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