mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/mman.h>
  11 #include <linux/slab.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/swap.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/namei.h>
  17 #include <linux/shm.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/random.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/mutex.h>
  29 #include <linux/capability.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/memcontrol.h>
  32
  33 #include <asm/pgtable.h>
  34 #include <asm/tlbflush.h>
  35 #include <linux/swapops.h>
  36 #include <linux/page_cgroup.h>
  37
  38 static DEFINE_SPINLOCK(swap_lock);
  39 static unsigned int nr_swapfiles;
  40 long nr_swap_pages;
  41 long total_swap_pages;
  42 static int swap_overflow;
  43 static int least_priority;
  44
  45 static const char Bad_file[] = "Bad swap file entry ";
  46 static const char Unused_file[] = "Unused swap file entry ";
  47 static const char Bad_offset[] = "Bad swap offset entry ";
  48 static const char Unused_offset[] = "Unused swap offset entry ";
  49
  50 static struct swap_list_t swap_list = {-1, -1};
  51
  52 static struct swap_info_struct swap_info[MAX_SWAPFILES];
  53
  54 static DEFINE_MUTEX(swapon_mutex);
  55
  56 /*
  57  * We need this because the bdev->unplug_fn can sleep and we cannot
  58  * hold swap_lock while calling the unplug_fn. And swap_lock
  59  * cannot be turned into a mutex.
  60  */
  61 static DECLARE_RWSEM(swap_unplug_sem);
  62
  63 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  64 {
  65         swp_entry_t entry;
  66
  67         down_read(&swap_unplug_sem);
  68         entry.val = page_private(page);
  69         if (PageSwapCache(page)) {
  70                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  71                 struct backing_dev_info *bdi;
  72
  73                 /*
  74                  * If the page is removed from swapcache from under us (with a
  75                  * racy try_to_unuse/swapoff) we need an additional reference
  76                  * count to avoid reading garbage from page_private(page) above.
  77                  * If the WARN_ON triggers during a swapoff it maybe the race
  78                  * condition and it's harmless. However if it triggers without
  79                  * swapoff it signals a problem.
  80                  */
  81                 WARN_ON(page_count(page) <= 1);
  82
  83                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  84                 blk_run_backing_dev(bdi, page);
  85         }
  86         up_read(&swap_unplug_sem);
  87 }
  88
  89 /*
  90  * swapon tell device that all the old swap contents can be discarded,
  91  * to allow the swap device to optimize its wear-levelling.
  92  */
  93 static int discard_swap(struct swap_info_struct *si)
  94 {
  95         struct swap_extent *se;
  96         int err = 0;
  97
  98         list_for_each_entry(se, &si->extent_list, list) {
  99                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
 100                 sector_t nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 101
 102                 if (se->start_page == 0) {
 103                         /* Do not discard the swap header page! */
 104                         start_block += 1 << (PAGE_SHIFT - 9);
 105                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
 106                         if (!nr_blocks)
 107                                 continue;
 108                 }
 109
 110                 err = blkdev_issue_discard(si->bdev, start_block,
 111                                                 nr_blocks, GFP_KERNEL);
 112                 if (err)
 113                         break;
 114
 115                 cond_resched();
 116         }
 117         return err;             /* That will often be -EOPNOTSUPP */
 118 }
 119
 120 /*
 121  * swap allocation tell device that a cluster of swap can now be discarded,
 122  * to allow the swap device to optimize its wear-levelling.
 123  */
 124 static void discard_swap_cluster(struct swap_info_struct *si,
 125                                  pgoff_t start_page, pgoff_t nr_pages)
 126 {
 127         struct swap_extent *se = si->curr_swap_extent;
 128         int found_extent = 0;
 129
 130         while (nr_pages) {
 131                 struct list_head *lh;
 132
 133                 if (se->start_page <= start_page &&
 134                     start_page < se->start_page + se->nr_pages) {
 135                         pgoff_t offset = start_page - se->start_page;
 136                         sector_t start_block = se->start_block + offset;
 137                         sector_t nr_blocks = se->nr_pages - offset;
 138
 139                         if (nr_blocks > nr_pages)
 140                                 nr_blocks = nr_pages;
 141                         start_page += nr_blocks;
 142                         nr_pages -= nr_blocks;
 143
 144                         if (!found_extent++)
 145                                 si->curr_swap_extent = se;
 146
 147                         start_block <<= PAGE_SHIFT - 9;
 148                         nr_blocks <<= PAGE_SHIFT - 9;
 149                         if (blkdev_issue_discard(si->bdev, start_block,
 150                                                         nr_blocks, GFP_NOIO))
 151                                 break;
 152                 }
 153
 154                 lh = se->list.next;
 155                 if (lh == &si->extent_list)
 156                         lh = lh->next;
 157                 se = list_entry(lh, struct swap_extent, list);
 158         }
 159 }
 160
 161 static int wait_for_discard(void *word)
 162 {
 163         schedule();
 164         return 0;
 165 }
 166
 167 #define SWAPFILE_CLUSTER        256
 168 #define LATENCY_LIMIT           256
 169
 170 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
 171 {
 172         unsigned long offset;
 173         unsigned long scan_base;
 174         unsigned long last_in_cluster = 0;
 175         int latency_ration = LATENCY_LIMIT;
 176         int found_free_cluster = 0;
 177
 178         /*
 179          * We try to cluster swap pages by allocating them sequentially
 180          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 181          * way, however, we resort to first-free allocation, starting
 182          * a new cluster.  This prevents us from scattering swap pages
 183          * all over the entire swap partition, so that we reduce
 184          * overall disk seek times between swap pages.  -- sct
 185          * But we do now try to find an empty cluster.  -Andrea
 186          * And we let swap pages go all over an SSD partition.  Hugh
 187          */
 188
 189         si->flags += SWP_SCANNING;
 190         scan_base = offset = si->cluster_next;
 191
 192         if (unlikely(!si->cluster_nr--)) {
 193                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 194                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
 195                         goto checks;
 196                 }
 197                 if (si->flags & SWP_DISCARDABLE) {
 198                         /*
 199                          * Start range check on racing allocations, in case
 200                          * they overlap the cluster we eventually decide on
 201                          * (we scan without swap_lock to allow preemption).
 202                          * It's hardly conceivable that cluster_nr could be
 203                          * wrapped during our scan, but don't depend on it.
 204                          */
 205                         if (si->lowest_alloc)
 206                                 goto checks;
 207                         si->lowest_alloc = si->max;
 208                         si->highest_alloc = 0;
 209                 }
 210                 spin_unlock(&swap_lock);
 211
 212                 /*
 213                  * If seek is expensive, start searching for new cluster from
 214                  * start of partition, to minimize the span of allocated swap.
 215                  * But if seek is cheap, search from our current position, so
 216                  * that swap is allocated from all over the partition: if the
 217                  * Flash Translation Layer only remaps within limited zones,
 218                  * we don't want to wear out the first zone too quickly.
 219                  */
 220                 if (!(si->flags & SWP_SOLIDSTATE))
 221                         scan_base = offset = si->lowest_bit;
 222                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 223
 224                 /* Locate the first empty (unaligned) cluster */
 225                 for (; last_in_cluster <= si->highest_bit; offset++) {
 226                         if (si->swap_map[offset])
 227                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 228                         else if (offset == last_in_cluster) {
 229                                 spin_lock(&swap_lock);
 230                                 offset -= SWAPFILE_CLUSTER - 1;
 231                                 si->cluster_next = offset;
 232                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 233                                 found_free_cluster = 1;
 234                                 goto checks;
 235                         }
 236                         if (unlikely(--latency_ration < 0)) {
 237                                 cond_resched();
 238                                 latency_ration = LATENCY_LIMIT;
 239                         }
 240                 }
 241
 242                 offset = si->lowest_bit;
 243                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 244
 245                 /* Locate the first empty (unaligned) cluster */
 246                 for (; last_in_cluster < scan_base; offset++) {
 247                         if (si->swap_map[offset])
 248                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 249                         else if (offset == last_in_cluster) {
 250                                 spin_lock(&swap_lock);
 251                                 offset -= SWAPFILE_CLUSTER - 1;
 252                                 si->cluster_next = offset;
 253                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 254                                 found_free_cluster = 1;
 255                                 goto checks;
 256                         }
 257                         if (unlikely(--latency_ration < 0)) {
 258                                 cond_resched();
 259                                 latency_ration = LATENCY_LIMIT;
 260                         }
 261                 }
 262
 263                 offset = scan_base;
 264                 spin_lock(&swap_lock);
 265                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 266                 si->lowest_alloc = 0;
 267         }
 268
 269 checks:
 270         if (!(si->flags & SWP_WRITEOK))
 271                 goto no_page;
 272         if (!si->highest_bit)
 273                 goto no_page;
 274         if (offset > si->highest_bit)
 275                 scan_base = offset = si->lowest_bit;
 276         if (si->swap_map[offset])
 277                 goto scan;
 278
 279         if (offset == si->lowest_bit)
 280                 si->lowest_bit++;
 281         if (offset == si->highest_bit)
 282                 si->highest_bit--;
 283         si->inuse_pages++;
 284         if (si->inuse_pages == si->pages) {
 285                 si->lowest_bit = si->max;
 286                 si->highest_bit = 0;
 287         }
 288         si->swap_map[offset] = 1;
 289         si->cluster_next = offset + 1;
 290         si->flags -= SWP_SCANNING;
 291
 292         if (si->lowest_alloc) {
 293                 /*
 294                  * Only set when SWP_DISCARDABLE, and there's a scan
 295                  * for a free cluster in progress or just completed.
 296                  */
 297                 if (found_free_cluster) {
 298                         /*
 299                          * To optimize wear-levelling, discard the
 300                          * old data of the cluster, taking care not to
 301                          * discard any of its pages that have already
 302                          * been allocated by racing tasks (offset has
 303                          * already stepped over any at the beginning).
 304                          */
 305                         if (offset < si->highest_alloc &&
 306                             si->lowest_alloc <= last_in_cluster)
 307                                 last_in_cluster = si->lowest_alloc - 1;
 308                         si->flags |= SWP_DISCARDING;
 309                         spin_unlock(&swap_lock);
 310
 311                         if (offset < last_in_cluster)
 312                                 discard_swap_cluster(si, offset,
 313                                         last_in_cluster - offset + 1);
 314
 315                         spin_lock(&swap_lock);
 316                         si->lowest_alloc = 0;
 317                         si->flags &= ~SWP_DISCARDING;
 318
 319                         smp_mb();       /* wake_up_bit advises this */
 320                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
 321
 322                 } else if (si->flags & SWP_DISCARDING) {
 323                         /*
 324                          * Delay using pages allocated by racing tasks
 325                          * until the whole discard has been issued. We
 326                          * could defer that delay until swap_writepage,
 327                          * but it's easier to keep this self-contained.
 328                          */
 329                         spin_unlock(&swap_lock);
 330                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
 331                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
 332                         spin_lock(&swap_lock);
 333                 } else {
 334                         /*
 335                          * Note pages allocated by racing tasks while
 336                          * scan for a free cluster is in progress, so
 337                          * that its final discard can exclude them.
 338                          */
 339                         if (offset < si->lowest_alloc)
 340                                 si->lowest_alloc = offset;
 341                         if (offset > si->highest_alloc)
 342                                 si->highest_alloc = offset;
 343                 }
 344         }
 345         return offset;
 346
 347 scan:
 348         spin_unlock(&swap_lock);
 349         while (++offset <= si->highest_bit) {
 350                 if (!si->swap_map[offset]) {
 351                         spin_lock(&swap_lock);
 352                         goto checks;
 353                 }
 354                 if (unlikely(--latency_ration < 0)) {
 355                         cond_resched();
 356                         latency_ration = LATENCY_LIMIT;
 357                 }
 358         }
 359         offset = si->lowest_bit;
 360         while (++offset < scan_base) {
 361                 if (!si->swap_map[offset]) {
 362                         spin_lock(&swap_lock);
 363                         goto checks;
 364                 }
 365                 if (unlikely(--latency_ration < 0)) {
 366                         cond_resched();
 367                         latency_ration = LATENCY_LIMIT;
 368                 }
 369         }
 370         spin_lock(&swap_lock);
 371
 372 no_page:
 373         si->flags -= SWP_SCANNING;
 374         return 0;
 375 }
 376
 377 swp_entry_t get_swap_page(void)
 378 {
 379         struct swap_info_struct *si;
 380         pgoff_t offset;
 381         int type, next;
 382         int wrapped = 0;
 383
 384         spin_lock(&swap_lock);
 385         if (nr_swap_pages <= 0)
 386                 goto noswap;
 387         nr_swap_pages--;
 388
 389         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 390                 si = swap_info + type;
 391                 next = si->next;
 392                 if (next < 0 ||
 393                     (!wrapped && si->prio != swap_info[next].prio)) {
 394                         next = swap_list.head;
 395                         wrapped++;
 396                 }
 397
 398                 if (!si->highest_bit)
 399                         continue;
 400                 if (!(si->flags & SWP_WRITEOK))
 401                         continue;
 402
 403                 swap_list.next = next;
 404                 offset = scan_swap_map(si);
 405                 if (offset) {
 406                         spin_unlock(&swap_lock);
 407                         return swp_entry(type, offset);
 408                 }
 409                 next = swap_list.next;
 410         }
 411
 412         nr_swap_pages++;
 413 noswap:
 414         spin_unlock(&swap_lock);
 415         return (swp_entry_t) {0};
 416 }
 417
 418 swp_entry_t get_swap_page_of_type(int type)
 419 {
 420         struct swap_info_struct *si;
 421         pgoff_t offset;
 422
 423         spin_lock(&swap_lock);
 424         si = swap_info + type;
 425         if (si->flags & SWP_WRITEOK) {
 426                 nr_swap_pages--;
 427                 offset = scan_swap_map(si);
 428                 if (offset) {
 429                         spin_unlock(&swap_lock);
 430                         return swp_entry(type, offset);
 431                 }
 432                 nr_swap_pages++;
 433         }
 434         spin_unlock(&swap_lock);
 435         return (swp_entry_t) {0};
 436 }
 437
 438 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 439 {
 440         struct swap_info_struct * p;
 441         unsigned long offset, type;
 442
 443         if (!entry.val)
 444                 goto out;
 445         type = swp_type(entry);
 446         if (type >= nr_swapfiles)
 447                 goto bad_nofile;
 448         p = & swap_info[type];
 449         if (!(p->flags & SWP_USED))
 450                 goto bad_device;
 451         offset = swp_offset(entry);
 452         if (offset >= p->max)
 453                 goto bad_offset;
 454         if (!p->swap_map[offset])
 455                 goto bad_free;
 456         spin_lock(&swap_lock);
 457         return p;
 458
 459 bad_free:
 460         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 461         goto out;
 462 bad_offset:
 463         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 464         goto out;
 465 bad_device:
 466         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 467         goto out;
 468 bad_nofile:
 469         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 470 out:
 471         return NULL;
 472 }
 473
 474 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 475 {
 476         int count = p->swap_map[offset];
 477
 478         if (count < SWAP_MAP_MAX) {
 479                 count--;
 480                 p->swap_map[offset] = count;
 481                 if (!count) {
 482                         if (offset < p->lowest_bit)
 483                                 p->lowest_bit = offset;
 484                         if (offset > p->highest_bit)
 485                                 p->highest_bit = offset;
 486                         if (p->prio > swap_info[swap_list.next].prio)
 487                                 swap_list.next = p - swap_info;
 488                         nr_swap_pages++;
 489                         p->inuse_pages--;
 490                 }
 491         }
 492         return count;
 493 }
 494
 495 /*
 496  * Caller has made sure that the swapdevice corresponding to entry
 497  * is still around or has not been recycled.
 498  */
 499 void swap_free(swp_entry_t entry)
 500 {
 501         struct swap_info_struct * p;
 502
 503         p = swap_info_get(entry);
 504         if (p) {
 505                 swap_entry_free(p, swp_offset(entry));
 506                 spin_unlock(&swap_lock);
 507         }
 508 }
 509
 510 /*
 511  * How many references to page are currently swapped out?
 512  */
 513 static inline int page_swapcount(struct page *page)
 514 {
 515         int count = 0;
 516         struct swap_info_struct *p;
 517         swp_entry_t entry;
 518
 519         entry.val = page_private(page);
 520         p = swap_info_get(entry);
 521         if (p) {
 522                 /* Subtract the 1 for the swap cache itself */
 523                 count = p->swap_map[swp_offset(entry)] - 1;
 524                 spin_unlock(&swap_lock);
 525         }
 526         return count;
 527 }
 528
 529 /*
 530  * We can write to an anon page without COW if there are no other references
 531  * to it.  And as a side-effect, free up its swap: because the old content
 532  * on disk will never be read, and seeking back there to write new content
 533  * later would only waste time away from clustering.
 534  */
 535 int reuse_swap_page(struct page *page)
 536 {
 537         int count;
 538
 539         VM_BUG_ON(!PageLocked(page));
 540         count = page_mapcount(page);
 541         if (count <= 1 && PageSwapCache(page)) {
 542                 count += page_swapcount(page);
 543                 if (count == 1 && !PageWriteback(page)) {
 544                         delete_from_swap_cache(page);
 545                         SetPageDirty(page);
 546                 }
 547         }
 548         return count == 1;
 549 }
 550
 551 /*
 552  * If swap is getting full, or if there are no more mappings of this page,
 553  * then try_to_free_swap is called to free its swap space.
 554  */
 555 int try_to_free_swap(struct page *page)
 556 {
 557         VM_BUG_ON(!PageLocked(page));
 558
 559         if (!PageSwapCache(page))
 560                 return 0;
 561         if (PageWriteback(page))
 562                 return 0;
 563         if (page_swapcount(page))
 564                 return 0;
 565
 566         delete_from_swap_cache(page);
 567         SetPageDirty(page);
 568         return 1;
 569 }
 570
 571 /*
 572  * Free the swap entry like above, but also try to
 573  * free the page cache entry if it is the last user.
 574  */
 575 int free_swap_and_cache(swp_entry_t entry)
 576 {
 577         struct swap_info_struct *p;
 578         struct page *page = NULL;
 579
 580         if (is_migration_entry(entry))
 581                 return 1;
 582
 583         p = swap_info_get(entry);
 584         if (p) {
 585                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 586                         page = find_get_page(&swapper_space, entry.val);
 587                         if (page && !trylock_page(page)) {
 588                                 page_cache_release(page);
 589                                 page = NULL;
 590                         }
 591                 }
 592                 spin_unlock(&swap_lock);
 593         }
 594         if (page) {
 595                 /*
 596                  * Not mapped elsewhere, or swap space full? Free it!
 597                  * Also recheck PageSwapCache now page is locked (above).
 598                  */
 599                 if (PageSwapCache(page) && !PageWriteback(page) &&
 600                                 (!page_mapped(page) || vm_swap_full())) {
 601                         delete_from_swap_cache(page);
 602                         SetPageDirty(page);
 603                 }
 604                 unlock_page(page);
 605                 page_cache_release(page);
 606         }
 607         return p != NULL;
 608 }
 609
 610 #ifdef CONFIG_HIBERNATION
 611 /*
 612  * Find the swap type that corresponds to given device (if any).
 613  *
 614  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 615  * from 0, in which the swap header is expected to be located.
 616  *
 617  * This is needed for the suspend to disk (aka swsusp).
 618  */
 619 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 620 {
 621         struct block_device *bdev = NULL;
 622         int i;
 623
 624         if (device)
 625                 bdev = bdget(device);
 626
 627         spin_lock(&swap_lock);
 628         for (i = 0; i < nr_swapfiles; i++) {
 629                 struct swap_info_struct *sis = swap_info + i;
 630
 631                 if (!(sis->flags & SWP_WRITEOK))
 632                         continue;
 633
 634                 if (!bdev) {
 635                         if (bdev_p)
 636                                 *bdev_p = sis->bdev;
 637
 638                         spin_unlock(&swap_lock);
 639                         return i;
 640                 }
 641                 if (bdev == sis->bdev) {
 642                         struct swap_extent *se;
 643
 644                         se = list_entry(sis->extent_list.next,
 645                                         struct swap_extent, list);
 646                         if (se->start_block == offset) {
 647                                 if (bdev_p)
 648                                         *bdev_p = sis->bdev;
 649
 650                                 spin_unlock(&swap_lock);
 651                                 bdput(bdev);
 652                                 return i;
 653                         }
 654                 }
 655         }
 656         spin_unlock(&swap_lock);
 657         if (bdev)
 658                 bdput(bdev);
 659
 660         return -ENODEV;
 661 }
 662
 663 /*
 664  * Return either the total number of swap pages of given type, or the number
 665  * of free pages of that type (depending on @free)
 666  *
 667  * This is needed for software suspend
 668  */
 669 unsigned int count_swap_pages(int type, int free)
 670 {
 671         unsigned int n = 0;
 672
 673         if (type < nr_swapfiles) {
 674                 spin_lock(&swap_lock);
 675                 if (swap_info[type].flags & SWP_WRITEOK) {
 676                         n = swap_info[type].pages;
 677                         if (free)
 678                                 n -= swap_info[type].inuse_pages;
 679                 }
 680                 spin_unlock(&swap_lock);
 681         }
 682         return n;
 683 }
 684 #endif
 685
 686 /*
 687  * No need to decide whether this PTE shares the swap entry with others,
 688  * just let do_wp_page work it out if a write is requested later - to
 689  * force COW, vm_page_prot omits write permission from any private vma.
 690  */
 691 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 692                 unsigned long addr, swp_entry_t entry, struct page *page)
 693 {
 694         struct mem_cgroup *ptr = NULL;
 695         spinlock_t *ptl;
 696         pte_t *pte;
 697         int ret = 1;
 698
 699         if (mem_cgroup_try_charge(vma->vm_mm, GFP_HIGHUSER_MOVABLE, &ptr))
 700                 ret = -ENOMEM;
 701
 702         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 703         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 704                 if (ret > 0)
 705                         mem_cgroup_cancel_charge_swapin(ptr);
 706                 ret = 0;
 707                 goto out;
 708         }
 709
 710         inc_mm_counter(vma->vm_mm, anon_rss);
 711         get_page(page);
 712         set_pte_at(vma->vm_mm, addr, pte,
 713                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 714         page_add_anon_rmap(page, vma, addr);
 715         mem_cgroup_commit_charge_swapin(page, ptr);
 716         swap_free(entry);
 717         /*
 718          * Move the page to the active list so it is not
 719          * immediately swapped out again after swapon.
 720          */
 721         activate_page(page);
 722 out:
 723         pte_unmap_unlock(pte, ptl);
 724         return ret;
 725 }
 726
 727 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 728                                 unsigned long addr, unsigned long end,
 729                                 swp_entry_t entry, struct page *page)
 730 {
 731         pte_t swp_pte = swp_entry_to_pte(entry);
 732         pte_t *pte;
 733         int ret = 0;
 734
 735         /*
 736          * We don't actually need pte lock while scanning for swp_pte: since
 737          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 738          * page table while we're scanning; though it could get zapped, and on
 739          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 740          * of unmatched parts which look like swp_pte, so unuse_pte must
 741          * recheck under pte lock.  Scanning without pte lock lets it be
 742          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 743          */
 744         pte = pte_offset_map(pmd, addr);
 745         do {
 746                 /*
 747                  * swapoff spends a _lot_ of time in this loop!
 748                  * Test inline before going to call unuse_pte.
 749                  */
 750                 if (unlikely(pte_same(*pte, swp_pte))) {
 751                         pte_unmap(pte);
 752                         ret = unuse_pte(vma, pmd, addr, entry, page);
 753                         if (ret)
 754                                 goto out;
 755                         pte = pte_offset_map(pmd, addr);
 756                 }
 757         } while (pte++, addr += PAGE_SIZE, addr != end);
 758         pte_unmap(pte - 1);
 759 out:
 760         return ret;
 761 }
 762
 763 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 764                                 unsigned long addr, unsigned long end,
 765                                 swp_entry_t entry, struct page *page)
 766 {
 767         pmd_t *pmd;
 768         unsigned long next;
 769         int ret;
 770
 771         pmd = pmd_offset(pud, addr);
 772         do {
 773                 next = pmd_addr_end(addr, end);
 774                 if (pmd_none_or_clear_bad(pmd))
 775                         continue;
 776                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 777                 if (ret)
 778                         return ret;
 779         } while (pmd++, addr = next, addr != end);
 780         return 0;
 781 }
 782
 783 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 784                                 unsigned long addr, unsigned long end,
 785                                 swp_entry_t entry, struct page *page)
 786 {
 787         pud_t *pud;
 788         unsigned long next;
 789         int ret;
 790
 791         pud = pud_offset(pgd, addr);
 792         do {
 793                 next = pud_addr_end(addr, end);
 794                 if (pud_none_or_clear_bad(pud))
 795                         continue;
 796                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 797                 if (ret)
 798                         return ret;
 799         } while (pud++, addr = next, addr != end);
 800         return 0;
 801 }
 802
 803 static int unuse_vma(struct vm_area_struct *vma,
 804                                 swp_entry_t entry, struct page *page)
 805 {
 806         pgd_t *pgd;
 807         unsigned long addr, end, next;
 808         int ret;
 809
 810         if (page->mapping) {
 811                 addr = page_address_in_vma(page, vma);
 812                 if (addr == -EFAULT)
 813                         return 0;
 814                 else
 815                         end = addr + PAGE_SIZE;
 816         } else {
 817                 addr = vma->vm_start;
 818                 end = vma->vm_end;
 819         }
 820
 821         pgd = pgd_offset(vma->vm_mm, addr);
 822         do {
 823                 next = pgd_addr_end(addr, end);
 824                 if (pgd_none_or_clear_bad(pgd))
 825                         continue;
 826                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 827                 if (ret)
 828                         return ret;
 829         } while (pgd++, addr = next, addr != end);
 830         return 0;
 831 }
 832
 833 static int unuse_mm(struct mm_struct *mm,
 834                                 swp_entry_t entry, struct page *page)
 835 {
 836         struct vm_area_struct *vma;
 837         int ret = 0;
 838
 839         if (!down_read_trylock(&mm->mmap_sem)) {
 840                 /*
 841                  * Activate page so shrink_inactive_list is unlikely to unmap
 842                  * its ptes while lock is dropped, so swapoff can make progress.
 843                  */
 844                 activate_page(page);
 845                 unlock_page(page);
 846                 down_read(&mm->mmap_sem);
 847                 lock_page(page);
 848         }
 849         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 850                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
 851                         break;
 852         }
 853         up_read(&mm->mmap_sem);
 854         return (ret < 0)? ret: 0;
 855 }
 856
 857 /*
 858  * Scan swap_map from current position to next entry still in use.
 859  * Recycle to start on reaching the end, returning 0 when empty.
 860  */
 861 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 862                                         unsigned int prev)
 863 {
 864         unsigned int max = si->max;
 865         unsigned int i = prev;
 866         int count;
 867
 868         /*
 869          * No need for swap_lock here: we're just looking
 870          * for whether an entry is in use, not modifying it; false
 871          * hits are okay, and sys_swapoff() has already prevented new
 872          * allocations from this area (while holding swap_lock).
 873          */
 874         for (;;) {
 875                 if (++i >= max) {
 876                         if (!prev) {
 877                                 i = 0;
 878                                 break;
 879                         }
 880                         /*
 881                          * No entries in use at top of swap_map,
 882                          * loop back to start and recheck there.
 883                          */
 884                         max = prev + 1;
 885                         prev = 0;
 886                         i = 1;
 887                 }
 888                 count = si->swap_map[i];
 889                 if (count && count != SWAP_MAP_BAD)
 890                         break;
 891         }
 892         return i;
 893 }
 894
 895 /*
 896  * We completely avoid races by reading each swap page in advance,
 897  * and then search for the process using it.  All the necessary
 898  * page table adjustments can then be made atomically.
 899  */
 900 static int try_to_unuse(unsigned int type)
 901 {
 902         struct swap_info_struct * si = &swap_info[type];
 903         struct mm_struct *start_mm;
 904         unsigned short *swap_map;
 905         unsigned short swcount;
 906         struct page *page;
 907         swp_entry_t entry;
 908         unsigned int i = 0;
 909         int retval = 0;
 910         int reset_overflow = 0;
 911         int shmem;
 912
 913         /*
 914          * When searching mms for an entry, a good strategy is to
 915          * start at the first mm we freed the previous entry from
 916          * (though actually we don't notice whether we or coincidence
 917          * freed the entry).  Initialize this start_mm with a hold.
 918          *
 919          * A simpler strategy would be to start at the last mm we
 920          * freed the previous entry from; but that would take less
 921          * advantage of mmlist ordering, which clusters forked mms
 922          * together, child after parent.  If we race with dup_mmap(), we
 923          * prefer to resolve parent before child, lest we miss entries
 924          * duplicated after we scanned child: using last mm would invert
 925          * that.  Though it's only a serious concern when an overflowed
 926          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 927          */
 928         start_mm = &init_mm;
 929         atomic_inc(&init_mm.mm_users);
 930
 931         /*
 932          * Keep on scanning until all entries have gone.  Usually,
 933          * one pass through swap_map is enough, but not necessarily:
 934          * there are races when an instance of an entry might be missed.
 935          */
 936         while ((i = find_next_to_unuse(si, i)) != 0) {
 937                 if (signal_pending(current)) {
 938                         retval = -EINTR;
 939                         break;
 940                 }
 941
 942                 /*
 943                  * Get a page for the entry, using the existing swap
 944                  * cache page if there is one.  Otherwise, get a clean
 945                  * page and read the swap into it.
 946                  */
 947                 swap_map = &si->swap_map[i];
 948                 entry = swp_entry(type, i);
 949                 page = read_swap_cache_async(entry,
 950                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
 951                 if (!page) {
 952                         /*
 953                          * Either swap_duplicate() failed because entry
 954                          * has been freed independently, and will not be
 955                          * reused since sys_swapoff() already disabled
 956                          * allocation from here, or alloc_page() failed.
 957                          */
 958                         if (!*swap_map)
 959                                 continue;
 960                         retval = -ENOMEM;
 961                         break;
 962                 }
 963
 964                 /*
 965                  * Don't hold on to start_mm if it looks like exiting.
 966                  */
 967                 if (atomic_read(&start_mm->mm_users) == 1) {
 968                         mmput(start_mm);
 969                         start_mm = &init_mm;
 970                         atomic_inc(&init_mm.mm_users);
 971                 }
 972
 973                 /*
 974                  * Wait for and lock page.  When do_swap_page races with
 975                  * try_to_unuse, do_swap_page can handle the fault much
 976                  * faster than try_to_unuse can locate the entry.  This
 977                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 978                  * defer to do_swap_page in such a case - in some tests,
 979                  * do_swap_page and try_to_unuse repeatedly compete.
 980                  */
 981                 wait_on_page_locked(page);
 982                 wait_on_page_writeback(page);
 983                 lock_page(page);
 984                 wait_on_page_writeback(page);
 985
 986                 /*
 987                  * Remove all references to entry.
 988                  * Whenever we reach init_mm, there's no address space
 989                  * to search, but use it as a reminder to search shmem.
 990                  */
 991                 shmem = 0;
 992                 swcount = *swap_map;
 993                 if (swcount > 1) {
 994                         if (start_mm == &init_mm)
 995                                 shmem = shmem_unuse(entry, page);
 996                         else
 997                                 retval = unuse_mm(start_mm, entry, page);
 998                 }
 999                 if (*swap_map > 1) {
1000                         int set_start_mm = (*swap_map >= swcount);
1001                         struct list_head *p = &start_mm->mmlist;
1002                         struct mm_struct *new_start_mm = start_mm;
1003                         struct mm_struct *prev_mm = start_mm;
1004                         struct mm_struct *mm;
1005
1006                         atomic_inc(&new_start_mm->mm_users);
1007                         atomic_inc(&prev_mm->mm_users);
1008                         spin_lock(&mmlist_lock);
1009                         while (*swap_map > 1 && !retval && !shmem &&
1010                                         (p = p->next) != &start_mm->mmlist) {
1011                                 mm = list_entry(p, struct mm_struct, mmlist);
1012                                 if (!atomic_inc_not_zero(&mm->mm_users))
1013                                         continue;
1014                                 spin_unlock(&mmlist_lock);
1015                                 mmput(prev_mm);
1016                                 prev_mm = mm;
1017
1018                                 cond_resched();
1019
1020                                 swcount = *swap_map;
1021                                 if (swcount <= 1)
1022                                         ;
1023                                 else if (mm == &init_mm) {
1024                                         set_start_mm = 1;
1025                                         shmem = shmem_unuse(entry, page);
1026                                 } else
1027                                         retval = unuse_mm(mm, entry, page);
1028                                 if (set_start_mm && *swap_map < swcount) {
1029                                         mmput(new_start_mm);
1030                                         atomic_inc(&mm->mm_users);
1031                                         new_start_mm = mm;
1032                                         set_start_mm = 0;
1033                                 }
1034                                 spin_lock(&mmlist_lock);
1035                         }
1036                         spin_unlock(&mmlist_lock);
1037                         mmput(prev_mm);
1038                         mmput(start_mm);
1039                         start_mm = new_start_mm;
1040                 }
1041                 if (shmem) {
1042                         /* page has already been unlocked and released */
1043                         if (shmem > 0)
1044                                 continue;
1045                         retval = shmem;
1046                         break;
1047                 }
1048                 if (retval) {
1049                         unlock_page(page);
1050                         page_cache_release(page);
1051                         break;
1052                 }
1053
1054                 /*
1055                  * How could swap count reach 0x7fff when the maximum
1056                  * pid is 0x7fff, and there's no way to repeat a swap
1057                  * page within an mm (except in shmem, where it's the
1058                  * shared object which takes the reference count)?
1059                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
1060                  *
1061                  * If that's wrong, then we should worry more about
1062                  * exit_mmap() and do_munmap() cases described above:
1063                  * we might be resetting SWAP_MAP_MAX too early here.
1064                  * We know "Undead"s can happen, they're okay, so don't
1065                  * report them; but do report if we reset SWAP_MAP_MAX.
1066                  */
1067                 if (*swap_map == SWAP_MAP_MAX) {
1068                         spin_lock(&swap_lock);
1069                         *swap_map = 1;
1070                         spin_unlock(&swap_lock);
1071                         reset_overflow = 1;
1072                 }
1073
1074                 /*
1075                  * If a reference remains (rare), we would like to leave
1076                  * the page in the swap cache; but try_to_unmap could
1077                  * then re-duplicate the entry once we drop page lock,
1078                  * so we might loop indefinitely; also, that page could
1079                  * not be swapped out to other storage meanwhile.  So:
1080                  * delete from cache even if there's another reference,
1081                  * after ensuring that the data has been saved to disk -
1082                  * since if the reference remains (rarer), it will be
1083                  * read from disk into another page.  Splitting into two
1084                  * pages would be incorrect if swap supported "shared
1085                  * private" pages, but they are handled by tmpfs files.
1086                  */
1087                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
1088                         struct writeback_control wbc = {
1089                                 .sync_mode = WB_SYNC_NONE,
1090                         };
1091
1092                         swap_writepage(page, &wbc);
1093                         lock_page(page);
1094                         wait_on_page_writeback(page);
1095                 }
1096
1097                 /*
1098                  * It is conceivable that a racing task removed this page from
1099                  * swap cache just before we acquired the page lock at the top,
1100                  * or while we dropped it in unuse_mm().  The page might even
1101                  * be back in swap cache on another swap area: that we must not
1102                  * delete, since it may not have been written out to swap yet.
1103                  */
1104                 if (PageSwapCache(page) &&
1105                     likely(page_private(page) == entry.val))
1106                         delete_from_swap_cache(page);
1107
1108                 /*
1109                  * So we could skip searching mms once swap count went
1110                  * to 1, we did not mark any present ptes as dirty: must
1111                  * mark page dirty so shrink_page_list will preserve it.
1112                  */
1113                 SetPageDirty(page);
1114                 unlock_page(page);
1115                 page_cache_release(page);
1116
1117                 /*
1118                  * Make sure that we aren't completely killing
1119                  * interactive performance.
1120                  */
1121                 cond_resched();
1122         }
1123
1124         mmput(start_mm);
1125         if (reset_overflow) {
1126                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1127                 swap_overflow = 0;
1128         }
1129         return retval;
1130 }
1131
1132 /*
1133  * After a successful try_to_unuse, if no swap is now in use, we know
1134  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1135  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1136  * added to the mmlist just after page_duplicate - before would be racy.
1137  */
1138 static void drain_mmlist(void)
1139 {
1140         struct list_head *p, *next;
1141         unsigned int i;
1142
1143         for (i = 0; i < nr_swapfiles; i++)
1144                 if (swap_info[i].inuse_pages)
1145                         return;
1146         spin_lock(&mmlist_lock);
1147         list_for_each_safe(p, next, &init_mm.mmlist)
1148                 list_del_init(p);
1149         spin_unlock(&mmlist_lock);
1150 }
1151
1152 /*
1153  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1154  * corresponds to page offset `offset'.
1155  */
1156 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1157 {
1158         struct swap_extent *se = sis->curr_swap_extent;
1159         struct swap_extent *start_se = se;
1160
1161         for ( ; ; ) {
1162                 struct list_head *lh;
1163
1164                 if (se->start_page <= offset &&
1165                                 offset < (se->start_page + se->nr_pages)) {
1166                         return se->start_block + (offset - se->start_page);
1167                 }
1168                 lh = se->list.next;
1169                 if (lh == &sis->extent_list)
1170                         lh = lh->next;
1171                 se = list_entry(lh, struct swap_extent, list);
1172                 sis->curr_swap_extent = se;
1173                 BUG_ON(se == start_se);         /* It *must* be present */
1174         }
1175 }
1176
1177 #ifdef CONFIG_HIBERNATION
1178 /*
1179  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1180  * corresponding to given index in swap_info (swap type).
1181  */
1182 sector_t swapdev_block(int swap_type, pgoff_t offset)
1183 {
1184         struct swap_info_struct *sis;
1185
1186         if (swap_type >= nr_swapfiles)
1187                 return 0;
1188
1189         sis = swap_info + swap_type;
1190         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1191 }
1192 #endif /* CONFIG_HIBERNATION */
1193
1194 /*
1195  * Free all of a swapdev's extent information
1196  */
1197 static void destroy_swap_extents(struct swap_info_struct *sis)
1198 {
1199         while (!list_empty(&sis->extent_list)) {
1200                 struct swap_extent *se;
1201
1202                 se = list_entry(sis->extent_list.next,
1203                                 struct swap_extent, list);
1204                 list_del(&se->list);
1205                 kfree(se);
1206         }
1207 }
1208
1209 /*
1210  * Add a block range (and the corresponding page range) into this swapdev's
1211  * extent list.  The extent list is kept sorted in page order.
1212  *
1213  * This function rather assumes that it is called in ascending page order.
1214  */
1215 static int
1216 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1217                 unsigned long nr_pages, sector_t start_block)
1218 {
1219         struct swap_extent *se;
1220         struct swap_extent *new_se;
1221         struct list_head *lh;
1222
1223         lh = sis->extent_list.prev;     /* The highest page extent */
1224         if (lh != &sis->extent_list) {
1225                 se = list_entry(lh, struct swap_extent, list);
1226                 BUG_ON(se->start_page + se->nr_pages != start_page);
1227                 if (se->start_block + se->nr_pages == start_block) {
1228                         /* Merge it */
1229                         se->nr_pages += nr_pages;
1230                         return 0;
1231                 }
1232         }
1233
1234         /*
1235          * No merge.  Insert a new extent, preserving ordering.
1236          */
1237         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1238         if (new_se == NULL)
1239                 return -ENOMEM;
1240         new_se->start_page = start_page;
1241         new_se->nr_pages = nr_pages;
1242         new_se->start_block = start_block;
1243
1244         list_add_tail(&new_se->list, &sis->extent_list);
1245         return 1;
1246 }
1247
1248 /*
1249  * A `swap extent' is a simple thing which maps a contiguous range of pages
1250  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1251  * is built at swapon time and is then used at swap_writepage/swap_readpage
1252  * time for locating where on disk a page belongs.
1253  *
1254  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1255  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1256  * swap files identically.
1257  *
1258  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1259  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1260  * swapfiles are handled *identically* after swapon time.
1261  *
1262  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1263  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1264  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1265  * requirements, they are simply tossed out - we will never use those blocks
1266  * for swapping.
1267  *
1268  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1269  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1270  * which will scribble on the fs.
1271  *
1272  * The amount of disk space which a single swap extent represents varies.
1273  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1274  * extents in the list.  To avoid much list walking, we cache the previous
1275  * search location in `curr_swap_extent', and start new searches from there.
1276  * This is extremely effective.  The average number of iterations in
1277  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1278  */
1279 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1280 {
1281         struct inode *inode;
1282         unsigned blocks_per_page;
1283         unsigned long page_no;
1284         unsigned blkbits;
1285         sector_t probe_block;
1286         sector_t last_block;
1287         sector_t lowest_block = -1;
1288         sector_t highest_block = 0;
1289         int nr_extents = 0;
1290         int ret;
1291
1292         inode = sis->swap_file->f_mapping->host;
1293         if (S_ISBLK(inode->i_mode)) {
1294                 ret = add_swap_extent(sis, 0, sis->max, 0);
1295                 *span = sis->pages;
1296                 goto done;
1297         }
1298
1299         blkbits = inode->i_blkbits;
1300         blocks_per_page = PAGE_SIZE >> blkbits;
1301
1302         /*
1303          * Map all the blocks into the extent list.  This code doesn't try
1304          * to be very smart.
1305          */
1306         probe_block = 0;
1307         page_no = 0;
1308         last_block = i_size_read(inode) >> blkbits;
1309         while ((probe_block + blocks_per_page) <= last_block &&
1310                         page_no < sis->max) {
1311                 unsigned block_in_page;
1312                 sector_t first_block;
1313
1314                 first_block = bmap(inode, probe_block);
1315                 if (first_block == 0)
1316                         goto bad_bmap;
1317
1318                 /*
1319                  * It must be PAGE_SIZE aligned on-disk
1320                  */
1321                 if (first_block & (blocks_per_page - 1)) {
1322                         probe_block++;
1323                         goto reprobe;
1324                 }
1325
1326                 for (block_in_page = 1; block_in_page < blocks_per_page;
1327                                         block_in_page++) {
1328                         sector_t block;
1329
1330                         block = bmap(inode, probe_block + block_in_page);
1331                         if (block == 0)
1332                                 goto bad_bmap;
1333                         if (block != first_block + block_in_page) {
1334                                 /* Discontiguity */
1335                                 probe_block++;
1336                                 goto reprobe;
1337                         }
1338                 }
1339
1340                 first_block >>= (PAGE_SHIFT - blkbits);
1341                 if (page_no) {  /* exclude the header page */
1342                         if (first_block < lowest_block)
1343                                 lowest_block = first_block;
1344                         if (first_block > highest_block)
1345                                 highest_block = first_block;
1346                 }
1347
1348                 /*
1349                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1350                  */
1351                 ret = add_swap_extent(sis, page_no, 1, first_block);
1352                 if (ret < 0)
1353                         goto out;
1354                 nr_extents += ret;
1355                 page_no++;
1356                 probe_block += blocks_per_page;
1357 reprobe:
1358                 continue;
1359         }
1360         ret = nr_extents;
1361         *span = 1 + highest_block - lowest_block;
1362         if (page_no == 0)
1363                 page_no = 1;    /* force Empty message */
1364         sis->max = page_no;
1365         sis->pages = page_no - 1;
1366         sis->highest_bit = page_no - 1;
1367 done:
1368         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1369                                         struct swap_extent, list);
1370         goto out;
1371 bad_bmap:
1372         printk(KERN_ERR "swapon: swapfile has holes\n");
1373         ret = -EINVAL;
1374 out:
1375         return ret;
1376 }
1377
1378 asmlinkage long sys_swapoff(const char __user * specialfile)
1379 {
1380         struct swap_info_struct * p = NULL;
1381         unsigned short *swap_map;
1382         struct file *swap_file, *victim;
1383         struct address_space *mapping;
1384         struct inode *inode;
1385         char * pathname;
1386         int i, type, prev;
1387         int err;
1388
1389         if (!capable(CAP_SYS_ADMIN))
1390                 return -EPERM;
1391
1392         pathname = getname(specialfile);
1393         err = PTR_ERR(pathname);
1394         if (IS_ERR(pathname))
1395                 goto out;
1396
1397         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1398         putname(pathname);
1399         err = PTR_ERR(victim);
1400         if (IS_ERR(victim))
1401                 goto out;
1402
1403         mapping = victim->f_mapping;
1404         prev = -1;
1405         spin_lock(&swap_lock);
1406         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1407                 p = swap_info + type;
1408                 if (p->flags & SWP_WRITEOK) {
1409                         if (p->swap_file->f_mapping == mapping)
1410                                 break;
1411                 }
1412                 prev = type;
1413         }
1414         if (type < 0) {
1415                 err = -EINVAL;
1416                 spin_unlock(&swap_lock);
1417                 goto out_dput;
1418         }
1419         if (!security_vm_enough_memory(p->pages))
1420                 vm_unacct_memory(p->pages);
1421         else {
1422                 err = -ENOMEM;
1423                 spin_unlock(&swap_lock);
1424                 goto out_dput;
1425         }
1426         if (prev < 0) {
1427                 swap_list.head = p->next;
1428         } else {
1429                 swap_info[prev].next = p->next;
1430         }
1431         if (type == swap_list.next) {
1432                 /* just pick something that's safe... */
1433                 swap_list.next = swap_list.head;
1434         }
1435         if (p->prio < 0) {
1436                 for (i = p->next; i >= 0; i = swap_info[i].next)
1437                         swap_info[i].prio = p->prio--;
1438                 least_priority++;
1439         }
1440         nr_swap_pages -= p->pages;
1441         total_swap_pages -= p->pages;
1442         p->flags &= ~SWP_WRITEOK;
1443         spin_unlock(&swap_lock);
1444
1445         current->flags |= PF_SWAPOFF;
1446         err = try_to_unuse(type);
1447         current->flags &= ~PF_SWAPOFF;
1448
1449         if (err) {
1450                 /* re-insert swap space back into swap_list */
1451                 spin_lock(&swap_lock);
1452                 if (p->prio < 0)
1453                         p->prio = --least_priority;
1454                 prev = -1;
1455                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1456                         if (p->prio >= swap_info[i].prio)
1457                                 break;
1458                         prev = i;
1459                 }
1460                 p->next = i;
1461                 if (prev < 0)
1462                         swap_list.head = swap_list.next = p - swap_info;
1463                 else
1464                         swap_info[prev].next = p - swap_info;
1465                 nr_swap_pages += p->pages;
1466                 total_swap_pages += p->pages;
1467                 p->flags |= SWP_WRITEOK;
1468                 spin_unlock(&swap_lock);
1469                 goto out_dput;
1470         }
1471
1472         /* wait for any unplug function to finish */
1473         down_write(&swap_unplug_sem);
1474         up_write(&swap_unplug_sem);
1475
1476         destroy_swap_extents(p);
1477         mutex_lock(&swapon_mutex);
1478         spin_lock(&swap_lock);
1479         drain_mmlist();
1480
1481         /* wait for anyone still in scan_swap_map */
1482         p->highest_bit = 0;             /* cuts scans short */
1483         while (p->flags >= SWP_SCANNING) {
1484                 spin_unlock(&swap_lock);
1485                 schedule_timeout_uninterruptible(1);
1486                 spin_lock(&swap_lock);
1487         }
1488
1489         swap_file = p->swap_file;
1490         p->swap_file = NULL;
1491         p->max = 0;
1492         swap_map = p->swap_map;
1493         p->swap_map = NULL;
1494         p->flags = 0;
1495         spin_unlock(&swap_lock);
1496         mutex_unlock(&swapon_mutex);
1497         vfree(swap_map);
1498         /* Destroy swap account informatin */
1499         swap_cgroup_swapoff(type);
1500
1501         inode = mapping->host;
1502         if (S_ISBLK(inode->i_mode)) {
1503                 struct block_device *bdev = I_BDEV(inode);
1504                 set_blocksize(bdev, p->old_block_size);
1505                 bd_release(bdev);
1506         } else {
1507                 mutex_lock(&inode->i_mutex);
1508                 inode->i_flags &= ~S_SWAPFILE;
1509                 mutex_unlock(&inode->i_mutex);
1510         }
1511         filp_close(swap_file, NULL);
1512         err = 0;
1513
1514 out_dput:
1515         filp_close(victim, NULL);
1516 out:
1517         return err;
1518 }
1519
1520 #ifdef CONFIG_PROC_FS
1521 /* iterator */
1522 static void *swap_start(struct seq_file *swap, loff_t *pos)
1523 {
1524         struct swap_info_struct *ptr = swap_info;
1525         int i;
1526         loff_t l = *pos;
1527
1528         mutex_lock(&swapon_mutex);
1529
1530         if (!l)
1531                 return SEQ_START_TOKEN;
1532
1533         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1534                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1535                         continue;
1536                 if (!--l)
1537                         return ptr;
1538         }
1539
1540         return NULL;
1541 }
1542
1543 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1544 {
1545         struct swap_info_struct *ptr;
1546         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1547
1548         if (v == SEQ_START_TOKEN)
1549                 ptr = swap_info;
1550         else {
1551                 ptr = v;
1552                 ptr++;
1553         }
1554
1555         for (; ptr < endptr; ptr++) {
1556                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1557                         continue;
1558                 ++*pos;
1559                 return ptr;
1560         }
1561
1562         return NULL;
1563 }
1564
1565 static void swap_stop(struct seq_file *swap, void *v)
1566 {
1567         mutex_unlock(&swapon_mutex);
1568 }
1569
1570 static int swap_show(struct seq_file *swap, void *v)
1571 {
1572         struct swap_info_struct *ptr = v;
1573         struct file *file;
1574         int len;
1575
1576         if (ptr == SEQ_START_TOKEN) {
1577                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1578                 return 0;
1579         }
1580
1581         file = ptr->swap_file;
1582         len = seq_path(swap, &file->f_path, " \t\n\\");
1583         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1584                         len < 40 ? 40 - len : 1, " ",
1585                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1586                                 "partition" : "file\t",
1587                         ptr->pages << (PAGE_SHIFT - 10),
1588                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1589                         ptr->prio);
1590         return 0;
1591 }
1592
1593 static const struct seq_operations swaps_op = {
1594         .start =        swap_start,
1595         .next =         swap_next,
1596         .stop =         swap_stop,
1597         .show =         swap_show
1598 };
1599
1600 static int swaps_open(struct inode *inode, struct file *file)
1601 {
1602         return seq_open(file, &swaps_op);
1603 }
1604
1605 static const struct file_operations proc_swaps_operations = {
1606         .open           = swaps_open,
1607         .read           = seq_read,
1608         .llseek         = seq_lseek,
1609         .release        = seq_release,
1610 };
1611
1612 static int __init procswaps_init(void)
1613 {
1614         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1615         return 0;
1616 }
1617 __initcall(procswaps_init);
1618 #endif /* CONFIG_PROC_FS */
1619
1620 #ifdef MAX_SWAPFILES_CHECK
1621 static int __init max_swapfiles_check(void)
1622 {
1623         MAX_SWAPFILES_CHECK();
1624         return 0;
1625 }
1626 late_initcall(max_swapfiles_check);
1627 #endif
1628
1629 /*
1630  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1631  *
1632  * The swapon system call
1633  */
1634 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1635 {
1636         struct swap_info_struct * p;
1637         char *name = NULL;
1638         struct block_device *bdev = NULL;
1639         struct file *swap_file = NULL;
1640         struct address_space *mapping;
1641         unsigned int type;
1642         int i, prev;
1643         int error;
1644         union swap_header *swap_header = NULL;
1645         unsigned int nr_good_pages = 0;
1646         int nr_extents = 0;
1647         sector_t span;
1648         unsigned long maxpages = 1;
1649         unsigned long swapfilepages;
1650         unsigned short *swap_map = NULL;
1651         struct page *page = NULL;
1652         struct inode *inode = NULL;
1653         int did_down = 0;
1654
1655         if (!capable(CAP_SYS_ADMIN))
1656                 return -EPERM;
1657         spin_lock(&swap_lock);
1658         p = swap_info;
1659         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1660                 if (!(p->flags & SWP_USED))
1661                         break;
1662         error = -EPERM;
1663         if (type >= MAX_SWAPFILES) {
1664                 spin_unlock(&swap_lock);
1665                 goto out;
1666         }
1667         if (type >= nr_swapfiles)
1668                 nr_swapfiles = type+1;
1669         memset(p, 0, sizeof(*p));
1670         INIT_LIST_HEAD(&p->extent_list);
1671         p->flags = SWP_USED;
1672         p->next = -1;
1673         spin_unlock(&swap_lock);
1674         name = getname(specialfile);
1675         error = PTR_ERR(name);
1676         if (IS_ERR(name)) {
1677                 name = NULL;
1678                 goto bad_swap_2;
1679         }
1680         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1681         error = PTR_ERR(swap_file);
1682         if (IS_ERR(swap_file)) {
1683                 swap_file = NULL;
1684                 goto bad_swap_2;
1685         }
1686
1687         p->swap_file = swap_file;
1688         mapping = swap_file->f_mapping;
1689         inode = mapping->host;
1690
1691         error = -EBUSY;
1692         for (i = 0; i < nr_swapfiles; i++) {
1693                 struct swap_info_struct *q = &swap_info[i];
1694
1695                 if (i == type || !q->swap_file)
1696                         continue;
1697                 if (mapping == q->swap_file->f_mapping)
1698                         goto bad_swap;
1699         }
1700
1701         error = -EINVAL;
1702         if (S_ISBLK(inode->i_mode)) {
1703                 bdev = I_BDEV(inode);
1704                 error = bd_claim(bdev, sys_swapon);
1705                 if (error < 0) {
1706                         bdev = NULL;
1707                         error = -EINVAL;
1708                         goto bad_swap;
1709                 }
1710                 p->old_block_size = block_size(bdev);
1711                 error = set_blocksize(bdev, PAGE_SIZE);
1712                 if (error < 0)
1713                         goto bad_swap;
1714                 p->bdev = bdev;
1715         } else if (S_ISREG(inode->i_mode)) {
1716                 p->bdev = inode->i_sb->s_bdev;
1717                 mutex_lock(&inode->i_mutex);
1718                 did_down = 1;
1719                 if (IS_SWAPFILE(inode)) {
1720                         error = -EBUSY;
1721                         goto bad_swap;
1722                 }
1723         } else {
1724                 goto bad_swap;
1725         }
1726
1727         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1728
1729         /*
1730          * Read the swap header.
1731          */
1732         if (!mapping->a_ops->readpage) {
1733                 error = -EINVAL;
1734                 goto bad_swap;
1735         }
1736         page = read_mapping_page(mapping, 0, swap_file);
1737         if (IS_ERR(page)) {
1738                 error = PTR_ERR(page);
1739                 goto bad_swap;
1740         }
1741         swap_header = kmap(page);
1742
1743         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1744                 printk(KERN_ERR "Unable to find swap-space signature\n");
1745                 error = -EINVAL;
1746                 goto bad_swap;
1747         }
1748
1749         /* swap partition endianess hack... */
1750         if (swab32(swap_header->info.version) == 1) {
1751                 swab32s(&swap_header->info.version);
1752                 swab32s(&swap_header->info.last_page);
1753                 swab32s(&swap_header->info.nr_badpages);
1754                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1755                         swab32s(&swap_header->info.badpages[i]);
1756         }
1757         /* Check the swap header's sub-version */
1758         if (swap_header->info.version != 1) {
1759                 printk(KERN_WARNING
1760                        "Unable to handle swap header version %d\n",
1761                        swap_header->info.version);
1762                 error = -EINVAL;
1763                 goto bad_swap;
1764         }
1765
1766         p->lowest_bit  = 1;
1767         p->cluster_next = 1;
1768
1769         /*
1770          * Find out how many pages are allowed for a single swap
1771          * device. There are two limiting factors: 1) the number of
1772          * bits for the swap offset in the swp_entry_t type and
1773          * 2) the number of bits in the a swap pte as defined by
1774          * the different architectures. In order to find the
1775          * largest possible bit mask a swap entry with swap type 0
1776          * and swap offset ~0UL is created, encoded to a swap pte,
1777          * decoded to a swp_entry_t again and finally the swap
1778          * offset is extracted. This will mask all the bits from
1779          * the initial ~0UL mask that can't be encoded in either
1780          * the swp_entry_t or the architecture definition of a
1781          * swap pte.
1782          */
1783         maxpages = swp_offset(pte_to_swp_entry(
1784                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1785         if (maxpages > swap_header->info.last_page)
1786                 maxpages = swap_header->info.last_page;
1787         p->highest_bit = maxpages - 1;
1788
1789         error = -EINVAL;
1790         if (!maxpages)
1791                 goto bad_swap;
1792         if (swapfilepages && maxpages > swapfilepages) {
1793                 printk(KERN_WARNING
1794                        "Swap area shorter than signature indicates\n");
1795                 goto bad_swap;
1796         }
1797         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1798                 goto bad_swap;
1799         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1800                 goto bad_swap;
1801
1802         /* OK, set up the swap map and apply the bad block list */
1803         swap_map = vmalloc(maxpages * sizeof(short));
1804         if (!swap_map) {
1805                 error = -ENOMEM;
1806                 goto bad_swap;
1807         }
1808
1809         memset(swap_map, 0, maxpages * sizeof(short));
1810         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1811                 int page_nr = swap_header->info.badpages[i];
1812                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1813                         error = -EINVAL;
1814                         goto bad_swap;
1815                 }
1816                 swap_map[page_nr] = SWAP_MAP_BAD;
1817         }
1818
1819         error = swap_cgroup_swapon(type, maxpages);
1820         if (error)
1821                 goto bad_swap;
1822
1823         nr_good_pages = swap_header->info.last_page -
1824                         swap_header->info.nr_badpages -
1825                         1 /* header page */;
1826
1827         if (nr_good_pages) {
1828                 swap_map[0] = SWAP_MAP_BAD;
1829                 p->max = maxpages;
1830                 p->pages = nr_good_pages;
1831                 nr_extents = setup_swap_extents(p, &span);
1832                 if (nr_extents < 0) {
1833                         error = nr_extents;
1834                         goto bad_swap;
1835                 }
1836                 nr_good_pages = p->pages;
1837         }
1838         if (!nr_good_pages) {
1839                 printk(KERN_WARNING "Empty swap-file\n");
1840                 error = -EINVAL;
1841                 goto bad_swap;
1842         }
1843
1844         if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1845                 p->flags |= SWP_SOLIDSTATE;
1846                 p->cluster_next = 1 + (random32() % p->highest_bit);
1847         }
1848         if (discard_swap(p) == 0)
1849                 p->flags |= SWP_DISCARDABLE;
1850
1851         mutex_lock(&swapon_mutex);
1852         spin_lock(&swap_lock);
1853         if (swap_flags & SWAP_FLAG_PREFER)
1854                 p->prio =
1855                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1856         else
1857                 p->prio = --least_priority;
1858         p->swap_map = swap_map;
1859         p->flags |= SWP_WRITEOK;
1860         nr_swap_pages += nr_good_pages;
1861         total_swap_pages += nr_good_pages;
1862
1863         printk(KERN_INFO "Adding %uk swap on %s.  "
1864                         "Priority:%d extents:%d across:%lluk %s%s\n",
1865                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1866                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1867                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1868                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1869
1870         /* insert swap space into swap_list: */
1871         prev = -1;
1872         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1873                 if (p->prio >= swap_info[i].prio) {
1874                         break;
1875                 }
1876                 prev = i;
1877         }
1878         p->next = i;
1879         if (prev < 0) {
1880                 swap_list.head = swap_list.next = p - swap_info;
1881         } else {
1882                 swap_info[prev].next = p - swap_info;
1883         }
1884         spin_unlock(&swap_lock);
1885         mutex_unlock(&swapon_mutex);
1886         error = 0;
1887         goto out;
1888 bad_swap:
1889         if (bdev) {
1890                 set_blocksize(bdev, p->old_block_size);
1891                 bd_release(bdev);
1892         }
1893         destroy_swap_extents(p);
1894         swap_cgroup_swapoff(type);
1895 bad_swap_2:
1896         spin_lock(&swap_lock);
1897         p->swap_file = NULL;
1898         p->flags = 0;
1899         spin_unlock(&swap_lock);
1900         vfree(swap_map);
1901         if (swap_file)
1902                 filp_close(swap_file, NULL);
1903 out:
1904         if (page && !IS_ERR(page)) {
1905                 kunmap(page);
1906                 page_cache_release(page);
1907         }
1908         if (name)
1909                 putname(name);
1910         if (did_down) {
1911                 if (!error)
1912                         inode->i_flags |= S_SWAPFILE;
1913                 mutex_unlock(&inode->i_mutex);
1914         }
1915         return error;
1916 }
1917
1918 void si_swapinfo(struct sysinfo *val)
1919 {
1920         unsigned int i;
1921         unsigned long nr_to_be_unused = 0;
1922
1923         spin_lock(&swap_lock);
1924         for (i = 0; i < nr_swapfiles; i++) {
1925                 if (!(swap_info[i].flags & SWP_USED) ||
1926                      (swap_info[i].flags & SWP_WRITEOK))
1927                         continue;
1928                 nr_to_be_unused += swap_info[i].inuse_pages;
1929         }
1930         val->freeswap = nr_swap_pages + nr_to_be_unused;
1931         val->totalswap = total_swap_pages + nr_to_be_unused;
1932         spin_unlock(&swap_lock);
1933 }
1934
1935 /*
1936  * Verify that a swap entry is valid and increment its swap map count.
1937  *
1938  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1939  * "permanent", but will be reclaimed by the next swapoff.
1940  */
1941 int swap_duplicate(swp_entry_t entry)
1942 {
1943         struct swap_info_struct * p;
1944         unsigned long offset, type;
1945         int result = 0;
1946
1947         if (is_migration_entry(entry))
1948                 return 1;
1949
1950         type = swp_type(entry);
1951         if (type >= nr_swapfiles)
1952                 goto bad_file;
1953         p = type + swap_info;
1954         offset = swp_offset(entry);
1955
1956         spin_lock(&swap_lock);
1957         if (offset < p->max && p->swap_map[offset]) {
1958                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1959                         p->swap_map[offset]++;
1960                         result = 1;
1961                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1962                         if (swap_overflow++ < 5)
1963                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1964                         p->swap_map[offset] = SWAP_MAP_MAX;
1965                         result = 1;
1966                 }
1967         }
1968         spin_unlock(&swap_lock);
1969 out:
1970         return result;
1971
1972 bad_file:
1973         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1974         goto out;
1975 }
1976
1977 struct swap_info_struct *
1978 get_swap_info_struct(unsigned type)
1979 {
1980         return &swap_info[type];
1981 }
1982
1983 /*
1984  * swap_lock prevents swap_map being freed. Don't grab an extra
1985  * reference on the swaphandle, it doesn't matter if it becomes unused.
1986  */
1987 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1988 {
1989         struct swap_info_struct *si;
1990         int our_page_cluster = page_cluster;
1991         pgoff_t target, toff;
1992         pgoff_t base, end;
1993         int nr_pages = 0;
1994
1995         if (!our_page_cluster)  /* no readahead */
1996                 return 0;
1997
1998         si = &swap_info[swp_type(entry)];
1999         target = swp_offset(entry);
2000         base = (target >> our_page_cluster) << our_page_cluster;
2001         end = base + (1 << our_page_cluster);
2002         if (!base)              /* first page is swap header */
2003                 base++;
2004
2005         spin_lock(&swap_lock);
2006         if (end > si->max)      /* don't go beyond end of map */
2007                 end = si->max;
2008
2009         /* Count contiguous allocated slots above our target */
2010         for (toff = target; ++toff < end; nr_pages++) {
2011                 /* Don't read in free or bad pages */
2012                 if (!si->swap_map[toff])
2013                         break;
2014                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2015                         break;
2016         }
2017         /* Count contiguous allocated slots below our target */
2018         for (toff = target; --toff >= base; nr_pages++) {
2019                 /* Don't read in free or bad pages */
2020                 if (!si->swap_map[toff])
2021                         break;
2022                 if (si->swap_map[toff] == SWAP_MAP_BAD)
2023                         break;
2024         }
2025         spin_unlock(&swap_lock);
2026
2027         /*
2028          * Indicate starting offset, and return number of pages to get:
2029          * if only 1, say 0, since there's then no readahead to be done.
2030          */
2031         *offset = ++toff;
2032         return nr_pages? ++nr_pages: 0;
2033 }