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/shmem_fs.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/ksm.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 #include <linux/poll.h>
  33 #include <linux/oom.h>
  34
  35 #include <asm/pgtable.h>
  36 #include <asm/tlbflush.h>
  37 #include <linux/swapops.h>
  38 #include <linux/page_cgroup.h>
  39
  40 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
  41                                  unsigned char);
  42 static void free_swap_count_continuations(struct swap_info_struct *);
  43 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
  44
  45 static DEFINE_SPINLOCK(swap_lock);
  46 static unsigned int nr_swapfiles;
  47 long nr_swap_pages;
  48 long total_swap_pages;
  49 static int least_priority;
  50
  51 static const char Bad_file[] = "Bad swap file entry ";
  52 static const char Unused_file[] = "Unused swap file entry ";
  53 static const char Bad_offset[] = "Bad swap offset entry ";
  54 static const char Unused_offset[] = "Unused swap offset entry ";
  55
  56 static struct swap_list_t swap_list = {-1, -1};
  57
  58 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
  59
  60 static DEFINE_MUTEX(swapon_mutex);
  61
  62 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
  63 /* Activity counter to indicate that a swapon or swapoff has occurred */
  64 static atomic_t proc_poll_event = ATOMIC_INIT(0);
  65
  66 static inline unsigned char swap_count(unsigned char ent)
  67 {
  68         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
  69 }
  70
  71 /* returns 1 if swap entry is freed */
  72 static int
  73 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
  74 {
  75         swp_entry_t entry = swp_entry(si->type, offset);
  76         struct page *page;
  77         int ret = 0;
  78
  79         page = find_get_page(&swapper_space, entry.val);
  80         if (!page)
  81                 return 0;
  82         /*
  83          * This function is called from scan_swap_map() and it's called
  84          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
  85          * We have to use trylock for avoiding deadlock. This is a special
  86          * case and you should use try_to_free_swap() with explicit lock_page()
  87          * in usual operations.
  88          */
  89         if (trylock_page(page)) {
  90                 ret = try_to_free_swap(page);
  91                 unlock_page(page);
  92         }
  93         page_cache_release(page);
  94         return ret;
  95 }
  96
  97 /*
  98  * swapon tell device that all the old swap contents can be discarded,
  99  * to allow the swap device to optimize its wear-levelling.
 100  */
 101 static int discard_swap(struct swap_info_struct *si)
 102 {
 103         struct swap_extent *se;
 104         sector_t start_block;
 105         sector_t nr_blocks;
 106         int err = 0;
 107
 108         /* Do not discard the swap header page! */
 109         se = &si->first_swap_extent;
 110         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
 111         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
 112         if (nr_blocks) {
 113                 err = blkdev_issue_discard(si->bdev, start_block,
 114                                 nr_blocks, GFP_KERNEL, 0);
 115                 if (err)
 116                         return err;
 117                 cond_resched();
 118         }
 119
 120         list_for_each_entry(se, &si->first_swap_extent.list, list) {
 121                 start_block = se->start_block << (PAGE_SHIFT - 9);
 122                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 123
 124                 err = blkdev_issue_discard(si->bdev, start_block,
 125                                 nr_blocks, GFP_KERNEL, 0);
 126                 if (err)
 127                         break;
 128
 129                 cond_resched();
 130         }
 131         return err;             /* That will often be -EOPNOTSUPP */
 132 }
 133
 134 /*
 135  * swap allocation tell device that a cluster of swap can now be discarded,
 136  * to allow the swap device to optimize its wear-levelling.
 137  */
 138 static void discard_swap_cluster(struct swap_info_struct *si,
 139                                  pgoff_t start_page, pgoff_t nr_pages)
 140 {
 141         struct swap_extent *se = si->curr_swap_extent;
 142         int found_extent = 0;
 143
 144         while (nr_pages) {
 145                 struct list_head *lh;
 146
 147                 if (se->start_page <= start_page &&
 148                     start_page < se->start_page + se->nr_pages) {
 149                         pgoff_t offset = start_page - se->start_page;
 150                         sector_t start_block = se->start_block + offset;
 151                         sector_t nr_blocks = se->nr_pages - offset;
 152
 153                         if (nr_blocks > nr_pages)
 154                                 nr_blocks = nr_pages;
 155                         start_page += nr_blocks;
 156                         nr_pages -= nr_blocks;
 157
 158                         if (!found_extent++)
 159                                 si->curr_swap_extent = se;
 160
 161                         start_block <<= PAGE_SHIFT - 9;
 162                         nr_blocks <<= PAGE_SHIFT - 9;
 163                         if (blkdev_issue_discard(si->bdev, start_block,
 164                                     nr_blocks, GFP_NOIO, 0))
 165                                 break;
 166                 }
 167
 168                 lh = se->list.next;
 169                 se = list_entry(lh, struct swap_extent, list);
 170         }
 171 }
 172
 173 static int wait_for_discard(void *word)
 174 {
 175         schedule();
 176         return 0;
 177 }
 178
 179 #define SWAPFILE_CLUSTER        256
 180 #define LATENCY_LIMIT           256
 181
 182 static unsigned long scan_swap_map(struct swap_info_struct *si,
 183                                    unsigned char usage)
 184 {
 185         unsigned long offset;
 186         unsigned long scan_base;
 187         unsigned long last_in_cluster = 0;
 188         int latency_ration = LATENCY_LIMIT;
 189         int found_free_cluster = 0;
 190
 191         /*
 192          * We try to cluster swap pages by allocating them sequentially
 193          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 194          * way, however, we resort to first-free allocation, starting
 195          * a new cluster.  This prevents us from scattering swap pages
 196          * all over the entire swap partition, so that we reduce
 197          * overall disk seek times between swap pages.  -- sct
 198          * But we do now try to find an empty cluster.  -Andrea
 199          * And we let swap pages go all over an SSD partition.  Hugh
 200          */
 201
 202         si->flags += SWP_SCANNING;
 203         scan_base = offset = si->cluster_next;
 204
 205         if (unlikely(!si->cluster_nr--)) {
 206                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 207                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
 208                         goto checks;
 209                 }
 210                 if (si->flags & SWP_DISCARDABLE) {
 211                         /*
 212                          * Start range check on racing allocations, in case
 213                          * they overlap the cluster we eventually decide on
 214                          * (we scan without swap_lock to allow preemption).
 215                          * It's hardly conceivable that cluster_nr could be
 216                          * wrapped during our scan, but don't depend on it.
 217                          */
 218                         if (si->lowest_alloc)
 219                                 goto checks;
 220                         si->lowest_alloc = si->max;
 221                         si->highest_alloc = 0;
 222                 }
 223                 spin_unlock(&swap_lock);
 224
 225                 /*
 226                  * If seek is expensive, start searching for new cluster from
 227                  * start of partition, to minimize the span of allocated swap.
 228                  * But if seek is cheap, search from our current position, so
 229                  * that swap is allocated from all over the partition: if the
 230                  * Flash Translation Layer only remaps within limited zones,
 231                  * we don't want to wear out the first zone too quickly.
 232                  */
 233                 if (!(si->flags & SWP_SOLIDSTATE))
 234                         scan_base = offset = si->lowest_bit;
 235                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 236
 237                 /* Locate the first empty (unaligned) cluster */
 238                 for (; last_in_cluster <= si->highest_bit; offset++) {
 239                         if (si->swap_map[offset])
 240                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 241                         else if (offset == last_in_cluster) {
 242                                 spin_lock(&swap_lock);
 243                                 offset -= SWAPFILE_CLUSTER - 1;
 244                                 si->cluster_next = offset;
 245                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 246                                 found_free_cluster = 1;
 247                                 goto checks;
 248                         }
 249                         if (unlikely(--latency_ration < 0)) {
 250                                 cond_resched();
 251                                 latency_ration = LATENCY_LIMIT;
 252                         }
 253                 }
 254
 255                 offset = si->lowest_bit;
 256                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 257
 258                 /* Locate the first empty (unaligned) cluster */
 259                 for (; last_in_cluster < scan_base; offset++) {
 260                         if (si->swap_map[offset])
 261                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 262                         else if (offset == last_in_cluster) {
 263                                 spin_lock(&swap_lock);
 264                                 offset -= SWAPFILE_CLUSTER - 1;
 265                                 si->cluster_next = offset;
 266                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 267                                 found_free_cluster = 1;
 268                                 goto checks;
 269                         }
 270                         if (unlikely(--latency_ration < 0)) {
 271                                 cond_resched();
 272                                 latency_ration = LATENCY_LIMIT;
 273                         }
 274                 }
 275
 276                 offset = scan_base;
 277                 spin_lock(&swap_lock);
 278                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 279                 si->lowest_alloc = 0;
 280         }
 281
 282 checks:
 283         if (!(si->flags & SWP_WRITEOK))
 284                 goto no_page;
 285         if (!si->highest_bit)
 286                 goto no_page;
 287         if (offset > si->highest_bit)
 288                 scan_base = offset = si->lowest_bit;
 289
 290         /* reuse swap entry of cache-only swap if not busy. */
 291         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 292                 int swap_was_freed;
 293                 spin_unlock(&swap_lock);
 294                 swap_was_freed = __try_to_reclaim_swap(si, offset);
 295                 spin_lock(&swap_lock);
 296                 /* entry was freed successfully, try to use this again */
 297                 if (swap_was_freed)
 298                         goto checks;
 299                 goto scan; /* check next one */
 300         }
 301
 302         if (si->swap_map[offset])
 303                 goto scan;
 304
 305         if (offset == si->lowest_bit)
 306                 si->lowest_bit++;
 307         if (offset == si->highest_bit)
 308                 si->highest_bit--;
 309         si->inuse_pages++;
 310         if (si->inuse_pages == si->pages) {
 311                 si->lowest_bit = si->max;
 312                 si->highest_bit = 0;
 313         }
 314         si->swap_map[offset] = usage;
 315         si->cluster_next = offset + 1;
 316         si->flags -= SWP_SCANNING;
 317
 318         if (si->lowest_alloc) {
 319                 /*
 320                  * Only set when SWP_DISCARDABLE, and there's a scan
 321                  * for a free cluster in progress or just completed.
 322                  */
 323                 if (found_free_cluster) {
 324                         /*
 325                          * To optimize wear-levelling, discard the
 326                          * old data of the cluster, taking care not to
 327                          * discard any of its pages that have already
 328                          * been allocated by racing tasks (offset has
 329                          * already stepped over any at the beginning).
 330                          */
 331                         if (offset < si->highest_alloc &&
 332                             si->lowest_alloc <= last_in_cluster)
 333                                 last_in_cluster = si->lowest_alloc - 1;
 334                         si->flags |= SWP_DISCARDING;
 335                         spin_unlock(&swap_lock);
 336
 337                         if (offset < last_in_cluster)
 338                                 discard_swap_cluster(si, offset,
 339                                         last_in_cluster - offset + 1);
 340
 341                         spin_lock(&swap_lock);
 342                         si->lowest_alloc = 0;
 343                         si->flags &= ~SWP_DISCARDING;
 344
 345                         smp_mb();       /* wake_up_bit advises this */
 346                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
 347
 348                 } else if (si->flags & SWP_DISCARDING) {
 349                         /*
 350                          * Delay using pages allocated by racing tasks
 351                          * until the whole discard has been issued. We
 352                          * could defer that delay until swap_writepage,
 353                          * but it's easier to keep this self-contained.
 354                          */
 355                         spin_unlock(&swap_lock);
 356                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
 357                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
 358                         spin_lock(&swap_lock);
 359                 } else {
 360                         /*
 361                          * Note pages allocated by racing tasks while
 362                          * scan for a free cluster is in progress, so
 363                          * that its final discard can exclude them.
 364                          */
 365                         if (offset < si->lowest_alloc)
 366                                 si->lowest_alloc = offset;
 367                         if (offset > si->highest_alloc)
 368                                 si->highest_alloc = offset;
 369                 }
 370         }
 371         return offset;
 372
 373 scan:
 374         spin_unlock(&swap_lock);
 375         while (++offset <= si->highest_bit) {
 376                 if (!si->swap_map[offset]) {
 377                         spin_lock(&swap_lock);
 378                         goto checks;
 379                 }
 380                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 381                         spin_lock(&swap_lock);
 382                         goto checks;
 383                 }
 384                 if (unlikely(--latency_ration < 0)) {
 385                         cond_resched();
 386                         latency_ration = LATENCY_LIMIT;
 387                 }
 388         }
 389         offset = si->lowest_bit;
 390         while (++offset < scan_base) {
 391                 if (!si->swap_map[offset]) {
 392                         spin_lock(&swap_lock);
 393                         goto checks;
 394                 }
 395                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 396                         spin_lock(&swap_lock);
 397                         goto checks;
 398                 }
 399                 if (unlikely(--latency_ration < 0)) {
 400                         cond_resched();
 401                         latency_ration = LATENCY_LIMIT;
 402                 }
 403         }
 404         spin_lock(&swap_lock);
 405
 406 no_page:
 407         si->flags -= SWP_SCANNING;
 408         return 0;
 409 }
 410
 411 swp_entry_t get_swap_page(void)
 412 {
 413         struct swap_info_struct *si;
 414         pgoff_t offset;
 415         int type, next;
 416         int wrapped = 0;
 417
 418         spin_lock(&swap_lock);
 419         if (nr_swap_pages <= 0)
 420                 goto noswap;
 421         nr_swap_pages--;
 422
 423         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 424                 si = swap_info[type];
 425                 next = si->next;
 426                 if (next < 0 ||
 427                     (!wrapped && si->prio != swap_info[next]->prio)) {
 428                         next = swap_list.head;
 429                         wrapped++;
 430                 }
 431
 432                 if (!si->highest_bit)
 433                         continue;
 434                 if (!(si->flags & SWP_WRITEOK))
 435                         continue;
 436
 437                 swap_list.next = next;
 438                 /* This is called for allocating swap entry for cache */
 439                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
 440                 if (offset) {
 441                         spin_unlock(&swap_lock);
 442                         return swp_entry(type, offset);
 443                 }
 444                 next = swap_list.next;
 445         }
 446
 447         nr_swap_pages++;
 448 noswap:
 449         spin_unlock(&swap_lock);
 450         return (swp_entry_t) {0};
 451 }
 452
 453 /* The only caller of this function is now susupend routine */
 454 swp_entry_t get_swap_page_of_type(int type)
 455 {
 456         struct swap_info_struct *si;
 457         pgoff_t offset;
 458
 459         spin_lock(&swap_lock);
 460         si = swap_info[type];
 461         if (si && (si->flags & SWP_WRITEOK)) {
 462                 nr_swap_pages--;
 463                 /* This is called for allocating swap entry, not cache */
 464                 offset = scan_swap_map(si, 1);
 465                 if (offset) {
 466                         spin_unlock(&swap_lock);
 467                         return swp_entry(type, offset);
 468                 }
 469                 nr_swap_pages++;
 470         }
 471         spin_unlock(&swap_lock);
 472         return (swp_entry_t) {0};
 473 }
 474
 475 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
 476 {
 477         struct swap_info_struct *p;
 478         unsigned long offset, type;
 479
 480         if (!entry.val)
 481                 goto out;
 482         type = swp_type(entry);
 483         if (type >= nr_swapfiles)
 484                 goto bad_nofile;
 485         p = swap_info[type];
 486         if (!(p->flags & SWP_USED))
 487                 goto bad_device;
 488         offset = swp_offset(entry);
 489         if (offset >= p->max)
 490                 goto bad_offset;
 491         if (!p->swap_map[offset])
 492                 goto bad_free;
 493         spin_lock(&swap_lock);
 494         return p;
 495
 496 bad_free:
 497         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 498         goto out;
 499 bad_offset:
 500         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 501         goto out;
 502 bad_device:
 503         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 504         goto out;
 505 bad_nofile:
 506         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 507 out:
 508         return NULL;
 509 }
 510
 511 static unsigned char swap_entry_free(struct swap_info_struct *p,
 512                                      swp_entry_t entry, unsigned char usage)
 513 {
 514         unsigned long offset = swp_offset(entry);
 515         unsigned char count;
 516         unsigned char has_cache;
 517
 518         count = p->swap_map[offset];
 519         has_cache = count & SWAP_HAS_CACHE;
 520         count &= ~SWAP_HAS_CACHE;
 521
 522         if (usage == SWAP_HAS_CACHE) {
 523                 VM_BUG_ON(!has_cache);
 524                 has_cache = 0;
 525         } else if (count == SWAP_MAP_SHMEM) {
 526                 /*
 527                  * Or we could insist on shmem.c using a special
 528                  * swap_shmem_free() and free_shmem_swap_and_cache()...
 529                  */
 530                 count = 0;
 531         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
 532                 if (count == COUNT_CONTINUED) {
 533                         if (swap_count_continued(p, offset, count))
 534                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
 535                         else
 536                                 count = SWAP_MAP_MAX;
 537                 } else
 538                         count--;
 539         }
 540
 541         if (!count)
 542                 mem_cgroup_uncharge_swap(entry);
 543
 544         usage = count | has_cache;
 545         p->swap_map[offset] = usage;
 546
 547         /* free if no reference */
 548         if (!usage) {
 549                 struct gendisk *disk = p->bdev->bd_disk;
 550                 if (offset < p->lowest_bit)
 551                         p->lowest_bit = offset;
 552                 if (offset > p->highest_bit)
 553                         p->highest_bit = offset;
 554                 if (swap_list.next >= 0 &&
 555                     p->prio > swap_info[swap_list.next]->prio)
 556                         swap_list.next = p->type;
 557                 nr_swap_pages++;
 558                 p->inuse_pages--;
 559                 if ((p->flags & SWP_BLKDEV) &&
 560                                 disk->fops->swap_slot_free_notify)
 561                         disk->fops->swap_slot_free_notify(p->bdev, offset);
 562         }
 563
 564         return usage;
 565 }
 566
 567 /*
 568  * Caller has made sure that the swapdevice corresponding to entry
 569  * is still around or has not been recycled.
 570  */
 571 void swap_free(swp_entry_t entry)
 572 {
 573         struct swap_info_struct *p;
 574
 575         p = swap_info_get(entry);
 576         if (p) {
 577                 swap_entry_free(p, entry, 1);
 578                 spin_unlock(&swap_lock);
 579         }
 580 }
 581
 582 /*
 583  * Called after dropping swapcache to decrease refcnt to swap entries.
 584  */
 585 void swapcache_free(swp_entry_t entry, struct page *page)
 586 {
 587         struct swap_info_struct *p;
 588         unsigned char count;
 589
 590         p = swap_info_get(entry);
 591         if (p) {
 592                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
 593                 if (page)
 594                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
 595                 spin_unlock(&swap_lock);
 596         }
 597 }
 598
 599 /*
 600  * How many references to page are currently swapped out?
 601  * This does not give an exact answer when swap count is continued,
 602  * but does include the high COUNT_CONTINUED flag to allow for that.
 603  */
 604 static inline int page_swapcount(struct page *page)
 605 {
 606         int count = 0;
 607         struct swap_info_struct *p;
 608         swp_entry_t entry;
 609
 610         entry.val = page_private(page);
 611         p = swap_info_get(entry);
 612         if (p) {
 613                 count = swap_count(p->swap_map[swp_offset(entry)]);
 614                 spin_unlock(&swap_lock);
 615         }
 616         return count;
 617 }
 618
 619 /*
 620  * We can write to an anon page without COW if there are no other references
 621  * to it.  And as a side-effect, free up its swap: because the old content
 622  * on disk will never be read, and seeking back there to write new content
 623  * later would only waste time away from clustering.
 624  */
 625 int reuse_swap_page(struct page *page)
 626 {
 627         int count;
 628
 629         VM_BUG_ON(!PageLocked(page));
 630         if (unlikely(PageKsm(page)))
 631                 return 0;
 632         count = page_mapcount(page);
 633         if (count <= 1 && PageSwapCache(page)) {
 634                 count += page_swapcount(page);
 635                 if (count == 1 && !PageWriteback(page)) {
 636                         delete_from_swap_cache(page);
 637                         SetPageDirty(page);
 638                 }
 639         }
 640         return count <= 1;
 641 }
 642
 643 /*
 644  * If swap is getting full, or if there are no more mappings of this page,
 645  * then try_to_free_swap is called to free its swap space.
 646  */
 647 int try_to_free_swap(struct page *page)
 648 {
 649         VM_BUG_ON(!PageLocked(page));
 650
 651         if (!PageSwapCache(page))
 652                 return 0;
 653         if (PageWriteback(page))
 654                 return 0;
 655         if (page_swapcount(page))
 656                 return 0;
 657
 658         /*
 659          * Once hibernation has begun to create its image of memory,
 660          * there's a danger that one of the calls to try_to_free_swap()
 661          * - most probably a call from __try_to_reclaim_swap() while
 662          * hibernation is allocating its own swap pages for the image,
 663          * but conceivably even a call from memory reclaim - will free
 664          * the swap from a page which has already been recorded in the
 665          * image as a clean swapcache page, and then reuse its swap for
 666          * another page of the image.  On waking from hibernation, the
 667          * original page might be freed under memory pressure, then
 668          * later read back in from swap, now with the wrong data.
 669          *
 670          * Hibration suspends storage while it is writing the image
 671          * to disk so check that here.
 672          */
 673         if (pm_suspended_storage())
 674                 return 0;
 675
 676         delete_from_swap_cache(page);
 677         SetPageDirty(page);
 678         return 1;
 679 }
 680
 681 /*
 682  * Free the swap entry like above, but also try to
 683  * free the page cache entry if it is the last user.
 684  */
 685 int free_swap_and_cache(swp_entry_t entry)
 686 {
 687         struct swap_info_struct *p;
 688         struct page *page = NULL;
 689
 690         if (non_swap_entry(entry))
 691                 return 1;
 692
 693         p = swap_info_get(entry);
 694         if (p) {
 695                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
 696                         page = find_get_page(&swapper_space, entry.val);
 697                         if (page && !trylock_page(page)) {
 698                                 page_cache_release(page);
 699                                 page = NULL;
 700                         }
 701                 }
 702                 spin_unlock(&swap_lock);
 703         }
 704         if (page) {
 705                 /*
 706                  * Not mapped elsewhere, or swap space full? Free it!
 707                  * Also recheck PageSwapCache now page is locked (above).
 708                  */
 709                 if (PageSwapCache(page) && !PageWriteback(page) &&
 710                                 (!page_mapped(page) || vm_swap_full())) {
 711                         delete_from_swap_cache(page);
 712                         SetPageDirty(page);
 713                 }
 714                 unlock_page(page);
 715                 page_cache_release(page);
 716         }
 717         return p != NULL;
 718 }
 719
 720 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 721 /**
 722  * mem_cgroup_count_swap_user - count the user of a swap entry
 723  * @ent: the swap entry to be checked
 724  * @pagep: the pointer for the swap cache page of the entry to be stored
 725  *
 726  * Returns the number of the user of the swap entry. The number is valid only
 727  * for swaps of anonymous pages.
 728  * If the entry is found on swap cache, the page is stored to pagep with
 729  * refcount of it being incremented.
 730  */
 731 int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
 732 {
 733         struct page *page;
 734         struct swap_info_struct *p;
 735         int count = 0;
 736
 737         page = find_get_page(&swapper_space, ent.val);
 738         if (page)
 739                 count += page_mapcount(page);
 740         p = swap_info_get(ent);
 741         if (p) {
 742                 count += swap_count(p->swap_map[swp_offset(ent)]);
 743                 spin_unlock(&swap_lock);
 744         }
 745
 746         *pagep = page;
 747         return count;
 748 }
 749 #endif
 750
 751 #ifdef CONFIG_HIBERNATION
 752 /*
 753  * Find the swap type that corresponds to given device (if any).
 754  *
 755  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 756  * from 0, in which the swap header is expected to be located.
 757  *
 758  * This is needed for the suspend to disk (aka swsusp).
 759  */
 760 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 761 {
 762         struct block_device *bdev = NULL;
 763         int type;
 764
 765         if (device)
 766                 bdev = bdget(device);
 767
 768         spin_lock(&swap_lock);
 769         for (type = 0; type < nr_swapfiles; type++) {
 770                 struct swap_info_struct *sis = swap_info[type];
 771
 772                 if (!(sis->flags & SWP_WRITEOK))
 773                         continue;
 774
 775                 if (!bdev) {
 776                         if (bdev_p)
 777                                 *bdev_p = bdgrab(sis->bdev);
 778
 779                         spin_unlock(&swap_lock);
 780                         return type;
 781                 }
 782                 if (bdev == sis->bdev) {
 783                         struct swap_extent *se = &sis->first_swap_extent;
 784
 785                         if (se->start_block == offset) {
 786                                 if (bdev_p)
 787                                         *bdev_p = bdgrab(sis->bdev);
 788
 789                                 spin_unlock(&swap_lock);
 790                                 bdput(bdev);
 791                                 return type;
 792                         }
 793                 }
 794         }
 795         spin_unlock(&swap_lock);
 796         if (bdev)
 797                 bdput(bdev);
 798
 799         return -ENODEV;
 800 }
 801
 802 /*
 803  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 804  * corresponding to given index in swap_info (swap type).
 805  */
 806 sector_t swapdev_block(int type, pgoff_t offset)
 807 {
 808         struct block_device *bdev;
 809
 810         if ((unsigned int)type >= nr_swapfiles)
 811                 return 0;
 812         if (!(swap_info[type]->flags & SWP_WRITEOK))
 813                 return 0;
 814         return map_swap_entry(swp_entry(type, offset), &bdev);
 815 }
 816
 817 /*
 818  * Return either the total number of swap pages of given type, or the number
 819  * of free pages of that type (depending on @free)
 820  *
 821  * This is needed for software suspend
 822  */
 823 unsigned int count_swap_pages(int type, int free)
 824 {
 825         unsigned int n = 0;
 826
 827         spin_lock(&swap_lock);
 828         if ((unsigned int)type < nr_swapfiles) {
 829                 struct swap_info_struct *sis = swap_info[type];
 830
 831                 if (sis->flags & SWP_WRITEOK) {
 832                         n = sis->pages;
 833                         if (free)
 834                                 n -= sis->inuse_pages;
 835                 }
 836         }
 837         spin_unlock(&swap_lock);
 838         return n;
 839 }
 840 #endif /* CONFIG_HIBERNATION */
 841
 842 /*
 843  * No need to decide whether this PTE shares the swap entry with others,
 844  * just let do_wp_page work it out if a write is requested later - to
 845  * force COW, vm_page_prot omits write permission from any private vma.
 846  */
 847 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 848                 unsigned long addr, swp_entry_t entry, struct page *page)
 849 {
 850         struct mem_cgroup *memcg;
 851         spinlock_t *ptl;
 852         pte_t *pte;
 853         int ret = 1;
 854
 855         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
 856                                          GFP_KERNEL, &memcg)) {
 857                 ret = -ENOMEM;
 858                 goto out_nolock;
 859         }
 860
 861         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 862         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 863                 if (ret > 0)
 864                         mem_cgroup_cancel_charge_swapin(memcg);
 865                 ret = 0;
 866                 goto out;
 867         }
 868
 869         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
 870         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
 871         get_page(page);
 872         set_pte_at(vma->vm_mm, addr, pte,
 873                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 874         page_add_anon_rmap(page, vma, addr);
 875         mem_cgroup_commit_charge_swapin(page, memcg);
 876         swap_free(entry);
 877         /*
 878          * Move the page to the active list so it is not
 879          * immediately swapped out again after swapon.
 880          */
 881         activate_page(page);
 882 out:
 883         pte_unmap_unlock(pte, ptl);
 884 out_nolock:
 885         return ret;
 886 }
 887
 888 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 889                                 unsigned long addr, unsigned long end,
 890                                 swp_entry_t entry, struct page *page)
 891 {
 892         pte_t swp_pte = swp_entry_to_pte(entry);
 893         pte_t *pte;
 894         int ret = 0;
 895
 896         /*
 897          * We don't actually need pte lock while scanning for swp_pte: since
 898          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 899          * page table while we're scanning; though it could get zapped, and on
 900          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 901          * of unmatched parts which look like swp_pte, so unuse_pte must
 902          * recheck under pte lock.  Scanning without pte lock lets it be
 903          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 904          */
 905         pte = pte_offset_map(pmd, addr);
 906         do {
 907                 /*
 908                  * swapoff spends a _lot_ of time in this loop!
 909                  * Test inline before going to call unuse_pte.
 910                  */
 911                 if (unlikely(pte_same(*pte, swp_pte))) {
 912                         pte_unmap(pte);
 913                         ret = unuse_pte(vma, pmd, addr, entry, page);
 914                         if (ret)
 915                                 goto out;
 916                         pte = pte_offset_map(pmd, addr);
 917                 }
 918         } while (pte++, addr += PAGE_SIZE, addr != end);
 919         pte_unmap(pte - 1);
 920 out:
 921         return ret;
 922 }
 923
 924 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 925                                 unsigned long addr, unsigned long end,
 926                                 swp_entry_t entry, struct page *page)
 927 {
 928         pmd_t *pmd;
 929         unsigned long next;
 930         int ret;
 931
 932         pmd = pmd_offset(pud, addr);
 933         do {
 934                 next = pmd_addr_end(addr, end);
 935                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 936                         continue;
 937                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 938                 if (ret)
 939                         return ret;
 940         } while (pmd++, addr = next, addr != end);
 941         return 0;
 942 }
 943
 944 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 945                                 unsigned long addr, unsigned long end,
 946                                 swp_entry_t entry, struct page *page)
 947 {
 948         pud_t *pud;
 949         unsigned long next;
 950         int ret;
 951
 952         pud = pud_offset(pgd, addr);
 953         do {
 954                 next = pud_addr_end(addr, end);
 955                 if (pud_none_or_clear_bad(pud))
 956                         continue;
 957                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 958                 if (ret)
 959                         return ret;
 960         } while (pud++, addr = next, addr != end);
 961         return 0;
 962 }
 963
 964 static int unuse_vma(struct vm_area_struct *vma,
 965                                 swp_entry_t entry, struct page *page)
 966 {
 967         pgd_t *pgd;
 968         unsigned long addr, end, next;
 969         int ret;
 970
 971         if (page_anon_vma(page)) {
 972                 addr = page_address_in_vma(page, vma);
 973                 if (addr == -EFAULT)
 974                         return 0;
 975                 else
 976                         end = addr + PAGE_SIZE;
 977         } else {
 978                 addr = vma->vm_start;
 979                 end = vma->vm_end;
 980         }
 981
 982         pgd = pgd_offset(vma->vm_mm, addr);
 983         do {
 984                 next = pgd_addr_end(addr, end);
 985                 if (pgd_none_or_clear_bad(pgd))
 986                         continue;
 987                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 988                 if (ret)
 989                         return ret;
 990         } while (pgd++, addr = next, addr != end);
 991         return 0;
 992 }
 993
 994 static int unuse_mm(struct mm_struct *mm,
 995                                 swp_entry_t entry, struct page *page)
 996 {
 997         struct vm_area_struct *vma;
 998         int ret = 0;
 999
1000         if (!down_read_trylock(&mm->mmap_sem)) {
1001                 /*
1002                  * Activate page so shrink_inactive_list is unlikely to unmap
1003                  * its ptes while lock is dropped, so swapoff can make progress.
1004                  */
1005                 activate_page(page);
1006                 unlock_page(page);
1007                 down_read(&mm->mmap_sem);
1008                 lock_page(page);
1009         }
1010         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1011                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1012                         break;
1013         }
1014         up_read(&mm->mmap_sem);
1015         return (ret < 0)? ret: 0;
1016 }
1017
1018 /*
1019  * Scan swap_map from current position to next entry still in use.
1020  * Recycle to start on reaching the end, returning 0 when empty.
1021  */
1022 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1023                                         unsigned int prev)
1024 {
1025         unsigned int max = si->max;
1026         unsigned int i = prev;
1027         unsigned char count;
1028
1029         /*
1030          * No need for swap_lock here: we're just looking
1031          * for whether an entry is in use, not modifying it; false
1032          * hits are okay, and sys_swapoff() has already prevented new
1033          * allocations from this area (while holding swap_lock).
1034          */
1035         for (;;) {
1036                 if (++i >= max) {
1037                         if (!prev) {
1038                                 i = 0;
1039                                 break;
1040                         }
1041                         /*
1042                          * No entries in use at top of swap_map,
1043                          * loop back to start and recheck there.
1044                          */
1045                         max = prev + 1;
1046                         prev = 0;
1047                         i = 1;
1048                 }
1049                 count = si->swap_map[i];
1050                 if (count && swap_count(count) != SWAP_MAP_BAD)
1051                         break;
1052         }
1053         return i;
1054 }
1055
1056 /*
1057  * We completely avoid races by reading each swap page in advance,
1058  * and then search for the process using it.  All the necessary
1059  * page table adjustments can then be made atomically.
1060  */
1061 static int try_to_unuse(unsigned int type)
1062 {
1063         struct swap_info_struct *si = swap_info[type];
1064         struct mm_struct *start_mm;
1065         unsigned char *swap_map;
1066         unsigned char swcount;
1067         struct page *page;
1068         swp_entry_t entry;
1069         unsigned int i = 0;
1070         int retval = 0;
1071
1072         /*
1073          * When searching mms for an entry, a good strategy is to
1074          * start at the first mm we freed the previous entry from
1075          * (though actually we don't notice whether we or coincidence
1076          * freed the entry).  Initialize this start_mm with a hold.
1077          *
1078          * A simpler strategy would be to start at the last mm we
1079          * freed the previous entry from; but that would take less
1080          * advantage of mmlist ordering, which clusters forked mms
1081          * together, child after parent.  If we race with dup_mmap(), we
1082          * prefer to resolve parent before child, lest we miss entries
1083          * duplicated after we scanned child: using last mm would invert
1084          * that.
1085          */
1086         start_mm = &init_mm;
1087         atomic_inc(&init_mm.mm_users);
1088
1089         /*
1090          * Keep on scanning until all entries have gone.  Usually,
1091          * one pass through swap_map is enough, but not necessarily:
1092          * there are races when an instance of an entry might be missed.
1093          */
1094         while ((i = find_next_to_unuse(si, i)) != 0) {
1095                 if (signal_pending(current)) {
1096                         retval = -EINTR;
1097                         break;
1098                 }
1099
1100                 /*
1101                  * Get a page for the entry, using the existing swap
1102                  * cache page if there is one.  Otherwise, get a clean
1103                  * page and read the swap into it.
1104                  */
1105                 swap_map = &si->swap_map[i];
1106                 entry = swp_entry(type, i);
1107                 page = read_swap_cache_async(entry,
1108                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1109                 if (!page) {
1110                         /*
1111                          * Either swap_duplicate() failed because entry
1112                          * has been freed independently, and will not be
1113                          * reused since sys_swapoff() already disabled
1114                          * allocation from here, or alloc_page() failed.
1115                          */
1116                         if (!*swap_map)
1117                                 continue;
1118                         retval = -ENOMEM;
1119                         break;
1120                 }
1121
1122                 /*
1123                  * Don't hold on to start_mm if it looks like exiting.
1124                  */
1125                 if (atomic_read(&start_mm->mm_users) == 1) {
1126                         mmput(start_mm);
1127                         start_mm = &init_mm;
1128                         atomic_inc(&init_mm.mm_users);
1129                 }
1130
1131                 /*
1132                  * Wait for and lock page.  When do_swap_page races with
1133                  * try_to_unuse, do_swap_page can handle the fault much
1134                  * faster than try_to_unuse can locate the entry.  This
1135                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1136                  * defer to do_swap_page in such a case - in some tests,
1137                  * do_swap_page and try_to_unuse repeatedly compete.
1138                  */
1139                 wait_on_page_locked(page);
1140                 wait_on_page_writeback(page);
1141                 lock_page(page);
1142                 wait_on_page_writeback(page);
1143
1144                 /*
1145                  * Remove all references to entry.
1146                  */
1147                 swcount = *swap_map;
1148                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1149                         retval = shmem_unuse(entry, page);
1150                         /* page has already been unlocked and released */
1151                         if (retval < 0)
1152                                 break;
1153                         continue;
1154                 }
1155                 if (swap_count(swcount) && start_mm != &init_mm)
1156                         retval = unuse_mm(start_mm, entry, page);
1157
1158                 if (swap_count(*swap_map)) {
1159                         int set_start_mm = (*swap_map >= swcount);
1160                         struct list_head *p = &start_mm->mmlist;
1161                         struct mm_struct *new_start_mm = start_mm;
1162                         struct mm_struct *prev_mm = start_mm;
1163                         struct mm_struct *mm;
1164
1165                         atomic_inc(&new_start_mm->mm_users);
1166                         atomic_inc(&prev_mm->mm_users);
1167                         spin_lock(&mmlist_lock);
1168                         while (swap_count(*swap_map) && !retval &&
1169                                         (p = p->next) != &start_mm->mmlist) {
1170                                 mm = list_entry(p, struct mm_struct, mmlist);
1171                                 if (!atomic_inc_not_zero(&mm->mm_users))
1172                                         continue;
1173                                 spin_unlock(&mmlist_lock);
1174                                 mmput(prev_mm);
1175                                 prev_mm = mm;
1176
1177                                 cond_resched();
1178
1179                                 swcount = *swap_map;
1180                                 if (!swap_count(swcount)) /* any usage ? */
1181                                         ;
1182                                 else if (mm == &init_mm)
1183                                         set_start_mm = 1;
1184                                 else
1185                                         retval = unuse_mm(mm, entry, page);
1186
1187                                 if (set_start_mm && *swap_map < swcount) {
1188                                         mmput(new_start_mm);
1189                                         atomic_inc(&mm->mm_users);
1190                                         new_start_mm = mm;
1191                                         set_start_mm = 0;
1192                                 }
1193                                 spin_lock(&mmlist_lock);
1194                         }
1195                         spin_unlock(&mmlist_lock);
1196                         mmput(prev_mm);
1197                         mmput(start_mm);
1198                         start_mm = new_start_mm;
1199                 }
1200                 if (retval) {
1201                         unlock_page(page);
1202                         page_cache_release(page);
1203                         break;
1204                 }
1205
1206                 /*
1207                  * If a reference remains (rare), we would like to leave
1208                  * the page in the swap cache; but try_to_unmap could
1209                  * then re-duplicate the entry once we drop page lock,
1210                  * so we might loop indefinitely; also, that page could
1211                  * not be swapped out to other storage meanwhile.  So:
1212                  * delete from cache even if there's another reference,
1213                  * after ensuring that the data has been saved to disk -
1214                  * since if the reference remains (rarer), it will be
1215                  * read from disk into another page.  Splitting into two
1216                  * pages would be incorrect if swap supported "shared
1217                  * private" pages, but they are handled by tmpfs files.
1218                  *
1219                  * Given how unuse_vma() targets one particular offset
1220                  * in an anon_vma, once the anon_vma has been determined,
1221                  * this splitting happens to be just what is needed to
1222                  * handle where KSM pages have been swapped out: re-reading
1223                  * is unnecessarily slow, but we can fix that later on.
1224                  */
1225                 if (swap_count(*swap_map) &&
1226                      PageDirty(page) && PageSwapCache(page)) {
1227                         struct writeback_control wbc = {
1228                                 .sync_mode = WB_SYNC_NONE,
1229                         };
1230
1231                         swap_writepage(page, &wbc);
1232                         lock_page(page);
1233                         wait_on_page_writeback(page);
1234                 }
1235
1236                 /*
1237                  * It is conceivable that a racing task removed this page from
1238                  * swap cache just before we acquired the page lock at the top,
1239                  * or while we dropped it in unuse_mm().  The page might even
1240                  * be back in swap cache on another swap area: that we must not
1241                  * delete, since it may not have been written out to swap yet.
1242                  */
1243                 if (PageSwapCache(page) &&
1244                     likely(page_private(page) == entry.val))
1245                         delete_from_swap_cache(page);
1246
1247                 /*
1248                  * So we could skip searching mms once swap count went
1249                  * to 1, we did not mark any present ptes as dirty: must
1250                  * mark page dirty so shrink_page_list will preserve it.
1251                  */
1252                 SetPageDirty(page);
1253                 unlock_page(page);
1254                 page_cache_release(page);
1255
1256                 /*
1257                  * Make sure that we aren't completely killing
1258                  * interactive performance.
1259                  */
1260                 cond_resched();
1261         }
1262
1263         mmput(start_mm);
1264         return retval;
1265 }
1266
1267 /*
1268  * After a successful try_to_unuse, if no swap is now in use, we know
1269  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1270  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1271  * added to the mmlist just after page_duplicate - before would be racy.
1272  */
1273 static void drain_mmlist(void)
1274 {
1275         struct list_head *p, *next;
1276         unsigned int type;
1277
1278         for (type = 0; type < nr_swapfiles; type++)
1279                 if (swap_info[type]->inuse_pages)
1280                         return;
1281         spin_lock(&mmlist_lock);
1282         list_for_each_safe(p, next, &init_mm.mmlist)
1283                 list_del_init(p);
1284         spin_unlock(&mmlist_lock);
1285 }
1286
1287 /*
1288  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1289  * corresponds to page offset for the specified swap entry.
1290  * Note that the type of this function is sector_t, but it returns page offset
1291  * into the bdev, not sector offset.
1292  */
1293 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1294 {
1295         struct swap_info_struct *sis;
1296         struct swap_extent *start_se;
1297         struct swap_extent *se;
1298         pgoff_t offset;
1299
1300         sis = swap_info[swp_type(entry)];
1301         *bdev = sis->bdev;
1302
1303         offset = swp_offset(entry);
1304         start_se = sis->curr_swap_extent;
1305         se = start_se;
1306
1307         for ( ; ; ) {
1308                 struct list_head *lh;
1309
1310                 if (se->start_page <= offset &&
1311                                 offset < (se->start_page + se->nr_pages)) {
1312                         return se->start_block + (offset - se->start_page);
1313                 }
1314                 lh = se->list.next;
1315                 se = list_entry(lh, struct swap_extent, list);
1316                 sis->curr_swap_extent = se;
1317                 BUG_ON(se == start_se);         /* It *must* be present */
1318         }
1319 }
1320
1321 /*
1322  * Returns the page offset into bdev for the specified page's swap entry.
1323  */
1324 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1325 {
1326         swp_entry_t entry;
1327         entry.val = page_private(page);
1328         return map_swap_entry(entry, bdev);
1329 }
1330
1331 /*
1332  * Free all of a swapdev's extent information
1333  */
1334 static void destroy_swap_extents(struct swap_info_struct *sis)
1335 {
1336         while (!list_empty(&sis->first_swap_extent.list)) {
1337                 struct swap_extent *se;
1338
1339                 se = list_entry(sis->first_swap_extent.list.next,
1340                                 struct swap_extent, list);
1341                 list_del(&se->list);
1342                 kfree(se);
1343         }
1344 }
1345
1346 /*
1347  * Add a block range (and the corresponding page range) into this swapdev's
1348  * extent list.  The extent list is kept sorted in page order.
1349  *
1350  * This function rather assumes that it is called in ascending page order.
1351  */
1352 static int
1353 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1354                 unsigned long nr_pages, sector_t start_block)
1355 {
1356         struct swap_extent *se;
1357         struct swap_extent *new_se;
1358         struct list_head *lh;
1359
1360         if (start_page == 0) {
1361                 se = &sis->first_swap_extent;
1362                 sis->curr_swap_extent = se;
1363                 se->start_page = 0;
1364                 se->nr_pages = nr_pages;
1365                 se->start_block = start_block;
1366                 return 1;
1367         } else {
1368                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1369                 se = list_entry(lh, struct swap_extent, list);
1370                 BUG_ON(se->start_page + se->nr_pages != start_page);
1371                 if (se->start_block + se->nr_pages == start_block) {
1372                         /* Merge it */
1373                         se->nr_pages += nr_pages;
1374                         return 0;
1375                 }
1376         }
1377
1378         /*
1379          * No merge.  Insert a new extent, preserving ordering.
1380          */
1381         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1382         if (new_se == NULL)
1383                 return -ENOMEM;
1384         new_se->start_page = start_page;
1385         new_se->nr_pages = nr_pages;
1386         new_se->start_block = start_block;
1387
1388         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1389         return 1;
1390 }
1391
1392 /*
1393  * A `swap extent' is a simple thing which maps a contiguous range of pages
1394  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1395  * is built at swapon time and is then used at swap_writepage/swap_readpage
1396  * time for locating where on disk a page belongs.
1397  *
1398  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1399  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1400  * swap files identically.
1401  *
1402  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1403  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1404  * swapfiles are handled *identically* after swapon time.
1405  *
1406  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1407  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1408  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1409  * requirements, they are simply tossed out - we will never use those blocks
1410  * for swapping.
1411  *
1412  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1413  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1414  * which will scribble on the fs.
1415  *
1416  * The amount of disk space which a single swap extent represents varies.
1417  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1418  * extents in the list.  To avoid much list walking, we cache the previous
1419  * search location in `curr_swap_extent', and start new searches from there.
1420  * This is extremely effective.  The average number of iterations in
1421  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1422  */
1423 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1424 {
1425         struct inode *inode;
1426         unsigned blocks_per_page;
1427         unsigned long page_no;
1428         unsigned blkbits;
1429         sector_t probe_block;
1430         sector_t last_block;
1431         sector_t lowest_block = -1;
1432         sector_t highest_block = 0;
1433         int nr_extents = 0;
1434         int ret;
1435
1436         inode = sis->swap_file->f_mapping->host;
1437         if (S_ISBLK(inode->i_mode)) {
1438                 ret = add_swap_extent(sis, 0, sis->max, 0);
1439                 *span = sis->pages;
1440                 goto out;
1441         }
1442
1443         blkbits = inode->i_blkbits;
1444         blocks_per_page = PAGE_SIZE >> blkbits;
1445
1446         /*
1447          * Map all the blocks into the extent list.  This code doesn't try
1448          * to be very smart.
1449          */
1450         probe_block = 0;
1451         page_no = 0;
1452         last_block = i_size_read(inode) >> blkbits;
1453         while ((probe_block + blocks_per_page) <= last_block &&
1454                         page_no < sis->max) {
1455                 unsigned block_in_page;
1456                 sector_t first_block;
1457
1458                 first_block = bmap(inode, probe_block);
1459                 if (first_block == 0)
1460                         goto bad_bmap;
1461
1462                 /*
1463                  * It must be PAGE_SIZE aligned on-disk
1464                  */
1465                 if (first_block & (blocks_per_page - 1)) {
1466                         probe_block++;
1467                         goto reprobe;
1468                 }
1469
1470                 for (block_in_page = 1; block_in_page < blocks_per_page;
1471                                         block_in_page++) {
1472                         sector_t block;
1473
1474                         block = bmap(inode, probe_block + block_in_page);
1475                         if (block == 0)
1476                                 goto bad_bmap;
1477                         if (block != first_block + block_in_page) {
1478                                 /* Discontiguity */
1479                                 probe_block++;
1480                                 goto reprobe;
1481                         }
1482                 }
1483
1484                 first_block >>= (PAGE_SHIFT - blkbits);
1485                 if (page_no) {  /* exclude the header page */
1486                         if (first_block < lowest_block)
1487                                 lowest_block = first_block;
1488                         if (first_block > highest_block)
1489                                 highest_block = first_block;
1490                 }
1491
1492                 /*
1493                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1494                  */
1495                 ret = add_swap_extent(sis, page_no, 1, first_block);
1496                 if (ret < 0)
1497                         goto out;
1498                 nr_extents += ret;
1499                 page_no++;
1500                 probe_block += blocks_per_page;
1501 reprobe:
1502                 continue;
1503         }
1504         ret = nr_extents;
1505         *span = 1 + highest_block - lowest_block;
1506         if (page_no == 0)
1507                 page_no = 1;    /* force Empty message */
1508         sis->max = page_no;
1509         sis->pages = page_no - 1;
1510         sis->highest_bit = page_no - 1;
1511 out:
1512         return ret;
1513 bad_bmap:
1514         printk(KERN_ERR "swapon: swapfile has holes\n");
1515         ret = -EINVAL;
1516         goto out;
1517 }
1518
1519 static void enable_swap_info(struct swap_info_struct *p, int prio,
1520                                 unsigned char *swap_map)
1521 {
1522         int i, prev;
1523
1524         spin_lock(&swap_lock);
1525         if (prio >= 0)
1526                 p->prio = prio;
1527         else
1528                 p->prio = --least_priority;
1529         p->swap_map = swap_map;
1530         p->flags |= SWP_WRITEOK;
1531         nr_swap_pages += p->pages;
1532         total_swap_pages += p->pages;
1533
1534         /* insert swap space into swap_list: */
1535         prev = -1;
1536         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1537                 if (p->prio >= swap_info[i]->prio)
1538                         break;
1539                 prev = i;
1540         }
1541         p->next = i;
1542         if (prev < 0)
1543                 swap_list.head = swap_list.next = p->type;
1544         else
1545                 swap_info[prev]->next = p->type;
1546         spin_unlock(&swap_lock);
1547 }
1548
1549 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1550 {
1551         struct swap_info_struct *p = NULL;
1552         unsigned char *swap_map;
1553         struct file *swap_file, *victim;
1554         struct address_space *mapping;
1555         struct inode *inode;
1556         char *pathname;
1557         int oom_score_adj;
1558         int i, type, prev;
1559         int err;
1560
1561         if (!capable(CAP_SYS_ADMIN))
1562                 return -EPERM;
1563
1564         BUG_ON(!current->mm);
1565
1566         pathname = getname(specialfile);
1567         err = PTR_ERR(pathname);
1568         if (IS_ERR(pathname))
1569                 goto out;
1570
1571         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1572         putname(pathname);
1573         err = PTR_ERR(victim);
1574         if (IS_ERR(victim))
1575                 goto out;
1576
1577         mapping = victim->f_mapping;
1578         prev = -1;
1579         spin_lock(&swap_lock);
1580         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1581                 p = swap_info[type];
1582                 if (p->flags & SWP_WRITEOK) {
1583                         if (p->swap_file->f_mapping == mapping)
1584                                 break;
1585                 }
1586                 prev = type;
1587         }
1588         if (type < 0) {
1589                 err = -EINVAL;
1590                 spin_unlock(&swap_lock);
1591                 goto out_dput;
1592         }
1593         if (!security_vm_enough_memory_mm(current->mm, p->pages))
1594                 vm_unacct_memory(p->pages);
1595         else {
1596                 err = -ENOMEM;
1597                 spin_unlock(&swap_lock);
1598                 goto out_dput;
1599         }
1600         if (prev < 0)
1601                 swap_list.head = p->next;
1602         else
1603                 swap_info[prev]->next = p->next;
1604         if (type == swap_list.next) {
1605                 /* just pick something that's safe... */
1606                 swap_list.next = swap_list.head;
1607         }
1608         if (p->prio < 0) {
1609                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1610                         swap_info[i]->prio = p->prio--;
1611                 least_priority++;
1612         }
1613         nr_swap_pages -= p->pages;
1614         total_swap_pages -= p->pages;
1615         p->flags &= ~SWP_WRITEOK;
1616         spin_unlock(&swap_lock);
1617
1618         oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1619         err = try_to_unuse(type);
1620         compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj);
1621
1622         if (err) {
1623                 /*
1624                  * reading p->prio and p->swap_map outside the lock is
1625                  * safe here because only sys_swapon and sys_swapoff
1626                  * change them, and there can be no other sys_swapon or
1627                  * sys_swapoff for this swap_info_struct at this point.
1628                  */
1629                 /* re-insert swap space back into swap_list */
1630                 enable_swap_info(p, p->prio, p->swap_map);
1631                 goto out_dput;
1632         }
1633
1634         destroy_swap_extents(p);
1635         if (p->flags & SWP_CONTINUED)
1636                 free_swap_count_continuations(p);
1637
1638         mutex_lock(&swapon_mutex);
1639         spin_lock(&swap_lock);
1640         drain_mmlist();
1641
1642         /* wait for anyone still in scan_swap_map */
1643         p->highest_bit = 0;             /* cuts scans short */
1644         while (p->flags >= SWP_SCANNING) {
1645                 spin_unlock(&swap_lock);
1646                 schedule_timeout_uninterruptible(1);
1647                 spin_lock(&swap_lock);
1648         }
1649
1650         swap_file = p->swap_file;
1651         p->swap_file = NULL;
1652         p->max = 0;
1653         swap_map = p->swap_map;
1654         p->swap_map = NULL;
1655         p->flags = 0;
1656         spin_unlock(&swap_lock);
1657         mutex_unlock(&swapon_mutex);
1658         vfree(swap_map);
1659         /* Destroy swap account informatin */
1660         swap_cgroup_swapoff(type);
1661
1662         inode = mapping->host;
1663         if (S_ISBLK(inode->i_mode)) {
1664                 struct block_device *bdev = I_BDEV(inode);
1665                 set_blocksize(bdev, p->old_block_size);
1666                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1667         } else {
1668                 mutex_lock(&inode->i_mutex);
1669                 inode->i_flags &= ~S_SWAPFILE;
1670                 mutex_unlock(&inode->i_mutex);
1671         }
1672         filp_close(swap_file, NULL);
1673         err = 0;
1674         atomic_inc(&proc_poll_event);
1675         wake_up_interruptible(&proc_poll_wait);
1676
1677 out_dput:
1678         filp_close(victim, NULL);
1679 out:
1680         return err;
1681 }
1682
1683 #ifdef CONFIG_PROC_FS
1684 static unsigned swaps_poll(struct file *file, poll_table *wait)
1685 {
1686         struct seq_file *seq = file->private_data;
1687
1688         poll_wait(file, &proc_poll_wait, wait);
1689
1690         if (seq->poll_event != atomic_read(&proc_poll_event)) {
1691                 seq->poll_event = atomic_read(&proc_poll_event);
1692                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1693         }
1694
1695         return POLLIN | POLLRDNORM;
1696 }
1697
1698 /* iterator */
1699 static void *swap_start(struct seq_file *swap, loff_t *pos)
1700 {
1701         struct swap_info_struct *si;
1702         int type;
1703         loff_t l = *pos;
1704
1705         mutex_lock(&swapon_mutex);
1706
1707         if (!l)
1708                 return SEQ_START_TOKEN;
1709
1710         for (type = 0; type < nr_swapfiles; type++) {
1711                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1712                 si = swap_info[type];
1713                 if (!(si->flags & SWP_USED) || !si->swap_map)
1714                         continue;
1715                 if (!--l)
1716                         return si;
1717         }
1718
1719         return NULL;
1720 }
1721
1722 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1723 {
1724         struct swap_info_struct *si = v;
1725         int type;
1726
1727         if (v == SEQ_START_TOKEN)
1728                 type = 0;
1729         else
1730                 type = si->type + 1;
1731
1732         for (; type < nr_swapfiles; type++) {
1733                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1734                 si = swap_info[type];
1735                 if (!(si->flags & SWP_USED) || !si->swap_map)
1736                         continue;
1737                 ++*pos;
1738                 return si;
1739         }
1740
1741         return NULL;
1742 }
1743
1744 static void swap_stop(struct seq_file *swap, void *v)
1745 {
1746         mutex_unlock(&swapon_mutex);
1747 }
1748
1749 static int swap_show(struct seq_file *swap, void *v)
1750 {
1751         struct swap_info_struct *si = v;
1752         struct file *file;
1753         int len;
1754
1755         if (si == SEQ_START_TOKEN) {
1756                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1757                 return 0;
1758         }
1759
1760         file = si->swap_file;
1761         len = seq_path(swap, &file->f_path, " \t\n\\");
1762         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1763                         len < 40 ? 40 - len : 1, " ",
1764                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1765                                 "partition" : "file\t",
1766                         si->pages << (PAGE_SHIFT - 10),
1767                         si->inuse_pages << (PAGE_SHIFT - 10),
1768                         si->prio);
1769         return 0;
1770 }
1771
1772 static const struct seq_operations swaps_op = {
1773         .start =        swap_start,
1774         .next =         swap_next,
1775         .stop =         swap_stop,
1776         .show =         swap_show
1777 };
1778
1779 static int swaps_open(struct inode *inode, struct file *file)
1780 {
1781         struct seq_file *seq;
1782         int ret;
1783
1784         ret = seq_open(file, &swaps_op);
1785         if (ret)
1786                 return ret;
1787
1788         seq = file->private_data;
1789         seq->poll_event = atomic_read(&proc_poll_event);
1790         return 0;
1791 }
1792
1793 static const struct file_operations proc_swaps_operations = {
1794         .open           = swaps_open,
1795         .read           = seq_read,
1796         .llseek         = seq_lseek,
1797         .release        = seq_release,
1798         .poll           = swaps_poll,
1799 };
1800
1801 static int __init procswaps_init(void)
1802 {
1803         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1804         return 0;
1805 }
1806 __initcall(procswaps_init);
1807 #endif /* CONFIG_PROC_FS */
1808
1809 #ifdef MAX_SWAPFILES_CHECK
1810 static int __init max_swapfiles_check(void)
1811 {
1812         MAX_SWAPFILES_CHECK();
1813         return 0;
1814 }
1815 late_initcall(max_swapfiles_check);
1816 #endif
1817
1818 static struct swap_info_struct *alloc_swap_info(void)
1819 {
1820         struct swap_info_struct *p;
1821         unsigned int type;
1822
1823         p = kzalloc(sizeof(*p), GFP_KERNEL);
1824         if (!p)
1825                 return ERR_PTR(-ENOMEM);
1826
1827         spin_lock(&swap_lock);
1828         for (type = 0; type < nr_swapfiles; type++) {
1829                 if (!(swap_info[type]->flags & SWP_USED))
1830                         break;
1831         }
1832         if (type >= MAX_SWAPFILES) {
1833                 spin_unlock(&swap_lock);
1834                 kfree(p);
1835                 return ERR_PTR(-EPERM);
1836         }
1837         if (type >= nr_swapfiles) {
1838                 p->type = type;
1839                 swap_info[type] = p;
1840                 /*
1841                  * Write swap_info[type] before nr_swapfiles, in case a
1842                  * racing procfs swap_start() or swap_next() is reading them.
1843                  * (We never shrink nr_swapfiles, we never free this entry.)
1844                  */
1845                 smp_wmb();
1846                 nr_swapfiles++;
1847         } else {
1848                 kfree(p);
1849                 p = swap_info[type];
1850                 /*
1851                  * Do not memset this entry: a racing procfs swap_next()
1852                  * would be relying on p->type to remain valid.
1853                  */
1854         }
1855         INIT_LIST_HEAD(&p->first_swap_extent.list);
1856         p->flags = SWP_USED;
1857         p->next = -1;
1858         spin_unlock(&swap_lock);
1859
1860         return p;
1861 }
1862
1863 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1864 {
1865         int error;
1866
1867         if (S_ISBLK(inode->i_mode)) {
1868                 p->bdev = bdgrab(I_BDEV(inode));
1869                 error = blkdev_get(p->bdev,
1870                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1871                                    sys_swapon);
1872                 if (error < 0) {
1873                         p->bdev = NULL;
1874                         return -EINVAL;
1875                 }
1876                 p->old_block_size = block_size(p->bdev);
1877                 error = set_blocksize(p->bdev, PAGE_SIZE);
1878                 if (error < 0)
1879                         return error;
1880                 p->flags |= SWP_BLKDEV;
1881         } else if (S_ISREG(inode->i_mode)) {
1882                 p->bdev = inode->i_sb->s_bdev;
1883                 mutex_lock(&inode->i_mutex);
1884                 if (IS_SWAPFILE(inode))
1885                         return -EBUSY;
1886         } else
1887                 return -EINVAL;
1888
1889         return 0;
1890 }
1891
1892 static unsigned long read_swap_header(struct swap_info_struct *p,
1893                                         union swap_header *swap_header,
1894                                         struct inode *inode)
1895 {
1896         int i;
1897         unsigned long maxpages;
1898         unsigned long swapfilepages;
1899
1900         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1901                 printk(KERN_ERR "Unable to find swap-space signature\n");
1902                 return 0;
1903         }
1904
1905         /* swap partition endianess hack... */
1906         if (swab32(swap_header->info.version) == 1) {
1907                 swab32s(&swap_header->info.version);
1908                 swab32s(&swap_header->info.last_page);
1909                 swab32s(&swap_header->info.nr_badpages);
1910                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1911                         swab32s(&swap_header->info.badpages[i]);
1912         }
1913         /* Check the swap header's sub-version */
1914         if (swap_header->info.version != 1) {
1915                 printk(KERN_WARNING
1916                        "Unable to handle swap header version %d\n",
1917                        swap_header->info.version);
1918                 return 0;
1919         }
1920
1921         p->lowest_bit  = 1;
1922         p->cluster_next = 1;
1923         p->cluster_nr = 0;
1924
1925         /*
1926          * Find out how many pages are allowed for a single swap
1927          * device. There are three limiting factors: 1) the number
1928          * of bits for the swap offset in the swp_entry_t type, and
1929          * 2) the number of bits in the swap pte as defined by the
1930          * the different architectures, and 3) the number of free bits
1931          * in an exceptional radix_tree entry. In order to find the
1932          * largest possible bit mask, a swap entry with swap type 0
1933          * and swap offset ~0UL is created, encoded to a swap pte,
1934          * decoded to a swp_entry_t again, and finally the swap
1935          * offset is extracted. This will mask all the bits from
1936          * the initial ~0UL mask that can't be encoded in either
1937          * the swp_entry_t or the architecture definition of a
1938          * swap pte.  Then the same is done for a radix_tree entry.
1939          */
1940         maxpages = swp_offset(pte_to_swp_entry(
1941                         swp_entry_to_pte(swp_entry(0, ~0UL))));
1942         maxpages = swp_offset(radix_to_swp_entry(
1943                         swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1944
1945         if (maxpages > swap_header->info.last_page) {
1946                 maxpages = swap_header->info.last_page + 1;
1947                 /* p->max is an unsigned int: don't overflow it */
1948                 if ((unsigned int)maxpages == 0)
1949                         maxpages = UINT_MAX;
1950         }
1951         p->highest_bit = maxpages - 1;
1952
1953         if (!maxpages)
1954                 return 0;
1955         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1956         if (swapfilepages && maxpages > swapfilepages) {
1957                 printk(KERN_WARNING
1958                        "Swap area shorter than signature indicates\n");
1959                 return 0;
1960         }
1961         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1962                 return 0;
1963         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1964                 return 0;
1965
1966         return maxpages;
1967 }
1968
1969 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1970                                         union swap_header *swap_header,
1971                                         unsigned char *swap_map,
1972                                         unsigned long maxpages,
1973                                         sector_t *span)
1974 {
1975         int i;
1976         unsigned int nr_good_pages;
1977         int nr_extents;
1978
1979         nr_good_pages = maxpages - 1;   /* omit header page */
1980
1981         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1982                 unsigned int page_nr = swap_header->info.badpages[i];
1983                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1984                         return -EINVAL;
1985                 if (page_nr < maxpages) {
1986                         swap_map[page_nr] = SWAP_MAP_BAD;
1987                         nr_good_pages--;
1988                 }
1989         }
1990
1991         if (nr_good_pages) {
1992                 swap_map[0] = SWAP_MAP_BAD;
1993                 p->max = maxpages;
1994                 p->pages = nr_good_pages;
1995                 nr_extents = setup_swap_extents(p, span);
1996                 if (nr_extents < 0)
1997                         return nr_extents;
1998                 nr_good_pages = p->pages;
1999         }
2000         if (!nr_good_pages) {
2001                 printk(KERN_WARNING "Empty swap-file\n");
2002                 return -EINVAL;
2003         }
2004
2005         return nr_extents;
2006 }
2007
2008 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2009 {
2010         struct swap_info_struct *p;
2011         char *name;
2012         struct file *swap_file = NULL;
2013         struct address_space *mapping;
2014         int i;
2015         int prio;
2016         int error;
2017         union swap_header *swap_header;
2018         int nr_extents;
2019         sector_t span;
2020         unsigned long maxpages;
2021         unsigned char *swap_map = NULL;
2022         struct page *page = NULL;
2023         struct inode *inode = NULL;
2024
2025         if (!capable(CAP_SYS_ADMIN))
2026                 return -EPERM;
2027
2028         p = alloc_swap_info();
2029         if (IS_ERR(p))
2030                 return PTR_ERR(p);
2031
2032         name = getname(specialfile);
2033         if (IS_ERR(name)) {
2034                 error = PTR_ERR(name);
2035                 name = NULL;
2036                 goto bad_swap;
2037         }
2038         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2039         if (IS_ERR(swap_file)) {
2040                 error = PTR_ERR(swap_file);
2041                 swap_file = NULL;
2042                 goto bad_swap;
2043         }
2044
2045         p->swap_file = swap_file;
2046         mapping = swap_file->f_mapping;
2047
2048         for (i = 0; i < nr_swapfiles; i++) {
2049                 struct swap_info_struct *q = swap_info[i];
2050
2051                 if (q == p || !q->swap_file)
2052                         continue;
2053                 if (mapping == q->swap_file->f_mapping) {
2054                         error = -EBUSY;
2055                         goto bad_swap;
2056                 }
2057         }
2058
2059         inode = mapping->host;
2060         /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2061         error = claim_swapfile(p, inode);
2062         if (unlikely(error))
2063                 goto bad_swap;
2064
2065         /*
2066          * Read the swap header.
2067          */
2068         if (!mapping->a_ops->readpage) {
2069                 error = -EINVAL;
2070                 goto bad_swap;
2071         }
2072         page = read_mapping_page(mapping, 0, swap_file);
2073         if (IS_ERR(page)) {
2074                 error = PTR_ERR(page);
2075                 goto bad_swap;
2076         }
2077         swap_header = kmap(page);
2078
2079         maxpages = read_swap_header(p, swap_header, inode);
2080         if (unlikely(!maxpages)) {
2081                 error = -EINVAL;
2082                 goto bad_swap;
2083         }
2084
2085         /* OK, set up the swap map and apply the bad block list */
2086         swap_map = vzalloc(maxpages);
2087         if (!swap_map) {
2088                 error = -ENOMEM;
2089                 goto bad_swap;
2090         }
2091
2092         error = swap_cgroup_swapon(p->type, maxpages);
2093         if (error)
2094                 goto bad_swap;
2095
2096         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2097                 maxpages, &span);
2098         if (unlikely(nr_extents < 0)) {
2099                 error = nr_extents;
2100                 goto bad_swap;
2101         }
2102
2103         if (p->bdev) {
2104                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2105                         p->flags |= SWP_SOLIDSTATE;
2106                         p->cluster_next = 1 + (random32() % p->highest_bit);
2107                 }
2108                 if ((swap_flags & SWAP_FLAG_DISCARD) && discard_swap(p) == 0)
2109                         p->flags |= SWP_DISCARDABLE;
2110         }
2111
2112         mutex_lock(&swapon_mutex);
2113         prio = -1;
2114         if (swap_flags & SWAP_FLAG_PREFER)
2115                 prio =
2116                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2117         enable_swap_info(p, prio, swap_map);
2118
2119         printk(KERN_INFO "Adding %uk swap on %s.  "
2120                         "Priority:%d extents:%d across:%lluk %s%s\n",
2121                 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2122                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2123                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2124                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2125
2126         mutex_unlock(&swapon_mutex);
2127         atomic_inc(&proc_poll_event);
2128         wake_up_interruptible(&proc_poll_wait);
2129
2130         if (S_ISREG(inode->i_mode))
2131                 inode->i_flags |= S_SWAPFILE;
2132         error = 0;
2133         goto out;
2134 bad_swap:
2135         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2136                 set_blocksize(p->bdev, p->old_block_size);
2137                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2138         }
2139         destroy_swap_extents(p);
2140         swap_cgroup_swapoff(p->type);
2141         spin_lock(&swap_lock);
2142         p->swap_file = NULL;
2143         p->flags = 0;
2144         spin_unlock(&swap_lock);
2145         vfree(swap_map);
2146         if (swap_file) {
2147                 if (inode && S_ISREG(inode->i_mode)) {
2148                         mutex_unlock(&inode->i_mutex);
2149                         inode = NULL;
2150                 }
2151                 filp_close(swap_file, NULL);
2152         }
2153 out:
2154         if (page && !IS_ERR(page)) {
2155                 kunmap(page);
2156                 page_cache_release(page);
2157         }
2158         if (name)
2159                 putname(name);
2160         if (inode && S_ISREG(inode->i_mode))
2161                 mutex_unlock(&inode->i_mutex);
2162         return error;
2163 }
2164
2165 void si_swapinfo(struct sysinfo *val)
2166 {
2167         unsigned int type;
2168         unsigned long nr_to_be_unused = 0;
2169
2170         spin_lock(&swap_lock);
2171         for (type = 0; type < nr_swapfiles; type++) {
2172                 struct swap_info_struct *si = swap_info[type];
2173
2174                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2175                         nr_to_be_unused += si->inuse_pages;
2176         }
2177         val->freeswap = nr_swap_pages + nr_to_be_unused;
2178         val->totalswap = total_swap_pages + nr_to_be_unused;
2179         spin_unlock(&swap_lock);
2180 }
2181
2182 /*
2183  * Verify that a swap entry is valid and increment its swap map count.
2184  *
2185  * Returns error code in following case.
2186  * - success -> 0
2187  * - swp_entry is invalid -> EINVAL
2188  * - swp_entry is migration entry -> EINVAL
2189  * - swap-cache reference is requested but there is already one. -> EEXIST
2190  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2191  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2192  */
2193 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2194 {
2195         struct swap_info_struct *p;
2196         unsigned long offset, type;
2197         unsigned char count;
2198         unsigned char has_cache;
2199         int err = -EINVAL;
2200
2201         if (non_swap_entry(entry))
2202                 goto out;
2203
2204         type = swp_type(entry);
2205         if (type >= nr_swapfiles)
2206                 goto bad_file;
2207         p = swap_info[type];
2208         offset = swp_offset(entry);
2209
2210         spin_lock(&swap_lock);
2211         if (unlikely(offset >= p->max))
2212                 goto unlock_out;
2213
2214         count = p->swap_map[offset];
2215         has_cache = count & SWAP_HAS_CACHE;
2216         count &= ~SWAP_HAS_CACHE;
2217         err = 0;
2218
2219         if (usage == SWAP_HAS_CACHE) {
2220
2221                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2222                 if (!has_cache && count)
2223                         has_cache = SWAP_HAS_CACHE;
2224                 else if (has_cache)             /* someone else added cache */
2225                         err = -EEXIST;
2226                 else                            /* no users remaining */
2227                         err = -ENOENT;
2228
2229         } else if (count || has_cache) {
2230
2231                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2232                         count += usage;
2233                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2234                         err = -EINVAL;
2235                 else if (swap_count_continued(p, offset, count))
2236                         count = COUNT_CONTINUED;
2237                 else
2238                         err = -ENOMEM;
2239         } else
2240                 err = -ENOENT;                  /* unused swap entry */
2241
2242         p->swap_map[offset] = count | has_cache;
2243
2244 unlock_out:
2245         spin_unlock(&swap_lock);
2246 out:
2247         return err;
2248
2249 bad_file:
2250         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2251         goto out;
2252 }
2253
2254 /*
2255  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2256  * (in which case its reference count is never incremented).
2257  */
2258 void swap_shmem_alloc(swp_entry_t entry)
2259 {
2260         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2261 }
2262
2263 /*
2264  * Increase reference count of swap entry by 1.
2265  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2266  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2267  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2268  * might occur if a page table entry has got corrupted.
2269  */
2270 int swap_duplicate(swp_entry_t entry)
2271 {
2272         int err = 0;
2273
2274         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2275                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2276         return err;
2277 }
2278
2279 /*
2280  * @entry: swap entry for which we allocate swap cache.
2281  *
2282  * Called when allocating swap cache for existing swap entry,
2283  * This can return error codes. Returns 0 at success.
2284  * -EBUSY means there is a swap cache.
2285  * Note: return code is different from swap_duplicate().
2286  */
2287 int swapcache_prepare(swp_entry_t entry)
2288 {
2289         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2290 }
2291
2292 /*
2293  * add_swap_count_continuation - called when a swap count is duplicated
2294  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2295  * page of the original vmalloc'ed swap_map, to hold the continuation count
2296  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2297  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2298  *
2299  * These continuation pages are seldom referenced: the common paths all work
2300  * on the original swap_map, only referring to a continuation page when the
2301  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2302  *
2303  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2304  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2305  * can be called after dropping locks.
2306  */
2307 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2308 {
2309         struct swap_info_struct *si;
2310         struct page *head;
2311         struct page *page;
2312         struct page *list_page;
2313         pgoff_t offset;
2314         unsigned char count;
2315
2316         /*
2317          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2318          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2319          */
2320         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2321
2322         si = swap_info_get(entry);
2323         if (!si) {
2324                 /*
2325                  * An acceptable race has occurred since the failing
2326                  * __swap_duplicate(): the swap entry has been freed,
2327                  * perhaps even the whole swap_map cleared for swapoff.
2328                  */
2329                 goto outer;
2330         }
2331
2332         offset = swp_offset(entry);
2333         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2334
2335         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2336                 /*
2337                  * The higher the swap count, the more likely it is that tasks
2338                  * will race to add swap count continuation: we need to avoid
2339                  * over-provisioning.
2340                  */
2341                 goto out;
2342         }
2343
2344         if (!page) {
2345                 spin_unlock(&swap_lock);
2346                 return -ENOMEM;
2347         }
2348
2349         /*
2350          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2351          * no architecture is using highmem pages for kernel pagetables: so it
2352          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2353          */
2354         head = vmalloc_to_page(si->swap_map + offset);
2355         offset &= ~PAGE_MASK;
2356
2357         /*
2358          * Page allocation does not initialize the page's lru field,
2359          * but it does always reset its private field.
2360          */
2361         if (!page_private(head)) {
2362                 BUG_ON(count & COUNT_CONTINUED);
2363                 INIT_LIST_HEAD(&head->lru);
2364                 set_page_private(head, SWP_CONTINUED);
2365                 si->flags |= SWP_CONTINUED;
2366         }
2367
2368         list_for_each_entry(list_page, &head->lru, lru) {
2369                 unsigned char *map;
2370
2371                 /*
2372                  * If the previous map said no continuation, but we've found
2373                  * a continuation page, free our allocation and use this one.
2374                  */
2375                 if (!(count & COUNT_CONTINUED))
2376                         goto out;
2377
2378                 map = kmap_atomic(list_page) + offset;
2379                 count = *map;
2380                 kunmap_atomic(map);
2381
2382                 /*
2383                  * If this continuation count now has some space in it,
2384                  * free our allocation and use this one.
2385                  */
2386                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2387                         goto out;
2388         }
2389
2390         list_add_tail(&page->lru, &head->lru);
2391         page = NULL;                    /* now it's attached, don't free it */
2392 out:
2393         spin_unlock(&swap_lock);
2394 outer:
2395         if (page)
2396                 __free_page(page);
2397         return 0;
2398 }
2399
2400 /*
2401  * swap_count_continued - when the original swap_map count is incremented
2402  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2403  * into, carry if so, or else fail until a new continuation page is allocated;
2404  * when the original swap_map count is decremented from 0 with continuation,
2405  * borrow from the continuation and report whether it still holds more.
2406  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2407  */
2408 static bool swap_count_continued(struct swap_info_struct *si,
2409                                  pgoff_t offset, unsigned char count)
2410 {
2411         struct page *head;
2412         struct page *page;
2413         unsigned char *map;
2414
2415         head = vmalloc_to_page(si->swap_map + offset);
2416         if (page_private(head) != SWP_CONTINUED) {
2417                 BUG_ON(count & COUNT_CONTINUED);
2418                 return false;           /* need to add count continuation */
2419         }
2420
2421         offset &= ~PAGE_MASK;
2422         page = list_entry(head->lru.next, struct page, lru);
2423         map = kmap_atomic(page) + offset;
2424
2425         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2426                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2427
2428         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2429                 /*
2430                  * Think of how you add 1 to 999
2431                  */
2432                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2433                         kunmap_atomic(map);
2434                         page = list_entry(page->lru.next, struct page, lru);
2435                         BUG_ON(page == head);
2436                         map = kmap_atomic(page) + offset;
2437                 }
2438                 if (*map == SWAP_CONT_MAX) {
2439                         kunmap_atomic(map);
2440                         page = list_entry(page->lru.next, struct page, lru);
2441                         if (page == head)
2442                                 return false;   /* add count continuation */
2443                         map = kmap_atomic(page) + offset;
2444 init_map:               *map = 0;               /* we didn't zero the page */
2445                 }
2446                 *map += 1;
2447                 kunmap_atomic(map);
2448                 page = list_entry(page->lru.prev, struct page, lru);
2449                 while (page != head) {
2450                         map = kmap_atomic(page) + offset;
2451                         *map = COUNT_CONTINUED;
2452                         kunmap_atomic(map);
2453                         page = list_entry(page->lru.prev, struct page, lru);
2454                 }
2455                 return true;                    /* incremented */
2456
2457         } else {                                /* decrementing */
2458                 /*
2459                  * Think of how you subtract 1 from 1000
2460                  */
2461                 BUG_ON(count != COUNT_CONTINUED);
2462                 while (*map == COUNT_CONTINUED) {
2463                         kunmap_atomic(map);
2464                         page = list_entry(page->lru.next, struct page, lru);
2465                         BUG_ON(page == head);
2466                         map = kmap_atomic(page) + offset;
2467                 }
2468                 BUG_ON(*map == 0);
2469                 *map -= 1;
2470                 if (*map == 0)
2471                         count = 0;
2472                 kunmap_atomic(map);
2473                 page = list_entry(page->lru.prev, struct page, lru);
2474                 while (page != head) {
2475                         map = kmap_atomic(page) + offset;
2476                         *map = SWAP_CONT_MAX | count;
2477                         count = COUNT_CONTINUED;
2478                         kunmap_atomic(map);
2479                         page = list_entry(page->lru.prev, struct page, lru);
2480                 }
2481                 return count == COUNT_CONTINUED;
2482         }
2483 }
2484
2485 /*
2486  * free_swap_count_continuations - swapoff free all the continuation pages
2487  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2488  */
2489 static void free_swap_count_continuations(struct swap_info_struct *si)
2490 {
2491         pgoff_t offset;
2492
2493         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2494                 struct page *head;
2495                 head = vmalloc_to_page(si->swap_map + offset);
2496                 if (page_private(head)) {
2497                         struct list_head *this, *next;
2498                         list_for_each_safe(this, next, &head->lru) {
2499                                 struct page *page;
2500                                 page = list_entry(this, struct page, lru);
2501                                 list_del(this);
2502                                 __free_page(page);
2503                         }
2504                 }
2505         }
2506 }