
2530                  * areas. This is because a future no-page fault on this VMA
2531                  * could insert a zeroed page instead of the data existing
2532                  * from the time of fork. This would look like data corruption
2533                  */
2534                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2535                         unmap_hugepage_range(iter_vma, address,
2536                                              address + huge_page_size(h), page);
2537         }
2538         mutex_unlock(&mapping->i_mmap_mutex);
2539
2540         return 1;
2541 }
2542
2543 /*
2544  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2545  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2546  * cannot race with other handlers or page migration.
2547  * Keep the pte_same checks anyway to make transition from the mutex easier.
2548  */
2549 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2550                         unsigned long address, pte_t *ptep, pte_t pte,
2551                         struct page *pagecache_page)
2552 {
2553         struct hstate *h = hstate_vma(vma);
2554         struct page *old_page, *new_page;
2555         int avoidcopy;
2556         int outside_reserve = 0;
2557         unsigned long mmun_start;       /* For mmu_notifiers */
2558         unsigned long mmun_end;         /* For mmu_notifiers */
2559
2560         old_page = pte_page(pte);
2561
2562 retry_avoidcopy:
2563         /* If no-one else is actually using this page, avoid the copy
2564          * and just make the page writable */
2565         avoidcopy = (page_mapcount(old_page) == 1);
2566         if (avoidcopy) {
2567                 if (PageAnon(old_page))
2568                         page_move_anon_rmap(old_page, vma, address);
2569                 set_huge_ptep_writable(vma, address, ptep);
2570                 return 0;
2571         }
2572
2573         /*
2574          * If the process that created a MAP_PRIVATE mapping is about to
2575          * perform a COW due to a shared page count, attempt to satisfy
2576          * the allocation without using the existing reserves. The pagecache
2577          * page is used to determine if the reserve at this address was
2578          * consumed or not. If reserves were used, a partial faulted mapping
2579          * at the time of fork() could consume its reserves on COW instead
2580          * of the full address range.
2581          */
2582         if (!(vma->vm_flags & VM_MAYSHARE) &&
2583                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2584                         old_page != pagecache_page)
2585                 outside_reserve = 1;
2586
2587         page_cache_get(old_page);
2588
2589         /* Drop page_table_lock as buddy allocator may be called */
2590         spin_unlock(&mm->page_table_lock);
2591         new_page = alloc_huge_page(vma, address, outside_reserve);
2592
2593         if (IS_ERR(new_page)) {
2594                 long err = PTR_ERR(new_page);
2595                 page_cache_release(old_page);
2596
2597                 /*
2598                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2599                  * it is due to references held by a child and an insufficient
2600                  * huge page pool. To guarantee the original mappers
2601                  * reliability, unmap the page from child processes. The child
2602                  * may get SIGKILLed if it later faults.
2603                  */
2604                 if (outside_reserve) {
2605                         BUG_ON(huge_pte_none(pte));
2606                         if (unmap_ref_private(mm, vma, old_page, address)) {
2607                                 BUG_ON(huge_pte_none(pte));
2608                                 spin_lock(&mm->page_table_lock);
2609                                 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2610                                 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2611                                         goto retry_avoidcopy;
2612                                 /*
2613                                  * race occurs while re-acquiring page_table_lock, and
2614                                  * our job is done.
2615                                  */
2616                                 return 0;
2617                         }
2618                         WARN_ON_ONCE(1);
2619                 }
2620
2621                 /* Caller expects lock to be held */
2622                 spin_lock(&mm->page_table_lock);
2623                 if (err == -ENOMEM)
2624                         return VM_FAULT_OOM;
2625                 else
2626                         return VM_FAULT_SIGBUS;
2627         }
2628
2629         /*
2630          * When the original hugepage is shared one, it does not have
2631          * anon_vma prepared.
2632          */
2633         if (unlikely(anon_vma_prepare(vma))) {
2634                 page_cache_release(new_page);
2635                 page_cache_release(old_page);
2636                 /* Caller expects lock to be held */
2637                 spin_lock(&mm->page_table_lock);
2638                 return VM_FAULT_OOM;
2639         }
2640
2641         copy_user_huge_page(new_page, old_page, address, vma,
2642                             pages_per_huge_page(h));
2643         __SetPageUptodate(new_page);
2644
2645         mmun_start = address & huge_page_mask(h);
2646         mmun_end = mmun_start + huge_page_size(h);
2647         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2648         /*
2649          * Retake the page_table_lock to check for racing updates
2650          * before the page tables are altered
2651          */
2652         spin_lock(&mm->page_table_lock);
2653         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2654         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2655                 /* Break COW */
2656                 huge_ptep_clear_flush(vma, address, ptep);
2657                 set_huge_pte_at(mm, address, ptep,
2658                                 make_huge_pte(vma, new_page, 1));
2659                 page_remove_rmap(old_page);
2660                 hugepage_add_new_anon_rmap(new_page, vma, address);
2661                 /* Make the old page be freed below */
2662                 new_page = old_page;
2663         }
2664         spin_unlock(&mm->page_table_lock);
2665         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2666         /* Caller expects lock to be held */
2667         spin_lock(&mm->page_table_lock);
2668         page_cache_release(new_page);
2669         page_cache_release(old_page);
2670         return 0;
2671 }
2672
2673 /* Return the pagecache page at a given address within a VMA */
2674 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2675                         struct vm_area_struct *vma, unsigned long address)
2676 {
2677         struct address_space *mapping;
2678         pgoff_t idx;
2679
2680         mapping = vma->vm_file->f_mapping;
2681         idx = vma_hugecache_offset(h, vma, address);
2682
2683         return find_lock_page(mapping, idx);
2684 }
2685
2686 /*
2687  * Return whether there is a pagecache page to back given address within VMA.
2688  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2689  */
2690 static bool hugetlbfs_pagecache_present(struct hstate *h,
2691                         struct vm_area_struct *vma, unsigned long address)
2692 {
2693         struct address_space *mapping;
2694         pgoff_t idx;
2695         struct page *page;
2696
2697         mapping = vma->vm_file->f_mapping;
2698         idx = vma_hugecache_offset(h, vma, address);
2699
2700         page = find_get_page(mapping, idx);
2701         if (page)
2702                 put_page(page);
2703         return page != NULL;
2704 }
2705
2706 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2707                         unsigned long address, pte_t *ptep, unsigned int flags)
2708 {
2709         struct hstate *h = hstate_vma(vma);
2710         int ret = VM_FAULT_SIGBUS;
2711         int anon_rmap = 0;
2712         pgoff_t idx;
2713         unsigned long size;
2714         struct page *page;
2715         struct address_space *mapping;
2716         pte_t new_pte;
2717
2718         /*
2719          * Currently, we are forced to kill the process in the event the
2720          * original mapper has unmapped pages from the child due to a failed
2721          * COW. Warn that such a situation has occurred as it may not be obvious
2722          */
2723         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2724                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2725                            current->pid);
2726                 return ret;
2727         }
2728
2729         mapping = vma->vm_file->f_mapping;
2730         idx = vma_hugecache_offset(h, vma, address);
2731
2732         /*
2733          * Use page lock to guard against racing truncation
2734          * before we get page_table_lock.
2735          */
2736 retry:
2737         page = find_lock_page(mapping, idx);
2738         if (!page) {
2739                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2740                 if (idx >= size)
2741                         goto out;
2742                 page = alloc_huge_page(vma, address, 0);
2743                 if (IS_ERR(page)) {
2744                         ret = PTR_ERR(page);
2745                         if (ret == -ENOMEM)
2746                                 ret = VM_FAULT_OOM;
2747                         else
2748                                 ret = VM_FAULT_SIGBUS;
2749                         goto out;
2750                 }
2751                 clear_huge_page(page, address, pages_per_huge_page(h));
2752                 __SetPageUptodate(page);
2753
2754                 if (vma->vm_flags & VM_MAYSHARE) {
2755                         int err;
2756                         struct inode *inode = mapping->host;
2757
2758                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2759                         if (err) {
2760                                 put_page(page);
2761                                 if (err == -EEXIST)
2762                                         goto retry;
2763                                 goto out;
2764                         }
2765
2766                         spin_lock(&inode->i_lock);
2767                         inode->i_blocks += blocks_per_huge_page(h);
2768                         spin_unlock(&inode->i_lock);
2769                 } else {
2770                         lock_page(page);
2771                         if (unlikely(anon_vma_prepare(vma))) {
2772                                 ret = VM_FAULT_OOM;
2773                                 goto backout_unlocked;
2774                         }
2775                         anon_rmap = 1;
2776                 }
2777         } else {
2778                 /*
2779                  * If memory error occurs between mmap() and fault, some process
2780                  * don't have hwpoisoned swap entry for errored virtual address.
2781                  * So we need to block hugepage fault by PG_hwpoison bit check.
2782                  */
2783                 if (unlikely(PageHWPoison(page))) {
2784                         ret = VM_FAULT_HWPOISON |
2785                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2786                         goto backout_unlocked;
2787                 }
2788         }
2789
2790         /*
2791          * If we are going to COW a private mapping later, we examine the
2792          * pending reservations for this page now. This will ensure that
2793          * any allocations necessary to record that reservation occur outside
2794          * the spinlock.
2795          */
2796         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2797                 if (vma_needs_reservation(h, vma, address) < 0) {
2798                         ret = VM_FAULT_OOM;
2799                         goto backout_unlocked;
2800                 }
2801
2802         spin_lock(&mm->page_table_lock);
2803         size = i_size_read(mapping->host) >> huge_page_shift(h);
2804         if (idx >= size)
2805                 goto backout;
2806
2807         ret = 0;
2808         if (!huge_pte_none(huge_ptep_get(ptep)))
2809                 goto backout;
2810
2811         if (anon_rmap)
2812                 hugepage_add_new_anon_rmap(page, vma, address);
2813         else
2814                 page_dup_rmap(page);
2815         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2816                                 && (vma->vm_flags & VM_SHARED)));
2817         set_huge_pte_at(mm, address, ptep, new_pte);
2818
2819         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2820                 /* Optimization, do the COW without a second fault */
2821                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2822         }
2823
2824         spin_unlock(&mm->page_table_lock);
2825         unlock_page(page);
2826 out:
2827         return ret;
2828
2829 backout:
2830         spin_unlock(&mm->page_table_lock);
2831 backout_unlocked:
2832         unlock_page(page);
2833         put_page(page);
2834         goto out;
2835 }
2836
2837 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2838                         unsigned long address, unsigned int flags)
2839 {
2840         pte_t *ptep;
2841         pte_t entry;
2842         int ret;
2843         struct page *page = NULL;
2844         struct page *pagecache_page = NULL;
2845         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2846         struct hstate *h = hstate_vma(vma);
2847
2848         address &= huge_page_mask(h);
2849
2850         ptep = huge_pte_offset(mm, address);
2851         if (ptep) {
2852                 entry = huge_ptep_get(ptep);
2853                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2854                         migration_entry_wait_huge(mm, ptep);
2855                         return 0;
2856                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2857                         return VM_FAULT_HWPOISON_LARGE |
2858                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2859         }
2860
2861         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2862         if (!ptep)
2863                 return VM_FAULT_OOM;
2864
2865         /*
2866          * Serialize hugepage allocation and instantiation, so that we don't
2867          * get spurious allocation failures if two CPUs race to instantiate
2868          * the same page in the page cache.
2869          */
2870         mutex_lock(&hugetlb_instantiation_mutex);
2871         entry = huge_ptep_get(ptep);
2872         if (huge_pte_none(entry)) {
2873                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2874                 goto out_mutex;
2875         }
2876
2877         ret = 0;
2878
2879         /*
2880          * If we are going to COW the mapping later, we examine the pending
2881          * reservations for this page now. This will ensure that any
2882          * allocations necessary to record that reservation occur outside the
2883          * spinlock. For private mappings, we also lookup the pagecache
2884          * page now as it is used to determine if a reservation has been
2885          * consumed.
2886          */
2887         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2888                 if (vma_needs_reservation(h, vma, address) < 0) {
2889                         ret = VM_FAULT_OOM;
2890                         goto out_mutex;
2891                 }
2892
2893                 if (!(vma->vm_flags & VM_MAYSHARE))
2894                         pagecache_page = hugetlbfs_pagecache_page(h,
2895                                                                 vma, address);
2896         }
2897
2898         /*
2899          * hugetlb_cow() requires page locks of pte_page(entry) and
2900          * pagecache_page, so here we need take the former one
2901          * when page != pagecache_page or !pagecache_page.
2902          * Note that locking order is always pagecache_page -> page,
2903          * so no worry about deadlock.
2904          */
2905         page = pte_page(entry);
2906         get_page(page);
2907         if (page != pagecache_page)
2908                 lock_page(page);
2909
2910         spin_lock(&mm->page_table_lock);
2911         /* Check for a racing update before calling hugetlb_cow */
2912         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2913                 goto out_page_table_lock;
2914
2915
2916         if (flags & FAULT_FLAG_WRITE) {
2917                 if (!huge_pte_write(entry)) {
2918                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2919                                                         pagecache_page);
2920                         goto out_page_table_lock;
2921                 }
2922                 entry = huge_pte_mkdirty(entry);
2923         }
2924         entry = pte_mkyoung(entry);
2925         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2926                                                 flags & FAULT_FLAG_WRITE))
2927                 update_mmu_cache(vma, address, ptep);
2928
2929 out_page_table_lock:
2930         spin_unlock(&mm->page_table_lock);
2931
2932         if (pagecache_page) {
2933                 unlock_page(pagecache_page);
2934                 put_page(pagecache_page);
2935         }
2936         if (page != pagecache_page)
2937                 unlock_page(page);
2938         put_page(page);
2939
2940 out_mutex:
2941         mutex_unlock(&hugetlb_instantiation_mutex);
2942
2943         return ret;
2944 }
2945
2946 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2947                          struct page **pages, struct vm_area_struct **vmas,
2948                          unsigned long *position, unsigned long *nr_pages,
2949                          long i, unsigned int flags)
2950 {
2951         unsigned long pfn_offset;
2952         unsigned long vaddr = *position;
2953         unsigned long remainder = *nr_pages;
2954         struct hstate *h = hstate_vma(vma);
2955
2956         spin_lock(&mm->page_table_lock);
2957         while (vaddr < vma->vm_end && remainder) {
2958                 pte_t *pte;
2959                 int absent;
2960                 struct page *page;
2961
2962                 /*
2963                  * Some archs (sparc64, sh*) have multiple pte_ts to
2964                  * each hugepage.  We have to make sure we get the
2965                  * first, for the page indexing below to work.
2966                  */
2967                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2968                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2969
2970                 /*
2971                  * When coredumping, it suits get_dump_page if we just return
2972                  * an error where there's an empty slot with no huge pagecache
2973                  * to back it.  This way, we avoid allocating a hugepage, and
2974                  * the sparse dumpfile avoids allocating disk blocks, but its
2975                  * huge holes still show up with zeroes where they need to be.
2976                  */
2977                 if (absent && (flags & FOLL_DUMP) &&
2978                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2979                         remainder = 0;
2980                         break;
2981                 }
2982
2983                 /*
2984                  * We need call hugetlb_fault for both hugepages under migration
2985                  * (in which case hugetlb_fault waits for the migration,) and
2986                  * hwpoisoned hugepages (in which case we need to prevent the
2987                  * caller from accessing to them.) In order to do this, we use
2988                  * here is_swap_pte instead of is_hugetlb_entry_migration and
2989                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2990                  * both cases, and because we can't follow correct pages
2991                  * directly from any kind of swap entries.
2992                  */
2993                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2994                     ((flags & FOLL_WRITE) &&
2995                       !huge_pte_write(huge_ptep_get(pte)))) {
2996                         int ret;
2997
2998                         spin_unlock(&mm->page_table_lock);
2999                         ret = hugetlb_fault(mm, vma, vaddr,
3000                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3001                         spin_lock(&mm->page_table_lock);
3002                         if (!(ret & VM_FAULT_ERROR))
3003                                 continue;
3004
3005                         remainder = 0;
3006                         break;
3007                 }
3008
3009                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3010                 page = pte_page(huge_ptep_get(pte));
3011 same_page:
3012                 if (pages) {
3013                         pages[i] = mem_map_offset(page, pfn_offset);
3014                         get_page(pages[i]);
3015                 }
3016
3017                 if (vmas)
3018                         vmas[i] = vma;
3019
3020                 vaddr += PAGE_SIZE;
3021                 ++pfn_offset;
3022                 --remainder;
3023                 ++i;
3024                 if (vaddr < vma->vm_end && remainder &&
3025                                 pfn_offset < pages_per_huge_page(h)) {
3026                         /*
3027                          * We use pfn_offset to avoid touching the pageframes
3028                          * of this compound page.
3029                          */
3030                         goto same_page;
3031                 }
3032         }
3033         spin_unlock(&mm->page_table_lock);
3034         *nr_pages = remainder;
3035         *position = vaddr;
3036
3037         return i ? i : -EFAULT;
3038 }
3039
3040 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3041                 unsigned long address, unsigned long end, pgprot_t newprot)
3042 {
3043         struct mm_struct *mm = vma->vm_mm;
3044         unsigned long start = address;
3045         pte_t *ptep;
3046         pte_t pte;
3047         struct hstate *h = hstate_vma(vma);
3048         unsigned long pages = 0;
3049
3050         BUG_ON(address >= end);
3051         flush_cache_range(vma, address, end);
3052
3053         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3054         spin_lock(&mm->page_table_lock);
3055         for (; address < end; address += huge_page_size(h)) {
3056                 ptep = huge_pte_offset(mm, address);
3057                 if (!ptep)
3058                         continue;
3059                 if (huge_pmd_unshare(mm, &address, ptep)) {
3060                         pages++;
3061                         continue;
3062                 }
3063                 if (!huge_pte_none(huge_ptep_get(ptep))) {
3064                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3065                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3066                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3067                         set_huge_pte_at(mm, address, ptep, pte);
3068                         pages++;
3069                 }
3070         }
3071         spin_unlock(&mm->page_table_lock);
3072         /*
3073          * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3074          * may have cleared our pud entry and done put_page on the page table:
3075          * once we release i_mmap_mutex, another task can do the final put_page
3076          * and that page table be reused and filled with junk.
3077          */
3078         flush_tlb_range(vma, start, end);
3079         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3080
3081         return pages << h->order;
3082 }
3083
3084 int hugetlb_reserve_pages(struct inode *inode,
3085                                         long from, long to,
3086                                         struct vm_area_struct *vma,
3087                                         vm_flags_t vm_flags)
3088 {
3089         long ret, chg;
3090         struct hstate *h = hstate_inode(inode);
3091         struct hugepage_subpool *spool = subpool_inode(inode);
3092
3093         /*
3094          * Only apply hugepage reservation if asked. At fault time, an
3095          * attempt will be made for VM_NORESERVE to allocate a page
3096          * without using reserves
3097          */
3098         if (vm_flags & VM_NORESERVE)
3099                 return 0;
3100
3101         /*
3102          * Shared mappings base their reservation on the number of pages that
3103          * are already allocated on behalf of the file. Private mappings need
3104          * to reserve the full area even if read-only as mprotect() may be
3105          * called to make the mapping read-write. Assume !vma is a shm mapping
3106          */
3107         if (!vma || vma->vm_flags & VM_MAYSHARE)
3108                 chg = region_chg(&inode->i_mapping->private_list, from, to);
3109         else {
3110                 struct resv_map *resv_map = resv_map_alloc();
3111                 if (!resv_map)
3112                         return -ENOMEM;
3113
3114                 chg = to - from;
3115
3116                 set_vma_resv_map(vma, resv_map);
3117                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3118         }
3119
3120         if (chg < 0) {
3121                 ret = chg;
3122                 goto out_err;
3123         }
3124
3125         /* There must be enough pages in the subpool for the mapping */
3126         if (hugepage_subpool_get_pages(spool, chg)) {
3127                 ret = -ENOSPC;
3128                 goto out_err;
3129         }
3130
3131         /*
3132          * Check enough hugepages are available for the reservation.
3133          * Hand the pages back to the subpool if there are not
3134          */
3135         ret = hugetlb_acct_memory(h, chg);
3136         if (ret < 0) {
3137                 hugepage_subpool_put_pages(spool, chg);
3138                 goto out_err;
3139         }
3140
3141         /*
3142          * Account for the reservations made. Shared mappings record regions
3143          * that have reservations as they are shared by multiple VMAs.
3144          * When the last VMA disappears, the region map says how much
3145          * the reservation was and the page cache tells how much of
3146          * the reservation was consumed. Private mappings are per-VMA and
3147          * only the consumed reservations are tracked. When the VMA
3148          * disappears, the original reservation is the VMA size and the
3149          * consumed reservations are stored in the map. Hence, nothing
3150          * else has to be done for private mappings here
3151          */
3152         if (!vma || vma->vm_flags & VM_MAYSHARE)
3153                 region_add(&inode->i_mapping->private_list, from, to);
3154         return 0;
3155 out_err:
3156         if (vma)
3157                 resv_map_put(vma);
3158         return ret;
3159 }
3160
3161 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3162 {
3163         struct hstate *h = hstate_inode(inode);
3164         long chg = region_truncate(&inode->i_mapping->private_list, offset);
3165         struct hugepage_subpool *spool = subpool_inode(inode);
3166
3167         spin_lock(&inode->i_lock);
3168         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3169         spin_unlock(&inode->i_lock);
3170
3171         hugepage_subpool_put_pages(spool, (chg - freed));
3172         hugetlb_acct_memory(h, -(chg - freed));
3173 }
3174
3175 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3176 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3177                                 struct vm_area_struct *vma,
3178                                 unsigned long addr, pgoff_t idx)
3179 {
3180         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3181                                 svma->vm_start;
3182         unsigned long sbase = saddr & PUD_MASK;
3183         unsigned long s_end = sbase + PUD_SIZE;
3184
3185         /* Allow segments to share if only one is marked locked */
3186         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3187         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3188
3189         /*
3190          * match the virtual addresses, permission and the alignment of the
3191          * page table page.
3192          */
3193         if (pmd_index(addr) != pmd_index(saddr) ||
3194             vm_flags != svm_flags ||
3195             sbase < svma->vm_start || svma->vm_end < s_end)
3196                 return 0;
3197
3198         return saddr;
3199 }
3200
3201 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3202 {
3203         unsigned long base = addr & PUD_MASK;
3204         unsigned long end = base + PUD_SIZE;
3205
3206         /*
3207          * check on proper vm_flags and page table alignment
3208          */
3209         if (vma->vm_flags & VM_MAYSHARE &&
3210             vma->vm_start <= base && end <= vma->vm_end)
3211                 return 1;
3212         return 0;
3213 }
3214
3215 /*
3216  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3217  * and returns the corresponding pte. While this is not necessary for the
3218  * !shared pmd case because we can allocate the pmd later as well, it makes the
3219  * code much cleaner. pmd allocation is essential for the shared case because
3220  * pud has to be populated inside the same i_mmap_mutex section - otherwise
3221  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3222  * bad pmd for sharing.
3223  */
3224 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3225 {
3226         struct vm_area_struct *vma = find_vma(mm, addr);
3227         struct address_space *mapping = vma->vm_file->f_mapping;
3228         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3229                         vma->vm_pgoff;
3230         struct vm_area_struct *svma;
3231         unsigned long saddr;
3232         pte_t *spte = NULL;
3233         pte_t *pte;
3234
3235         if (!vma_shareable(vma, addr))
3236                 return (pte_t *)pmd_alloc(mm, pud, addr);
3237
3238         mutex_lock(&mapping->i_mmap_mutex);
3239         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3240                 if (svma == vma)
3241                         continue;
3242
3243                 saddr = page_table_shareable(svma, vma, addr, idx);
3244                 if (saddr) {
3245                         spte = huge_pte_offset(svma->vm_mm, saddr);
3246                         if (spte) {
3247                                 get_page(virt_to_page(spte));
3248                                 break;
3249                         }
3250                 }
3251         }
3252
3253         if (!spte)
3254                 goto out;
3255
3256         spin_lock(&mm->page_table_lock);
3257         if (pud_none(*pud))
3258                 pud_populate(mm, pud,
3259                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3260         else
3261                 put_page(virt_to_page(spte));
3262         spin_unlock(&mm->page_table_lock);
3263 out:
3264         pte = (pte_t *)pmd_alloc(mm, pud, addr);
3265         mutex_unlock(&mapping->i_mmap_mutex);
3266         return pte;
3267 }
3268
3269 /*
3270  * unmap huge page backed by shared pte.
3271  *
3272  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3273  * indicated by page_count > 1, unmap is achieved by clearing pud and
3274  * decrementing the ref count. If count == 1, the pte page is not shared.
3275  *
3276  * called with vma->vm_mm->page_table_lock held.
3277  *
3278  * returns: 1 successfully unmapped a shared pte page
3279  *          0 the underlying pte page is not shared, or it is the last user
3280  */
3281 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3282 {
3283         pgd_t *pgd = pgd_offset(mm, *addr);
3284         pud_t *pud = pud_offset(pgd, *addr);
3285
3286         BUG_ON(page_count(virt_to_page(ptep)) == 0);
3287         if (page_count(virt_to_page(ptep)) == 1)
3288                 return 0;
3289
3290         pud_clear(pud);
3291         put_page(virt_to_page(ptep));
3292         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3293         return 1;
3294 }
3295 #define want_pmd_share()        (1)
3296 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3297 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3298 {
3299         return NULL;
3300 }
3301 #define want_pmd_share()        (0)
3302 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3303
3304 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3305 pte_t *huge_pte_alloc(struct mm_struct *mm,
3306                         unsigned long addr, unsigned long sz)
3307 {
3308         pgd_t *pgd;
3309         pud_t *pud;
3310         pte_t *pte = NULL;
3311
3312         pgd = pgd_offset(mm, addr);
3313         pud = pud_alloc(mm, pgd, addr);
3314         if (pud) {
3315                 if (sz == PUD_SIZE) {
3316                         pte = (pte_t *)pud;
3317                 } else {
3318                         BUG_ON(sz != PMD_SIZE);
3319                         if (want_pmd_share() && pud_none(*pud))
3320                                 pte = huge_pmd_share(mm, addr, pud);
3321                         else
3322                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3323                 }
3324         }
3325         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3326
3327         return pte;
3328 }
3329
3330 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3331 {
3332         pgd_t *pgd;
3333         pud_t *pud;
3334         pmd_t *pmd = NULL;
3335
3336         pgd = pgd_offset(mm, addr);
3337         if (pgd_present(*pgd)) {
3338                 pud = pud_offset(pgd, addr);
3339                 if (pud_present(*pud)) {
3340                         if (pud_huge(*pud))
3341                                 return (pte_t *)pud;
3342                         pmd = pmd_offset(pud, addr);
3343                 }
3344         }
3345         return (pte_t *) pmd;
3346 }
3347
3348 struct page *
3349 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3350                 pmd_t *pmd, int write)
3351 {
3352         struct page *page;
3353
3354         page = pte_page(*(pte_t *)pmd);
3355         if (page)
3356                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3357         return page;
3358 }
3359
3360 struct page *
3361 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3362                 pud_t *pud, int write)
3363 {
3364         struct page *page;
3365
3366         page = pte_page(*(pte_t *)pud);
3367         if (page)
3368                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3369         return page;
3370 }
3371
3372 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3373
3374 /* Can be overriden by architectures */
3375 __attribute__((weak)) struct page *
3376 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3377                pud_t *pud, int write)
3378 {
3379         BUG();
3380         return NULL;
3381 }
3382
3383 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3384
3385 #ifdef CONFIG_MEMORY_FAILURE
3386
3387 /* Should be called in hugetlb_lock */
3388 static int is_hugepage_on_freelist(struct page *hpage)
3389 {
3390         struct page *page;
3391         struct page *tmp;
3392         struct hstate *h = page_hstate(hpage);
3393         int nid = page_to_nid(hpage);
3394
3395         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3396                 if (page == hpage)
3397                         return 1;
3398         return 0;
3399 }
3400
3401 /*
3402  * This function is called from memory failure code.
3403  * Assume the caller holds page lock of the head page.
3404  */
3405 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3406 {
3407         struct hstate *h = page_hstate(hpage);
3408         int nid = page_to_nid(hpage);
3409         int ret = -EBUSY;
3410
3411         spin_lock(&hugetlb_lock);
3412         if (is_hugepage_on_freelist(hpage)) {
3413                 /*
3414                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
3415                  * but dangling hpage->lru can trigger list-debug warnings
3416                  * (this happens when we call unpoison_memory() on it),
3417                  * so let it point to itself with list_del_init().
3418                  */
3419                 list_del_init(&hpage->lru);
3420                 set_page_refcounted(hpage);
3421                 h->free_huge_pages--;
3422                 h->free_huge_pages_node[nid]--;
3423                 ret = 0;
3424         }
3425         spin_unlock(&hugetlb_lock);
3426         return ret;
3427 }
3428 #endif