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 /* For reference count accounting in swap_map */
  57 /* enum for swap_map[] handling. internal use only */
  58 enum {
  59         SWAP_MAP = 0,   /* ops for reference from swap users */
  60         SWAP_CACHE,     /* ops for reference from swap cache */
  61 };
  62
  63 static inline int swap_count(unsigned short ent)
  64 {
  65         return ent & SWAP_COUNT_MASK;
  66 }
  67
  68 static inline bool swap_has_cache(unsigned short ent)
  69 {
  70         return !!(ent & SWAP_HAS_CACHE);
  71 }
  72
  73 static inline unsigned short encode_swapmap(int count, bool has_cache)
  74 {
  75         unsigned short ret = count;
  76
  77         if (has_cache)
  78                 return SWAP_HAS_CACHE | ret;
  79         return ret;
  80 }
  81
  82
  83 /*
  84  * We need this because the bdev->unplug_fn can sleep and we cannot
  85  * hold swap_lock while calling the unplug_fn. And swap_lock
  86  * cannot be turned into a mutex.
  87  */
  88 static DECLARE_RWSEM(swap_unplug_sem);
  89
  90 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  91 {
  92         swp_entry_t entry;
  93
  94         down_read(&swap_unplug_sem);
  95         entry.val = page_private(page);
  96         if (PageSwapCache(page)) {
  97                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  98                 struct backing_dev_info *bdi;
  99
 100                 /*
 101                  * If the page is removed from swapcache from under us (with a
 102                  * racy try_to_unuse/swapoff) we need an additional reference
 103                  * count to avoid reading garbage from page_private(page) above.
 104                  * If the WARN_ON triggers during a swapoff it maybe the race
 105                  * condition and it's harmless. However if it triggers without
 106                  * swapoff it signals a problem.
 107                  */
 108                 WARN_ON(page_count(page) <= 1);
 109
 110                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
 111                 blk_run_backing_dev(bdi, page);
 112         }
 113         up_read(&swap_unplug_sem);
 114 }
 115
 116 /*
 117  * swapon tell device that all the old swap contents can be discarded,
 118  * to allow the swap device to optimize its wear-levelling.
 119  */
 120 static int discard_swap(struct swap_info_struct *si)
 121 {
 122         struct swap_extent *se;
 123         int err = 0;
 124
 125         list_for_each_entry(se, &si->extent_list, list) {
 126                 sector_t start_block = se->start_block << (PAGE_SHIFT - 9);
 127                 sector_t nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 128
 129                 if (se->start_page == 0) {
 130                         /* Do not discard the swap header page! */
 131                         start_block += 1 << (PAGE_SHIFT - 9);
 132                         nr_blocks -= 1 << (PAGE_SHIFT - 9);
 133                         if (!nr_blocks)
 134                                 continue;
 135                 }
 136
 137                 err = blkdev_issue_discard(si->bdev, start_block,
 138                                                 nr_blocks, GFP_KERNEL);
 139                 if (err)
 140                         break;
 141
 142                 cond_resched();
 143         }
 144         return err;             /* That will often be -EOPNOTSUPP */
 145 }
 146
 147 /*
 148  * swap allocation tell device that a cluster of swap can now be discarded,
 149  * to allow the swap device to optimize its wear-levelling.
 150  */
 151 static void discard_swap_cluster(struct swap_info_struct *si,
 152                                  pgoff_t start_page, pgoff_t nr_pages)
 153 {
 154         struct swap_extent *se = si->curr_swap_extent;
 155         int found_extent = 0;
 156
 157         while (nr_pages) {
 158                 struct list_head *lh;
 159
 160                 if (se->start_page <= start_page &&
 161                     start_page < se->start_page + se->nr_pages) {
 162                         pgoff_t offset = start_page - se->start_page;
 163                         sector_t start_block = se->start_block + offset;
 164                         sector_t nr_blocks = se->nr_pages - offset;
 165
 166                         if (nr_blocks > nr_pages)
 167                                 nr_blocks = nr_pages;
 168                         start_page += nr_blocks;
 169                         nr_pages -= nr_blocks;
 170
 171                         if (!found_extent++)
 172                                 si->curr_swap_extent = se;
 173
 174                         start_block <<= PAGE_SHIFT - 9;
 175                         nr_blocks <<= PAGE_SHIFT - 9;
 176                         if (blkdev_issue_discard(si->bdev, start_block,
 177                                                         nr_blocks, GFP_NOIO))
 178                                 break;
 179                 }
 180
 181                 lh = se->list.next;
 182                 if (lh == &si->extent_list)
 183                         lh = lh->next;
 184                 se = list_entry(lh, struct swap_extent, list);
 185         }
 186 }
 187
 188 static int wait_for_discard(void *word)
 189 {
 190         schedule();
 191         return 0;
 192 }
 193
 194 #define SWAPFILE_CLUSTER        256
 195 #define LATENCY_LIMIT           256
 196
 197 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
 198                                           int cache)
 199 {
 200         unsigned long offset;
 201         unsigned long scan_base;
 202         unsigned long last_in_cluster = 0;
 203         int latency_ration = LATENCY_LIMIT;
 204         int found_free_cluster = 0;
 205
 206         /*
 207          * We try to cluster swap pages by allocating them sequentially
 208          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 209          * way, however, we resort to first-free allocation, starting
 210          * a new cluster.  This prevents us from scattering swap pages
 211          * all over the entire swap partition, so that we reduce
 212          * overall disk seek times between swap pages.  -- sct
 213          * But we do now try to find an empty cluster.  -Andrea
 214          * And we let swap pages go all over an SSD partition.  Hugh
 215          */
 216
 217         si->flags += SWP_SCANNING;
 218         scan_base = offset = si->cluster_next;
 219
 220         if (unlikely(!si->cluster_nr--)) {
 221                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 222                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
 223                         goto checks;
 224                 }
 225                 if (si->flags & SWP_DISCARDABLE) {
 226                         /*
 227                          * Start range check on racing allocations, in case
 228                          * they overlap the cluster we eventually decide on
 229                          * (we scan without swap_lock to allow preemption).
 230                          * It's hardly conceivable that cluster_nr could be
 231                          * wrapped during our scan, but don't depend on it.
 232                          */
 233                         if (si->lowest_alloc)
 234                                 goto checks;
 235                         si->lowest_alloc = si->max;
 236                         si->highest_alloc = 0;
 237                 }
 238                 spin_unlock(&swap_lock);
 239
 240                 /*
 241                  * If seek is expensive, start searching for new cluster from
 242                  * start of partition, to minimize the span of allocated swap.
 243                  * But if seek is cheap, search from our current position, so
 244                  * that swap is allocated from all over the partition: if the
 245                  * Flash Translation Layer only remaps within limited zones,
 246                  * we don't want to wear out the first zone too quickly.
 247                  */
 248                 if (!(si->flags & SWP_SOLIDSTATE))
 249                         scan_base = offset = si->lowest_bit;
 250                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 251
 252                 /* Locate the first empty (unaligned) cluster */
 253                 for (; last_in_cluster <= si->highest_bit; offset++) {
 254                         if (si->swap_map[offset])
 255                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 256                         else if (offset == last_in_cluster) {
 257                                 spin_lock(&swap_lock);
 258                                 offset -= SWAPFILE_CLUSTER - 1;
 259                                 si->cluster_next = offset;
 260                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 261                                 found_free_cluster = 1;
 262                                 goto checks;
 263                         }
 264                         if (unlikely(--latency_ration < 0)) {
 265                                 cond_resched();
 266                                 latency_ration = LATENCY_LIMIT;
 267                         }
 268                 }
 269
 270                 offset = si->lowest_bit;
 271                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 272
 273                 /* Locate the first empty (unaligned) cluster */
 274                 for (; last_in_cluster < scan_base; offset++) {
 275                         if (si->swap_map[offset])
 276                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 277                         else if (offset == last_in_cluster) {
 278                                 spin_lock(&swap_lock);
 279                                 offset -= SWAPFILE_CLUSTER - 1;
 280                                 si->cluster_next = offset;
 281                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 282                                 found_free_cluster = 1;
 283                                 goto checks;
 284                         }
 285                         if (unlikely(--latency_ration < 0)) {
 286                                 cond_resched();
 287                                 latency_ration = LATENCY_LIMIT;
 288                         }
 289                 }
 290
 291                 offset = scan_base;
 292                 spin_lock(&swap_lock);
 293                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 294                 si->lowest_alloc = 0;
 295         }
 296
 297 checks:
 298         if (!(si->flags & SWP_WRITEOK))
 299                 goto no_page;
 300         if (!si->highest_bit)
 301                 goto no_page;
 302         if (offset > si->highest_bit)
 303                 scan_base = offset = si->lowest_bit;
 304         if (si->swap_map[offset])
 305                 goto scan;
 306
 307         if (offset == si->lowest_bit)
 308                 si->lowest_bit++;
 309         if (offset == si->highest_bit)
 310                 si->highest_bit--;
 311         si->inuse_pages++;
 312         if (si->inuse_pages == si->pages) {
 313                 si->lowest_bit = si->max;
 314                 si->highest_bit = 0;
 315         }
 316         if (cache == SWAP_CACHE) /* at usual swap-out via vmscan.c */
 317                 si->swap_map[offset] = encode_swapmap(0, true);
 318         else /* at suspend */
 319                 si->swap_map[offset] = encode_swapmap(1, false);
 320         si->cluster_next = offset + 1;
 321         si->flags -= SWP_SCANNING;
 322
 323         if (si->lowest_alloc) {
 324                 /*
 325                  * Only set when SWP_DISCARDABLE, and there's a scan
 326                  * for a free cluster in progress or just completed.
 327                  */
 328                 if (found_free_cluster) {
 329                         /*
 330                          * To optimize wear-levelling, discard the
 331                          * old data of the cluster, taking care not to
 332                          * discard any of its pages that have already
 333                          * been allocated by racing tasks (offset has
 334                          * already stepped over any at the beginning).
 335                          */
 336                         if (offset < si->highest_alloc &&
 337                             si->lowest_alloc <= last_in_cluster)
 338                                 last_in_cluster = si->lowest_alloc - 1;
 339                         si->flags |= SWP_DISCARDING;
 340                         spin_unlock(&swap_lock);
 341
 342                         if (offset < last_in_cluster)
 343                                 discard_swap_cluster(si, offset,
 344                                         last_in_cluster - offset + 1);
 345
 346                         spin_lock(&swap_lock);
 347                         si->lowest_alloc = 0;
 348                         si->flags &= ~SWP_DISCARDING;
 349
 350                         smp_mb();       /* wake_up_bit advises this */
 351                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
 352
 353                 } else if (si->flags & SWP_DISCARDING) {
 354                         /*
 355                          * Delay using pages allocated by racing tasks
 356                          * until the whole discard has been issued. We
 357                          * could defer that delay until swap_writepage,
 358                          * but it's easier to keep this self-contained.
 359                          */
 360                         spin_unlock(&swap_lock);
 361                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
 362                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
 363                         spin_lock(&swap_lock);
 364                 } else {
 365                         /*
 366                          * Note pages allocated by racing tasks while
 367                          * scan for a free cluster is in progress, so
 368                          * that its final discard can exclude them.
 369                          */
 370                         if (offset < si->lowest_alloc)
 371                                 si->lowest_alloc = offset;
 372                         if (offset > si->highest_alloc)
 373                                 si->highest_alloc = offset;
 374                 }
 375         }
 376         return offset;
 377
 378 scan:
 379         spin_unlock(&swap_lock);
 380         while (++offset <= si->highest_bit) {
 381                 if (!si->swap_map[offset]) {
 382                         spin_lock(&swap_lock);
 383                         goto checks;
 384                 }
 385                 if (unlikely(--latency_ration < 0)) {
 386                         cond_resched();
 387                         latency_ration = LATENCY_LIMIT;
 388                 }
 389         }
 390         offset = si->lowest_bit;
 391         while (++offset < scan_base) {
 392                 if (!si->swap_map[offset]) {
 393                         spin_lock(&swap_lock);
 394                         goto checks;
 395                 }
 396                 if (unlikely(--latency_ration < 0)) {
 397                         cond_resched();
 398                         latency_ration = LATENCY_LIMIT;
 399                 }
 400         }
 401         spin_lock(&swap_lock);
 402
 403 no_page:
 404         si->flags -= SWP_SCANNING;
 405         return 0;
 406 }
 407
 408 swp_entry_t get_swap_page(void)
 409 {
 410         struct swap_info_struct *si;
 411         pgoff_t offset;
 412         int type, next;
 413         int wrapped = 0;
 414
 415         spin_lock(&swap_lock);
 416         if (nr_swap_pages <= 0)
 417                 goto noswap;
 418         nr_swap_pages--;
 419
 420         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 421                 si = swap_info + type;
 422                 next = si->next;
 423                 if (next < 0 ||
 424                     (!wrapped && si->prio != swap_info[next].prio)) {
 425                         next = swap_list.head;
 426                         wrapped++;
 427                 }
 428
 429                 if (!si->highest_bit)
 430                         continue;
 431                 if (!(si->flags & SWP_WRITEOK))
 432                         continue;
 433
 434                 swap_list.next = next;
 435                 /* This is called for allocating swap entry for cache */
 436                 offset = scan_swap_map(si, SWAP_CACHE);
 437                 if (offset) {
 438                         spin_unlock(&swap_lock);
 439                         return swp_entry(type, offset);
 440                 }
 441                 next = swap_list.next;
 442         }
 443
 444         nr_swap_pages++;
 445 noswap:
 446         spin_unlock(&swap_lock);
 447         return (swp_entry_t) {0};
 448 }
 449
 450 /* The only caller of this function is now susupend routine */
 451 swp_entry_t get_swap_page_of_type(int type)
 452 {
 453         struct swap_info_struct *si;
 454         pgoff_t offset;
 455
 456         spin_lock(&swap_lock);
 457         si = swap_info + type;
 458         if (si->flags & SWP_WRITEOK) {
 459                 nr_swap_pages--;
 460                 /* This is called for allocating swap entry, not cache */
 461                 offset = scan_swap_map(si, SWAP_MAP);
 462                 if (offset) {
 463                         spin_unlock(&swap_lock);
 464                         return swp_entry(type, offset);
 465                 }
 466                 nr_swap_pages++;
 467         }
 468         spin_unlock(&swap_lock);
 469         return (swp_entry_t) {0};
 470 }
 471
 472 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 473 {
 474         struct swap_info_struct * p;
 475         unsigned long offset, type;
 476
 477         if (!entry.val)
 478                 goto out;
 479         type = swp_type(entry);
 480         if (type >= nr_swapfiles)
 481                 goto bad_nofile;
 482         p = & swap_info[type];
 483         if (!(p->flags & SWP_USED))
 484                 goto bad_device;
 485         offset = swp_offset(entry);
 486         if (offset >= p->max)
 487                 goto bad_offset;
 488         if (!p->swap_map[offset])
 489                 goto bad_free;
 490         spin_lock(&swap_lock);
 491         return p;
 492
 493 bad_free:
 494         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 495         goto out;
 496 bad_offset:
 497         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 498         goto out;
 499 bad_device:
 500         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 501         goto out;
 502 bad_nofile:
 503         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 504 out:
 505         return NULL;
 506 }
 507
 508 static int swap_entry_free(struct swap_info_struct *p,
 509                            swp_entry_t ent, int cache)
 510 {
 511         unsigned long offset = swp_offset(ent);
 512         int count = swap_count(p->swap_map[offset]);
 513         bool has_cache;
 514
 515         has_cache = swap_has_cache(p->swap_map[offset]);
 516
 517         if (cache == SWAP_MAP) { /* dropping usage count of swap */
 518                 if (count < SWAP_MAP_MAX) {
 519                         count--;
 520                         p->swap_map[offset] = encode_swapmap(count, has_cache);
 521                 }
 522         } else { /* dropping swap cache flag */
 523                 VM_BUG_ON(!has_cache);
 524                 p->swap_map[offset] = encode_swapmap(count, false);
 525
 526         }
 527         /* return code. */
 528         count = p->swap_map[offset];
 529         /* free if no reference */
 530         if (!count) {
 531                 if (offset < p->lowest_bit)
 532                         p->lowest_bit = offset;
 533                 if (offset > p->highest_bit)
 534                         p->highest_bit = offset;
 535                 if (p->prio > swap_info[swap_list.next].prio)
 536                         swap_list.next = p - swap_info;
 537                 nr_swap_pages++;
 538                 p->inuse_pages--;
 539                 mem_cgroup_uncharge_swap(ent);
 540         }
 541         return count;
 542 }
 543
 544 /*
 545  * Caller has made sure that the swapdevice corresponding to entry
 546  * is still around or has not been recycled.
 547  */
 548 void swap_free(swp_entry_t entry)
 549 {
 550         struct swap_info_struct * p;
 551
 552         p = swap_info_get(entry);
 553         if (p) {
 554                 swap_entry_free(p, entry, SWAP_MAP);
 555                 spin_unlock(&swap_lock);
 556         }
 557 }
 558
 559 /*
 560  * Called after dropping swapcache to decrease refcnt to swap entries.
 561  */
 562 void swapcache_free(swp_entry_t entry, struct page *page)
 563 {
 564         struct swap_info_struct *p;
 565
 566         if (page)
 567                 mem_cgroup_uncharge_swapcache(page, entry);
 568         p = swap_info_get(entry);
 569         if (p) {
 570                 swap_entry_free(p, entry, SWAP_CACHE);
 571                 spin_unlock(&swap_lock);
 572         }
 573         return;
 574 }
 575
 576 /*
 577  * How many references to page are currently swapped out?
 578  */
 579 static inline int page_swapcount(struct page *page)
 580 {
 581         int count = 0;
 582         struct swap_info_struct *p;
 583         swp_entry_t entry;
 584
 585         entry.val = page_private(page);
 586         p = swap_info_get(entry);
 587         if (p) {
 588                 count = swap_count(p->swap_map[swp_offset(entry)]);
 589                 spin_unlock(&swap_lock);
 590         }
 591         return count;
 592 }
 593
 594 /*
 595  * We can write to an anon page without COW if there are no other references
 596  * to it.  And as a side-effect, free up its swap: because the old content
 597  * on disk will never be read, and seeking back there to write new content
 598  * later would only waste time away from clustering.
 599  */
 600 int reuse_swap_page(struct page *page)
 601 {
 602         int count;
 603
 604         VM_BUG_ON(!PageLocked(page));
 605         count = page_mapcount(page);
 606         if (count <= 1 && PageSwapCache(page)) {
 607                 count += page_swapcount(page);
 608                 if (count == 1 && !PageWriteback(page)) {
 609                         delete_from_swap_cache(page);
 610                         SetPageDirty(page);
 611                 }
 612         }
 613         return count == 1;
 614 }
 615
 616 /*
 617  * If swap is getting full, or if there are no more mappings of this page,
 618  * then try_to_free_swap is called to free its swap space.
 619  */
 620 int try_to_free_swap(struct page *page)
 621 {
 622         VM_BUG_ON(!PageLocked(page));
 623
 624         if (!PageSwapCache(page))
 625                 return 0;
 626         if (PageWriteback(page))
 627                 return 0;
 628         if (page_swapcount(page))
 629                 return 0;
 630
 631         delete_from_swap_cache(page);
 632         SetPageDirty(page);
 633         return 1;
 634 }
 635
 636 /*
 637  * Free the swap entry like above, but also try to
 638  * free the page cache entry if it is the last user.
 639  */
 640 int free_swap_and_cache(swp_entry_t entry)
 641 {
 642         struct swap_info_struct *p;
 643         struct page *page = NULL;
 644
 645         if (is_migration_entry(entry))
 646                 return 1;
 647
 648         p = swap_info_get(entry);
 649         if (p) {
 650                 if (swap_entry_free(p, entry, SWAP_MAP) == SWAP_HAS_CACHE) {
 651                         page = find_get_page(&swapper_space, entry.val);
 652                         if (page && !trylock_page(page)) {
 653                                 page_cache_release(page);
 654                                 page = NULL;
 655                         }
 656                 }
 657                 spin_unlock(&swap_lock);
 658         }
 659         if (page) {
 660                 /*
 661                  * Not mapped elsewhere, or swap space full? Free it!
 662                  * Also recheck PageSwapCache now page is locked (above).
 663                  */
 664                 if (PageSwapCache(page) && !PageWriteback(page) &&
 665                                 (!page_mapped(page) || vm_swap_full())) {
 666                         delete_from_swap_cache(page);
 667                         SetPageDirty(page);
 668                 }
 669                 unlock_page(page);
 670                 page_cache_release(page);
 671         }
 672         return p != NULL;
 673 }
 674
 675 #ifdef CONFIG_HIBERNATION
 676 /*
 677  * Find the swap type that corresponds to given device (if any).
 678  *
 679  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 680  * from 0, in which the swap header is expected to be located.
 681  *
 682  * This is needed for the suspend to disk (aka swsusp).
 683  */
 684 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 685 {
 686         struct block_device *bdev = NULL;
 687         int i;
 688
 689         if (device)
 690                 bdev = bdget(device);
 691
 692         spin_lock(&swap_lock);
 693         for (i = 0; i < nr_swapfiles; i++) {
 694                 struct swap_info_struct *sis = swap_info + i;
 695
 696                 if (!(sis->flags & SWP_WRITEOK))
 697                         continue;
 698
 699                 if (!bdev) {
 700                         if (bdev_p)
 701                                 *bdev_p = bdget(sis->bdev->bd_dev);
 702
 703                         spin_unlock(&swap_lock);
 704                         return i;
 705                 }
 706                 if (bdev == sis->bdev) {
 707                         struct swap_extent *se;
 708
 709                         se = list_entry(sis->extent_list.next,
 710                                         struct swap_extent, list);
 711                         if (se->start_block == offset) {
 712                                 if (bdev_p)
 713                                         *bdev_p = bdget(sis->bdev->bd_dev);
 714
 715                                 spin_unlock(&swap_lock);
 716                                 bdput(bdev);
 717                                 return i;
 718                         }
 719                 }
 720         }
 721         spin_unlock(&swap_lock);
 722         if (bdev)
 723                 bdput(bdev);
 724
 725         return -ENODEV;
 726 }
 727
 728 /*
 729  * Return either the total number of swap pages of given type, or the number
 730  * of free pages of that type (depending on @free)
 731  *
 732  * This is needed for software suspend
 733  */
 734 unsigned int count_swap_pages(int type, int free)
 735 {
 736         unsigned int n = 0;
 737
 738         if (type < nr_swapfiles) {
 739                 spin_lock(&swap_lock);
 740                 if (swap_info[type].flags & SWP_WRITEOK) {
 741                         n = swap_info[type].pages;
 742                         if (free)
 743                                 n -= swap_info[type].inuse_pages;
 744                 }
 745                 spin_unlock(&swap_lock);
 746         }
 747         return n;
 748 }
 749 #endif
 750
 751 /*
 752  * No need to decide whether this PTE shares the swap entry with others,
 753  * just let do_wp_page work it out if a write is requested later - to
 754  * force COW, vm_page_prot omits write permission from any private vma.
 755  */
 756 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 757                 unsigned long addr, swp_entry_t entry, struct page *page)
 758 {
 759         struct mem_cgroup *ptr = NULL;
 760         spinlock_t *ptl;
 761         pte_t *pte;
 762         int ret = 1;
 763
 764         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
 765                 ret = -ENOMEM;
 766                 goto out_nolock;
 767         }
 768
 769         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 770         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 771                 if (ret > 0)
 772                         mem_cgroup_cancel_charge_swapin(ptr);
 773                 ret = 0;
 774                 goto out;
 775         }
 776
 777         inc_mm_counter(vma->vm_mm, anon_rss);
 778         get_page(page);
 779         set_pte_at(vma->vm_mm, addr, pte,
 780                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 781         page_add_anon_rmap(page, vma, addr);
 782         mem_cgroup_commit_charge_swapin(page, ptr);
 783         swap_free(entry);
 784         /*
 785          * Move the page to the active list so it is not
 786          * immediately swapped out again after swapon.
 787          */
 788         activate_page(page);
 789 out:
 790         pte_unmap_unlock(pte, ptl);
 791 out_nolock:
 792         return ret;
 793 }
 794
 795 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 796                                 unsigned long addr, unsigned long end,
 797                                 swp_entry_t entry, struct page *page)
 798 {
 799         pte_t swp_pte = swp_entry_to_pte(entry);
 800         pte_t *pte;
 801         int ret = 0;
 802
 803         /*
 804          * We don't actually need pte lock while scanning for swp_pte: since
 805          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 806          * page table while we're scanning; though it could get zapped, and on
 807          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 808          * of unmatched parts which look like swp_pte, so unuse_pte must
 809          * recheck under pte lock.  Scanning without pte lock lets it be
 810          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 811          */
 812         pte = pte_offset_map(pmd, addr);
 813         do {
 814                 /*
 815                  * swapoff spends a _lot_ of time in this loop!
 816                  * Test inline before going to call unuse_pte.
 817                  */
 818                 if (unlikely(pte_same(*pte, swp_pte))) {
 819                         pte_unmap(pte);
 820                         ret = unuse_pte(vma, pmd, addr, entry, page);
 821                         if (ret)
 822                                 goto out;
 823                         pte = pte_offset_map(pmd, addr);
 824                 }
 825         } while (pte++, addr += PAGE_SIZE, addr != end);
 826         pte_unmap(pte - 1);
 827 out:
 828         return ret;
 829 }
 830
 831 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 832                                 unsigned long addr, unsigned long end,
 833                                 swp_entry_t entry, struct page *page)
 834 {
 835         pmd_t *pmd;
 836         unsigned long next;
 837         int ret;
 838
 839         pmd = pmd_offset(pud, addr);
 840         do {
 841                 next = pmd_addr_end(addr, end);
 842                 if (pmd_none_or_clear_bad(pmd))
 843                         continue;
 844                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 845                 if (ret)
 846                         return ret;
 847         } while (pmd++, addr = next, addr != end);
 848         return 0;
 849 }
 850
 851 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 852                                 unsigned long addr, unsigned long end,
 853                                 swp_entry_t entry, struct page *page)
 854 {
 855         pud_t *pud;
 856         unsigned long next;
 857         int ret;
 858
 859         pud = pud_offset(pgd, addr);
 860         do {
 861                 next = pud_addr_end(addr, end);
 862                 if (pud_none_or_clear_bad(pud))
 863                         continue;
 864                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 865                 if (ret)
 866                         return ret;
 867         } while (pud++, addr = next, addr != end);
 868         return 0;
 869 }
 870
 871 static int unuse_vma(struct vm_area_struct *vma,
 872                                 swp_entry_t entry, struct page *page)
 873 {
 874         pgd_t *pgd;
 875         unsigned long addr, end, next;
 876         int ret;
 877
 878         if (page->mapping) {
 879                 addr = page_address_in_vma(page, vma);
 880                 if (addr == -EFAULT)
 881                         return 0;
 882                 else
 883                         end = addr + PAGE_SIZE;
 884         } else {
 885                 addr = vma->vm_start;
 886                 end = vma->vm_end;
 887         }
 888
 889         pgd = pgd_offset(vma->vm_mm, addr);
 890         do {
 891                 next = pgd_addr_end(addr, end);
 892                 if (pgd_none_or_clear_bad(pgd))
 893                         continue;
 894                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 895                 if (ret)
 896                         return ret;
 897         } while (pgd++, addr = next, addr != end);
 898         return 0;
 899 }
 900
 901 static int unuse_mm(struct mm_struct *mm,
 902                                 swp_entry_t entry, struct page *page)
 903 {
 904         struct vm_area_struct *vma;
 905         int ret = 0;
 906
 907         if (!down_read_trylock(&mm->mmap_sem)) {
 908                 /*
 909                  * Activate page so shrink_inactive_list is unlikely to unmap
 910                  * its ptes while lock is dropped, so swapoff can make progress.
 911                  */
 912                 activate_page(page);
 913                 unlock_page(page);
 914                 down_read(&mm->mmap_sem);
 915                 lock_page(page);
 916         }
 917         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 918                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
 919                         break;
 920         }
 921         up_read(&mm->mmap_sem);
 922         return (ret < 0)? ret: 0;
 923 }
 924
 925 /*
 926  * Scan swap_map from current position to next entry still in use.
 927  * Recycle to start on reaching the end, returning 0 when empty.
 928  */
 929 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 930                                         unsigned int prev)
 931 {
 932         unsigned int max = si->max;
 933         unsigned int i = prev;
 934         int count;
 935
 936         /*
 937          * No need for swap_lock here: we're just looking
 938          * for whether an entry is in use, not modifying it; false
 939          * hits are okay, and sys_swapoff() has already prevented new
 940          * allocations from this area (while holding swap_lock).
 941          */
 942         for (;;) {
 943                 if (++i >= max) {
 944                         if (!prev) {
 945                                 i = 0;
 946                                 break;
 947                         }
 948                         /*
 949                          * No entries in use at top of swap_map,
 950                          * loop back to start and recheck there.
 951                          */
 952                         max = prev + 1;
 953                         prev = 0;
 954                         i = 1;
 955                 }
 956                 count = si->swap_map[i];
 957                 if (count && swap_count(count) != SWAP_MAP_BAD)
 958                         break;
 959         }
 960         return i;
 961 }
 962
 963 /*
 964  * We completely avoid races by reading each swap page in advance,
 965  * and then search for the process using it.  All the necessary
 966  * page table adjustments can then be made atomically.
 967  */
 968 static int try_to_unuse(unsigned int type)
 969 {
 970         struct swap_info_struct * si = &swap_info[type];
 971         struct mm_struct *start_mm;
 972         unsigned short *swap_map;
 973         unsigned short swcount;
 974         struct page *page;
 975         swp_entry_t entry;
 976         unsigned int i = 0;
 977         int retval = 0;
 978         int reset_overflow = 0;
 979         int shmem;
 980
 981         /*
 982          * When searching mms for an entry, a good strategy is to
 983          * start at the first mm we freed the previous entry from
 984          * (though actually we don't notice whether we or coincidence
 985          * freed the entry).  Initialize this start_mm with a hold.
 986          *
 987          * A simpler strategy would be to start at the last mm we
 988          * freed the previous entry from; but that would take less
 989          * advantage of mmlist ordering, which clusters forked mms
 990          * together, child after parent.  If we race with dup_mmap(), we
 991          * prefer to resolve parent before child, lest we miss entries
 992          * duplicated after we scanned child: using last mm would invert
 993          * that.  Though it's only a serious concern when an overflowed
 994          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 995          */
 996         start_mm = &init_mm;
 997         atomic_inc(&init_mm.mm_users);
 998
 999         /*
1000          * Keep on scanning until all entries have gone.  Usually,
1001          * one pass through swap_map is enough, but not necessarily:
1002          * there are races when an instance of an entry might be missed.
1003          */
1004         while ((i = find_next_to_unuse(si, i)) != 0) {
1005                 if (signal_pending(current)) {
1006                         retval = -EINTR;
1007                         break;
1008                 }
1009
1010                 /*
1011                  * Get a page for the entry, using the existing swap
1012                  * cache page if there is one.  Otherwise, get a clean
1013                  * page and read the swap into it.
1014                  */
1015                 swap_map = &si->swap_map[i];
1016                 entry = swp_entry(type, i);
1017                 page = read_swap_cache_async(entry,
1018                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1019                 if (!page) {
1020                         /*
1021                          * Either swap_duplicate() failed because entry
1022                          * has been freed independently, and will not be
1023                          * reused since sys_swapoff() already disabled
1024                          * allocation from here, or alloc_page() failed.
1025                          */
1026                         if (!*swap_map)
1027                                 continue;
1028                         retval = -ENOMEM;
1029                         break;
1030                 }
1031
1032                 /*
1033                  * Don't hold on to start_mm if it looks like exiting.
1034                  */
1035                 if (atomic_read(&start_mm->mm_users) == 1) {
1036                         mmput(start_mm);
1037                         start_mm = &init_mm;
1038                         atomic_inc(&init_mm.mm_users);
1039                 }
1040
1041                 /*
1042                  * Wait for and lock page.  When do_swap_page races with
1043                  * try_to_unuse, do_swap_page can handle the fault much
1044                  * faster than try_to_unuse can locate the entry.  This
1045                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1046                  * defer to do_swap_page in such a case - in some tests,
1047                  * do_swap_page and try_to_unuse repeatedly compete.
1048                  */
1049                 wait_on_page_locked(page);
1050                 wait_on_page_writeback(page);
1051                 lock_page(page);
1052                 wait_on_page_writeback(page);
1053
1054                 /*
1055                  * Remove all references to entry.
1056                  * Whenever we reach init_mm, there's no address space
1057                  * to search, but use it as a reminder to search shmem.
1058                  */
1059                 shmem = 0;
1060                 swcount = *swap_map;
1061                 if (swap_count(swcount)) {
1062                         if (start_mm == &init_mm)
1063                                 shmem = shmem_unuse(entry, page);
1064                         else
1065                                 retval = unuse_mm(start_mm, entry, page);
1066                 }
1067                 if (swap_count(*swap_map)) {
1068                         int set_start_mm = (*swap_map >= swcount);
1069                         struct list_head *p = &start_mm->mmlist;
1070                         struct mm_struct *new_start_mm = start_mm;
1071                         struct mm_struct *prev_mm = start_mm;
1072                         struct mm_struct *mm;
1073
1074                         atomic_inc(&new_start_mm->mm_users);
1075                         atomic_inc(&prev_mm->mm_users);
1076                         spin_lock(&mmlist_lock);
1077                         while (swap_count(*swap_map) && !retval && !shmem &&
1078                                         (p = p->next) != &start_mm->mmlist) {
1079                                 mm = list_entry(p, struct mm_struct, mmlist);
1080                                 if (!atomic_inc_not_zero(&mm->mm_users))
1081                                         continue;
1082                                 spin_unlock(&mmlist_lock);
1083                                 mmput(prev_mm);
1084                                 prev_mm = mm;
1085
1086                                 cond_resched();
1087
1088                                 swcount = *swap_map;
1089                                 if (!swap_count(swcount)) /* any usage ? */
1090                                         ;
1091                                 else if (mm == &init_mm) {
1092                                         set_start_mm = 1;
1093                                         shmem = shmem_unuse(entry, page);
1094                                 } else
1095                                         retval = unuse_mm(mm, entry, page);
1096
1097                                 if (set_start_mm &&
1098                                     swap_count(*swap_map) < swcount) {
1099                                         mmput(new_start_mm);
1100                                         atomic_inc(&mm->mm_users);
1101                                         new_start_mm = mm;
1102                                         set_start_mm = 0;
1103                                 }
1104                                 spin_lock(&mmlist_lock);
1105                         }
1106                         spin_unlock(&mmlist_lock);
1107                         mmput(prev_mm);
1108                         mmput(start_mm);
1109                         start_mm = new_start_mm;
1110                 }
1111                 if (shmem) {
1112                         /* page has already been unlocked and released */
1113                         if (shmem > 0)
1114                                 continue;
1115                         retval = shmem;
1116                         break;
1117                 }
1118                 if (retval) {
1119                         unlock_page(page);
1120                         page_cache_release(page);
1121                         break;
1122                 }
1123
1124                 /*
1125                  * How could swap count reach 0x7ffe ?
1126                  * There's no way to repeat a swap page within an mm
1127                  * (except in shmem, where it's the shared object which takes
1128                  * the reference count)?
1129                  * We believe SWAP_MAP_MAX cannot occur.(if occur, unsigned
1130                  * short is too small....)
1131                  * If that's wrong, then we should worry more about
1132                  * exit_mmap() and do_munmap() cases described above:
1133                  * we might be resetting SWAP_MAP_MAX too early here.
1134                  * We know "Undead"s can happen, they're okay, so don't
1135                  * report them; but do report if we reset SWAP_MAP_MAX.
1136                  */
1137                 /* We might release the lock_page() in unuse_mm(). */
1138                 if (!PageSwapCache(page) || page_private(page) != entry.val)
1139                         goto retry;
1140
1141                 if (swap_count(*swap_map) == SWAP_MAP_MAX) {
1142                         spin_lock(&swap_lock);
1143                         *swap_map = encode_swapmap(0, true);
1144                         spin_unlock(&swap_lock);
1145                         reset_overflow = 1;
1146                 }
1147
1148                 /*
1149                  * If a reference remains (rare), we would like to leave
1150                  * the page in the swap cache; but try_to_unmap could
1151                  * then re-duplicate the entry once we drop page lock,
1152                  * so we might loop indefinitely; also, that page could
1153                  * not be swapped out to other storage meanwhile.  So:
1154                  * delete from cache even if there's another reference,
1155                  * after ensuring that the data has been saved to disk -
1156                  * since if the reference remains (rarer), it will be
1157                  * read from disk into another page.  Splitting into two
1158                  * pages would be incorrect if swap supported "shared
1159                  * private" pages, but they are handled by tmpfs files.
1160                  */
1161                 if (swap_count(*swap_map) &&
1162                      PageDirty(page) && PageSwapCache(page)) {
1163                         struct writeback_control wbc = {
1164                                 .sync_mode = WB_SYNC_NONE,
1165                         };
1166
1167                         swap_writepage(page, &wbc);
1168                         lock_page(page);
1169                         wait_on_page_writeback(page);
1170                 }
1171
1172                 /*
1173                  * It is conceivable that a racing task removed this page from
1174                  * swap cache just before we acquired the page lock at the top,
1175                  * or while we dropped it in unuse_mm().  The page might even
1176                  * be back in swap cache on another swap area: that we must not
1177                  * delete, since it may not have been written out to swap yet.
1178                  */
1179                 if (PageSwapCache(page) &&
1180                     likely(page_private(page) == entry.val))
1181                         delete_from_swap_cache(page);
1182
1183                 /*
1184                  * So we could skip searching mms once swap count went
1185                  * to 1, we did not mark any present ptes as dirty: must
1186                  * mark page dirty so shrink_page_list will preserve it.
1187                  */
1188                 SetPageDirty(page);
1189 retry:
1190                 unlock_page(page);
1191                 page_cache_release(page);
1192
1193                 /*
1194                  * Make sure that we aren't completely killing
1195                  * interactive performance.
1196                  */
1197                 cond_resched();
1198         }
1199
1200         mmput(start_mm);
1201         if (reset_overflow) {
1202                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1203                 swap_overflow = 0;
1204         }
1205         return retval;
1206 }
1207
1208 /*
1209  * After a successful try_to_unuse, if no swap is now in use, we know
1210  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1211  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1212  * added to the mmlist just after page_duplicate - before would be racy.
1213  */
1214 static void drain_mmlist(void)
1215 {
1216         struct list_head *p, *next;
1217         unsigned int i;
1218
1219         for (i = 0; i < nr_swapfiles; i++)
1220                 if (swap_info[i].inuse_pages)
1221                         return;
1222         spin_lock(&mmlist_lock);
1223         list_for_each_safe(p, next, &init_mm.mmlist)
1224                 list_del_init(p);
1225         spin_unlock(&mmlist_lock);
1226 }
1227
1228 /*
1229  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1230  * corresponds to page offset `offset'.
1231  */
1232 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
1233 {
1234         struct swap_extent *se = sis->curr_swap_extent;
1235         struct swap_extent *start_se = se;
1236
1237         for ( ; ; ) {
1238                 struct list_head *lh;
1239
1240                 if (se->start_page <= offset &&
1241                                 offset < (se->start_page + se->nr_pages)) {
1242                         return se->start_block + (offset - se->start_page);
1243                 }
1244                 lh = se->list.next;
1245                 if (lh == &sis->extent_list)
1246                         lh = lh->next;
1247                 se = list_entry(lh, struct swap_extent, list);
1248                 sis->curr_swap_extent = se;
1249                 BUG_ON(se == start_se);         /* It *must* be present */
1250         }
1251 }
1252
1253 #ifdef CONFIG_HIBERNATION
1254 /*
1255  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1256  * corresponding to given index in swap_info (swap type).
1257  */
1258 sector_t swapdev_block(int swap_type, pgoff_t offset)
1259 {
1260         struct swap_info_struct *sis;
1261
1262         if (swap_type >= nr_swapfiles)
1263                 return 0;
1264
1265         sis = swap_info + swap_type;
1266         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1267 }
1268 #endif /* CONFIG_HIBERNATION */
1269
1270 /*
1271  * Free all of a swapdev's extent information
1272  */
1273 static void destroy_swap_extents(struct swap_info_struct *sis)
1274 {
1275         while (!list_empty(&sis->extent_list)) {
1276                 struct swap_extent *se;
1277
1278                 se = list_entry(sis->extent_list.next,
1279                                 struct swap_extent, list);
1280                 list_del(&se->list);
1281                 kfree(se);
1282         }
1283 }
1284
1285 /*
1286  * Add a block range (and the corresponding page range) into this swapdev's
1287  * extent list.  The extent list is kept sorted in page order.
1288  *
1289  * This function rather assumes that it is called in ascending page order.
1290  */
1291 static int
1292 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1293                 unsigned long nr_pages, sector_t start_block)
1294 {
1295         struct swap_extent *se;
1296         struct swap_extent *new_se;
1297         struct list_head *lh;
1298
1299         lh = sis->extent_list.prev;     /* The highest page extent */
1300         if (lh != &sis->extent_list) {
1301                 se = list_entry(lh, struct swap_extent, list);
1302                 BUG_ON(se->start_page + se->nr_pages != start_page);
1303                 if (se->start_block + se->nr_pages == start_block) {
1304                         /* Merge it */
1305                         se->nr_pages += nr_pages;
1306                         return 0;
1307                 }
1308         }
1309
1310         /*
1311          * No merge.  Insert a new extent, preserving ordering.
1312          */
1313         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1314         if (new_se == NULL)
1315                 return -ENOMEM;
1316         new_se->start_page = start_page;
1317         new_se->nr_pages = nr_pages;
1318         new_se->start_block = start_block;
1319
1320         list_add_tail(&new_se->list, &sis->extent_list);
1321         return 1;
1322 }
1323
1324 /*
1325  * A `swap extent' is a simple thing which maps a contiguous range of pages
1326  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1327  * is built at swapon time and is then used at swap_writepage/swap_readpage
1328  * time for locating where on disk a page belongs.
1329  *
1330  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1331  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1332  * swap files identically.
1333  *
1334  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1335  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1336  * swapfiles are handled *identically* after swapon time.
1337  *
1338  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1339  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1340  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1341  * requirements, they are simply tossed out - we will never use those blocks
1342  * for swapping.
1343  *
1344  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1345  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1346  * which will scribble on the fs.
1347  *
1348  * The amount of disk space which a single swap extent represents varies.
1349  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1350  * extents in the list.  To avoid much list walking, we cache the previous
1351  * search location in `curr_swap_extent', and start new searches from there.
1352  * This is extremely effective.  The average number of iterations in
1353  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1354  */
1355 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1356 {
1357         struct inode *inode;
1358         unsigned blocks_per_page;
1359         unsigned long page_no;
1360         unsigned blkbits;
1361         sector_t probe_block;
1362         sector_t last_block;
1363         sector_t lowest_block = -1;
1364         sector_t highest_block = 0;
1365         int nr_extents = 0;
1366         int ret;
1367
1368         inode = sis->swap_file->f_mapping->host;
1369         if (S_ISBLK(inode->i_mode)) {
1370                 ret = add_swap_extent(sis, 0, sis->max, 0);
1371                 *span = sis->pages;
1372                 goto done;
1373         }
1374
1375         blkbits = inode->i_blkbits;
1376         blocks_per_page = PAGE_SIZE >> blkbits;
1377
1378         /*
1379          * Map all the blocks into the extent list.  This code doesn't try
1380          * to be very smart.
1381          */
1382         probe_block = 0;
1383         page_no = 0;
1384         last_block = i_size_read(inode) >> blkbits;
1385         while ((probe_block + blocks_per_page) <= last_block &&
1386                         page_no < sis->max) {
1387                 unsigned block_in_page;
1388                 sector_t first_block;
1389
1390                 first_block = bmap(inode, probe_block);
1391                 if (first_block == 0)
1392                         goto bad_bmap;
1393
1394                 /*
1395                  * It must be PAGE_SIZE aligned on-disk
1396                  */
1397                 if (first_block & (blocks_per_page - 1)) {
1398                         probe_block++;
1399                         goto reprobe;
1400                 }
1401
1402                 for (block_in_page = 1; block_in_page < blocks_per_page;
1403                                         block_in_page++) {
1404                         sector_t block;
1405
1406                         block = bmap(inode, probe_block + block_in_page);
1407                         if (block == 0)
1408                                 goto bad_bmap;
1409                         if (block != first_block + block_in_page) {
1410                                 /* Discontiguity */
1411                                 probe_block++;
1412                                 goto reprobe;
1413                         }
1414                 }
1415
1416                 first_block >>= (PAGE_SHIFT - blkbits);
1417                 if (page_no) {  /* exclude the header page */
1418                         if (first_block < lowest_block)
1419                                 lowest_block = first_block;
1420                         if (first_block > highest_block)
1421                                 highest_block = first_block;
1422                 }
1423
1424                 /*
1425                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1426                  */
1427                 ret = add_swap_extent(sis, page_no, 1, first_block);
1428                 if (ret < 0)
1429                         goto out;
1430                 nr_extents += ret;
1431                 page_no++;
1432                 probe_block += blocks_per_page;
1433 reprobe:
1434                 continue;
1435         }
1436         ret = nr_extents;
1437         *span = 1 + highest_block - lowest_block;
1438         if (page_no == 0)
1439                 page_no = 1;    /* force Empty message */
1440         sis->max = page_no;
1441         sis->pages = page_no - 1;
1442         sis->highest_bit = page_no - 1;
1443 done:
1444         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1445                                         struct swap_extent, list);
1446         goto out;
1447 bad_bmap:
1448         printk(KERN_ERR "swapon: swapfile has holes\n");
1449         ret = -EINVAL;
1450 out:
1451         return ret;
1452 }
1453
1454 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1455 {
1456         struct swap_info_struct * p = NULL;
1457         unsigned short *swap_map;
1458         struct file *swap_file, *victim;
1459         struct address_space *mapping;
1460         struct inode *inode;
1461         char * pathname;
1462         int i, type, prev;
1463         int err;
1464
1465         if (!capable(CAP_SYS_ADMIN))
1466                 return -EPERM;
1467
1468         pathname = getname(specialfile);
1469         err = PTR_ERR(pathname);
1470         if (IS_ERR(pathname))
1471                 goto out;
1472
1473         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1474         putname(pathname);
1475         err = PTR_ERR(victim);
1476         if (IS_ERR(victim))
1477                 goto out;
1478
1479         mapping = victim->f_mapping;
1480         prev = -1;
1481         spin_lock(&swap_lock);
1482         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1483                 p = swap_info + type;
1484                 if (p->flags & SWP_WRITEOK) {
1485                         if (p->swap_file->f_mapping == mapping)
1486                                 break;
1487                 }
1488                 prev = type;
1489         }
1490         if (type < 0) {
1491                 err = -EINVAL;
1492                 spin_unlock(&swap_lock);
1493                 goto out_dput;
1494         }
1495         if (!security_vm_enough_memory(p->pages))
1496                 vm_unacct_memory(p->pages);
1497         else {
1498                 err = -ENOMEM;
1499                 spin_unlock(&swap_lock);
1500                 goto out_dput;
1501         }
1502         if (prev < 0) {
1503                 swap_list.head = p->next;
1504         } else {
1505                 swap_info[prev].next = p->next;
1506         }
1507         if (type == swap_list.next) {
1508                 /* just pick something that's safe... */
1509                 swap_list.next = swap_list.head;
1510         }
1511         if (p->prio < 0) {
1512                 for (i = p->next; i >= 0; i = swap_info[i].next)
1513                         swap_info[i].prio = p->prio--;
1514                 least_priority++;
1515         }
1516         nr_swap_pages -= p->pages;
1517         total_swap_pages -= p->pages;
1518         p->flags &= ~SWP_WRITEOK;
1519         spin_unlock(&swap_lock);
1520
1521         current->flags |= PF_SWAPOFF;
1522         err = try_to_unuse(type);
1523         current->flags &= ~PF_SWAPOFF;
1524
1525         if (err) {
1526                 /* re-insert swap space back into swap_list */
1527                 spin_lock(&swap_lock);
1528                 if (p->prio < 0)
1529                         p->prio = --least_priority;
1530                 prev = -1;
1531                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1532                         if (p->prio >= swap_info[i].prio)
1533                                 break;
1534                         prev = i;
1535                 }
1536                 p->next = i;
1537                 if (prev < 0)
1538                         swap_list.head = swap_list.next = p - swap_info;
1539                 else
1540                         swap_info[prev].next = p - swap_info;
1541                 nr_swap_pages += p->pages;
1542                 total_swap_pages += p->pages;
1543                 p->flags |= SWP_WRITEOK;
1544                 spin_unlock(&swap_lock);
1545                 goto out_dput;
1546         }
1547
1548         /* wait for any unplug function to finish */
1549         down_write(&swap_unplug_sem);
1550         up_write(&swap_unplug_sem);
1551
1552         destroy_swap_extents(p);
1553         mutex_lock(&swapon_mutex);
1554         spin_lock(&swap_lock);
1555         drain_mmlist();
1556
1557         /* wait for anyone still in scan_swap_map */
1558         p->highest_bit = 0;             /* cuts scans short */
1559         while (p->flags >= SWP_SCANNING) {
1560                 spin_unlock(&swap_lock);
1561                 schedule_timeout_uninterruptible(1);
1562                 spin_lock(&swap_lock);
1563         }
1564
1565         swap_file = p->swap_file;
1566         p->swap_file = NULL;
1567         p->max = 0;
1568         swap_map = p->swap_map;
1569         p->swap_map = NULL;
1570         p->flags = 0;
1571         spin_unlock(&swap_lock);
1572         mutex_unlock(&swapon_mutex);
1573         vfree(swap_map);
1574         /* Destroy swap account informatin */
1575         swap_cgroup_swapoff(type);
1576
1577         inode = mapping->host;
1578         if (S_ISBLK(inode->i_mode)) {
1579                 struct block_device *bdev = I_BDEV(inode);
1580                 set_blocksize(bdev, p->old_block_size);
1581                 bd_release(bdev);
1582         } else {
1583                 mutex_lock(&inode->i_mutex);
1584                 inode->i_flags &= ~S_SWAPFILE;
1585                 mutex_unlock(&inode->i_mutex);
1586         }
1587         filp_close(swap_file, NULL);
1588         err = 0;
1589
1590 out_dput:
1591         filp_close(victim, NULL);
1592 out:
1593         return err;
1594 }
1595
1596 #ifdef CONFIG_PROC_FS
1597 /* iterator */
1598 static void *swap_start(struct seq_file *swap, loff_t *pos)
1599 {
1600         struct swap_info_struct *ptr = swap_info;
1601         int i;
1602         loff_t l = *pos;
1603
1604         mutex_lock(&swapon_mutex);
1605
1606         if (!l)
1607                 return SEQ_START_TOKEN;
1608
1609         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1610                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1611                         continue;
1612                 if (!--l)
1613                         return ptr;
1614         }
1615
1616         return NULL;
1617 }
1618
1619 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1620 {
1621         struct swap_info_struct *ptr;
1622         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1623
1624         if (v == SEQ_START_TOKEN)
1625                 ptr = swap_info;
1626         else {
1627                 ptr = v;
1628                 ptr++;
1629         }
1630
1631         for (; ptr < endptr; ptr++) {
1632                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1633                         continue;
1634                 ++*pos;
1635                 return ptr;
1636         }
1637
1638         return NULL;
1639 }
1640
1641 static void swap_stop(struct seq_file *swap, void *v)
1642 {
1643         mutex_unlock(&swapon_mutex);
1644 }
1645
1646 static int swap_show(struct seq_file *swap, void *v)
1647 {
1648         struct swap_info_struct *ptr = v;
1649         struct file *file;
1650         int len;
1651
1652         if (ptr == SEQ_START_TOKEN) {
1653                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1654                 return 0;
1655         }
1656
1657         file = ptr->swap_file;
1658         len = seq_path(swap, &file->f_path, " \t\n\\");
1659         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1660                         len < 40 ? 40 - len : 1, " ",
1661                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1662                                 "partition" : "file\t",
1663                         ptr->pages << (PAGE_SHIFT - 10),
1664                         ptr->inuse_pages << (PAGE_SHIFT - 10),
1665                         ptr->prio);
1666         return 0;
1667 }
1668
1669 static const struct seq_operations swaps_op = {
1670         .start =        swap_start,
1671         .next =         swap_next,
1672         .stop =         swap_stop,
1673         .show =         swap_show
1674 };
1675
1676 static int swaps_open(struct inode *inode, struct file *file)
1677 {
1678         return seq_open(file, &swaps_op);
1679 }
1680
1681 static const struct file_operations proc_swaps_operations = {
1682         .open           = swaps_open,
1683         .read           = seq_read,
1684         .llseek         = seq_lseek,
1685         .release        = seq_release,
1686 };
1687
1688 static int __init procswaps_init(void)
1689 {
1690         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1691         return 0;
1692 }
1693 __initcall(procswaps_init);
1694 #endif /* CONFIG_PROC_FS */
1695
1696 #ifdef MAX_SWAPFILES_CHECK
1697 static int __init max_swapfiles_check(void)
1698 {
1699         MAX_SWAPFILES_CHECK();
1700         return 0;
1701 }
1702 late_initcall(max_swapfiles_check);
1703 #endif
1704
1705 /*
1706  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1707  *
1708  * The swapon system call
1709  */
1710 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1711 {
1712         struct swap_info_struct * p;
1713         char *name = NULL;
1714         struct block_device *bdev = NULL;
1715         struct file *swap_file = NULL;
1716         struct address_space *mapping;
1717         unsigned int type;
1718         int i, prev;
1719         int error;
1720         union swap_header *swap_header = NULL;
1721         unsigned int nr_good_pages = 0;
1722         int nr_extents = 0;
1723         sector_t span;
1724         unsigned long maxpages = 1;
1725         unsigned long swapfilepages;
1726         unsigned short *swap_map = NULL;
1727         struct page *page = NULL;
1728         struct inode *inode = NULL;
1729         int did_down = 0;
1730
1731         if (!capable(CAP_SYS_ADMIN))
1732                 return -EPERM;
1733         spin_lock(&swap_lock);
1734         p = swap_info;
1735         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1736                 if (!(p->flags & SWP_USED))
1737                         break;
1738         error = -EPERM;
1739         if (type >= MAX_SWAPFILES) {
1740                 spin_unlock(&swap_lock);
1741                 goto out;
1742         }
1743         if (type >= nr_swapfiles)
1744                 nr_swapfiles = type+1;
1745         memset(p, 0, sizeof(*p));
1746         INIT_LIST_HEAD(&p->extent_list);
1747         p->flags = SWP_USED;
1748         p->next = -1;
1749         spin_unlock(&swap_lock);
1750         name = getname(specialfile);
1751         error = PTR_ERR(name);
1752         if (IS_ERR(name)) {
1753                 name = NULL;
1754                 goto bad_swap_2;
1755         }
1756         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1757         error = PTR_ERR(swap_file);
1758         if (IS_ERR(swap_file)) {
1759                 swap_file = NULL;
1760                 goto bad_swap_2;
1761         }
1762
1763         p->swap_file = swap_file;
1764         mapping = swap_file->f_mapping;
1765         inode = mapping->host;
1766
1767         error = -EBUSY;
1768         for (i = 0; i < nr_swapfiles; i++) {
1769                 struct swap_info_struct *q = &swap_info[i];
1770
1771                 if (i == type || !q->swap_file)
1772                         continue;
1773                 if (mapping == q->swap_file->f_mapping)
1774                         goto bad_swap;
1775         }
1776
1777         error = -EINVAL;
1778         if (S_ISBLK(inode->i_mode)) {
1779                 bdev = I_BDEV(inode);
1780                 error = bd_claim(bdev, sys_swapon);
1781                 if (error < 0) {
1782                         bdev = NULL;
1783                         error = -EINVAL;
1784                         goto bad_swap;
1785                 }
1786                 p->old_block_size = block_size(bdev);
1787                 error = set_blocksize(bdev, PAGE_SIZE);
1788                 if (error < 0)
1789                         goto bad_swap;
1790                 p->bdev = bdev;
1791         } else if (S_ISREG(inode->i_mode)) {
1792                 p->bdev = inode->i_sb->s_bdev;
1793                 mutex_lock(&inode->i_mutex);
1794                 did_down = 1;
1795                 if (IS_SWAPFILE(inode)) {
1796                         error = -EBUSY;
1797                         goto bad_swap;
1798                 }
1799         } else {
1800                 goto bad_swap;
1801         }
1802
1803         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1804
1805         /*
1806          * Read the swap header.
1807          */
1808         if (!mapping->a_ops->readpage) {
1809                 error = -EINVAL;
1810                 goto bad_swap;
1811         }
1812         page = read_mapping_page(mapping, 0, swap_file);
1813         if (IS_ERR(page)) {
1814                 error = PTR_ERR(page);
1815                 goto bad_swap;
1816         }
1817         swap_header = kmap(page);
1818
1819         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1820                 printk(KERN_ERR "Unable to find swap-space signature\n");
1821                 error = -EINVAL;
1822                 goto bad_swap;
1823         }
1824
1825         /* swap partition endianess hack... */
1826         if (swab32(swap_header->info.version) == 1) {
1827                 swab32s(&swap_header->info.version);
1828                 swab32s(&swap_header->info.last_page);
1829                 swab32s(&swap_header->info.nr_badpages);
1830                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1831                         swab32s(&swap_header->info.badpages[i]);
1832         }
1833         /* Check the swap header's sub-version */
1834         if (swap_header->info.version != 1) {
1835                 printk(KERN_WARNING
1836                        "Unable to handle swap header version %d\n",
1837                        swap_header->info.version);
1838                 error = -EINVAL;
1839                 goto bad_swap;
1840         }
1841
1842         p->lowest_bit  = 1;
1843         p->cluster_next = 1;
1844
1845         /*
1846          * Find out how many pages are allowed for a single swap
1847          * device. There are two limiting factors: 1) the number of
1848          * bits for the swap offset in the swp_entry_t type and
1849          * 2) the number of bits in the a swap pte as defined by
1850          * the different architectures. In order to find the
1851          * largest possible bit mask a swap entry with swap type 0
1852          * and swap offset ~0UL is created, encoded to a swap pte,
1853          * decoded to a swp_entry_t again and finally the swap
1854          * offset is extracted. This will mask all the bits from
1855          * the initial ~0UL mask that can't be encoded in either
1856          * the swp_entry_t or the architecture definition of a
1857          * swap pte.
1858          */
1859         maxpages = swp_offset(pte_to_swp_entry(
1860                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1861         if (maxpages > swap_header->info.last_page)
1862                 maxpages = swap_header->info.last_page;
1863         p->highest_bit = maxpages - 1;
1864
1865         error = -EINVAL;
1866         if (!maxpages)
1867                 goto bad_swap;
1868         if (swapfilepages && maxpages > swapfilepages) {
1869                 printk(KERN_WARNING
1870                        "Swap area shorter than signature indicates\n");
1871                 goto bad_swap;
1872         }
1873         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1874                 goto bad_swap;
1875         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1876                 goto bad_swap;
1877
1878         /* OK, set up the swap map and apply the bad block list */
1879         swap_map = vmalloc(maxpages * sizeof(short));
1880         if (!swap_map) {
1881                 error = -ENOMEM;
1882                 goto bad_swap;
1883         }
1884
1885         memset(swap_map, 0, maxpages * sizeof(short));
1886         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1887                 int page_nr = swap_header->info.badpages[i];
1888                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1889                         error = -EINVAL;
1890                         goto bad_swap;
1891                 }
1892                 swap_map[page_nr] = SWAP_MAP_BAD;
1893         }
1894
1895         error = swap_cgroup_swapon(type, maxpages);
1896         if (error)
1897                 goto bad_swap;
1898
1899         nr_good_pages = swap_header->info.last_page -
1900                         swap_header->info.nr_badpages -
1901                         1 /* header page */;
1902
1903         if (nr_good_pages) {
1904                 swap_map[0] = SWAP_MAP_BAD;
1905                 p->max = maxpages;
1906                 p->pages = nr_good_pages;
1907                 nr_extents = setup_swap_extents(p, &span);
1908                 if (nr_extents < 0) {
1909                         error = nr_extents;
1910                         goto bad_swap;
1911                 }
1912                 nr_good_pages = p->pages;
1913         }
1914         if (!nr_good_pages) {
1915                 printk(KERN_WARNING "Empty swap-file\n");
1916                 error = -EINVAL;
1917                 goto bad_swap;
1918         }
1919
1920         if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1921                 p->flags |= SWP_SOLIDSTATE;
1922                 p->cluster_next = 1 + (random32() % p->highest_bit);
1923         }
1924         if (discard_swap(p) == 0)
1925                 p->flags |= SWP_DISCARDABLE;
1926
1927         mutex_lock(&swapon_mutex);
1928         spin_lock(&swap_lock);
1929         if (swap_flags & SWAP_FLAG_PREFER)
1930                 p->prio =
1931                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1932         else
1933                 p->prio = --least_priority;
1934         p->swap_map = swap_map;
1935         p->flags |= SWP_WRITEOK;
1936         nr_swap_pages += nr_good_pages;
1937         total_swap_pages += nr_good_pages;
1938
1939         printk(KERN_INFO "Adding %uk swap on %s.  "
1940                         "Priority:%d extents:%d across:%lluk %s%s\n",
1941                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1942                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
1943                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
1944                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
1945
1946         /* insert swap space into swap_list: */
1947         prev = -1;
1948         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1949                 if (p->prio >= swap_info[i].prio) {
1950                         break;
1951                 }
1952                 prev = i;
1953         }
1954         p->next = i;
1955         if (prev < 0) {
1956                 swap_list.head = swap_list.next = p - swap_info;
1957         } else {
1958                 swap_info[prev].next = p - swap_info;
1959         }
1960         spin_unlock(&swap_lock);
1961         mutex_unlock(&swapon_mutex);
1962         error = 0;
1963         goto out;
1964 bad_swap:
1965         if (bdev) {
1966                 set_blocksize(bdev, p->old_block_size);
1967                 bd_release(bdev);
1968         }
1969         destroy_swap_extents(p);
1970         swap_cgroup_swapoff(type);
1971 bad_swap_2:
1972         spin_lock(&swap_lock);
1973         p->swap_file = NULL;
1974         p->flags = 0;
1975         spin_unlock(&swap_lock);
1976         vfree(swap_map);
1977         if (swap_file)
1978                 filp_close(swap_file, NULL);
1979 out:
1980         if (page && !IS_ERR(page)) {
1981                 kunmap(page);
1982                 page_cache_release(page);
1983         }
1984         if (name)
1985                 putname(name);
1986         if (did_down) {
1987                 if (!error)
1988                         inode->i_flags |= S_SWAPFILE;
1989                 mutex_unlock(&inode->i_mutex);
1990         }
1991         return error;
1992 }
1993
1994 void si_swapinfo(struct sysinfo *val)
1995 {
1996         unsigned int i;
1997         unsigned long nr_to_be_unused = 0;
1998
1999         spin_lock(&swap_lock);
2000         for (i = 0; i < nr_swapfiles; i++) {
2001                 if (!(swap_info[i].flags & SWP_USED) ||
2002                      (swap_info[i].flags & SWP_WRITEOK))
2003                         continue;
2004                 nr_to_be_unused += swap_info[i].inuse_pages;
2005         }
2006         val->freeswap = nr_swap_pages + nr_to_be_unused;
2007         val->totalswap = total_swap_pages + nr_to_be_unused;
2008         spin_unlock(&swap_lock);
2009 }
2010
2011 /*
2012  * Verify that a swap entry is valid and increment its swap map count.
2013  *
2014  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
2015  * "permanent", but will be reclaimed by the next swapoff.
2016  * Returns error code in following case.
2017  * - success -> 0
2018  * - swp_entry is invalid -> EINVAL
2019  * - swp_entry is migration entry -> EINVAL
2020  * - swap-cache reference is requested but there is already one. -> EEXIST
2021  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2022  */
2023 static int __swap_duplicate(swp_entry_t entry, bool cache)
2024 {
2025         struct swap_info_struct * p;
2026         unsigned long offset, type;
2027         int result = -EINVAL;
2028         int count;
2029         bool has_cache;
2030
2031         if (is_migration_entry(entry))
2032                 return -EINVAL;
2033
2034         type = swp_type(entry);
2035         if (type >= nr_swapfiles)
2036                 goto bad_file;
2037         p = type + swap_info;
2038         offset = swp_offset(entry);
2039
2040         spin_lock(&swap_lock);
2041
2042         if (unlikely(offset >= p->max))
2043                 goto unlock_out;
2044
2045         count = swap_count(p->swap_map[offset]);
2046         has_cache = swap_has_cache(p->swap_map[offset]);
2047
2048         if (cache == SWAP_CACHE) { /* called for swapcache/swapin-readahead */
2049
2050                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2051                 if (!has_cache && count) {
2052                         p->swap_map[offset] = encode_swapmap(count, true);
2053                         result = 0;
2054                 } else if (has_cache) /* someone added cache */
2055                         result = -EEXIST;
2056                 else if (!count) /* no users */
2057                         result = -ENOENT;
2058
2059         } else if (count || has_cache) {
2060                 if (count < SWAP_MAP_MAX - 1) {
2061                         p->swap_map[offset] = encode_swapmap(count + 1,
2062                                                              has_cache);
2063                         result = 0;
2064                 } else if (count <= SWAP_MAP_MAX) {
2065                         if (swap_overflow++ < 5)
2066                                 printk(KERN_WARNING
2067                                        "swap_dup: swap entry overflow\n");
2068                         p->swap_map[offset] = encode_swapmap(SWAP_MAP_MAX,
2069                                                               has_cache);
2070                         result = 0;
2071                 }
2072         } else
2073                 result = -ENOENT; /* unused swap entry */
2074 unlock_out:
2075         spin_unlock(&swap_lock);
2076 out:
2077         return result;
2078
2079 bad_file:
2080         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2081         goto out;
2082 }
2083 /*
2084  * increase reference count of swap entry by 1.
2085  */
2086 void swap_duplicate(swp_entry_t entry)
2087 {
2088         __swap_duplicate(entry, SWAP_MAP);
2089 }
2090
2091 /*
2092  * @entry: swap entry for which we allocate swap cache.
2093  *
2094  * Called when allocating swap cache for exising swap entry,
2095  * This can return error codes. Returns 0 at success.
2096  * -EBUSY means there is a swap cache.
2097  * Note: return code is different from swap_duplicate().
2098  */
2099 int swapcache_prepare(swp_entry_t entry)
2100 {
2101         return __swap_duplicate(entry, SWAP_CACHE);
2102 }
2103
2104
2105 struct swap_info_struct *
2106 get_swap_info_struct(unsigned type)
2107 {
2108         return &swap_info[type];
2109 }
2110
2111 /*
2112  * swap_lock prevents swap_map being freed. Don't grab an extra
2113  * reference on the swaphandle, it doesn't matter if it becomes unused.
2114  */
2115 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2116 {
2117         struct swap_info_struct *si;
2118         int our_page_cluster = page_cluster;
2119         pgoff_t target, toff;
2120         pgoff_t base, end;
2121         int nr_pages = 0;
2122
2123         if (!our_page_cluster)  /* no readahead */
2124                 return 0;
2125
2126         si = &swap_info[swp_type(entry)];
2127         target = swp_offset(entry);
2128         base = (target >> our_page_cluster) << our_page_cluster;
2129         end = base + (1 << our_page_cluster);
2130         if (!base)              /* first page is swap header */
2131                 base++;
2132
2133         spin_lock(&swap_lock);
2134         if (end > si->max)      /* don't go beyond end of map */
2135                 end = si->max;
2136
2137         /* Count contiguous allocated slots above our target */
2138         for (toff = target; ++toff < end; nr_pages++) {
2139                 /* Don't read in free or bad pages */
2140                 if (!si->swap_map[toff])
2141                         break;
2142                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2143                         break;
2144         }
2145         /* Count contiguous allocated slots below our target */
2146         for (toff = target; --toff >= base; nr_pages++) {
2147                 /* Don't read in free or bad pages */
2148                 if (!si->swap_map[toff])
2149                         break;
2150                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2151                         break;
2152         }
2153         spin_unlock(&swap_lock);
2154
2155         /*
2156          * Indicate starting offset, and return number of pages to get:
2157          * if only 1, say 0, since there's then no readahead to be done.
2158          */
2159         *offset = ++toff;
2160         return nr_pages? ++nr_pages: 0;
2161 }