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 #include <linux/page-isolation.h>
  25
  26 #include <asm/page.h>
  27 #include <asm/pgtable.h>
  28 #include <asm/tlb.h>
  29
  30 #include <linux/io.h>
  31 #include <linux/hugetlb.h>
  32 #include <linux/hugetlb_cgroup.h>
  33 #include <linux/node.h>
  34 #include "internal.h"
  35
  36 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  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, hugepage_activelist, nr_huge_pages,
  52  * free_huge_pages, and surplus_huge_pages.
  53  */
  54 DEFINE_SPINLOCK(hugetlb_lock);
  55
  56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
  57 {
  58         bool free = (spool->count == 0) && (spool->used_hpages == 0);
  59
  60         spin_unlock(&spool->lock);
  61
  62         /* If no pages are used, and no other handles to the subpool
  63          * remain, free the subpool the subpool remain */
  64         if (free)
  65                 kfree(spool);
  66 }
  67
  68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
  69 {
  70         struct hugepage_subpool *spool;
  71
  72         spool = kmalloc(sizeof(*spool), GFP_KERNEL);
  73         if (!spool)
  74                 return NULL;
  75
  76         spin_lock_init(&spool->lock);
  77         spool->count = 1;
  78         spool->max_hpages = nr_blocks;
  79         spool->used_hpages = 0;
  80
  81         return spool;
  82 }
  83
  84 void hugepage_put_subpool(struct hugepage_subpool *spool)
  85 {
  86         spin_lock(&spool->lock);
  87         BUG_ON(!spool->count);
  88         spool->count--;
  89         unlock_or_release_subpool(spool);
  90 }
  91
  92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
  93                                       long delta)
  94 {
  95         int ret = 0;
  96
  97         if (!spool)
  98                 return 0;
  99
 100         spin_lock(&spool->lock);
 101         if ((spool->used_hpages + delta) <= spool->max_hpages) {
 102                 spool->used_hpages += delta;
 103         } else {
 104                 ret = -ENOMEM;
 105         }
 106         spin_unlock(&spool->lock);
 107
 108         return ret;
 109 }
 110
 111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
 112                                        long delta)
 113 {
 114         if (!spool)
 115                 return;
 116
 117         spin_lock(&spool->lock);
 118         spool->used_hpages -= delta;
 119         /* If hugetlbfs_put_super couldn't free spool due to
 120         * an outstanding quota reference, free it now. */
 121         unlock_or_release_subpool(spool);
 122 }
 123
 124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
 125 {
 126         return HUGETLBFS_SB(inode->i_sb)->spool;
 127 }
 128
 129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
 130 {
 131         return subpool_inode(file_inode(vma->vm_file));
 132 }
 133
 134 /*
 135  * Region tracking -- allows tracking of reservations and instantiated pages
 136  *                    across the pages in a mapping.
 137  *
 138  * The region data structures are protected by a combination of the mmap_sem
 139  * and the hugetlb_instantiation_mutex.  To access or modify a region the caller
 140  * must either hold the mmap_sem for write, or the mmap_sem for read and
 141  * the hugetlb_instantiation_mutex:
 142  *
 143  *      down_write(&mm->mmap_sem);
 144  * or
 145  *      down_read(&mm->mmap_sem);
 146  *      mutex_lock(&hugetlb_instantiation_mutex);
 147  */
 148 struct file_region {
 149         struct list_head link;
 150         long from;
 151         long to;
 152 };
 153
 154 static long region_add(struct list_head *head, long f, long t)
 155 {
 156         struct file_region *rg, *nrg, *trg;
 157
 158         /* Locate the region we are either in or before. */
 159         list_for_each_entry(rg, head, link)
 160                 if (f <= rg->to)
 161                         break;
 162
 163         /* Round our left edge to the current segment if it encloses us. */
 164         if (f > rg->from)
 165                 f = rg->from;
 166
 167         /* Check for and consume any regions we now overlap with. */
 168         nrg = rg;
 169         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 170                 if (&rg->link == head)
 171                         break;
 172                 if (rg->from > t)
 173                         break;
 174
 175                 /* If this area reaches higher then extend our area to
 176                  * include it completely.  If this is not the first area
 177                  * which we intend to reuse, free it. */
 178                 if (rg->to > t)
 179                         t = rg->to;
 180                 if (rg != nrg) {
 181                         list_del(&rg->link);
 182                         kfree(rg);
 183                 }
 184         }
 185         nrg->from = f;
 186         nrg->to = t;
 187         return 0;
 188 }
 189
 190 static long region_chg(struct list_head *head, long f, long t)
 191 {
 192         struct file_region *rg, *nrg;
 193         long chg = 0;
 194
 195         /* Locate the region we are before or in. */
 196         list_for_each_entry(rg, head, link)
 197                 if (f <= rg->to)
 198                         break;
 199
 200         /* If we are below the current region then a new region is required.
 201          * Subtle, allocate a new region at the position but make it zero
 202          * size such that we can guarantee to record the reservation. */
 203         if (&rg->link == head || t < rg->from) {
 204                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 205                 if (!nrg)
 206                         return -ENOMEM;
 207                 nrg->from = f;
 208                 nrg->to   = f;
 209                 INIT_LIST_HEAD(&nrg->link);
 210                 list_add(&nrg->link, rg->link.prev);
 211
 212                 return t - f;
 213         }
 214
 215         /* Round our left edge to the current segment if it encloses us. */
 216         if (f > rg->from)
 217                 f = rg->from;
 218         chg = t - f;
 219
 220         /* Check for and consume any regions we now overlap with. */
 221         list_for_each_entry(rg, rg->link.prev, link) {
 222                 if (&rg->link == head)
 223                         break;
 224                 if (rg->from > t)
 225                         return chg;
 226
 227                 /* We overlap with this area, if it extends further than
 228                  * us then we must extend ourselves.  Account for its
 229                  * existing reservation. */
 230                 if (rg->to > t) {
 231                         chg += rg->to - t;
 232                         t = rg->to;
 233                 }
 234                 chg -= rg->to - rg->from;
 235         }
 236         return chg;
 237 }
 238
 239 static long region_truncate(struct list_head *head, long end)
 240 {
 241         struct file_region *rg, *trg;
 242         long chg = 0;
 243
 244         /* Locate the region we are either in or before. */
 245         list_for_each_entry(rg, head, link)
 246                 if (end <= rg->to)
 247                         break;
 248         if (&rg->link == head)
 249                 return 0;
 250
 251         /* If we are in the middle of a region then adjust it. */
 252         if (end > rg->from) {
 253                 chg = rg->to - end;
 254                 rg->to = end;
 255                 rg = list_entry(rg->link.next, typeof(*rg), link);
 256         }
 257
 258         /* Drop any remaining regions. */
 259         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 260                 if (&rg->link == head)
 261                         break;
 262                 chg += rg->to - rg->from;
 263                 list_del(&rg->link);
 264                 kfree(rg);
 265         }
 266         return chg;
 267 }
 268
 269 static long region_count(struct list_head *head, long f, long t)
 270 {
 271         struct file_region *rg;
 272         long chg = 0;
 273
 274         /* Locate each segment we overlap with, and count that overlap. */
 275         list_for_each_entry(rg, head, link) {
 276                 long seg_from;
 277                 long seg_to;
 278
 279                 if (rg->to <= f)
 280                         continue;
 281                 if (rg->from >= t)
 282                         break;
 283
 284                 seg_from = max(rg->from, f);
 285                 seg_to = min(rg->to, t);
 286
 287                 chg += seg_to - seg_from;
 288         }
 289
 290         return chg;
 291 }
 292
 293 /*
 294  * Convert the address within this vma to the page offset within
 295  * the mapping, in pagecache page units; huge pages here.
 296  */
 297 static pgoff_t vma_hugecache_offset(struct hstate *h,
 298                         struct vm_area_struct *vma, unsigned long address)
 299 {
 300         return ((address - vma->vm_start) >> huge_page_shift(h)) +
 301                         (vma->vm_pgoff >> huge_page_order(h));
 302 }
 303
 304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
 305                                      unsigned long address)
 306 {
 307         return vma_hugecache_offset(hstate_vma(vma), vma, address);
 308 }
 309
 310 /*
 311  * Return the size of the pages allocated when backing a VMA. In the majority
 312  * cases this will be same size as used by the page table entries.
 313  */
 314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 315 {
 316         struct hstate *hstate;
 317
 318         if (!is_vm_hugetlb_page(vma))
 319                 return PAGE_SIZE;
 320
 321         hstate = hstate_vma(vma);
 322
 323         return 1UL << huge_page_shift(hstate);
 324 }
 325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
 326
 327 /*
 328  * Return the page size being used by the MMU to back a VMA. In the majority
 329  * of cases, the page size used by the kernel matches the MMU size. On
 330  * architectures where it differs, an architecture-specific version of this
 331  * function is required.
 332  */
 333 #ifndef vma_mmu_pagesize
 334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 335 {
 336         return vma_kernel_pagesize(vma);
 337 }
 338 #endif
 339
 340 /*
 341  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 342  * bits of the reservation map pointer, which are always clear due to
 343  * alignment.
 344  */
 345 #define HPAGE_RESV_OWNER    (1UL << 0)
 346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
 347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 348
 349 /*
 350  * These helpers are used to track how many pages are reserved for
 351  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 352  * is guaranteed to have their future faults succeed.
 353  *
 354  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 355  * the reserve counters are updated with the hugetlb_lock held. It is safe
 356  * to reset the VMA at fork() time as it is not in use yet and there is no
 357  * chance of the global counters getting corrupted as a result of the values.
 358  *
 359  * The private mapping reservation is represented in a subtly different
 360  * manner to a shared mapping.  A shared mapping has a region map associated
 361  * with the underlying file, this region map represents the backing file
 362  * pages which have ever had a reservation assigned which this persists even
 363  * after the page is instantiated.  A private mapping has a region map
 364  * associated with the original mmap which is attached to all VMAs which
 365  * reference it, this region map represents those offsets which have consumed
 366  * reservation ie. where pages have been instantiated.
 367  */
 368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 369 {
 370         return (unsigned long)vma->vm_private_data;
 371 }
 372
 373 static void set_vma_private_data(struct vm_area_struct *vma,
 374                                                         unsigned long value)
 375 {
 376         vma->vm_private_data = (void *)value;
 377 }
 378
 379 struct resv_map {
 380         struct kref refs;
 381         struct list_head regions;
 382 };
 383
 384 static struct resv_map *resv_map_alloc(void)
 385 {
 386         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 387         if (!resv_map)
 388                 return NULL;
 389
 390         kref_init(&resv_map->refs);
 391         INIT_LIST_HEAD(&resv_map->regions);
 392
 393         return resv_map;
 394 }
 395
 396 static void resv_map_release(struct kref *ref)
 397 {
 398         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 399
 400         /* Clear out any active regions before we release the map. */
 401         region_truncate(&resv_map->regions, 0);
 402         kfree(resv_map);
 403 }
 404
 405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 406 {
 407         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 408         if (!(vma->vm_flags & VM_MAYSHARE))
 409                 return (struct resv_map *)(get_vma_private_data(vma) &
 410                                                         ~HPAGE_RESV_MASK);
 411         return NULL;
 412 }
 413
 414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 415 {
 416         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 417         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 418
 419         set_vma_private_data(vma, (get_vma_private_data(vma) &
 420                                 HPAGE_RESV_MASK) | (unsigned long)map);
 421 }
 422
 423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 424 {
 425         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 426         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 427
 428         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 429 }
 430
 431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 432 {
 433         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 434
 435         return (get_vma_private_data(vma) & flag) != 0;
 436 }
 437
 438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 439 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 440 {
 441         VM_BUG_ON(!is_vm_hugetlb_page(vma));
 442         if (!(vma->vm_flags & VM_MAYSHARE))
 443                 vma->vm_private_data = (void *)0;
 444 }
 445
 446 /* Returns true if the VMA has associated reserve pages */
 447 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
 448 {
 449         if (vma->vm_flags & VM_NORESERVE) {
 450                 /*
 451                  * This address is already reserved by other process(chg == 0),
 452                  * so, we should decrement reserved count. Without decrementing,
 453                  * reserve count remains after releasing inode, because this
 454                  * allocated page will go into page cache and is regarded as
 455                  * coming from reserved pool in releasing step.  Currently, we
 456                  * don't have any other solution to deal with this situation
 457                  * properly, so add work-around here.
 458                  */
 459                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
 460                         return 1;
 461                 else
 462                         return 0;
 463         }
 464
 465         /* Shared mappings always use reserves */
 466         if (vma->vm_flags & VM_MAYSHARE)
 467                 return 1;
 468
 469         /*
 470          * Only the process that called mmap() has reserves for
 471          * private mappings.
 472          */
 473         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 474                 return 1;
 475
 476         return 0;
 477 }
 478
 479 static void enqueue_huge_page(struct hstate *h, struct page *page)
 480 {
 481         int nid = page_to_nid(page);
 482         list_move(&page->lru, &h->hugepage_freelists[nid]);
 483         h->free_huge_pages++;
 484         h->free_huge_pages_node[nid]++;
 485 }
 486
 487 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 488 {
 489         struct page *page;
 490
 491         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
 492                 if (!is_migrate_isolate_page(page))
 493                         break;
 494         /*
 495          * if 'non-isolated free hugepage' not found on the list,
 496          * the allocation fails.
 497          */
 498         if (&h->hugepage_freelists[nid] == &page->lru)
 499                 return NULL;
 500         list_move(&page->lru, &h->hugepage_activelist);
 501         set_page_refcounted(page);
 502         h->free_huge_pages--;
 503         h->free_huge_pages_node[nid]--;
 504         return page;
 505 }
 506
 507 /* Movability of hugepages depends on migration support. */
 508 static inline gfp_t htlb_alloc_mask(struct hstate *h)
 509 {
 510         if (hugepages_treat_as_movable || hugepage_migration_support(h))
 511                 return GFP_HIGHUSER_MOVABLE;
 512         else
 513                 return GFP_HIGHUSER;
 514 }
 515
 516 static struct page *dequeue_huge_page_vma(struct hstate *h,
 517                                 struct vm_area_struct *vma,
 518                                 unsigned long address, int avoid_reserve,
 519                                 long chg)
 520 {
 521         struct page *page = NULL;
 522         struct mempolicy *mpol;
 523         nodemask_t *nodemask;
 524         struct zonelist *zonelist;
 525         struct zone *zone;
 526         struct zoneref *z;
 527         unsigned int cpuset_mems_cookie;
 528
 529         /*
 530          * A child process with MAP_PRIVATE mappings created by their parent
 531          * have no page reserves. This check ensures that reservations are
 532          * not "stolen". The child may still get SIGKILLed
 533          */
 534         if (!vma_has_reserves(vma, chg) &&
 535                         h->free_huge_pages - h->resv_huge_pages == 0)
 536                 goto err;
 537
 538         /* If reserves cannot be used, ensure enough pages are in the pool */
 539         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 540                 goto err;
 541
 542 retry_cpuset:
 543         cpuset_mems_cookie = get_mems_allowed();
 544         zonelist = huge_zonelist(vma, address,
 545                                         htlb_alloc_mask(h), &mpol, &nodemask);
 546
 547         for_each_zone_zonelist_nodemask(zone, z, zonelist,
 548                                                 MAX_NR_ZONES - 1, nodemask) {
 549                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
 550                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
 551                         if (page) {
 552                                 if (avoid_reserve)
 553                                         break;
 554                                 if (!vma_has_reserves(vma, chg))
 555                                         break;
 556
 557                                 SetPagePrivate(page);
 558                                 h->resv_huge_pages--;
 559                                 break;
 560                         }
 561                 }
 562         }
 563
 564         mpol_cond_put(mpol);
 565         if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
 566                 goto retry_cpuset;
 567         return page;
 568
 569 err:
 570         return NULL;
 571 }
 572
 573 static void update_and_free_page(struct hstate *h, struct page *page)
 574 {
 575         int i;
 576
 577         VM_BUG_ON(h->order >= MAX_ORDER);
 578
 579         h->nr_huge_pages--;
 580         h->nr_huge_pages_node[page_to_nid(page)]--;
 581         for (i = 0; i < pages_per_huge_page(h); i++) {
 582                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
 583                                 1 << PG_referenced | 1 << PG_dirty |
 584                                 1 << PG_active | 1 << PG_reserved |
 585                                 1 << PG_private | 1 << PG_writeback);
 586         }
 587         VM_BUG_ON(hugetlb_cgroup_from_page(page));
 588         set_compound_page_dtor(page, NULL);
 589         set_page_refcounted(page);
 590         arch_release_hugepage(page);
 591         __free_pages(page, huge_page_order(h));
 592 }
 593
 594 struct hstate *size_to_hstate(unsigned long size)
 595 {
 596         struct hstate *h;
 597
 598         for_each_hstate(h) {
 599                 if (huge_page_size(h) == size)
 600                         return h;
 601         }
 602         return NULL;
 603 }
 604
 605 static void free_huge_page(struct page *page)
 606 {
 607         /*
 608          * Can't pass hstate in here because it is called from the
 609          * compound page destructor.
 610          */
 611         struct hstate *h = page_hstate(page);
 612         int nid = page_to_nid(page);
 613         struct hugepage_subpool *spool =
 614                 (struct hugepage_subpool *)page_private(page);
 615         bool restore_reserve;
 616
 617         set_page_private(page, 0);
 618         page->mapping = NULL;
 619         BUG_ON(page_count(page));
 620         BUG_ON(page_mapcount(page));
 621         restore_reserve = PagePrivate(page);
 622         ClearPagePrivate(page);
 623
 624         spin_lock(&hugetlb_lock);
 625         hugetlb_cgroup_uncharge_page(hstate_index(h),
 626                                      pages_per_huge_page(h), page);
 627         if (restore_reserve)
 628                 h->resv_huge_pages++;
 629
 630         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 631                 /* remove the page from active list */
 632                 list_del(&page->lru);
 633                 update_and_free_page(h, page);
 634                 h->surplus_huge_pages--;
 635                 h->surplus_huge_pages_node[nid]--;
 636         } else {
 637                 arch_clear_hugepage_flags(page);
 638                 enqueue_huge_page(h, page);
 639         }
 640         spin_unlock(&hugetlb_lock);
 641         hugepage_subpool_put_pages(spool, 1);
 642 }
 643
 644 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 645 {
 646         INIT_LIST_HEAD(&page->lru);
 647         set_compound_page_dtor(page, free_huge_page);
 648         spin_lock(&hugetlb_lock);
 649         set_hugetlb_cgroup(page, NULL);
 650         h->nr_huge_pages++;
 651         h->nr_huge_pages_node[nid]++;
 652         spin_unlock(&hugetlb_lock);
 653         put_page(page); /* free it into the hugepage allocator */
 654 }
 655
 656 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
 657 {
 658         int i;
 659         int nr_pages = 1 << order;
 660         struct page *p = page + 1;
 661
 662         /* we rely on prep_new_huge_page to set the destructor */
 663         set_compound_order(page, order);
 664         __SetPageHead(page);
 665         __ClearPageReserved(page);
 666         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
 667                 __SetPageTail(p);
 668                 /*
 669                  * For gigantic hugepages allocated through bootmem at
 670                  * boot, it's safer to be consistent with the not-gigantic
 671                  * hugepages and clear the PG_reserved bit from all tail pages
 672                  * too.  Otherwse drivers using get_user_pages() to access tail
 673                  * pages may get the reference counting wrong if they see
 674                  * PG_reserved set on a tail page (despite the head page not
 675                  * having PG_reserved set).  Enforcing this consistency between
 676                  * head and tail pages allows drivers to optimize away a check
 677                  * on the head page when they need know if put_page() is needed
 678                  * after get_user_pages().
 679                  */
 680                 __ClearPageReserved(p);
 681                 set_page_count(p, 0);
 682                 p->first_page = page;
 683         }
 684 }
 685
 686 /*
 687  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
 688  * transparent huge pages.  See the PageTransHuge() documentation for more
 689  * details.
 690  */
 691 int PageHuge(struct page *page)
 692 {
 693         compound_page_dtor *dtor;
 694
 695         if (!PageCompound(page))
 696                 return 0;
 697
 698         page = compound_head(page);
 699         dtor = get_compound_page_dtor(page);
 700
 701         return dtor == free_huge_page;
 702 }
 703 EXPORT_SYMBOL_GPL(PageHuge);
 704
 705 /*
 706  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
 707  * normal or transparent huge pages.
 708  */
 709 int PageHeadHuge(struct page *page_head)
 710 {
 711         compound_page_dtor *dtor;
 712
 713         if (!PageHead(page_head))
 714                 return 0;
 715
 716         dtor = get_compound_page_dtor(page_head);
 717
 718         return dtor == free_huge_page;
 719 }
 720 EXPORT_SYMBOL_GPL(PageHeadHuge);
 721
 722 pgoff_t __basepage_index(struct page *page)
 723 {
 724         struct page *page_head = compound_head(page);
 725         pgoff_t index = page_index(page_head);
 726         unsigned long compound_idx;
 727
 728         if (!PageHuge(page_head))
 729                 return page_index(page);
 730
 731         if (compound_order(page_head) >= MAX_ORDER)
 732                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 733         else
 734                 compound_idx = page - page_head;
 735
 736         return (index << compound_order(page_head)) + compound_idx;
 737 }
 738
 739 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 740 {
 741         struct page *page;
 742
 743         if (h->order >= MAX_ORDER)
 744                 return NULL;
 745
 746         page = alloc_pages_exact_node(nid,
 747                 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
 748                                                 __GFP_REPEAT|__GFP_NOWARN,
 749                 huge_page_order(h));
 750         if (page) {
 751                 if (arch_prepare_hugepage(page)) {
 752                         __free_pages(page, huge_page_order(h));
 753                         return NULL;
 754                 }
 755                 prep_new_huge_page(h, page, nid);
 756         }
 757
 758         return page;
 759 }
 760
 761 /*
 762  * common helper functions for hstate_next_node_to_{alloc|free}.
 763  * We may have allocated or freed a huge page based on a different
 764  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
 765  * be outside of *nodes_allowed.  Ensure that we use an allowed
 766  * node for alloc or free.
 767  */
 768 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 769 {
 770         nid = next_node(nid, *nodes_allowed);
 771         if (nid == MAX_NUMNODES)
 772                 nid = first_node(*nodes_allowed);
 773         VM_BUG_ON(nid >= MAX_NUMNODES);
 774
 775         return nid;
 776 }
 777
 778 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
 779 {
 780         if (!node_isset(nid, *nodes_allowed))
 781                 nid = next_node_allowed(nid, nodes_allowed);
 782         return nid;
 783 }
 784
 785 /*
 786  * returns the previously saved node ["this node"] from which to
 787  * allocate a persistent huge page for the pool and advance the
 788  * next node from which to allocate, handling wrap at end of node
 789  * mask.
 790  */
 791 static int hstate_next_node_to_alloc(struct hstate *h,
 792                                         nodemask_t *nodes_allowed)
 793 {
 794         int nid;
 795
 796         VM_BUG_ON(!nodes_allowed);
 797
 798         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
 799         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
 800
 801         return nid;
 802 }
 803
 804 /*
 805  * helper for free_pool_huge_page() - return the previously saved
 806  * node ["this node"] from which to free a huge page.  Advance the
 807  * next node id whether or not we find a free huge page to free so
 808  * that the next attempt to free addresses the next node.
 809  */
 810 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
 811 {
 812         int nid;
 813
 814         VM_BUG_ON(!nodes_allowed);
 815
 816         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
 817         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
 818
 819         return nid;
 820 }
 821
 822 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
 823         for (nr_nodes = nodes_weight(*mask);                            \
 824                 nr_nodes > 0 &&                                         \
 825                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
 826                 nr_nodes--)
 827
 828 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
 829         for (nr_nodes = nodes_weight(*mask);                            \
 830                 nr_nodes > 0 &&                                         \
 831                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
 832                 nr_nodes--)
 833
 834 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
 835 {
 836         struct page *page;
 837         int nr_nodes, node;
 838         int ret = 0;
 839
 840         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
 841                 page = alloc_fresh_huge_page_node(h, node);
 842                 if (page) {
 843                         ret = 1;
 844                         break;
 845                 }
 846         }
 847
 848         if (ret)
 849                 count_vm_event(HTLB_BUDDY_PGALLOC);
 850         else
 851                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 852
 853         return ret;
 854 }
 855
 856 /*
 857  * Free huge page from pool from next node to free.
 858  * Attempt to keep persistent huge pages more or less
 859  * balanced over allowed nodes.
 860  * Called with hugetlb_lock locked.
 861  */
 862 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
 863                                                          bool acct_surplus)
 864 {
 865         int nr_nodes, node;
 866         int ret = 0;
 867
 868         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
 869                 /*
 870                  * If we're returning unused surplus pages, only examine
 871                  * nodes with surplus pages.
 872                  */
 873                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
 874                     !list_empty(&h->hugepage_freelists[node])) {
 875                         struct page *page =
 876                                 list_entry(h->hugepage_freelists[node].next,
 877                                           struct page, lru);
 878                         list_del(&page->lru);
 879                         h->free_huge_pages--;
 880                         h->free_huge_pages_node[node]--;
 881                         if (acct_surplus) {
 882                                 h->surplus_huge_pages--;
 883                                 h->surplus_huge_pages_node[node]--;
 884                         }
 885                         update_and_free_page(h, page);
 886                         ret = 1;
 887                         break;
 888                 }
 889         }
 890
 891         return ret;
 892 }
 893
 894 /*
 895  * Dissolve a given free hugepage into free buddy pages. This function does
 896  * nothing for in-use (including surplus) hugepages.
 897  */
 898 static void dissolve_free_huge_page(struct page *page)
 899 {
 900         spin_lock(&hugetlb_lock);
 901         if (PageHuge(page) && !page_count(page)) {
 902                 struct hstate *h = page_hstate(page);
 903                 int nid = page_to_nid(page);
 904                 list_del(&page->lru);
 905                 h->free_huge_pages--;
 906                 h->free_huge_pages_node[nid]--;
 907                 update_and_free_page(h, page);
 908         }
 909         spin_unlock(&hugetlb_lock);
 910 }
 911
 912 /*
 913  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
 914  * make specified memory blocks removable from the system.
 915  * Note that start_pfn should aligned with (minimum) hugepage size.
 916  */
 917 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
 918 {
 919         unsigned int order = 8 * sizeof(void *);
 920         unsigned long pfn;
 921         struct hstate *h;
 922
 923         /* Set scan step to minimum hugepage size */
 924         for_each_hstate(h)
 925                 if (order > huge_page_order(h))
 926                         order = huge_page_order(h);
 927         VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
 928         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
 929                 dissolve_free_huge_page(pfn_to_page(pfn));
 930 }
 931
 932 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 933 {
 934         struct page *page;
 935         unsigned int r_nid;
 936
 937         if (h->order >= MAX_ORDER)
 938                 return NULL;
 939
 940         /*
 941          * Assume we will successfully allocate the surplus page to
 942          * prevent racing processes from causing the surplus to exceed
 943          * overcommit
 944          *
 945          * This however introduces a different race, where a process B
 946          * tries to grow the static hugepage pool while alloc_pages() is
 947          * called by process A. B will only examine the per-node
 948          * counters in determining if surplus huge pages can be
 949          * converted to normal huge pages in adjust_pool_surplus(). A
 950          * won't be able to increment the per-node counter, until the
 951          * lock is dropped by B, but B doesn't drop hugetlb_lock until
 952          * no more huge pages can be converted from surplus to normal
 953          * state (and doesn't try to convert again). Thus, we have a
 954          * case where a surplus huge page exists, the pool is grown, and
 955          * the surplus huge page still exists after, even though it
 956          * should just have been converted to a normal huge page. This
 957          * does not leak memory, though, as the hugepage will be freed
 958          * once it is out of use. It also does not allow the counters to
 959          * go out of whack in adjust_pool_surplus() as we don't modify
 960          * the node values until we've gotten the hugepage and only the
 961          * per-node value is checked there.
 962          */
 963         spin_lock(&hugetlb_lock);
 964         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 965                 spin_unlock(&hugetlb_lock);
 966                 return NULL;
 967         } else {
 968                 h->nr_huge_pages++;
 969                 h->surplus_huge_pages++;
 970         }
 971         spin_unlock(&hugetlb_lock);
 972
 973         if (nid == NUMA_NO_NODE)
 974                 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
 975                                    __GFP_REPEAT|__GFP_NOWARN,
 976                                    huge_page_order(h));
 977         else
 978                 page = alloc_pages_exact_node(nid,
 979                         htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
 980                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
 981
 982         if (page && arch_prepare_hugepage(page)) {
 983                 __free_pages(page, huge_page_order(h));
 984                 page = NULL;
 985         }
 986
 987         spin_lock(&hugetlb_lock);
 988         if (page) {
 989                 INIT_LIST_HEAD(&page->lru);
 990                 r_nid = page_to_nid(page);
 991                 set_compound_page_dtor(page, free_huge_page);
 992                 set_hugetlb_cgroup(page, NULL);
 993                 /*
 994                  * We incremented the global counters already
 995                  */
 996                 h->nr_huge_pages_node[r_nid]++;
 997                 h->surplus_huge_pages_node[r_nid]++;
 998                 __count_vm_event(HTLB_BUDDY_PGALLOC);
 999         } else {
1000                 h->nr_huge_pages--;
1001                 h->surplus_huge_pages--;
1002                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1003         }
1004         spin_unlock(&hugetlb_lock);
1005
1006         return page;
1007 }
1008
1009 /*
1010  * This allocation function is useful in the context where vma is irrelevant.
1011  * E.g. soft-offlining uses this function because it only cares physical
1012  * address of error page.
1013  */
1014 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1015 {
1016         struct page *page = NULL;
1017
1018         spin_lock(&hugetlb_lock);
1019         if (h->free_huge_pages - h->resv_huge_pages > 0)
1020                 page = dequeue_huge_page_node(h, nid);
1021         spin_unlock(&hugetlb_lock);
1022
1023         if (!page)
1024                 page = alloc_buddy_huge_page(h, nid);
1025
1026         return page;
1027 }
1028
1029 /*
1030  * Increase the hugetlb pool such that it can accommodate a reservation
1031  * of size 'delta'.
1032  */
1033 static int gather_surplus_pages(struct hstate *h, int delta)
1034 {
1035         struct list_head surplus_list;
1036         struct page *page, *tmp;
1037         int ret, i;
1038         int needed, allocated;
1039         bool alloc_ok = true;
1040
1041         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1042         if (needed <= 0) {
1043                 h->resv_huge_pages += delta;
1044                 return 0;
1045         }
1046
1047         allocated = 0;
1048         INIT_LIST_HEAD(&surplus_list);
1049
1050         ret = -ENOMEM;
1051 retry:
1052         spin_unlock(&hugetlb_lock);
1053         for (i = 0; i < needed; i++) {
1054                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1055                 if (!page) {
1056                         alloc_ok = false;
1057                         break;
1058                 }
1059                 list_add(&page->lru, &surplus_list);
1060         }
1061         allocated += i;
1062
1063         /*
1064          * After retaking hugetlb_lock, we need to recalculate 'needed'
1065          * because either resv_huge_pages or free_huge_pages may have changed.
1066          */
1067         spin_lock(&hugetlb_lock);
1068         needed = (h->resv_huge_pages + delta) -
1069                         (h->free_huge_pages + allocated);
1070         if (needed > 0) {
1071                 if (alloc_ok)
1072                         goto retry;
1073                 /*
1074                  * We were not able to allocate enough pages to
1075                  * satisfy the entire reservation so we free what
1076                  * we've allocated so far.
1077                  */
1078                 goto free;
1079         }
1080         /*
1081          * The surplus_list now contains _at_least_ the number of extra pages
1082          * needed to accommodate the reservation.  Add the appropriate number
1083          * of pages to the hugetlb pool and free the extras back to the buddy
1084          * allocator.  Commit the entire reservation here to prevent another
1085          * process from stealing the pages as they are added to the pool but
1086          * before they are reserved.
1087          */
1088         needed += allocated;
1089         h->resv_huge_pages += delta;
1090         ret = 0;
1091
1092         /* Free the needed pages to the hugetlb pool */
1093         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1094                 if ((--needed) < 0)
1095                         break;
1096                 /*
1097                  * This page is now managed by the hugetlb allocator and has
1098                  * no users -- drop the buddy allocator's reference.
1099                  */
1100                 put_page_testzero(page);
1101                 VM_BUG_ON(page_count(page));
1102                 enqueue_huge_page(h, page);
1103         }
1104 free:
1105         spin_unlock(&hugetlb_lock);
1106
1107         /* Free unnecessary surplus pages to the buddy allocator */
1108         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1109                 put_page(page);
1110         spin_lock(&hugetlb_lock);
1111
1112         return ret;
1113 }
1114
1115 /*
1116  * When releasing a hugetlb pool reservation, any surplus pages that were
1117  * allocated to satisfy the reservation must be explicitly freed if they were
1118  * never used.
1119  * Called with hugetlb_lock held.
1120  */
1121 static void return_unused_surplus_pages(struct hstate *h,
1122                                         unsigned long unused_resv_pages)
1123 {
1124         unsigned long nr_pages;
1125
1126         /* Uncommit the reservation */
1127         h->resv_huge_pages -= unused_resv_pages;
1128
1129         /* Cannot return gigantic pages currently */
1130         if (h->order >= MAX_ORDER)
1131                 return;
1132
1133         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1134
1135         /*
1136          * We want to release as many surplus pages as possible, spread
1137          * evenly across all nodes with memory. Iterate across these nodes
1138          * until we can no longer free unreserved surplus pages. This occurs
1139          * when the nodes with surplus pages have no free pages.
1140          * free_pool_huge_page() will balance the the freed pages across the
1141          * on-line nodes with memory and will handle the hstate accounting.
1142          */
1143         while (nr_pages--) {
1144                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1145                         break;
1146         }
1147 }
1148
1149 /*
1150  * Determine if the huge page at addr within the vma has an associated
1151  * reservation.  Where it does not we will need to logically increase
1152  * reservation and actually increase subpool usage before an allocation
1153  * can occur.  Where any new reservation would be required the
1154  * reservation change is prepared, but not committed.  Once the page
1155  * has been allocated from the subpool and instantiated the change should
1156  * be committed via vma_commit_reservation.  No action is required on
1157  * failure.
1158  */
1159 static long vma_needs_reservation(struct hstate *h,
1160                         struct vm_area_struct *vma, unsigned long addr)
1161 {
1162         struct address_space *mapping = vma->vm_file->f_mapping;
1163         struct inode *inode = mapping->host;
1164
1165         if (vma->vm_flags & VM_MAYSHARE) {
1166                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1167                 return region_chg(&inode->i_mapping->private_list,
1168                                                         idx, idx + 1);
1169
1170         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1171                 return 1;
1172
1173         } else  {
1174                 long err;
1175                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1176                 struct resv_map *resv = vma_resv_map(vma);
1177
1178                 err = region_chg(&resv->regions, idx, idx + 1);
1179                 if (err < 0)
1180                         return err;
1181                 return 0;
1182         }
1183 }
1184 static void vma_commit_reservation(struct hstate *h,
1185                         struct vm_area_struct *vma, unsigned long addr)
1186 {
1187         struct address_space *mapping = vma->vm_file->f_mapping;
1188         struct inode *inode = mapping->host;
1189
1190         if (vma->vm_flags & VM_MAYSHARE) {
1191                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1192                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1193
1194         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1195                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1196                 struct resv_map *resv = vma_resv_map(vma);
1197
1198                 /* Mark this page used in the map. */
1199                 region_add(&resv->regions, idx, idx + 1);
1200         }
1201 }
1202
1203 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1204                                     unsigned long addr, int avoid_reserve)
1205 {
1206         struct hugepage_subpool *spool = subpool_vma(vma);
1207         struct hstate *h = hstate_vma(vma);
1208         struct page *page;
1209         long chg;
1210         int ret, idx;
1211         struct hugetlb_cgroup *h_cg;
1212
1213         idx = hstate_index(h);
1214         /*
1215          * Processes that did not create the mapping will have no
1216          * reserves and will not have accounted against subpool
1217          * limit. Check that the subpool limit can be made before
1218          * satisfying the allocation MAP_NORESERVE mappings may also
1219          * need pages and subpool limit allocated allocated if no reserve
1220          * mapping overlaps.
1221          */
1222         chg = vma_needs_reservation(h, vma, addr);
1223         if (chg < 0)
1224                 return ERR_PTR(-ENOMEM);
1225         if (chg || avoid_reserve)
1226                 if (hugepage_subpool_get_pages(spool, 1))
1227                         return ERR_PTR(-ENOSPC);
1228
1229         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1230         if (ret) {
1231                 if (chg || avoid_reserve)
1232                         hugepage_subpool_put_pages(spool, 1);
1233                 return ERR_PTR(-ENOSPC);
1234         }
1235         spin_lock(&hugetlb_lock);
1236         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1237         if (!page) {
1238                 spin_unlock(&hugetlb_lock);
1239                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1240                 if (!page) {
1241                         hugetlb_cgroup_uncharge_cgroup(idx,
1242                                                        pages_per_huge_page(h),
1243                                                        h_cg);
1244                         if (chg || avoid_reserve)
1245                                 hugepage_subpool_put_pages(spool, 1);
1246                         return ERR_PTR(-ENOSPC);
1247                 }
1248                 spin_lock(&hugetlb_lock);
1249                 list_move(&page->lru, &h->hugepage_activelist);
1250                 /* Fall through */
1251         }
1252         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1253         spin_unlock(&hugetlb_lock);
1254
1255         set_page_private(page, (unsigned long)spool);
1256
1257         vma_commit_reservation(h, vma, addr);
1258         return page;
1259 }
1260
1261 /*
1262  * alloc_huge_page()'s wrapper which simply returns the page if allocation
1263  * succeeds, otherwise NULL. This function is called from new_vma_page(),
1264  * where no ERR_VALUE is expected to be returned.
1265  */
1266 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1267                                 unsigned long addr, int avoid_reserve)
1268 {
1269         struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1270         if (IS_ERR(page))
1271                 page = NULL;
1272         return page;
1273 }
1274
1275 int __weak alloc_bootmem_huge_page(struct hstate *h)
1276 {
1277         struct huge_bootmem_page *m;
1278         int nr_nodes, node;
1279
1280         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1281                 void *addr;
1282
1283                 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1284                                 huge_page_size(h), huge_page_size(h), 0);
1285
1286                 if (addr) {
1287                         /*
1288                          * Use the beginning of the huge page to store the
1289                          * huge_bootmem_page struct (until gather_bootmem
1290                          * puts them into the mem_map).
1291                          */
1292                         m = addr;
1293                         goto found;
1294                 }
1295         }
1296         return 0;
1297
1298 found:
1299         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1300         /* Put them into a private list first because mem_map is not up yet */
1301         list_add(&m->list, &huge_boot_pages);
1302         m->hstate = h;
1303         return 1;
1304 }
1305
1306 static void prep_compound_huge_page(struct page *page, int order)
1307 {
1308         if (unlikely(order > (MAX_ORDER - 1)))
1309                 prep_compound_gigantic_page(page, order);
1310         else
1311                 prep_compound_page(page, order);
1312 }
1313
1314 /* Put bootmem huge pages into the standard lists after mem_map is up */
1315 static void __init gather_bootmem_prealloc(void)
1316 {
1317         struct huge_bootmem_page *m;
1318
1319         list_for_each_entry(m, &huge_boot_pages, list) {
1320                 struct hstate *h = m->hstate;
1321                 struct page *page;
1322
1323 #ifdef CONFIG_HIGHMEM
1324                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1325                 free_bootmem_late((unsigned long)m,
1326                                   sizeof(struct huge_bootmem_page));
1327 #else
1328                 page = virt_to_page(m);
1329 #endif
1330                 WARN_ON(page_count(page) != 1);
1331                 prep_compound_huge_page(page, h->order);
1332                 WARN_ON(PageReserved(page));
1333                 prep_new_huge_page(h, page, page_to_nid(page));
1334                 /*
1335                  * If we had gigantic hugepages allocated at boot time, we need
1336                  * to restore the 'stolen' pages to totalram_pages in order to
1337                  * fix confusing memory reports from free(1) and another
1338                  * side-effects, like CommitLimit going negative.
1339                  */
1340                 if (h->order > (MAX_ORDER - 1))
1341                         adjust_managed_page_count(page, 1 << h->order);
1342         }
1343 }
1344
1345 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1346 {
1347         unsigned long i;
1348
1349         for (i = 0; i < h->max_huge_pages; ++i) {
1350                 if (h->order >= MAX_ORDER) {
1351                         if (!alloc_bootmem_huge_page(h))
1352                                 break;
1353                 } else if (!alloc_fresh_huge_page(h,
1354                                          &node_states[N_MEMORY]))
1355                         break;
1356         }
1357         h->max_huge_pages = i;
1358 }
1359
1360 static void __init hugetlb_init_hstates(void)
1361 {
1362         struct hstate *h;
1363
1364         for_each_hstate(h) {
1365                 /* oversize hugepages were init'ed in early boot */
1366                 if (h->order < MAX_ORDER)
1367                         hugetlb_hstate_alloc_pages(h);
1368         }
1369 }
1370
1371 static char * __init memfmt(char *buf, unsigned long n)
1372 {
1373         if (n >= (1UL << 30))
1374                 sprintf(buf, "%lu GB", n >> 30);
1375         else if (n >= (1UL << 20))
1376                 sprintf(buf, "%lu MB", n >> 20);
1377         else
1378                 sprintf(buf, "%lu KB", n >> 10);
1379         return buf;
1380 }
1381
1382 static void __init report_hugepages(void)
1383 {
1384         struct hstate *h;
1385
1386         for_each_hstate(h) {
1387                 char buf[32];
1388                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1389                         memfmt(buf, huge_page_size(h)),
1390                         h->free_huge_pages);
1391         }
1392 }
1393
1394 #ifdef CONFIG_HIGHMEM
1395 static void try_to_free_low(struct hstate *h, unsigned long count,
1396                                                 nodemask_t *nodes_allowed)
1397 {
1398         int i;
1399
1400         if (h->order >= MAX_ORDER)
1401                 return;
1402
1403         for_each_node_mask(i, *nodes_allowed) {
1404                 struct page *page, *next;
1405                 struct list_head *freel = &h->hugepage_freelists[i];
1406                 list_for_each_entry_safe(page, next, freel, lru) {
1407                         if (count >= h->nr_huge_pages)
1408                                 return;
1409                         if (PageHighMem(page))
1410                                 continue;
1411                         list_del(&page->lru);
1412                         update_and_free_page(h, page);
1413                         h->free_huge_pages--;
1414                         h->free_huge_pages_node[page_to_nid(page)]--;
1415                 }
1416         }
1417 }
1418 #else
1419 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1420                                                 nodemask_t *nodes_allowed)
1421 {
1422 }
1423 #endif
1424
1425 /*
1426  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1427  * balanced by operating on them in a round-robin fashion.
1428  * Returns 1 if an adjustment was made.
1429  */
1430 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1431                                 int delta)
1432 {
1433         int nr_nodes, node;
1434
1435         VM_BUG_ON(delta != -1 && delta != 1);
1436
1437         if (delta < 0) {
1438                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1439                         if (h->surplus_huge_pages_node[node])
1440                                 goto found;
1441                 }
1442         } else {
1443                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1444                         if (h->surplus_huge_pages_node[node] <
1445                                         h->nr_huge_pages_node[node])
1446                                 goto found;
1447                 }
1448         }
1449         return 0;
1450
1451 found:
1452         h->surplus_huge_pages += delta;
1453         h->surplus_huge_pages_node[node] += delta;
1454         return 1;
1455 }
1456
1457 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1458 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1459                                                 nodemask_t *nodes_allowed)
1460 {
1461         unsigned long min_count, ret;
1462
1463         if (h->order >= MAX_ORDER)
1464                 return h->max_huge_pages;
1465
1466         /*
1467          * Increase the pool size
1468          * First take pages out of surplus state.  Then make up the
1469          * remaining difference by allocating fresh huge pages.
1470          *
1471          * We might race with alloc_buddy_huge_page() here and be unable
1472          * to convert a surplus huge page to a normal huge page. That is
1473          * not critical, though, it just means the overall size of the
1474          * pool might be one hugepage larger than it needs to be, but
1475          * within all the constraints specified by the sysctls.
1476          */
1477         spin_lock(&hugetlb_lock);
1478         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1479                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1480                         break;
1481         }
1482
1483         while (count > persistent_huge_pages(h)) {
1484                 /*
1485                  * If this allocation races such that we no longer need the
1486                  * page, free_huge_page will handle it by freeing the page
1487                  * and reducing the surplus.
1488                  */
1489                 spin_unlock(&hugetlb_lock);
1490                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1491                 spin_lock(&hugetlb_lock);
1492                 if (!ret)
1493                         goto out;
1494
1495                 /* Bail for signals. Probably ctrl-c from user */
1496                 if (signal_pending(current))
1497                         goto out;
1498         }
1499
1500         /*
1501          * Decrease the pool size
1502          * First return free pages to the buddy allocator (being careful
1503          * to keep enough around to satisfy reservations).  Then place
1504          * pages into surplus state as needed so the pool will shrink
1505          * to the desired size as pages become free.
1506          *
1507          * By placing pages into the surplus state independent of the
1508          * overcommit value, we are allowing the surplus pool size to
1509          * exceed overcommit. There are few sane options here. Since
1510          * alloc_buddy_huge_page() is checking the global counter,
1511          * though, we'll note that we're not allowed to exceed surplus
1512          * and won't grow the pool anywhere else. Not until one of the
1513          * sysctls are changed, or the surplus pages go out of use.
1514          */
1515         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1516         min_count = max(count, min_count);
1517         try_to_free_low(h, min_count, nodes_allowed);
1518         while (min_count < persistent_huge_pages(h)) {
1519                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1520                         break;
1521         }
1522         while (count < persistent_huge_pages(h)) {
1523                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1524                         break;
1525         }
1526 out:
1527         ret = persistent_huge_pages(h);
1528         spin_unlock(&hugetlb_lock);
1529         return ret;
1530 }
1531
1532 #define HSTATE_ATTR_RO(_name) \
1533         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1534
1535 #define HSTATE_ATTR(_name) \
1536         static struct kobj_attribute _name##_attr = \
1537                 __ATTR(_name, 0644, _name##_show, _name##_store)
1538
1539 static struct kobject *hugepages_kobj;
1540 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1541
1542 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1543
1544 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1545 {
1546         int i;
1547
1548         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1549                 if (hstate_kobjs[i] == kobj) {
1550                         if (nidp)
1551                                 *nidp = NUMA_NO_NODE;
1552                         return &hstates[i];
1553                 }
1554
1555         return kobj_to_node_hstate(kobj, nidp);
1556 }
1557
1558 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1559                                         struct kobj_attribute *attr, char *buf)
1560 {
1561         struct hstate *h;
1562         unsigned long nr_huge_pages;
1563         int nid;
1564
1565         h = kobj_to_hstate(kobj, &nid);
1566         if (nid == NUMA_NO_NODE)
1567                 nr_huge_pages = h->nr_huge_pages;
1568         else
1569                 nr_huge_pages = h->nr_huge_pages_node[nid];
1570
1571         return sprintf(buf, "%lu\n", nr_huge_pages);
1572 }
1573
1574 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1575                         struct kobject *kobj, struct kobj_attribute *attr,
1576                         const char *buf, size_t len)
1577 {
1578         int err;
1579         int nid;
1580         unsigned long count;
1581         struct hstate *h;
1582         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1583
1584         err = kstrtoul(buf, 10, &count);
1585         if (err)
1586                 goto out;
1587
1588         h = kobj_to_hstate(kobj, &nid);
1589         if (h->order >= MAX_ORDER) {
1590                 err = -EINVAL;
1591                 goto out;
1592         }
1593
1594         if (nid == NUMA_NO_NODE) {
1595                 /*
1596                  * global hstate attribute
1597                  */
1598                 if (!(obey_mempolicy &&
1599                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1600                         NODEMASK_FREE(nodes_allowed);
1601                         nodes_allowed = &node_states[N_MEMORY];
1602                 }
1603         } else if (nodes_allowed) {
1604                 /*
1605                  * per node hstate attribute: adjust count to global,
1606                  * but restrict alloc/free to the specified node.
1607                  */
1608                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1609                 init_nodemask_of_node(nodes_allowed, nid);
1610         } else
1611                 nodes_allowed = &node_states[N_MEMORY];
1612
1613         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1614
1615         if (nodes_allowed != &node_states[N_MEMORY])
1616                 NODEMASK_FREE(nodes_allowed);
1617
1618         return len;
1619 out:
1620         NODEMASK_FREE(nodes_allowed);
1621         return err;
1622 }
1623
1624 static ssize_t nr_hugepages_show(struct kobject *kobj,
1625                                        struct kobj_attribute *attr, char *buf)
1626 {
1627         return nr_hugepages_show_common(kobj, attr, buf);
1628 }
1629
1630 static ssize_t nr_hugepages_store(struct kobject *kobj,
1631                struct kobj_attribute *attr, const char *buf, size_t len)
1632 {
1633         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1634 }
1635 HSTATE_ATTR(nr_hugepages);
1636
1637 #ifdef CONFIG_NUMA
1638
1639 /*
1640  * hstate attribute for optionally mempolicy-based constraint on persistent
1641  * huge page alloc/free.
1642  */
1643 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1644                                        struct kobj_attribute *attr, char *buf)
1645 {
1646         return nr_hugepages_show_common(kobj, attr, buf);
1647 }
1648
1649 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1650                struct kobj_attribute *attr, const char *buf, size_t len)
1651 {
1652         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1653 }
1654 HSTATE_ATTR(nr_hugepages_mempolicy);
1655 #endif
1656
1657
1658 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1659                                         struct kobj_attribute *attr, char *buf)
1660 {
1661         struct hstate *h = kobj_to_hstate(kobj, NULL);
1662         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1663 }
1664
1665 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1666                 struct kobj_attribute *attr, const char *buf, size_t count)
1667 {
1668         int err;
1669         unsigned long input;
1670         struct hstate *h = kobj_to_hstate(kobj, NULL);
1671
1672         if (h->order >= MAX_ORDER)
1673                 return -EINVAL;
1674
1675         err = kstrtoul(buf, 10, &input);
1676         if (err)
1677                 return err;
1678
1679         spin_lock(&hugetlb_lock);
1680         h->nr_overcommit_huge_pages = input;
1681         spin_unlock(&hugetlb_lock);
1682
1683         return count;
1684 }
1685 HSTATE_ATTR(nr_overcommit_hugepages);
1686
1687 static ssize_t free_hugepages_show(struct kobject *kobj,
1688                                         struct kobj_attribute *attr, char *buf)
1689 {
1690         struct hstate *h;
1691         unsigned long free_huge_pages;
1692         int nid;
1693
1694         h = kobj_to_hstate(kobj, &nid);
1695         if (nid == NUMA_NO_NODE)
1696                 free_huge_pages = h->free_huge_pages;
1697         else
1698                 free_huge_pages = h->free_huge_pages_node[nid];
1699
1700         return sprintf(buf, "%lu\n", free_huge_pages);
1701 }
1702 HSTATE_ATTR_RO(free_hugepages);
1703
1704 static ssize_t resv_hugepages_show(struct kobject *kobj,
1705                                         struct kobj_attribute *attr, char *buf)
1706 {
1707         struct hstate *h = kobj_to_hstate(kobj, NULL);
1708         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1709 }
1710 HSTATE_ATTR_RO(resv_hugepages);
1711
1712 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1713                                         struct kobj_attribute *attr, char *buf)
1714 {
1715         struct hstate *h;
1716         unsigned long surplus_huge_pages;
1717         int nid;
1718
1719         h = kobj_to_hstate(kobj, &nid);
1720         if (nid == NUMA_NO_NODE)
1721                 surplus_huge_pages = h->surplus_huge_pages;
1722         else
1723                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1724
1725         return sprintf(buf, "%lu\n", surplus_huge_pages);
1726 }
1727 HSTATE_ATTR_RO(surplus_hugepages);
1728
1729 static struct attribute *hstate_attrs[] = {
1730         &nr_hugepages_attr.attr,
1731         &nr_overcommit_hugepages_attr.attr,
1732         &free_hugepages_attr.attr,
1733         &resv_hugepages_attr.attr,
1734         &surplus_hugepages_attr.attr,
1735 #ifdef CONFIG_NUMA
1736         &nr_hugepages_mempolicy_attr.attr,
1737 #endif
1738         NULL,
1739 };
1740
1741 static struct attribute_group hstate_attr_group = {
1742         .attrs = hstate_attrs,
1743 };
1744
1745 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1746                                     struct kobject **hstate_kobjs,
1747                                     struct attribute_group *hstate_attr_group)
1748 {
1749         int retval;
1750         int hi = hstate_index(h);
1751
1752         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1753         if (!hstate_kobjs[hi])
1754                 return -ENOMEM;
1755
1756         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1757         if (retval)
1758                 kobject_put(hstate_kobjs[hi]);
1759
1760         return retval;
1761 }
1762
1763 static void __init hugetlb_sysfs_init(void)
1764 {
1765         struct hstate *h;
1766         int err;
1767
1768         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1769         if (!hugepages_kobj)
1770                 return;
1771
1772         for_each_hstate(h) {
1773                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1774                                          hstate_kobjs, &hstate_attr_group);
1775                 if (err)
1776                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
1777         }
1778 }
1779
1780 #ifdef CONFIG_NUMA
1781
1782 /*
1783  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1784  * with node devices in node_devices[] using a parallel array.  The array
1785  * index of a node device or _hstate == node id.
1786  * This is here to avoid any static dependency of the node device driver, in
1787  * the base kernel, on the hugetlb module.
1788  */
1789 struct node_hstate {
1790         struct kobject          *hugepages_kobj;
1791         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1792 };
1793 struct node_hstate node_hstates[MAX_NUMNODES];
1794
1795 /*
1796  * A subset of global hstate attributes for node devices
1797  */
1798 static struct attribute *per_node_hstate_attrs[] = {
1799         &nr_hugepages_attr.attr,
1800         &free_hugepages_attr.attr,
1801         &surplus_hugepages_attr.attr,
1802         NULL,
1803 };
1804
1805 static struct attribute_group per_node_hstate_attr_group = {
1806         .attrs = per_node_hstate_attrs,
1807 };
1808
1809 /*
1810  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1811  * Returns node id via non-NULL nidp.
1812  */
1813 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1814 {
1815         int nid;
1816
1817         for (nid = 0; nid < nr_node_ids; nid++) {
1818                 struct node_hstate *nhs = &node_hstates[nid];
1819                 int i;
1820                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1821                         if (nhs->hstate_kobjs[i] == kobj) {
1822                                 if (nidp)
1823                                         *nidp = nid;
1824                                 return &hstates[i];
1825                         }
1826         }
1827
1828         BUG();
1829         return NULL;
1830 }
1831
1832 /*
1833  * Unregister hstate attributes from a single node device.
1834  * No-op if no hstate attributes attached.
1835  */
1836 static void hugetlb_unregister_node(struct node *node)
1837 {
1838         struct hstate *h;
1839         struct node_hstate *nhs = &node_hstates[node->dev.id];
1840
1841         if (!nhs->hugepages_kobj)
1842                 return;         /* no hstate attributes */
1843
1844         for_each_hstate(h) {
1845                 int idx = hstate_index(h);
1846                 if (nhs->hstate_kobjs[idx]) {
1847                         kobject_put(nhs->hstate_kobjs[idx]);
1848                         nhs->hstate_kobjs[idx] = NULL;
1849                 }
1850         }
1851
1852         kobject_put(nhs->hugepages_kobj);
1853         nhs->hugepages_kobj = NULL;
1854 }
1855
1856 /*
1857  * hugetlb module exit:  unregister hstate attributes from node devices
1858  * that have them.
1859  */
1860 static void hugetlb_unregister_all_nodes(void)
1861 {
1862         int nid;
1863
1864         /*
1865          * disable node device registrations.
1866          */
1867         register_hugetlbfs_with_node(NULL, NULL);
1868
1869         /*
1870          * remove hstate attributes from any nodes that have them.
1871          */
1872         for (nid = 0; nid < nr_node_ids; nid++)
1873                 hugetlb_unregister_node(node_devices[nid]);
1874 }
1875
1876 /*
1877  * Register hstate attributes for a single node device.
1878  * No-op if attributes already registered.
1879  */
1880 static void hugetlb_register_node(struct node *node)
1881 {
1882         struct hstate *h;
1883         struct node_hstate *nhs = &node_hstates[node->dev.id];
1884         int err;
1885
1886         if (nhs->hugepages_kobj)
1887                 return;         /* already allocated */
1888
1889         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1890                                                         &node->dev.kobj);
1891         if (!nhs->hugepages_kobj)
1892                 return;
1893
1894         for_each_hstate(h) {
1895                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1896                                                 nhs->hstate_kobjs,
1897                                                 &per_node_hstate_attr_group);
1898                 if (err) {
1899                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1900                                 h->name, node->dev.id);
1901                         hugetlb_unregister_node(node);
1902                         break;
1903                 }
1904         }
1905 }
1906
1907 /*
1908  * hugetlb init time:  register hstate attributes for all registered node
1909  * devices of nodes that have memory.  All on-line nodes should have
1910  * registered their associated device by this time.
1911  */
1912 static void hugetlb_register_all_nodes(void)
1913 {
1914         int nid;
1915
1916         for_each_node_state(nid, N_MEMORY) {
1917                 struct node *node = node_devices[nid];
1918                 if (node->dev.id == nid)
1919                         hugetlb_register_node(node);
1920         }
1921
1922         /*
1923          * Let the node device driver know we're here so it can
1924          * [un]register hstate attributes on node hotplug.
1925          */
1926         register_hugetlbfs_with_node(hugetlb_register_node,
1927                                      hugetlb_unregister_node);
1928 }
1929 #else   /* !CONFIG_NUMA */
1930
1931 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1932 {
1933         BUG();
1934         if (nidp)
1935                 *nidp = -1;
1936         return NULL;
1937 }
1938
1939 static void hugetlb_unregister_all_nodes(void) { }
1940
1941 static void hugetlb_register_all_nodes(void) { }
1942
1943 #endif
1944
1945 static void __exit hugetlb_exit(void)
1946 {
1947         struct hstate *h;
1948
1949         hugetlb_unregister_all_nodes();
1950
1951         for_each_hstate(h) {
1952                 kobject_put(hstate_kobjs[hstate_index(h)]);
1953         }
1954
1955         kobject_put(hugepages_kobj);
1956 }
1957 module_exit(hugetlb_exit);
1958
1959 static int __init hugetlb_init(void)
1960 {
1961         /* Some platform decide whether they support huge pages at boot
1962          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1963          * there is no such support
1964          */
1965         if (HPAGE_SHIFT == 0)
1966                 return 0;
1967
1968         if (!size_to_hstate(default_hstate_size)) {
1969                 default_hstate_size = HPAGE_SIZE;
1970                 if (!size_to_hstate(default_hstate_size))
1971                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1972         }
1973         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1974         if (default_hstate_max_huge_pages)
1975                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1976
1977         hugetlb_init_hstates();
1978         gather_bootmem_prealloc();
1979         report_hugepages();
1980
1981         hugetlb_sysfs_init();
1982         hugetlb_register_all_nodes();
1983         hugetlb_cgroup_file_init();
1984
1985         return 0;
1986 }
1987 module_init(hugetlb_init);
1988
1989 /* Should be called on processing a hugepagesz=... option */
1990 void __init hugetlb_add_hstate(unsigned order)
1991 {
1992         struct hstate *h;
1993         unsigned long i;
1994
1995         if (size_to_hstate(PAGE_SIZE << order)) {
1996                 pr_warning("hugepagesz= specified twice, ignoring\n");
1997                 return;
1998         }
1999         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2000         BUG_ON(order == 0);
2001         h = &hstates[hugetlb_max_hstate++];
2002         h->order = order;
2003         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2004         h->nr_huge_pages = 0;
2005         h->free_huge_pages = 0;
2006         for (i = 0; i < MAX_NUMNODES; ++i)
2007                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2008         INIT_LIST_HEAD(&h->hugepage_activelist);
2009         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2010         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2011         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2012                                         huge_page_size(h)/1024);
2013
2014         parsed_hstate = h;
2015 }
2016
2017 static int __init hugetlb_nrpages_setup(char *s)
2018 {
2019         unsigned long *mhp;
2020         static unsigned long *last_mhp;
2021
2022         /*
2023          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2024          * so this hugepages= parameter goes to the "default hstate".
2025          */
2026         if (!hugetlb_max_hstate)
2027                 mhp = &default_hstate_max_huge_pages;
2028         else
2029                 mhp = &parsed_hstate->max_huge_pages;
2030
2031         if (mhp == last_mhp) {
2032                 pr_warning("hugepages= specified twice without "
2033                            "interleaving hugepagesz=, ignoring\n");
2034                 return 1;
2035         }
2036
2037         if (sscanf(s, "%lu", mhp) <= 0)
2038                 *mhp = 0;
2039
2040         /*
2041          * Global state is always initialized later in hugetlb_init.
2042          * But we need to allocate >= MAX_ORDER hstates here early to still
2043          * use the bootmem allocator.
2044          */
2045         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2046                 hugetlb_hstate_alloc_pages(parsed_hstate);
2047
2048         last_mhp = mhp;
2049
2050         return 1;
2051 }
2052 __setup("hugepages=", hugetlb_nrpages_setup);
2053
2054 static int __init hugetlb_default_setup(char *s)
2055 {
2056         default_hstate_size = memparse(s, &s);
2057         return 1;
2058 }
2059 __setup("default_hugepagesz=", hugetlb_default_setup);
2060
2061 static unsigned int cpuset_mems_nr(unsigned int *array)
2062 {
2063         int node;
2064         unsigned int nr = 0;
2065
2066         for_each_node_mask(node, cpuset_current_mems_allowed)
2067                 nr += array[node];
2068
2069         return nr;
2070 }
2071
2072 #ifdef CONFIG_SYSCTL
2073 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2074                          struct ctl_table *table, int write,
2075                          void __user *buffer, size_t *length, loff_t *ppos)
2076 {
2077         struct hstate *h = &default_hstate;
2078         unsigned long tmp;
2079         int ret;
2080
2081         tmp = h->max_huge_pages;
2082
2083         if (write && h->order >= MAX_ORDER)
2084                 return -EINVAL;
2085
2086         table->data = &tmp;
2087         table->maxlen = sizeof(unsigned long);
2088         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2089         if (ret)
2090                 goto out;
2091
2092         if (write) {
2093                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2094                                                 GFP_KERNEL | __GFP_NORETRY);
2095                 if (!(obey_mempolicy &&
2096                                init_nodemask_of_mempolicy(nodes_allowed))) {
2097                         NODEMASK_FREE(nodes_allowed);
2098                         nodes_allowed = &node_states[N_MEMORY];
2099                 }
2100                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2101
2102                 if (nodes_allowed != &node_states[N_MEMORY])
2103                         NODEMASK_FREE(nodes_allowed);
2104         }
2105 out:
2106         return ret;
2107 }
2108
2109 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2110                           void __user *buffer, size_t *length, loff_t *ppos)
2111 {
2112
2113         return hugetlb_sysctl_handler_common(false, table, write,
2114                                                         buffer, length, ppos);
2115 }
2116
2117 #ifdef CONFIG_NUMA
2118 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2119                           void __user *buffer, size_t *length, loff_t *ppos)
2120 {
2121         return hugetlb_sysctl_handler_common(true, table, write,
2122                                                         buffer, length, ppos);
2123 }
2124 #endif /* CONFIG_NUMA */
2125
2126 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2127                         void __user *buffer,
2128                         size_t *length, loff_t *ppos)
2129 {
2130         struct hstate *h = &default_hstate;
2131         unsigned long tmp;
2132         int ret;
2133
2134         tmp = h->nr_overcommit_huge_pages;
2135
2136         if (write && h->order >= MAX_ORDER)
2137                 return -EINVAL;
2138
2139         table->data = &tmp;
2140         table->maxlen = sizeof(unsigned long);
2141         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2142         if (ret)
2143                 goto out;
2144
2145         if (write) {
2146                 spin_lock(&hugetlb_lock);
2147                 h->nr_overcommit_huge_pages = tmp;
2148                 spin_unlock(&hugetlb_lock);
2149         }
2150 out:
2151         return ret;
2152 }
2153
2154 #endif /* CONFIG_SYSCTL */
2155
2156 void hugetlb_report_meminfo(struct seq_file *m)
2157 {
2158         struct hstate *h = &default_hstate;
2159         seq_printf(m,
2160                         "HugePages_Total:   %5lu\n"
2161                         "HugePages_Free:    %5lu\n"
2162                         "HugePages_Rsvd:    %5lu\n"
2163                         "HugePages_Surp:    %5lu\n"
2164                         "Hugepagesize:   %8lu kB\n",
2165                         h->nr_huge_pages,
2166                         h->free_huge_pages,
2167                         h->resv_huge_pages,
2168                         h->surplus_huge_pages,
2169                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2170 }
2171
2172 int hugetlb_report_node_meminfo(int nid, char *buf)
2173 {
2174         struct hstate *h = &default_hstate;
2175         return sprintf(buf,
2176                 "Node %d HugePages_Total: %5u\n"
2177                 "Node %d HugePages_Free:  %5u\n"
2178                 "Node %d HugePages_Surp:  %5u\n",
2179                 nid, h->nr_huge_pages_node[nid],
2180                 nid, h->free_huge_pages_node[nid],
2181                 nid, h->surplus_huge_pages_node[nid]);
2182 }
2183
2184 void hugetlb_show_meminfo(void)
2185 {
2186         struct hstate *h;
2187         int nid;
2188
2189         for_each_node_state(nid, N_MEMORY)
2190                 for_each_hstate(h)
2191                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2192                                 nid,
2193                                 h->nr_huge_pages_node[nid],
2194                                 h->free_huge_pages_node[nid],
2195                                 h->surplus_huge_pages_node[nid],
2196                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2197 }
2198
2199 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2200 unsigned long hugetlb_total_pages(void)
2201 {
2202         struct hstate *h;
2203         unsigned long nr_total_pages = 0;
2204
2205         for_each_hstate(h)
2206                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2207         return nr_total_pages;
2208 }
2209
2210 static int hugetlb_acct_memory(struct hstate *h, long delta)
2211 {
2212         int ret = -ENOMEM;
2213
2214         spin_lock(&hugetlb_lock);
2215         /*
2216          * When cpuset is configured, it breaks the strict hugetlb page
2217          * reservation as the accounting is done on a global variable. Such
2218          * reservation is completely rubbish in the presence of cpuset because
2219          * the reservation is not checked against page availability for the
2220          * current cpuset. Application can still potentially OOM'ed by kernel
2221          * with lack of free htlb page in cpuset that the task is in.
2222          * Attempt to enforce strict accounting with cpuset is almost
2223          * impossible (or too ugly) because cpuset is too fluid that
2224          * task or memory node can be dynamically moved between cpusets.
2225          *
2226          * The change of semantics for shared hugetlb mapping with cpuset is
2227          * undesirable. However, in order to preserve some of the semantics,
2228          * we fall back to check against current free page availability as
2229          * a best attempt and hopefully to minimize the impact of changing
2230          * semantics that cpuset has.
2231          */
2232         if (delta > 0) {
2233                 if (gather_surplus_pages(h, delta) < 0)
2234                         goto out;
2235
2236                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2237                         return_unused_surplus_pages(h, delta);
2238                         goto out;
2239                 }
2240         }
2241
2242         ret = 0;
2243         if (delta < 0)
2244                 return_unused_surplus_pages(h, (unsigned long) -delta);
2245
2246 out:
2247         spin_unlock(&hugetlb_lock);
2248         return ret;
2249 }
2250
2251 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2252 {
2253         struct resv_map *resv = vma_resv_map(vma);
2254
2255         /*
2256          * This new VMA should share its siblings reservation map if present.
2257          * The VMA will only ever have a valid reservation map pointer where
2258          * it is being copied for another still existing VMA.  As that VMA
2259          * has a reference to the reservation map it cannot disappear until
2260          * after this open call completes.  It is therefore safe to take a
2261          * new reference here without additional locking.
2262          */
2263         if (resv)
2264                 kref_get(&resv->refs);
2265 }
2266
2267 static void resv_map_put(struct vm_area_struct *vma)
2268 {
2269         struct resv_map *resv = vma_resv_map(vma);
2270
2271         if (!resv)
2272                 return;
2273         kref_put(&resv->refs, resv_map_release);
2274 }
2275
2276 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2277 {
2278         struct hstate *h = hstate_vma(vma);
2279         struct resv_map *resv = vma_resv_map(vma);
2280         struct hugepage_subpool *spool = subpool_vma(vma);
2281         unsigned long reserve;
2282         unsigned long start;
2283         unsigned long end;
2284
2285         if (resv) {
2286                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2287                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2288
2289                 reserve = (end - start) -
2290                         region_count(&resv->regions, start, end);
2291
2292                 resv_map_put(vma);
2293
2294                 if (reserve) {
2295                         hugetlb_acct_memory(h, -reserve);
2296                         hugepage_subpool_put_pages(spool, reserve);
2297                 }
2298         }
2299 }
2300
2301 /*
2302  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2303  * handle_mm_fault() to try to instantiate regular-sized pages in the
2304  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2305  * this far.
2306  */
2307 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2308 {
2309         BUG();
2310         return 0;
2311 }
2312
2313 const struct vm_operations_struct hugetlb_vm_ops = {
2314         .fault = hugetlb_vm_op_fault,
2315         .open = hugetlb_vm_op_open,
2316         .close = hugetlb_vm_op_close,
2317 };
2318
2319 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2320                                 int writable)
2321 {
2322         pte_t entry;
2323
2324         if (writable) {
2325                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2326                                          vma->vm_page_prot)));
2327         } else {
2328                 entry = huge_pte_wrprotect(mk_huge_pte(page,
2329                                            vma->vm_page_prot));
2330         }
2331         entry = pte_mkyoung(entry);
2332         entry = pte_mkhuge(entry);
2333         entry = arch_make_huge_pte(entry, vma, page, writable);
2334
2335         return entry;
2336 }
2337
2338 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2339                                    unsigned long address, pte_t *ptep)
2340 {
2341         pte_t entry;
2342
2343         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2344         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2345                 update_mmu_cache(vma, address, ptep);
2346 }
2347
2348
2349 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2350                             struct vm_area_struct *vma)
2351 {
2352         pte_t *src_pte, *dst_pte, entry;
2353         struct page *ptepage;
2354         unsigned long addr;
2355         int cow;
2356         struct hstate *h = hstate_vma(vma);
2357         unsigned long sz = huge_page_size(h);
2358
2359         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2360
2361         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2362                 spinlock_t *src_ptl, *dst_ptl;
2363                 src_pte = huge_pte_offset(src, addr);
2364                 if (!src_pte)
2365                         continue;
2366                 dst_pte = huge_pte_alloc(dst, addr, sz);
2367                 if (!dst_pte)
2368                         goto nomem;
2369
2370                 /* If the pagetables are shared don't copy or take references */
2371                 if (dst_pte == src_pte)
2372                         continue;
2373
2374                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2375                 src_ptl = huge_pte_lockptr(h, src, src_pte);
2376                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2377                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2378                         if (cow)
2379                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2380                         entry = huge_ptep_get(src_pte);
2381                         ptepage = pte_page(entry);
2382                         get_page(ptepage);
2383                         page_dup_rmap(ptepage);
2384                         set_huge_pte_at(dst, addr, dst_pte, entry);
2385                 }
2386                 spin_unlock(src_ptl);
2387                 spin_unlock(dst_ptl);
2388         }
2389         return 0;
2390
2391 nomem:
2392         return -ENOMEM;
2393 }
2394
2395 static int is_hugetlb_entry_migration(pte_t pte)
2396 {
2397         swp_entry_t swp;
2398
2399         if (huge_pte_none(pte) || pte_present(pte))
2400                 return 0;
2401         swp = pte_to_swp_entry(pte);
2402         if (non_swap_entry(swp) && is_migration_entry(swp))
2403                 return 1;
2404         else
2405                 return 0;
2406 }
2407
2408 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2409 {
2410         swp_entry_t swp;
2411
2412         if (huge_pte_none(pte) || pte_present(pte))
2413                 return 0;
2414         swp = pte_to_swp_entry(pte);
2415         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2416                 return 1;
2417         else
2418                 return 0;
2419 }
2420
2421 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2422                             unsigned long start, unsigned long end,
2423                             struct page *ref_page)
2424 {
2425         int force_flush = 0;
2426         struct mm_struct *mm = vma->vm_mm;
2427         unsigned long address;
2428         pte_t *ptep;
2429         pte_t pte;
2430         spinlock_t *ptl;
2431         struct page *page;
2432         struct hstate *h = hstate_vma(vma);
2433         unsigned long sz = huge_page_size(h);
2434         const unsigned long mmun_start = start; /* For mmu_notifiers */
2435         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2436
2437         WARN_ON(!is_vm_hugetlb_page(vma));
2438         BUG_ON(start & ~huge_page_mask(h));
2439         BUG_ON(end & ~huge_page_mask(h));
2440
2441         tlb_start_vma(tlb, vma);
2442         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2443 again:
2444         for (address = start; address < end; address += sz) {
2445                 ptep = huge_pte_offset(mm, address);
2446                 if (!ptep)
2447                         continue;
2448
2449                 ptl = huge_pte_lock(h, mm, ptep);
2450                 if (huge_pmd_unshare(mm, &address, ptep))
2451                         goto unlock;
2452
2453                 pte = huge_ptep_get(ptep);
2454                 if (huge_pte_none(pte))
2455                         goto unlock;
2456
2457                 /*
2458                  * HWPoisoned hugepage is already unmapped and dropped reference
2459                  */
2460                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2461                         huge_pte_clear(mm, address, ptep);
2462                         goto unlock;
2463                 }
2464
2465                 page = pte_page(pte);
2466                 /*
2467                  * If a reference page is supplied, it is because a specific
2468                  * page is being unmapped, not a range. Ensure the page we
2469                  * are about to unmap is the actual page of interest.
2470                  */
2471                 if (ref_page) {
2472                         if (page != ref_page)
2473                                 goto unlock;
2474
2475                         /*
2476                          * Mark the VMA as having unmapped its page so that
2477                          * future faults in this VMA will fail rather than
2478                          * looking like data was lost
2479                          */
2480                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2481                 }
2482
2483                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2484                 tlb_remove_tlb_entry(tlb, ptep, address);
2485                 if (huge_pte_dirty(pte))
2486                         set_page_dirty(page);
2487
2488                 page_remove_rmap(page);
2489                 force_flush = !__tlb_remove_page(tlb, page);
2490                 if (force_flush) {
2491                         spin_unlock(ptl);
2492                         break;
2493                 }
2494                 /* Bail out after unmapping reference page if supplied */
2495                 if (ref_page) {
2496                         spin_unlock(ptl);
2497                         break;
2498                 }
2499 unlock:
2500                 spin_unlock(ptl);
2501         }
2502         /*
2503          * mmu_gather ran out of room to batch pages, we break out of
2504          * the PTE lock to avoid doing the potential expensive TLB invalidate
2505          * and page-free while holding it.
2506          */
2507         if (force_flush) {
2508                 force_flush = 0;
2509                 tlb_flush_mmu(tlb);
2510                 if (address < end && !ref_page)
2511                         goto again;
2512         }
2513         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2514         tlb_end_vma(tlb, vma);
2515 }
2516
2517 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2518                           struct vm_area_struct *vma, unsigned long start,
2519                           unsigned long end, struct page *ref_page)
2520 {
2521         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2522
2523         /*
2524          * Clear this flag so that x86's huge_pmd_share page_table_shareable
2525          * test will fail on a vma being torn down, and not grab a page table
2526          * on its way out.  We're lucky that the flag has such an appropriate
2527          * name, and can in fact be safely cleared here. We could clear it
2528          * before the __unmap_hugepage_range above, but all that's necessary
2529          * is to clear it before releasing the i_mmap_mutex. This works
2530          * because in the context this is called, the VMA is about to be
2531          * destroyed and the i_mmap_mutex is held.
2532          */
2533         vma->vm_flags &= ~VM_MAYSHARE;
2534 }
2535
2536 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2537                           unsigned long end, struct page *ref_page)
2538 {
2539         struct mm_struct *mm;
2540         struct mmu_gather tlb;
2541
2542         mm = vma->vm_mm;
2543
2544         tlb_gather_mmu(&tlb, mm, start, end);
2545         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2546         tlb_finish_mmu(&tlb, start, end);
2547 }
2548
2549 /*
2550  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2551  * mappping it owns the reserve page for. The intention is to unmap the page
2552  * from other VMAs and let the children be SIGKILLed if they are faulting the
2553  * same region.
2554  */
2555 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2556                                 struct page *page, unsigned long address)
2557 {
2558         struct hstate *h = hstate_vma(vma);
2559         struct vm_area_struct *iter_vma;
2560         struct address_space *mapping;
2561         pgoff_t pgoff;
2562
2563         /*
2564          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2565          * from page cache lookup which is in HPAGE_SIZE units.
2566          */
2567         address = address & huge_page_mask(h);
2568         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2569                         vma->vm_pgoff;
2570         mapping = file_inode(vma->vm_file)->i_mapping;
2571
2572         /*
2573          * Take the mapping lock for the duration of the table walk. As
2574          * this mapping should be shared between all the VMAs,
2575          * __unmap_hugepage_range() is called as the lock is already held
2576          */
2577         mutex_lock(&mapping->i_mmap_mutex);
2578         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2579                 /* Do not unmap the current VMA */
2580                 if (iter_vma == vma)
2581                         continue;
2582
2583                 /*
2584                  * Unmap the page from other VMAs without their own reserves.
2585                  * They get marked to be SIGKILLed if they fault in these
2586                  * areas. This is because a future no-page fault on this VMA
2587                  * could insert a zeroed page instead of the data existing
2588                  * from the time of fork. This would look like data corruption
2589                  */
2590                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2591                         unmap_hugepage_range(iter_vma, address,
2592                                              address + huge_page_size(h), page);
2593         }
2594         mutex_unlock(&mapping->i_mmap_mutex);
2595
2596         return 1;
2597 }
2598
2599 /*
2600  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2601  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2602  * cannot race with other handlers or page migration.
2603  * Keep the pte_same checks anyway to make transition from the mutex easier.
2604  */
2605 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2606                         unsigned long address, pte_t *ptep, pte_t pte,
2607                         struct page *pagecache_page, spinlock_t *ptl)
2608 {
2609         struct hstate *h = hstate_vma(vma);
2610         struct page *old_page, *new_page;
2611         int outside_reserve = 0;
2612         unsigned long mmun_start;       /* For mmu_notifiers */
2613         unsigned long mmun_end;         /* For mmu_notifiers */
2614
2615         old_page = pte_page(pte);
2616
2617 retry_avoidcopy:
2618         /* If no-one else is actually using this page, avoid the copy
2619          * and just make the page writable */
2620         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2621                 page_move_anon_rmap(old_page, vma, address);
2622                 set_huge_ptep_writable(vma, address, ptep);
2623                 return 0;
2624         }
2625
2626         /*
2627          * If the process that created a MAP_PRIVATE mapping is about to
2628          * perform a COW due to a shared page count, attempt to satisfy
2629          * the allocation without using the existing reserves. The pagecache
2630          * page is used to determine if the reserve at this address was
2631          * consumed or not. If reserves were used, a partial faulted mapping
2632          * at the time of fork() could consume its reserves on COW instead
2633          * of the full address range.
2634          */
2635         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2636                         old_page != pagecache_page)
2637                 outside_reserve = 1;
2638
2639         page_cache_get(old_page);
2640
2641         /* Drop page table lock as buddy allocator may be called */
2642         spin_unlock(ptl);
2643         new_page = alloc_huge_page(vma, address, outside_reserve);
2644
2645         if (IS_ERR(new_page)) {
2646                 long err = PTR_ERR(new_page);
2647                 page_cache_release(old_page);
2648
2649                 /*
2650                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2651                  * it is due to references held by a child and an insufficient
2652                  * huge page pool. To guarantee the original mappers
2653                  * reliability, unmap the page from child processes. The child
2654                  * may get SIGKILLed if it later faults.
2655                  */
2656                 if (outside_reserve) {
2657                         BUG_ON(huge_pte_none(pte));
2658                         if (unmap_ref_private(mm, vma, old_page, address)) {
2659                                 BUG_ON(huge_pte_none(pte));
2660                                 spin_lock(ptl);
2661                                 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2662                                 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2663                                         goto retry_avoidcopy;
2664                                 /*
2665                                  * race occurs while re-acquiring page table
2666                                  * lock, and our job is done.
2667                                  */
2668                                 return 0;
2669                         }
2670                         WARN_ON_ONCE(1);
2671                 }
2672
2673                 /* Caller expects lock to be held */
2674                 spin_lock(ptl);
2675                 if (err == -ENOMEM)
2676                         return VM_FAULT_OOM;
2677                 else
2678                         return VM_FAULT_SIGBUS;
2679         }
2680
2681         /*
2682          * When the original hugepage is shared one, it does not have
2683          * anon_vma prepared.
2684          */
2685         if (unlikely(anon_vma_prepare(vma))) {
2686                 page_cache_release(new_page);
2687                 page_cache_release(old_page);
2688                 /* Caller expects lock to be held */
2689                 spin_lock(ptl);
2690                 return VM_FAULT_OOM;
2691         }
2692
2693         copy_user_huge_page(new_page, old_page, address, vma,
2694                             pages_per_huge_page(h));
2695         __SetPageUptodate(new_page);
2696
2697         mmun_start = address & huge_page_mask(h);
2698         mmun_end = mmun_start + huge_page_size(h);
2699         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2700         /*
2701          * Retake the page table lock to check for racing updates
2702          * before the page tables are altered
2703          */
2704         spin_lock(ptl);
2705         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2706         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2707                 ClearPagePrivate(new_page);
2708
2709                 /* Break COW */
2710                 huge_ptep_clear_flush(vma, address, ptep);
2711                 set_huge_pte_at(mm, address, ptep,
2712                                 make_huge_pte(vma, new_page, 1));
2713                 page_remove_rmap(old_page);
2714                 hugepage_add_new_anon_rmap(new_page, vma, address);
2715                 /* Make the old page be freed below */
2716                 new_page = old_page;
2717         }
2718         spin_unlock(ptl);
2719         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2720         page_cache_release(new_page);
2721         page_cache_release(old_page);
2722
2723         /* Caller expects lock to be held */
2724         spin_lock(ptl);
2725         return 0;
2726 }
2727
2728 /* Return the pagecache page at a given address within a VMA */
2729 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2730                         struct vm_area_struct *vma, unsigned long address)
2731 {
2732         struct address_space *mapping;
2733         pgoff_t idx;
2734
2735         mapping = vma->vm_file->f_mapping;
2736         idx = vma_hugecache_offset(h, vma, address);
2737
2738         return find_lock_page(mapping, idx);
2739 }
2740
2741 /*
2742  * Return whether there is a pagecache page to back given address within VMA.
2743  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2744  */
2745 static bool hugetlbfs_pagecache_present(struct hstate *h,
2746                         struct vm_area_struct *vma, unsigned long address)
2747 {
2748         struct address_space *mapping;
2749         pgoff_t idx;
2750         struct page *page;
2751
2752         mapping = vma->vm_file->f_mapping;
2753         idx = vma_hugecache_offset(h, vma, address);
2754
2755         page = find_get_page(mapping, idx);
2756         if (page)
2757                 put_page(page);
2758         return page != NULL;
2759 }
2760
2761 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2762                         unsigned long address, pte_t *ptep, unsigned int flags)
2763 {
2764         struct hstate *h = hstate_vma(vma);
2765         int ret = VM_FAULT_SIGBUS;
2766         int anon_rmap = 0;
2767         pgoff_t idx;
2768         unsigned long size;
2769         struct page *page;
2770         struct address_space *mapping;
2771         pte_t new_pte;
2772         spinlock_t *ptl;
2773
2774         /*
2775          * Currently, we are forced to kill the process in the event the
2776          * original mapper has unmapped pages from the child due to a failed
2777          * COW. Warn that such a situation has occurred as it may not be obvious
2778          */
2779         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2780                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2781                            current->pid);
2782                 return ret;
2783         }
2784
2785         mapping = vma->vm_file->f_mapping;
2786         idx = vma_hugecache_offset(h, vma, address);
2787
2788         /*
2789          * Use page lock to guard against racing truncation
2790          * before we get page_table_lock.
2791          */
2792 retry:
2793         page = find_lock_page(mapping, idx);
2794         if (!page) {
2795                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2796                 if (idx >= size)
2797                         goto out;
2798                 page = alloc_huge_page(vma, address, 0);
2799                 if (IS_ERR(page)) {
2800                         ret = PTR_ERR(page);
2801                         if (ret == -ENOMEM)
2802                                 ret = VM_FAULT_OOM;
2803                         else
2804                                 ret = VM_FAULT_SIGBUS;
2805                         goto out;
2806                 }
2807                 clear_huge_page(page, address, pages_per_huge_page(h));
2808                 __SetPageUptodate(page);
2809
2810                 if (vma->vm_flags & VM_MAYSHARE) {
2811                         int err;
2812                         struct inode *inode = mapping->host;
2813
2814                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2815                         if (err) {
2816                                 put_page(page);
2817                                 if (err == -EEXIST)
2818                                         goto retry;
2819                                 goto out;
2820                         }
2821                         ClearPagePrivate(page);
2822
2823                         spin_lock(&inode->i_lock);
2824                         inode->i_blocks += blocks_per_huge_page(h);
2825                         spin_unlock(&inode->i_lock);
2826                 } else {
2827                         lock_page(page);
2828                         if (unlikely(anon_vma_prepare(vma))) {
2829                                 ret = VM_FAULT_OOM;
2830                                 goto backout_unlocked;
2831                         }
2832                         anon_rmap = 1;
2833                 }
2834         } else {
2835                 /*
2836                  * If memory error occurs between mmap() and fault, some process
2837                  * don't have hwpoisoned swap entry for errored virtual address.
2838                  * So we need to block hugepage fault by PG_hwpoison bit check.
2839                  */
2840                 if (unlikely(PageHWPoison(page))) {
2841                         ret = VM_FAULT_HWPOISON |
2842                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2843                         goto backout_unlocked;
2844                 }
2845         }
2846
2847         /*
2848          * If we are going to COW a private mapping later, we examine the
2849          * pending reservations for this page now. This will ensure that
2850          * any allocations necessary to record that reservation occur outside
2851          * the spinlock.
2852          */
2853         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2854                 if (vma_needs_reservation(h, vma, address) < 0) {
2855                         ret = VM_FAULT_OOM;
2856                         goto backout_unlocked;
2857                 }
2858
2859         ptl = huge_pte_lockptr(h, mm, ptep);
2860         spin_lock(ptl);
2861         size = i_size_read(mapping->host) >> huge_page_shift(h);
2862         if (idx >= size)
2863                 goto backout;
2864
2865         ret = 0;
2866         if (!huge_pte_none(huge_ptep_get(ptep)))
2867                 goto backout;
2868
2869         if (anon_rmap) {
2870                 ClearPagePrivate(page);
2871                 hugepage_add_new_anon_rmap(page, vma, address);
2872         }
2873         else
2874                 page_dup_rmap(page);
2875         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2876                                 && (vma->vm_flags & VM_SHARED)));
2877         set_huge_pte_at(mm, address, ptep, new_pte);
2878
2879         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2880                 /* Optimization, do the COW without a second fault */
2881                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
2882         }
2883
2884         spin_unlock(ptl);
2885         unlock_page(page);
2886 out:
2887         return ret;
2888
2889 backout:
2890         spin_unlock(ptl);
2891 backout_unlocked:
2892         unlock_page(page);
2893         put_page(page);
2894         goto out;
2895 }
2896
2897 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2898                         unsigned long address, unsigned int flags)
2899 {
2900         pte_t *ptep;
2901         pte_t entry;
2902         spinlock_t *ptl;
2903         int ret;
2904         struct page *page = NULL;
2905         struct page *pagecache_page = NULL;
2906         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2907         struct hstate *h = hstate_vma(vma);
2908
2909         address &= huge_page_mask(h);
2910
2911         ptep = huge_pte_offset(mm, address);
2912         if (ptep) {
2913                 entry = huge_ptep_get(ptep);
2914                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2915                         migration_entry_wait_huge(vma, mm, ptep);
2916                         return 0;
2917                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2918                         return VM_FAULT_HWPOISON_LARGE |
2919                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2920         }
2921
2922         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2923         if (!ptep)
2924                 return VM_FAULT_OOM;
2925
2926         /*
2927          * Serialize hugepage allocation and instantiation, so that we don't
2928          * get spurious allocation failures if two CPUs race to instantiate
2929          * the same page in the page cache.
2930          */
2931         mutex_lock(&hugetlb_instantiation_mutex);
2932         entry = huge_ptep_get(ptep);
2933         if (huge_pte_none(entry)) {
2934                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2935                 goto out_mutex;
2936         }
2937
2938         ret = 0;
2939
2940         /*
2941          * If we are going to COW the mapping later, we examine the pending
2942          * reservations for this page now. This will ensure that any
2943          * allocations necessary to record that reservation occur outside the
2944          * spinlock. For private mappings, we also lookup the pagecache
2945          * page now as it is used to determine if a reservation has been
2946          * consumed.
2947          */
2948         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2949                 if (vma_needs_reservation(h, vma, address) < 0) {
2950                         ret = VM_FAULT_OOM;
2951                         goto out_mutex;
2952                 }
2953
2954                 if (!(vma->vm_flags & VM_MAYSHARE))
2955                         pagecache_page = hugetlbfs_pagecache_page(h,
2956                                                                 vma, address);
2957         }
2958
2959         /*
2960          * hugetlb_cow() requires page locks of pte_page(entry) and
2961          * pagecache_page, so here we need take the former one
2962          * when page != pagecache_page or !pagecache_page.
2963          * Note that locking order is always pagecache_page -> page,
2964          * so no worry about deadlock.
2965          */
2966         page = pte_page(entry);
2967         get_page(page);
2968         if (page != pagecache_page)
2969                 lock_page(page);
2970
2971         ptl = huge_pte_lockptr(h, mm, ptep);
2972         spin_lock(ptl);
2973         /* Check for a racing update before calling hugetlb_cow */
2974         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2975                 goto out_ptl;
2976
2977
2978         if (flags & FAULT_FLAG_WRITE) {
2979                 if (!huge_pte_write(entry)) {
2980                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2981                                         pagecache_page, ptl);
2982                         goto out_ptl;
2983                 }
2984                 entry = huge_pte_mkdirty(entry);
2985         }
2986         entry = pte_mkyoung(entry);
2987         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2988                                                 flags & FAULT_FLAG_WRITE))
2989                 update_mmu_cache(vma, address, ptep);
2990
2991 out_ptl:
2992         spin_unlock(ptl);
2993
2994         if (pagecache_page) {
2995                 unlock_page(pagecache_page);
2996                 put_page(pagecache_page);
2997         }
2998         if (page != pagecache_page)
2999                 unlock_page(page);
3000         put_page(page);
3001
3002 out_mutex:
3003         mutex_unlock(&hugetlb_instantiation_mutex);
3004
3005         return ret;
3006 }
3007
3008 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3009                          struct page **pages, struct vm_area_struct **vmas,
3010                          unsigned long *position, unsigned long *nr_pages,
3011                          long i, unsigned int flags)
3012 {
3013         unsigned long pfn_offset;
3014         unsigned long vaddr = *position;
3015         unsigned long remainder = *nr_pages;
3016         struct hstate *h = hstate_vma(vma);
3017
3018         while (vaddr < vma->vm_end && remainder) {
3019                 pte_t *pte;
3020                 spinlock_t *ptl = NULL;
3021                 int absent;
3022                 struct page *page;
3023
3024                 /*
3025                  * Some archs (sparc64, sh*) have multiple pte_ts to
3026                  * each hugepage.  We have to make sure we get the
3027                  * first, for the page indexing below to work.
3028                  *
3029                  * Note that page table lock is not held when pte is null.
3030                  */
3031                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3032                 if (pte)
3033                         ptl = huge_pte_lock(h, mm, pte);
3034                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3035
3036                 /*
3037                  * When coredumping, it suits get_dump_page if we just return
3038                  * an error where there's an empty slot with no huge pagecache
3039                  * to back it.  This way, we avoid allocating a hugepage, and
3040                  * the sparse dumpfile avoids allocating disk blocks, but its
3041                  * huge holes still show up with zeroes where they need to be.
3042                  */
3043                 if (absent && (flags & FOLL_DUMP) &&
3044                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3045                         if (pte)
3046                                 spin_unlock(ptl);
3047                         remainder = 0;
3048                         break;
3049                 }
3050
3051                 /*
3052                  * We need call hugetlb_fault for both hugepages under migration
3053                  * (in which case hugetlb_fault waits for the migration,) and
3054                  * hwpoisoned hugepages (in which case we need to prevent the
3055                  * caller from accessing to them.) In order to do this, we use
3056                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3057                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3058                  * both cases, and because we can't follow correct pages
3059                  * directly from any kind of swap entries.
3060                  */
3061                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3062                     ((flags & FOLL_WRITE) &&
3063                       !huge_pte_write(huge_ptep_get(pte)))) {
3064                         int ret;
3065
3066                         if (pte)
3067                                 spin_unlock(ptl);
3068                         ret = hugetlb_fault(mm, vma, vaddr,
3069                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3070                         if (!(ret & VM_FAULT_ERROR))
3071                                 continue;
3072
3073                         remainder = 0;
3074                         break;
3075                 }
3076
3077                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3078                 page = pte_page(huge_ptep_get(pte));
3079 same_page:
3080                 if (pages) {
3081                         pages[i] = mem_map_offset(page, pfn_offset);
3082                         get_page(pages[i]);
3083                 }
3084
3085                 if (vmas)
3086                         vmas[i] = vma;
3087
3088                 vaddr += PAGE_SIZE;
3089                 ++pfn_offset;
3090                 --remainder;
3091                 ++i;
3092                 if (vaddr < vma->vm_end && remainder &&
3093                                 pfn_offset < pages_per_huge_page(h)) {
3094                         /*
3095                          * We use pfn_offset to avoid touching the pageframes
3096                          * of this compound page.
3097                          */
3098                         goto same_page;
3099                 }
3100                 spin_unlock(ptl);
3101         }
3102         *nr_pages = remainder;
3103         *position = vaddr;
3104
3105         return i ? i : -EFAULT;
3106 }
3107
3108 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3109                 unsigned long address, unsigned long end, pgprot_t newprot)
3110 {
3111         struct mm_struct *mm = vma->vm_mm;
3112         unsigned long start = address;
3113         pte_t *ptep;
3114         pte_t pte;
3115         struct hstate *h = hstate_vma(vma);
3116         unsigned long pages = 0;
3117
3118         BUG_ON(address >= end);
3119         flush_cache_range(vma, address, end);
3120
3121         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3122         for (; address < end; address += huge_page_size(h)) {
3123                 spinlock_t *ptl;
3124                 ptep = huge_pte_offset(mm, address);
3125                 if (!ptep)
3126                         continue;
3127                 ptl = huge_pte_lock(h, mm, ptep);
3128                 if (huge_pmd_unshare(mm, &address, ptep)) {
3129                         pages++;
3130                         spin_unlock(ptl);
3131                         continue;
3132                 }
3133                 if (!huge_pte_none(huge_ptep_get(ptep))) {
3134                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3135                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3136                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3137                         set_huge_pte_at(mm, address, ptep, pte);
3138                         pages++;
3139                 }
3140                 spin_unlock(ptl);
3141         }
3142         /*
3143          * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3144          * may have cleared our pud entry and done put_page on the page table:
3145          * once we release i_mmap_mutex, another task can do the final put_page
3146          * and that page table be reused and filled with junk.
3147          */
3148         flush_tlb_range(vma, start, end);
3149         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3150
3151         return pages << h->order;
3152 }
3153
3154 int hugetlb_reserve_pages(struct inode *inode,
3155                                         long from, long to,
3156                                         struct vm_area_struct *vma,
3157                                         vm_flags_t vm_flags)
3158 {
3159         long ret, chg;
3160         struct hstate *h = hstate_inode(inode);
3161         struct hugepage_subpool *spool = subpool_inode(inode);
3162
3163         /*
3164          * Only apply hugepage reservation if asked. At fault time, an
3165          * attempt will be made for VM_NORESERVE to allocate a page
3166          * without using reserves
3167          */
3168         if (vm_flags & VM_NORESERVE)
3169                 return 0;
3170
3171         /*
3172          * Shared mappings base their reservation on the number of pages that
3173          * are already allocated on behalf of the file. Private mappings need
3174          * to reserve the full area even if read-only as mprotect() may be
3175          * called to make the mapping read-write. Assume !vma is a shm mapping
3176          */
3177         if (!vma || vma->vm_flags & VM_MAYSHARE)
3178                 chg = region_chg(&inode->i_mapping->private_list, from, to);
3179         else {
3180                 struct resv_map *resv_map = resv_map_alloc();
3181                 if (!resv_map)
3182                         return -ENOMEM;
3183
3184                 chg = to - from;
3185
3186                 set_vma_resv_map(vma, resv_map);
3187                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3188         }
3189
3190         if (chg < 0) {
3191                 ret = chg;
3192                 goto out_err;
3193         }
3194
3195         /* There must be enough pages in the subpool for the mapping */
3196         if (hugepage_subpool_get_pages(spool, chg)) {
3197                 ret = -ENOSPC;
3198                 goto out_err;
3199         }
3200
3201         /*
3202          * Check enough hugepages are available for the reservation.
3203          * Hand the pages back to the subpool if there are not
3204          */
3205         ret = hugetlb_acct_memory(h, chg);
3206         if (ret < 0) {
3207                 hugepage_subpool_put_pages(spool, chg);
3208                 goto out_err;
3209         }
3210
3211         /*
3212          * Account for the reservations made. Shared mappings record regions
3213          * that have reservations as they are shared by multiple VMAs.
3214          * When the last VMA disappears, the region map says how much
3215          * the reservation was and the page cache tells how much of
3216          * the reservation was consumed. Private mappings are per-VMA and
3217          * only the consumed reservations are tracked. When the VMA
3218          * disappears, the original reservation is the VMA size and the
3219          * consumed reservations are stored in the map. Hence, nothing
3220          * else has to be done for private mappings here
3221          */
3222         if (!vma || vma->vm_flags & VM_MAYSHARE)
3223                 region_add(&inode->i_mapping->private_list, from, to);
3224         return 0;
3225 out_err:
3226         if (vma)
3227                 resv_map_put(vma);
3228         return ret;
3229 }
3230
3231 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3232 {
3233         struct hstate *h = hstate_inode(inode);
3234         long chg = region_truncate(&inode->i_mapping->private_list, offset);
3235         struct hugepage_subpool *spool = subpool_inode(inode);
3236
3237         spin_lock(&inode->i_lock);
3238         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3239         spin_unlock(&inode->i_lock);
3240
3241         hugepage_subpool_put_pages(spool, (chg - freed));
3242         hugetlb_acct_memory(h, -(chg - freed));
3243 }
3244
3245 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3246 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3247                                 struct vm_area_struct *vma,
3248                                 unsigned long addr, pgoff_t idx)
3249 {
3250         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3251                                 svma->vm_start;
3252         unsigned long sbase = saddr & PUD_MASK;
3253         unsigned long s_end = sbase + PUD_SIZE;
3254
3255         /* Allow segments to share if only one is marked locked */
3256         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3257         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3258
3259         /*
3260          * match the virtual addresses, permission and the alignment of the
3261          * page table page.
3262          */
3263         if (pmd_index(addr) != pmd_index(saddr) ||
3264             vm_flags != svm_flags ||
3265             sbase < svma->vm_start || svma->vm_end < s_end)
3266                 return 0;
3267
3268         return saddr;
3269 }
3270
3271 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3272 {
3273         unsigned long base = addr & PUD_MASK;
3274         unsigned long end = base + PUD_SIZE;
3275
3276         /*
3277          * check on proper vm_flags and page table alignment
3278          */
3279         if (vma->vm_flags & VM_MAYSHARE &&
3280             vma->vm_start <= base && end <= vma->vm_end)
3281                 return 1;
3282         return 0;
3283 }
3284
3285 /*
3286  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3287  * and returns the corresponding pte. While this is not necessary for the
3288  * !shared pmd case because we can allocate the pmd later as well, it makes the
3289  * code much cleaner. pmd allocation is essential for the shared case because
3290  * pud has to be populated inside the same i_mmap_mutex section - otherwise
3291  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3292  * bad pmd for sharing.
3293  */
3294 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3295 {
3296         struct vm_area_struct *vma = find_vma(mm, addr);
3297         struct address_space *mapping = vma->vm_file->f_mapping;
3298         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3299                         vma->vm_pgoff;
3300         struct vm_area_struct *svma;
3301         unsigned long saddr;
3302         pte_t *spte = NULL;
3303         pte_t *pte;
3304         spinlock_t *ptl;
3305
3306         if (!vma_shareable(vma, addr))
3307                 return (pte_t *)pmd_alloc(mm, pud, addr);
3308
3309         mutex_lock(&mapping->i_mmap_mutex);
3310         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3311                 if (svma == vma)
3312                         continue;
3313
3314                 saddr = page_table_shareable(svma, vma, addr, idx);
3315                 if (saddr) {
3316                         spte = huge_pte_offset(svma->vm_mm, saddr);
3317                         if (spte) {
3318                                 get_page(virt_to_page(spte));
3319                                 break;
3320                         }
3321                 }
3322         }
3323
3324         if (!spte)
3325                 goto out;
3326
3327         ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3328         spin_lock(ptl);
3329         if (pud_none(*pud))
3330                 pud_populate(mm, pud,
3331                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3332         else
3333                 put_page(virt_to_page(spte));
3334         spin_unlock(ptl);
3335 out:
3336         pte = (pte_t *)pmd_alloc(mm, pud, addr);
3337         mutex_unlock(&mapping->i_mmap_mutex);
3338         return pte;
3339 }
3340
3341 /*
3342  * unmap huge page backed by shared pte.
3343  *
3344  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3345  * indicated by page_count > 1, unmap is achieved by clearing pud and
3346  * decrementing the ref count. If count == 1, the pte page is not shared.
3347  *
3348  * called with page table lock held.
3349  *
3350  * returns: 1 successfully unmapped a shared pte page
3351  *          0 the underlying pte page is not shared, or it is the last user
3352  */
3353 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3354 {
3355         pgd_t *pgd = pgd_offset(mm, *addr);
3356         pud_t *pud = pud_offset(pgd, *addr);
3357
3358         BUG_ON(page_count(virt_to_page(ptep)) == 0);
3359         if (page_count(virt_to_page(ptep)) == 1)
3360                 return 0;
3361
3362         pud_clear(pud);
3363         put_page(virt_to_page(ptep));
3364         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3365         return 1;
3366 }
3367 #define want_pmd_share()        (1)
3368 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3369 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3370 {
3371         return NULL;
3372 }
3373 #define want_pmd_share()        (0)
3374 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3375
3376 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3377 pte_t *huge_pte_alloc(struct mm_struct *mm,
3378                         unsigned long addr, unsigned long sz)
3379 {
3380         pgd_t *pgd;
3381         pud_t *pud;
3382         pte_t *pte = NULL;
3383
3384         pgd = pgd_offset(mm, addr);
3385         pud = pud_alloc(mm, pgd, addr);
3386         if (pud) {
3387                 if (sz == PUD_SIZE) {
3388                         pte = (pte_t *)pud;
3389                 } else {
3390                         BUG_ON(sz != PMD_SIZE);
3391                         if (want_pmd_share() && pud_none(*pud))
3392                                 pte = huge_pmd_share(mm, addr, pud);
3393                         else
3394                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3395                 }
3396         }
3397         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3398
3399         return pte;
3400 }
3401
3402 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3403 {
3404         pgd_t *pgd;
3405         pud_t *pud;
3406         pmd_t *pmd = NULL;
3407
3408         pgd = pgd_offset(mm, addr);
3409         if (pgd_present(*pgd)) {
3410                 pud = pud_offset(pgd, addr);
3411                 if (pud_present(*pud)) {
3412                         if (pud_huge(*pud))
3413                                 return (pte_t *)pud;
3414                         pmd = pmd_offset(pud, addr);
3415                 }
3416         }
3417         return (pte_t *) pmd;
3418 }
3419
3420 struct page *
3421 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3422                 pmd_t *pmd, int write)
3423 {
3424         struct page *page;
3425
3426         page = pte_page(*(pte_t *)pmd);
3427         if (page)
3428                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3429         return page;
3430 }
3431
3432 struct page *
3433 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3434                 pud_t *pud, int write)
3435 {
3436         struct page *page;
3437
3438         page = pte_page(*(pte_t *)pud);
3439         if (page)
3440                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3441         return page;
3442 }
3443
3444 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3445
3446 /* Can be overriden by architectures */
3447 __attribute__((weak)) struct page *
3448 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3449                pud_t *pud, int write)
3450 {
3451         BUG();
3452         return NULL;
3453 }
3454
3455 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3456
3457 #ifdef CONFIG_MEMORY_FAILURE
3458
3459 /* Should be called in hugetlb_lock */
3460 static int is_hugepage_on_freelist(struct page *hpage)
3461 {
3462         struct page *page;
3463         struct page *tmp;
3464         struct hstate *h = page_hstate(hpage);
3465         int nid = page_to_nid(hpage);
3466
3467         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3468                 if (page == hpage)
3469                         return 1;
3470         return 0;
3471 }
3472
3473 /*
3474  * This function is called from memory failure code.
3475  * Assume the caller holds page lock of the head page.
3476  */
3477 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3478 {
3479         struct hstate *h = page_hstate(hpage);
3480         int nid = page_to_nid(hpage);
3481         int ret = -EBUSY;
3482
3483         spin_lock(&hugetlb_lock);
3484         if (is_hugepage_on_freelist(hpage)) {
3485                 /*
3486                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
3487                  * but dangling hpage->lru can trigger list-debug warnings
3488                  * (this happens when we call unpoison_memory() on it),
3489                  * so let it point to itself with list_del_init().
3490                  */
3491                 list_del_init(&hpage->lru);
3492                 set_page_refcounted(hpage);
3493                 h->free_huge_pages--;
3494                 h->free_huge_pages_node[nid]--;
3495                 ret = 0;
3496         }
3497         spin_unlock(&hugetlb_lock);
3498         return ret;
3499 }
3500 #endif
3501
3502 bool isolate_huge_page(struct page *page, struct list_head *list)
3503 {
3504         VM_BUG_ON(!PageHead(page));
3505         if (!get_page_unless_zero(page))
3506                 return false;
3507         spin_lock(&hugetlb_lock);
3508         list_move_tail(&page->lru, list);
3509         spin_unlock(&hugetlb_lock);
3510         return true;
3511 }
3512
3513 void putback_active_hugepage(struct page *page)
3514 {
3515         VM_BUG_ON(!PageHead(page));
3516         spin_lock(&hugetlb_lock);
3517         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3518         spin_unlock(&hugetlb_lock);
3519         put_page(page);
3520 }
3521
3522 bool is_hugepage_active(struct page *page)
3523 {
3524         VM_BUG_ON(!PageHuge(page));
3525         /*
3526          * This function can be called for a tail page because the caller,
3527          * scan_movable_pages, scans through a given pfn-range which typically
3528          * covers one memory block. In systems using gigantic hugepage (1GB
3529          * for x86_64,) a hugepage is larger than a memory block, and we don't
3530          * support migrating such large hugepages for now, so return false
3531          * when called for tail pages.
3532          */
3533         if (PageTail(page))
3534                 return false;
3535         /*
3536          * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3537          * so we should return false for them.
3538          */
3539         if (unlikely(PageHWPoison(page)))
3540                 return false;
3541         return page_count(page) > 0;
3542 }