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 #include <linux/export.h>
  37
  38 #include <asm/pgtable.h>
  39 #include <asm/tlbflush.h>
  40 #include <linux/swapops.h>
  41 #include <linux/page_cgroup.h>
  42
  43 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
  44                                  unsigned char);
  45 static void free_swap_count_continuations(struct swap_info_struct *);
  46 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
  47
  48 DEFINE_SPINLOCK(swap_lock);
  49 static unsigned int nr_swapfiles;
  50 atomic_long_t nr_swap_pages;
  51 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
  52 long total_swap_pages;
  53 static int least_priority;
  54 static atomic_t highest_priority_index = ATOMIC_INIT(-1);
  55
  56 static const char Bad_file[] = "Bad swap file entry ";
  57 static const char Unused_file[] = "Unused swap file entry ";
  58 static const char Bad_offset[] = "Bad swap offset entry ";
  59 static const char Unused_offset[] = "Unused swap offset entry ";
  60
  61 struct swap_list_t swap_list = {-1, -1};
  62
  63 struct swap_info_struct *swap_info[MAX_SWAPFILES];
  64
  65 static DEFINE_MUTEX(swapon_mutex);
  66
  67 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
  68 /* Activity counter to indicate that a swapon or swapoff has occurred */
  69 static atomic_t proc_poll_event = ATOMIC_INIT(0);
  70
  71 static inline unsigned char swap_count(unsigned char ent)
  72 {
  73         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
  74 }
  75
  76 /* returns 1 if swap entry is freed */
  77 static int
  78 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
  79 {
  80         swp_entry_t entry = swp_entry(si->type, offset);
  81         struct page *page;
  82         int ret = 0;
  83
  84         page = find_get_page(swap_address_space(entry), entry.val);
  85         if (!page)
  86                 return 0;
  87         /*
  88          * This function is called from scan_swap_map() and it's called
  89          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
  90          * We have to use trylock for avoiding deadlock. This is a special
  91          * case and you should use try_to_free_swap() with explicit lock_page()
  92          * in usual operations.
  93          */
  94         if (trylock_page(page)) {
  95                 ret = try_to_free_swap(page);
  96                 unlock_page(page);
  97         }
  98         page_cache_release(page);
  99         return ret;
 100 }
 101
 102 /*
 103  * swapon tell device that all the old swap contents can be discarded,
 104  * to allow the swap device to optimize its wear-levelling.
 105  */
 106 static int discard_swap(struct swap_info_struct *si)
 107 {
 108         struct swap_extent *se;
 109         sector_t start_block;
 110         sector_t nr_blocks;
 111         int err = 0;
 112
 113         /* Do not discard the swap header page! */
 114         se = &si->first_swap_extent;
 115         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
 116         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
 117         if (nr_blocks) {
 118                 err = blkdev_issue_discard(si->bdev, start_block,
 119                                 nr_blocks, GFP_KERNEL, 0);
 120                 if (err)
 121                         return err;
 122                 cond_resched();
 123         }
 124
 125         list_for_each_entry(se, &si->first_swap_extent.list, list) {
 126                 start_block = se->start_block << (PAGE_SHIFT - 9);
 127                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 128
 129                 err = blkdev_issue_discard(si->bdev, start_block,
 130                                 nr_blocks, GFP_KERNEL, 0);
 131                 if (err)
 132                         break;
 133
 134                 cond_resched();
 135         }
 136         return err;             /* That will often be -EOPNOTSUPP */
 137 }
 138
 139 /*
 140  * swap allocation tell device that a cluster of swap can now be discarded,
 141  * to allow the swap device to optimize its wear-levelling.
 142  */
 143 static void discard_swap_cluster(struct swap_info_struct *si,
 144                                  pgoff_t start_page, pgoff_t nr_pages)
 145 {
 146         struct swap_extent *se = si->curr_swap_extent;
 147         int found_extent = 0;
 148
 149         while (nr_pages) {
 150                 struct list_head *lh;
 151
 152                 if (se->start_page <= start_page &&
 153                     start_page < se->start_page + se->nr_pages) {
 154                         pgoff_t offset = start_page - se->start_page;
 155                         sector_t start_block = se->start_block + offset;
 156                         sector_t nr_blocks = se->nr_pages - offset;
 157
 158                         if (nr_blocks > nr_pages)
 159                                 nr_blocks = nr_pages;
 160                         start_page += nr_blocks;
 161                         nr_pages -= nr_blocks;
 162
 163                         if (!found_extent++)
 164                                 si->curr_swap_extent = se;
 165
 166                         start_block <<= PAGE_SHIFT - 9;
 167                         nr_blocks <<= PAGE_SHIFT - 9;
 168                         if (blkdev_issue_discard(si->bdev, start_block,
 169                                     nr_blocks, GFP_NOIO, 0))
 170                                 break;
 171                 }
 172
 173                 lh = se->list.next;
 174                 se = list_entry(lh, struct swap_extent, list);
 175         }
 176 }
 177
 178 static int wait_for_discard(void *word)
 179 {
 180         schedule();
 181         return 0;
 182 }
 183
 184 #define SWAPFILE_CLUSTER        256
 185 #define LATENCY_LIMIT           256
 186
 187 static unsigned long scan_swap_map(struct swap_info_struct *si,
 188                                    unsigned char usage)
 189 {
 190         unsigned long offset;
 191         unsigned long scan_base;
 192         unsigned long last_in_cluster = 0;
 193         int latency_ration = LATENCY_LIMIT;
 194         int found_free_cluster = 0;
 195
 196         /*
 197          * We try to cluster swap pages by allocating them sequentially
 198          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 199          * way, however, we resort to first-free allocation, starting
 200          * a new cluster.  This prevents us from scattering swap pages
 201          * all over the entire swap partition, so that we reduce
 202          * overall disk seek times between swap pages.  -- sct
 203          * But we do now try to find an empty cluster.  -Andrea
 204          * And we let swap pages go all over an SSD partition.  Hugh
 205          */
 206
 207         si->flags += SWP_SCANNING;
 208         scan_base = offset = si->cluster_next;
 209
 210         if (unlikely(!si->cluster_nr--)) {
 211                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 212                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
 213                         goto checks;
 214                 }
 215                 if (si->flags & SWP_PAGE_DISCARD) {
 216                         /*
 217                          * Start range check on racing allocations, in case
 218                          * they overlap the cluster we eventually decide on
 219                          * (we scan without swap_lock to allow preemption).
 220                          * It's hardly conceivable that cluster_nr could be
 221                          * wrapped during our scan, but don't depend on it.
 222                          */
 223                         if (si->lowest_alloc)
 224                                 goto checks;
 225                         si->lowest_alloc = si->max;
 226                         si->highest_alloc = 0;
 227                 }
 228                 spin_unlock(&si->lock);
 229
 230                 /*
 231                  * If seek is expensive, start searching for new cluster from
 232                  * start of partition, to minimize the span of allocated swap.
 233                  * But if seek is cheap, search from our current position, so
 234                  * that swap is allocated from all over the partition: if the
 235                  * Flash Translation Layer only remaps within limited zones,
 236                  * we don't want to wear out the first zone too quickly.
 237                  */
 238                 if (!(si->flags & SWP_SOLIDSTATE))
 239                         scan_base = offset = si->lowest_bit;
 240                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 241
 242                 /* Locate the first empty (unaligned) cluster */
 243                 for (; last_in_cluster <= si->highest_bit; offset++) {
 244                         if (si->swap_map[offset])
 245                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 246                         else if (offset == last_in_cluster) {
 247                                 spin_lock(&si->lock);
 248                                 offset -= SWAPFILE_CLUSTER - 1;
 249                                 si->cluster_next = offset;
 250                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 251                                 found_free_cluster = 1;
 252                                 goto checks;
 253                         }
 254                         if (unlikely(--latency_ration < 0)) {
 255                                 cond_resched();
 256                                 latency_ration = LATENCY_LIMIT;
 257                         }
 258                 }
 259
 260                 offset = si->lowest_bit;
 261                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 262
 263                 /* Locate the first empty (unaligned) cluster */
 264                 for (; last_in_cluster < scan_base; offset++) {
 265                         if (si->swap_map[offset])
 266                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 267                         else if (offset == last_in_cluster) {
 268                                 spin_lock(&si->lock);
 269                                 offset -= SWAPFILE_CLUSTER - 1;
 270                                 si->cluster_next = offset;
 271                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 272                                 found_free_cluster = 1;
 273                                 goto checks;
 274                         }
 275                         if (unlikely(--latency_ration < 0)) {
 276                                 cond_resched();
 277                                 latency_ration = LATENCY_LIMIT;
 278                         }
 279                 }
 280
 281                 offset = scan_base;
 282                 spin_lock(&si->lock);
 283                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 284                 si->lowest_alloc = 0;
 285         }
 286
 287 checks:
 288         if (!(si->flags & SWP_WRITEOK))
 289                 goto no_page;
 290         if (!si->highest_bit)
 291                 goto no_page;
 292         if (offset > si->highest_bit)
 293                 scan_base = offset = si->lowest_bit;
 294
 295         /* reuse swap entry of cache-only swap if not busy. */
 296         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 297                 int swap_was_freed;
 298                 spin_unlock(&si->lock);
 299                 swap_was_freed = __try_to_reclaim_swap(si, offset);
 300                 spin_lock(&si->lock);
 301                 /* entry was freed successfully, try to use this again */
 302                 if (swap_was_freed)
 303                         goto checks;
 304                 goto scan; /* check next one */
 305         }
 306
 307         if (si->swap_map[offset])
 308                 goto scan;
 309
 310         if (offset == si->lowest_bit)
 311                 si->lowest_bit++;
 312         if (offset == si->highest_bit)
 313                 si->highest_bit--;
 314         si->inuse_pages++;
 315         if (si->inuse_pages == si->pages) {
 316                 si->lowest_bit = si->max;
 317                 si->highest_bit = 0;
 318         }
 319         si->swap_map[offset] = usage;
 320         si->cluster_next = offset + 1;
 321         si->flags -= SWP_SCANNING;
 322
 323         if (si->lowest_alloc) {
 324                 /*
 325                  * Only set when SWP_PAGE_DISCARD, and there's a scan
 326                  * for a free cluster in progress or just completed.
 327                  */
 328                 if (found_free_cluster) {
 329                         /*
 330                          * To optimize wear-levelling, discard the
 331                          * old data of the cluster, taking care not to
 332                          * discard any of its pages that have already
 333                          * been allocated by racing tasks (offset has
 334                          * already stepped over any at the beginning).
 335                          */
 336                         if (offset < si->highest_alloc &&
 337                             si->lowest_alloc <= last_in_cluster)
 338                                 last_in_cluster = si->lowest_alloc - 1;
 339                         si->flags |= SWP_DISCARDING;
 340                         spin_unlock(&si->lock);
 341
 342                         if (offset < last_in_cluster)
 343                                 discard_swap_cluster(si, offset,
 344                                         last_in_cluster - offset + 1);
 345
 346                         spin_lock(&si->lock);
 347                         si->lowest_alloc = 0;
 348                         si->flags &= ~SWP_DISCARDING;
 349
 350                         smp_mb();       /* wake_up_bit advises this */
 351                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
 352
 353                 } else if (si->flags & SWP_DISCARDING) {
 354                         /*
 355                          * Delay using pages allocated by racing tasks
 356                          * until the whole discard has been issued. We
 357                          * could defer that delay until swap_writepage,
 358                          * but it's easier to keep this self-contained.
 359                          */
 360                         spin_unlock(&si->lock);
 361                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
 362                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
 363                         spin_lock(&si->lock);
 364                 } else {
 365                         /*
 366                          * Note pages allocated by racing tasks while
 367                          * scan for a free cluster is in progress, so
 368                          * that its final discard can exclude them.
 369                          */
 370                         if (offset < si->lowest_alloc)
 371                                 si->lowest_alloc = offset;
 372                         if (offset > si->highest_alloc)
 373                                 si->highest_alloc = offset;
 374                 }
 375         }
 376         return offset;
 377
 378 scan:
 379         spin_unlock(&si->lock);
 380         while (++offset <= si->highest_bit) {
 381                 if (!si->swap_map[offset]) {
 382                         spin_lock(&si->lock);
 383                         goto checks;
 384                 }
 385                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 386                         spin_lock(&si->lock);
 387                         goto checks;
 388                 }
 389                 if (unlikely(--latency_ration < 0)) {
 390                         cond_resched();
 391                         latency_ration = LATENCY_LIMIT;
 392                 }
 393         }
 394         offset = si->lowest_bit;
 395         while (++offset < scan_base) {
 396                 if (!si->swap_map[offset]) {
 397                         spin_lock(&si->lock);
 398                         goto checks;
 399                 }
 400                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 401                         spin_lock(&si->lock);
 402                         goto checks;
 403                 }
 404                 if (unlikely(--latency_ration < 0)) {
 405                         cond_resched();
 406                         latency_ration = LATENCY_LIMIT;
 407                 }
 408         }
 409         spin_lock(&si->lock);
 410
 411 no_page:
 412         si->flags -= SWP_SCANNING;
 413         return 0;
 414 }
 415
 416 swp_entry_t get_swap_page(void)
 417 {
 418         struct swap_info_struct *si;
 419         pgoff_t offset;
 420         int type, next;
 421         int wrapped = 0;
 422         int hp_index;
 423
 424         spin_lock(&swap_lock);
 425         if (atomic_long_read(&nr_swap_pages) <= 0)
 426                 goto noswap;
 427         atomic_long_dec(&nr_swap_pages);
 428
 429         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 430                 hp_index = atomic_xchg(&highest_priority_index, -1);
 431                 /*
 432                  * highest_priority_index records current highest priority swap
 433                  * type which just frees swap entries. If its priority is
 434                  * higher than that of swap_list.next swap type, we use it.  It
 435                  * isn't protected by swap_lock, so it can be an invalid value
 436                  * if the corresponding swap type is swapoff. We double check
 437                  * the flags here. It's even possible the swap type is swapoff
 438                  * and swapon again and its priority is changed. In such rare
 439                  * case, low prority swap type might be used, but eventually
 440                  * high priority swap will be used after several rounds of
 441                  * swap.
 442                  */
 443                 if (hp_index != -1 && hp_index != type &&
 444                     swap_info[type]->prio < swap_info[hp_index]->prio &&
 445                     (swap_info[hp_index]->flags & SWP_WRITEOK)) {
 446                         type = hp_index;
 447                         swap_list.next = type;
 448                 }
 449
 450                 si = swap_info[type];
 451                 next = si->next;
 452                 if (next < 0 ||
 453                     (!wrapped && si->prio != swap_info[next]->prio)) {
 454                         next = swap_list.head;
 455                         wrapped++;
 456                 }
 457
 458                 spin_lock(&si->lock);
 459                 if (!si->highest_bit) {
 460                         spin_unlock(&si->lock);
 461                         continue;
 462                 }
 463                 if (!(si->flags & SWP_WRITEOK)) {
 464                         spin_unlock(&si->lock);
 465                         continue;
 466                 }
 467
 468                 swap_list.next = next;
 469
 470                 spin_unlock(&swap_lock);
 471                 /* This is called for allocating swap entry for cache */
 472                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
 473                 spin_unlock(&si->lock);
 474                 if (offset)
 475                         return swp_entry(type, offset);
 476                 spin_lock(&swap_lock);
 477                 next = swap_list.next;
 478         }
 479
 480         atomic_long_inc(&nr_swap_pages);
 481 noswap:
 482         spin_unlock(&swap_lock);
 483         return (swp_entry_t) {0};
 484 }
 485
 486 /* The only caller of this function is now susupend routine */
 487 swp_entry_t get_swap_page_of_type(int type)
 488 {
 489         struct swap_info_struct *si;
 490         pgoff_t offset;
 491
 492         si = swap_info[type];
 493         spin_lock(&si->lock);
 494         if (si && (si->flags & SWP_WRITEOK)) {
 495                 atomic_long_dec(&nr_swap_pages);
 496                 /* This is called for allocating swap entry, not cache */
 497                 offset = scan_swap_map(si, 1);
 498                 if (offset) {
 499                         spin_unlock(&si->lock);
 500                         return swp_entry(type, offset);
 501                 }
 502                 atomic_long_inc(&nr_swap_pages);
 503         }
 504         spin_unlock(&si->lock);
 505         return (swp_entry_t) {0};
 506 }
 507
 508 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
 509 {
 510         struct swap_info_struct *p;
 511         unsigned long offset, type;
 512
 513         if (!entry.val)
 514                 goto out;
 515         type = swp_type(entry);
 516         if (type >= nr_swapfiles)
 517                 goto bad_nofile;
 518         p = swap_info[type];
 519         if (!(p->flags & SWP_USED))
 520                 goto bad_device;
 521         offset = swp_offset(entry);
 522         if (offset >= p->max)
 523                 goto bad_offset;
 524         if (!p->swap_map[offset])
 525                 goto bad_free;
 526         spin_lock(&p->lock);
 527         return p;
 528
 529 bad_free:
 530         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 531         goto out;
 532 bad_offset:
 533         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 534         goto out;
 535 bad_device:
 536         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 537         goto out;
 538 bad_nofile:
 539         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 540 out:
 541         return NULL;
 542 }
 543
 544 /*
 545  * This swap type frees swap entry, check if it is the highest priority swap
 546  * type which just frees swap entry. get_swap_page() uses
 547  * highest_priority_index to search highest priority swap type. The
 548  * swap_info_struct.lock can't protect us if there are multiple swap types
 549  * active, so we use atomic_cmpxchg.
 550  */
 551 static void set_highest_priority_index(int type)
 552 {
 553         int old_hp_index, new_hp_index;
 554
 555         do {
 556                 old_hp_index = atomic_read(&highest_priority_index);
 557                 if (old_hp_index != -1 &&
 558                         swap_info[old_hp_index]->prio >= swap_info[type]->prio)
 559                         break;
 560                 new_hp_index = type;
 561         } while (atomic_cmpxchg(&highest_priority_index,
 562                 old_hp_index, new_hp_index) != old_hp_index);
 563 }
 564
 565 static unsigned char swap_entry_free(struct swap_info_struct *p,
 566                                      swp_entry_t entry, unsigned char usage)
 567 {
 568         unsigned long offset = swp_offset(entry);
 569         unsigned char count;
 570         unsigned char has_cache;
 571
 572         count = p->swap_map[offset];
 573         has_cache = count & SWAP_HAS_CACHE;
 574         count &= ~SWAP_HAS_CACHE;
 575
 576         if (usage == SWAP_HAS_CACHE) {
 577                 VM_BUG_ON(!has_cache);
 578                 has_cache = 0;
 579         } else if (count == SWAP_MAP_SHMEM) {
 580                 /*
 581                  * Or we could insist on shmem.c using a special
 582                  * swap_shmem_free() and free_shmem_swap_and_cache()...
 583                  */
 584                 count = 0;
 585         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
 586                 if (count == COUNT_CONTINUED) {
 587                         if (swap_count_continued(p, offset, count))
 588                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
 589                         else
 590                                 count = SWAP_MAP_MAX;
 591                 } else
 592                         count--;
 593         }
 594
 595         if (!count)
 596                 mem_cgroup_uncharge_swap(entry);
 597
 598         usage = count | has_cache;
 599         p->swap_map[offset] = usage;
 600
 601         /* free if no reference */
 602         if (!usage) {
 603                 if (offset < p->lowest_bit)
 604                         p->lowest_bit = offset;
 605                 if (offset > p->highest_bit)
 606                         p->highest_bit = offset;
 607                 set_highest_priority_index(p->type);
 608                 atomic_long_inc(&nr_swap_pages);
 609                 p->inuse_pages--;
 610                 frontswap_invalidate_page(p->type, offset);
 611                 if (p->flags & SWP_BLKDEV) {
 612                         struct gendisk *disk = p->bdev->bd_disk;
 613                         if (disk->fops->swap_slot_free_notify)
 614                                 disk->fops->swap_slot_free_notify(p->bdev,
 615                                                                   offset);
 616                 }
 617         }
 618
 619         return usage;
 620 }
 621
 622 /*
 623  * Caller has made sure that the swapdevice corresponding to entry
 624  * is still around or has not been recycled.
 625  */
 626 void swap_free(swp_entry_t entry)
 627 {
 628         struct swap_info_struct *p;
 629
 630         p = swap_info_get(entry);
 631         if (p) {
 632                 swap_entry_free(p, entry, 1);
 633                 spin_unlock(&p->lock);
 634         }
 635 }
 636
 637 /*
 638  * Called after dropping swapcache to decrease refcnt to swap entries.
 639  */
 640 void swapcache_free(swp_entry_t entry, struct page *page)
 641 {
 642         struct swap_info_struct *p;
 643         unsigned char count;
 644
 645         p = swap_info_get(entry);
 646         if (p) {
 647                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
 648                 if (page)
 649                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
 650                 spin_unlock(&p->lock);
 651         }
 652 }
 653
 654 /*
 655  * How many references to page are currently swapped out?
 656  * This does not give an exact answer when swap count is continued,
 657  * but does include the high COUNT_CONTINUED flag to allow for that.
 658  */
 659 int page_swapcount(struct page *page)
 660 {
 661         int count = 0;
 662         struct swap_info_struct *p;
 663         swp_entry_t entry;
 664
 665         entry.val = page_private(page);
 666         p = swap_info_get(entry);
 667         if (p) {
 668                 count = swap_count(p->swap_map[swp_offset(entry)]);
 669                 spin_unlock(&p->lock);
 670         }
 671         return count;
 672 }
 673
 674 /*
 675  * We can write to an anon page without COW if there are no other references
 676  * to it.  And as a side-effect, free up its swap: because the old content
 677  * on disk will never be read, and seeking back there to write new content
 678  * later would only waste time away from clustering.
 679  */
 680 int reuse_swap_page(struct page *page)
 681 {
 682         int count;
 683
 684         VM_BUG_ON(!PageLocked(page));
 685         if (unlikely(PageKsm(page)))
 686                 return 0;
 687         count = page_mapcount(page);
 688         if (count <= 1 && PageSwapCache(page)) {
 689                 count += page_swapcount(page);
 690                 if (count == 1 && !PageWriteback(page)) {
 691                         delete_from_swap_cache(page);
 692                         SetPageDirty(page);
 693                 }
 694         }
 695         return count <= 1;
 696 }
 697
 698 /*
 699  * If swap is getting full, or if there are no more mappings of this page,
 700  * then try_to_free_swap is called to free its swap space.
 701  */
 702 int try_to_free_swap(struct page *page)
 703 {
 704         VM_BUG_ON(!PageLocked(page));
 705
 706         if (!PageSwapCache(page))
 707                 return 0;
 708         if (PageWriteback(page))
 709                 return 0;
 710         if (page_swapcount(page))
 711                 return 0;
 712
 713         /*
 714          * Once hibernation has begun to create its image of memory,
 715          * there's a danger that one of the calls to try_to_free_swap()
 716          * - most probably a call from __try_to_reclaim_swap() while
 717          * hibernation is allocating its own swap pages for the image,
 718          * but conceivably even a call from memory reclaim - will free
 719          * the swap from a page which has already been recorded in the
 720          * image as a clean swapcache page, and then reuse its swap for
 721          * another page of the image.  On waking from hibernation, the
 722          * original page might be freed under memory pressure, then
 723          * later read back in from swap, now with the wrong data.
 724          *
 725          * Hibration suspends storage while it is writing the image
 726          * to disk so check that here.
 727          */
 728         if (pm_suspended_storage())
 729                 return 0;
 730
 731         delete_from_swap_cache(page);
 732         SetPageDirty(page);
 733         return 1;
 734 }
 735
 736 /*
 737  * Free the swap entry like above, but also try to
 738  * free the page cache entry if it is the last user.
 739  */
 740 int free_swap_and_cache(swp_entry_t entry)
 741 {
 742         struct swap_info_struct *p;
 743         struct page *page = NULL;
 744
 745         if (non_swap_entry(entry))
 746                 return 1;
 747
 748         p = swap_info_get(entry);
 749         if (p) {
 750                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
 751                         page = find_get_page(swap_address_space(entry),
 752                                                 entry.val);
 753                         if (page && !trylock_page(page)) {
 754                                 page_cache_release(page);
 755                                 page = NULL;
 756                         }
 757                 }
 758                 spin_unlock(&p->lock);
 759         }
 760         if (page) {
 761                 /*
 762                  * Not mapped elsewhere, or swap space full? Free it!
 763                  * Also recheck PageSwapCache now page is locked (above).
 764                  */
 765                 if (PageSwapCache(page) && !PageWriteback(page) &&
 766                                 (!page_mapped(page) || vm_swap_full())) {
 767                         delete_from_swap_cache(page);
 768                         SetPageDirty(page);
 769                 }
 770                 unlock_page(page);
 771                 page_cache_release(page);
 772         }
 773         return p != NULL;
 774 }
 775
 776 #ifdef CONFIG_HIBERNATION
 777 /*
 778  * Find the swap type that corresponds to given device (if any).
 779  *
 780  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 781  * from 0, in which the swap header is expected to be located.
 782  *
 783  * This is needed for the suspend to disk (aka swsusp).
 784  */
 785 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 786 {
 787         struct block_device *bdev = NULL;
 788         int type;
 789
 790         if (device)
 791                 bdev = bdget(device);
 792
 793         spin_lock(&swap_lock);
 794         for (type = 0; type < nr_swapfiles; type++) {
 795                 struct swap_info_struct *sis = swap_info[type];
 796
 797                 if (!(sis->flags & SWP_WRITEOK))
 798                         continue;
 799
 800                 if (!bdev) {
 801                         if (bdev_p)
 802                                 *bdev_p = bdgrab(sis->bdev);
 803
 804                         spin_unlock(&swap_lock);
 805                         return type;
 806                 }
 807                 if (bdev == sis->bdev) {
 808                         struct swap_extent *se = &sis->first_swap_extent;
 809
 810                         if (se->start_block == offset) {
 811                                 if (bdev_p)
 812                                         *bdev_p = bdgrab(sis->bdev);
 813
 814                                 spin_unlock(&swap_lock);
 815                                 bdput(bdev);
 816                                 return type;
 817                         }
 818                 }
 819         }
 820         spin_unlock(&swap_lock);
 821         if (bdev)
 822                 bdput(bdev);
 823
 824         return -ENODEV;
 825 }
 826
 827 /*
 828  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 829  * corresponding to given index in swap_info (swap type).
 830  */
 831 sector_t swapdev_block(int type, pgoff_t offset)
 832 {
 833         struct block_device *bdev;
 834
 835         if ((unsigned int)type >= nr_swapfiles)
 836                 return 0;
 837         if (!(swap_info[type]->flags & SWP_WRITEOK))
 838                 return 0;
 839         return map_swap_entry(swp_entry(type, offset), &bdev);
 840 }
 841
 842 /*
 843  * Return either the total number of swap pages of given type, or the number
 844  * of free pages of that type (depending on @free)
 845  *
 846  * This is needed for software suspend
 847  */
 848 unsigned int count_swap_pages(int type, int free)
 849 {
 850         unsigned int n = 0;
 851
 852         spin_lock(&swap_lock);
 853         if ((unsigned int)type < nr_swapfiles) {
 854                 struct swap_info_struct *sis = swap_info[type];
 855
 856                 spin_lock(&sis->lock);
 857                 if (sis->flags & SWP_WRITEOK) {
 858                         n = sis->pages;
 859                         if (free)
 860                                 n -= sis->inuse_pages;
 861                 }
 862                 spin_unlock(&sis->lock);
 863         }
 864         spin_unlock(&swap_lock);
 865         return n;
 866 }
 867 #endif /* CONFIG_HIBERNATION */
 868
 869 /*
 870  * No need to decide whether this PTE shares the swap entry with others,
 871  * just let do_wp_page work it out if a write is requested later - to
 872  * force COW, vm_page_prot omits write permission from any private vma.
 873  */
 874 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 875                 unsigned long addr, swp_entry_t entry, struct page *page)
 876 {
 877         struct page *swapcache;
 878         struct mem_cgroup *memcg;
 879         spinlock_t *ptl;
 880         pte_t *pte;
 881         int ret = 1;
 882
 883         swapcache = page;
 884         page = ksm_might_need_to_copy(page, vma, addr);
 885         if (unlikely(!page))
 886                 return -ENOMEM;
 887
 888         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
 889                                          GFP_KERNEL, &memcg)) {
 890                 ret = -ENOMEM;
 891                 goto out_nolock;
 892         }
 893
 894         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 895         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 896                 mem_cgroup_cancel_charge_swapin(memcg);
 897                 ret = 0;
 898                 goto out;
 899         }
 900
 901         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
 902         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
 903         get_page(page);
 904         set_pte_at(vma->vm_mm, addr, pte,
 905                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 906         if (page == swapcache)
 907                 page_add_anon_rmap(page, vma, addr);
 908         else /* ksm created a completely new copy */
 909                 page_add_new_anon_rmap(page, vma, addr);
 910         mem_cgroup_commit_charge_swapin(page, memcg);
 911         swap_free(entry);
 912         /*
 913          * Move the page to the active list so it is not
 914          * immediately swapped out again after swapon.
 915          */
 916         activate_page(page);
 917 out:
 918         pte_unmap_unlock(pte, ptl);
 919 out_nolock:
 920         if (page != swapcache) {
 921                 unlock_page(page);
 922                 put_page(page);
 923         }
 924         return ret;
 925 }
 926
 927 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 928                                 unsigned long addr, unsigned long end,
 929                                 swp_entry_t entry, struct page *page)
 930 {
 931         pte_t swp_pte = swp_entry_to_pte(entry);
 932         pte_t *pte;
 933         int ret = 0;
 934
 935         /*
 936          * We don't actually need pte lock while scanning for swp_pte: since
 937          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 938          * page table while we're scanning; though it could get zapped, and on
 939          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 940          * of unmatched parts which look like swp_pte, so unuse_pte must
 941          * recheck under pte lock.  Scanning without pte lock lets it be
 942          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 943          */
 944         pte = pte_offset_map(pmd, addr);
 945         do {
 946                 /*
 947                  * swapoff spends a _lot_ of time in this loop!
 948                  * Test inline before going to call unuse_pte.
 949                  */
 950                 if (unlikely(pte_same(*pte, swp_pte))) {
 951                         pte_unmap(pte);
 952                         ret = unuse_pte(vma, pmd, addr, entry, page);
 953                         if (ret)
 954                                 goto out;
 955                         pte = pte_offset_map(pmd, addr);
 956                 }
 957         } while (pte++, addr += PAGE_SIZE, addr != end);
 958         pte_unmap(pte - 1);
 959 out:
 960         return ret;
 961 }
 962
 963 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 964                                 unsigned long addr, unsigned long end,
 965                                 swp_entry_t entry, struct page *page)
 966 {
 967         pmd_t *pmd;
 968         unsigned long next;
 969         int ret;
 970
 971         pmd = pmd_offset(pud, addr);
 972         do {
 973                 next = pmd_addr_end(addr, end);
 974                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 975                         continue;
 976                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 977                 if (ret)
 978                         return ret;
 979         } while (pmd++, addr = next, addr != end);
 980         return 0;
 981 }
 982
 983 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 984                                 unsigned long addr, unsigned long end,
 985                                 swp_entry_t entry, struct page *page)
 986 {
 987         pud_t *pud;
 988         unsigned long next;
 989         int ret;
 990
 991         pud = pud_offset(pgd, addr);
 992         do {
 993                 next = pud_addr_end(addr, end);
 994                 if (pud_none_or_clear_bad(pud))
 995                         continue;
 996                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 997                 if (ret)
 998                         return ret;
 999         } while (pud++, addr = next, addr != end);
1000         return 0;
1001 }
1002
1003 static int unuse_vma(struct vm_area_struct *vma,
1004                                 swp_entry_t entry, struct page *page)
1005 {
1006         pgd_t *pgd;
1007         unsigned long addr, end, next;
1008         int ret;
1009
1010         if (page_anon_vma(page)) {
1011                 addr = page_address_in_vma(page, vma);
1012                 if (addr == -EFAULT)
1013                         return 0;
1014                 else
1015                         end = addr + PAGE_SIZE;
1016         } else {
1017                 addr = vma->vm_start;
1018                 end = vma->vm_end;
1019         }
1020
1021         pgd = pgd_offset(vma->vm_mm, addr);
1022         do {
1023                 next = pgd_addr_end(addr, end);
1024                 if (pgd_none_or_clear_bad(pgd))
1025                         continue;
1026                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1027                 if (ret)
1028                         return ret;
1029         } while (pgd++, addr = next, addr != end);
1030         return 0;
1031 }
1032
1033 static int unuse_mm(struct mm_struct *mm,
1034                                 swp_entry_t entry, struct page *page)
1035 {
1036         struct vm_area_struct *vma;
1037         int ret = 0;
1038
1039         if (!down_read_trylock(&mm->mmap_sem)) {
1040                 /*
1041                  * Activate page so shrink_inactive_list is unlikely to unmap
1042                  * its ptes while lock is dropped, so swapoff can make progress.
1043                  */
1044                 activate_page(page);
1045                 unlock_page(page);
1046                 down_read(&mm->mmap_sem);
1047                 lock_page(page);
1048         }
1049         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1050                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1051                         break;
1052         }
1053         up_read(&mm->mmap_sem);
1054         return (ret < 0)? ret: 0;
1055 }
1056
1057 /*
1058  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1059  * from current position to next entry still in use.
1060  * Recycle to start on reaching the end, returning 0 when empty.
1061  */
1062 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1063                                         unsigned int prev, bool frontswap)
1064 {
1065         unsigned int max = si->max;
1066         unsigned int i = prev;
1067         unsigned char count;
1068
1069         /*
1070          * No need for swap_lock here: we're just looking
1071          * for whether an entry is in use, not modifying it; false
1072          * hits are okay, and sys_swapoff() has already prevented new
1073          * allocations from this area (while holding swap_lock).
1074          */
1075         for (;;) {
1076                 if (++i >= max) {
1077                         if (!prev) {
1078                                 i = 0;
1079                                 break;
1080                         }
1081                         /*
1082                          * No entries in use at top of swap_map,
1083                          * loop back to start and recheck there.
1084                          */
1085                         max = prev + 1;
1086                         prev = 0;
1087                         i = 1;
1088                 }
1089                 if (frontswap) {
1090                         if (frontswap_test(si, i))
1091                                 break;
1092                         else
1093                                 continue;
1094                 }
1095                 count = si->swap_map[i];
1096                 if (count && swap_count(count) != SWAP_MAP_BAD)
1097                         break;
1098         }
1099         return i;
1100 }
1101
1102 /*
1103  * We completely avoid races by reading each swap page in advance,
1104  * and then search for the process using it.  All the necessary
1105  * page table adjustments can then be made atomically.
1106  *
1107  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1108  * pages_to_unuse==0 means all pages; ignored if frontswap is false
1109  */
1110 int try_to_unuse(unsigned int type, bool frontswap,
1111                  unsigned long pages_to_unuse)
1112 {
1113         struct swap_info_struct *si = swap_info[type];
1114         struct mm_struct *start_mm;
1115         unsigned char *swap_map;
1116         unsigned char swcount;
1117         struct page *page;
1118         swp_entry_t entry;
1119         unsigned int i = 0;
1120         int retval = 0;
1121
1122         /*
1123          * When searching mms for an entry, a good strategy is to
1124          * start at the first mm we freed the previous entry from
1125          * (though actually we don't notice whether we or coincidence
1126          * freed the entry).  Initialize this start_mm with a hold.
1127          *
1128          * A simpler strategy would be to start at the last mm we
1129          * freed the previous entry from; but that would take less
1130          * advantage of mmlist ordering, which clusters forked mms
1131          * together, child after parent.  If we race with dup_mmap(), we
1132          * prefer to resolve parent before child, lest we miss entries
1133          * duplicated after we scanned child: using last mm would invert
1134          * that.
1135          */
1136         start_mm = &init_mm;
1137         atomic_inc(&init_mm.mm_users);
1138
1139         /*
1140          * Keep on scanning until all entries have gone.  Usually,
1141          * one pass through swap_map is enough, but not necessarily:
1142          * there are races when an instance of an entry might be missed.
1143          */
1144         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1145                 if (signal_pending(current)) {
1146                         retval = -EINTR;
1147                         break;
1148                 }
1149
1150                 /*
1151                  * Get a page for the entry, using the existing swap
1152                  * cache page if there is one.  Otherwise, get a clean
1153                  * page and read the swap into it.
1154                  */
1155                 swap_map = &si->swap_map[i];
1156                 entry = swp_entry(type, i);
1157                 page = read_swap_cache_async(entry,
1158                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1159                 if (!page) {
1160                         /*
1161                          * Either swap_duplicate() failed because entry
1162                          * has been freed independently, and will not be
1163                          * reused since sys_swapoff() already disabled
1164                          * allocation from here, or alloc_page() failed.
1165                          */
1166                         if (!*swap_map)
1167                                 continue;
1168                         retval = -ENOMEM;
1169                         break;
1170                 }
1171
1172                 /*
1173                  * Don't hold on to start_mm if it looks like exiting.
1174                  */
1175                 if (atomic_read(&start_mm->mm_users) == 1) {
1176                         mmput(start_mm);
1177                         start_mm = &init_mm;
1178                         atomic_inc(&init_mm.mm_users);
1179                 }
1180
1181                 /*
1182                  * Wait for and lock page.  When do_swap_page races with
1183                  * try_to_unuse, do_swap_page can handle the fault much
1184                  * faster than try_to_unuse can locate the entry.  This
1185                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1186                  * defer to do_swap_page in such a case - in some tests,
1187                  * do_swap_page and try_to_unuse repeatedly compete.
1188                  */
1189                 wait_on_page_locked(page);
1190                 wait_on_page_writeback(page);
1191                 lock_page(page);
1192                 wait_on_page_writeback(page);
1193
1194                 /*
1195                  * Remove all references to entry.
1196                  */
1197                 swcount = *swap_map;
1198                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1199                         retval = shmem_unuse(entry, page);
1200                         /* page has already been unlocked and released */
1201                         if (retval < 0)
1202                                 break;
1203                         continue;
1204                 }
1205                 if (swap_count(swcount) && start_mm != &init_mm)
1206                         retval = unuse_mm(start_mm, entry, page);
1207
1208                 if (swap_count(*swap_map)) {
1209                         int set_start_mm = (*swap_map >= swcount);
1210                         struct list_head *p = &start_mm->mmlist;
1211                         struct mm_struct *new_start_mm = start_mm;
1212                         struct mm_struct *prev_mm = start_mm;
1213                         struct mm_struct *mm;
1214
1215                         atomic_inc(&new_start_mm->mm_users);
1216                         atomic_inc(&prev_mm->mm_users);
1217                         spin_lock(&mmlist_lock);
1218                         while (swap_count(*swap_map) && !retval &&
1219                                         (p = p->next) != &start_mm->mmlist) {
1220                                 mm = list_entry(p, struct mm_struct, mmlist);
1221                                 if (!atomic_inc_not_zero(&mm->mm_users))
1222                                         continue;
1223                                 spin_unlock(&mmlist_lock);
1224                                 mmput(prev_mm);
1225                                 prev_mm = mm;
1226
1227                                 cond_resched();
1228
1229                                 swcount = *swap_map;
1230                                 if (!swap_count(swcount)) /* any usage ? */
1231                                         ;
1232                                 else if (mm == &init_mm)
1233                                         set_start_mm = 1;
1234                                 else
1235                                         retval = unuse_mm(mm, entry, page);
1236
1237                                 if (set_start_mm && *swap_map < swcount) {
1238                                         mmput(new_start_mm);
1239                                         atomic_inc(&mm->mm_users);
1240                                         new_start_mm = mm;
1241                                         set_start_mm = 0;
1242                                 }
1243                                 spin_lock(&mmlist_lock);
1244                         }
1245                         spin_unlock(&mmlist_lock);
1246                         mmput(prev_mm);
1247                         mmput(start_mm);
1248                         start_mm = new_start_mm;
1249                 }
1250                 if (retval) {
1251                         unlock_page(page);
1252                         page_cache_release(page);
1253                         break;
1254                 }
1255
1256                 /*
1257                  * If a reference remains (rare), we would like to leave
1258                  * the page in the swap cache; but try_to_unmap could
1259                  * then re-duplicate the entry once we drop page lock,
1260                  * so we might loop indefinitely; also, that page could
1261                  * not be swapped out to other storage meanwhile.  So:
1262                  * delete from cache even if there's another reference,
1263                  * after ensuring that the data has been saved to disk -
1264                  * since if the reference remains (rarer), it will be
1265                  * read from disk into another page.  Splitting into two
1266                  * pages would be incorrect if swap supported "shared
1267                  * private" pages, but they are handled by tmpfs files.
1268                  *
1269                  * Given how unuse_vma() targets one particular offset
1270                  * in an anon_vma, once the anon_vma has been determined,
1271                  * this splitting happens to be just what is needed to
1272                  * handle where KSM pages have been swapped out: re-reading
1273                  * is unnecessarily slow, but we can fix that later on.
1274                  */
1275                 if (swap_count(*swap_map) &&
1276                      PageDirty(page) && PageSwapCache(page)) {
1277                         struct writeback_control wbc = {
1278                                 .sync_mode = WB_SYNC_NONE,
1279                         };
1280
1281                         swap_writepage(page, &wbc);
1282                         lock_page(page);
1283                         wait_on_page_writeback(page);
1284                 }
1285
1286                 /*
1287                  * It is conceivable that a racing task removed this page from
1288                  * swap cache just before we acquired the page lock at the top,
1289                  * or while we dropped it in unuse_mm().  The page might even
1290                  * be back in swap cache on another swap area: that we must not
1291                  * delete, since it may not have been written out to swap yet.
1292                  */
1293                 if (PageSwapCache(page) &&
1294                     likely(page_private(page) == entry.val))
1295                         delete_from_swap_cache(page);
1296
1297                 /*
1298                  * So we could skip searching mms once swap count went
1299                  * to 1, we did not mark any present ptes as dirty: must
1300                  * mark page dirty so shrink_page_list will preserve it.
1301                  */
1302                 SetPageDirty(page);
1303                 unlock_page(page);
1304                 page_cache_release(page);
1305
1306                 /*
1307                  * Make sure that we aren't completely killing
1308                  * interactive performance.
1309                  */
1310                 cond_resched();
1311                 if (frontswap && pages_to_unuse > 0) {
1312                         if (!--pages_to_unuse)
1313                                 break;
1314                 }
1315         }
1316
1317         mmput(start_mm);
1318         return retval;
1319 }
1320
1321 /*
1322  * After a successful try_to_unuse, if no swap is now in use, we know
1323  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1324  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1325  * added to the mmlist just after page_duplicate - before would be racy.
1326  */
1327 static void drain_mmlist(void)
1328 {
1329         struct list_head *p, *next;
1330         unsigned int type;
1331
1332         for (type = 0; type < nr_swapfiles; type++)
1333                 if (swap_info[type]->inuse_pages)
1334                         return;
1335         spin_lock(&mmlist_lock);
1336         list_for_each_safe(p, next, &init_mm.mmlist)
1337                 list_del_init(p);
1338         spin_unlock(&mmlist_lock);
1339 }
1340
1341 /*
1342  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1343  * corresponds to page offset for the specified swap entry.
1344  * Note that the type of this function is sector_t, but it returns page offset
1345  * into the bdev, not sector offset.
1346  */
1347 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1348 {
1349         struct swap_info_struct *sis;
1350         struct swap_extent *start_se;
1351         struct swap_extent *se;
1352         pgoff_t offset;
1353
1354         sis = swap_info[swp_type(entry)];
1355         *bdev = sis->bdev;
1356
1357         offset = swp_offset(entry);
1358         start_se = sis->curr_swap_extent;
1359         se = start_se;
1360
1361         for ( ; ; ) {
1362                 struct list_head *lh;
1363
1364                 if (se->start_page <= offset &&
1365                                 offset < (se->start_page + se->nr_pages)) {
1366                         return se->start_block + (offset - se->start_page);
1367                 }
1368                 lh = se->list.next;
1369                 se = list_entry(lh, struct swap_extent, list);
1370                 sis->curr_swap_extent = se;
1371                 BUG_ON(se == start_se);         /* It *must* be present */
1372         }
1373 }
1374
1375 /*
1376  * Returns the page offset into bdev for the specified page's swap entry.
1377  */
1378 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1379 {
1380         swp_entry_t entry;
1381         entry.val = page_private(page);
1382         return map_swap_entry(entry, bdev);
1383 }
1384
1385 /*
1386  * Free all of a swapdev's extent information
1387  */
1388 static void destroy_swap_extents(struct swap_info_struct *sis)
1389 {
1390         while (!list_empty(&sis->first_swap_extent.list)) {
1391                 struct swap_extent *se;
1392
1393                 se = list_entry(sis->first_swap_extent.list.next,
1394                                 struct swap_extent, list);
1395                 list_del(&se->list);
1396                 kfree(se);
1397         }
1398
1399         if (sis->flags & SWP_FILE) {
1400                 struct file *swap_file = sis->swap_file;
1401                 struct address_space *mapping = swap_file->f_mapping;
1402
1403                 sis->flags &= ~SWP_FILE;
1404                 mapping->a_ops->swap_deactivate(swap_file);
1405         }
1406 }
1407
1408 /*
1409  * Add a block range (and the corresponding page range) into this swapdev's
1410  * extent list.  The extent list is kept sorted in page order.
1411  *
1412  * This function rather assumes that it is called in ascending page order.
1413  */
1414 int
1415 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1416                 unsigned long nr_pages, sector_t start_block)
1417 {
1418         struct swap_extent *se;
1419         struct swap_extent *new_se;
1420         struct list_head *lh;
1421
1422         if (start_page == 0) {
1423                 se = &sis->first_swap_extent;
1424                 sis->curr_swap_extent = se;
1425                 se->start_page = 0;
1426                 se->nr_pages = nr_pages;
1427                 se->start_block = start_block;
1428                 return 1;
1429         } else {
1430                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1431                 se = list_entry(lh, struct swap_extent, list);
1432                 BUG_ON(se->start_page + se->nr_pages != start_page);
1433                 if (se->start_block + se->nr_pages == start_block) {
1434                         /* Merge it */
1435                         se->nr_pages += nr_pages;
1436                         return 0;
1437                 }
1438         }
1439
1440         /*
1441          * No merge.  Insert a new extent, preserving ordering.
1442          */
1443         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1444         if (new_se == NULL)
1445                 return -ENOMEM;
1446         new_se->start_page = start_page;
1447         new_se->nr_pages = nr_pages;
1448         new_se->start_block = start_block;
1449
1450         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1451         return 1;
1452 }
1453
1454 /*
1455  * A `swap extent' is a simple thing which maps a contiguous range of pages
1456  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1457  * is built at swapon time and is then used at swap_writepage/swap_readpage
1458  * time for locating where on disk a page belongs.
1459  *
1460  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1461  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1462  * swap files identically.
1463  *
1464  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1465  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1466  * swapfiles are handled *identically* after swapon time.
1467  *
1468  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1469  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1470  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1471  * requirements, they are simply tossed out - we will never use those blocks
1472  * for swapping.
1473  *
1474  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1475  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1476  * which will scribble on the fs.
1477  *
1478  * The amount of disk space which a single swap extent represents varies.
1479  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1480  * extents in the list.  To avoid much list walking, we cache the previous
1481  * search location in `curr_swap_extent', and start new searches from there.
1482  * This is extremely effective.  The average number of iterations in
1483  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1484  */
1485 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1486 {
1487         struct file *swap_file = sis->swap_file;
1488         struct address_space *mapping = swap_file->f_mapping;
1489         struct inode *inode = mapping->host;
1490         int ret;
1491
1492         if (S_ISBLK(inode->i_mode)) {
1493                 ret = add_swap_extent(sis, 0, sis->max, 0);
1494                 *span = sis->pages;
1495                 return ret;
1496         }
1497
1498         if (mapping->a_ops->swap_activate) {
1499                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1500                 if (!ret) {
1501                         sis->flags |= SWP_FILE;
1502                         ret = add_swap_extent(sis, 0, sis->max, 0);
1503                         *span = sis->pages;
1504                 }
1505                 return ret;
1506         }
1507
1508         return generic_swapfile_activate(sis, swap_file, span);
1509 }
1510
1511 static void _enable_swap_info(struct swap_info_struct *p, int prio,
1512                                 unsigned char *swap_map)
1513 {
1514         int i, prev;
1515
1516         if (prio >= 0)
1517                 p->prio = prio;
1518         else
1519                 p->prio = --least_priority;
1520         p->swap_map = swap_map;
1521         p->flags |= SWP_WRITEOK;
1522         atomic_long_add(p->pages, &nr_swap_pages);
1523         total_swap_pages += p->pages;
1524
1525         /* insert swap space into swap_list: */
1526         prev = -1;
1527         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1528                 if (p->prio >= swap_info[i]->prio)
1529                         break;
1530                 prev = i;
1531         }
1532         p->next = i;
1533         if (prev < 0)
1534                 swap_list.head = swap_list.next = p->type;
1535         else
1536                 swap_info[prev]->next = p->type;
1537 }
1538
1539 static void enable_swap_info(struct swap_info_struct *p, int prio,
1540                                 unsigned char *swap_map,
1541                                 unsigned long *frontswap_map)
1542 {
1543         frontswap_init(p->type, frontswap_map);
1544         spin_lock(&swap_lock);
1545         spin_lock(&p->lock);
1546          _enable_swap_info(p, prio, swap_map);
1547         spin_unlock(&p->lock);
1548         spin_unlock(&swap_lock);
1549 }
1550
1551 static void reinsert_swap_info(struct swap_info_struct *p)
1552 {
1553         spin_lock(&swap_lock);
1554         spin_lock(&p->lock);
1555         _enable_swap_info(p, p->prio, p->swap_map);
1556         spin_unlock(&p->lock);
1557         spin_unlock(&swap_lock);
1558 }
1559
1560 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1561 {
1562         struct swap_info_struct *p = NULL;
1563         unsigned char *swap_map;
1564         unsigned long *frontswap_map;
1565         struct file *swap_file, *victim;
1566         struct address_space *mapping;
1567         struct inode *inode;
1568         struct filename *pathname;
1569         int i, type, prev;
1570         int err;
1571
1572         if (!capable(CAP_SYS_ADMIN))
1573                 return -EPERM;
1574
1575         BUG_ON(!current->mm);
1576
1577         pathname = getname(specialfile);
1578         if (IS_ERR(pathname))
1579                 return PTR_ERR(pathname);
1580
1581         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
1582         err = PTR_ERR(victim);
1583         if (IS_ERR(victim))
1584                 goto out;
1585
1586         mapping = victim->f_mapping;
1587         prev = -1;
1588         spin_lock(&swap_lock);
1589         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1590                 p = swap_info[type];
1591                 if (p->flags & SWP_WRITEOK) {
1592                         if (p->swap_file->f_mapping == mapping)
1593                                 break;
1594                 }
1595                 prev = type;
1596         }
1597         if (type < 0) {
1598                 err = -EINVAL;
1599                 spin_unlock(&swap_lock);
1600                 goto out_dput;
1601         }
1602         if (!security_vm_enough_memory_mm(current->mm, p->pages))
1603                 vm_unacct_memory(p->pages);
1604         else {
1605                 err = -ENOMEM;
1606                 spin_unlock(&swap_lock);
1607                 goto out_dput;
1608         }
1609         if (prev < 0)
1610                 swap_list.head = p->next;
1611         else
1612                 swap_info[prev]->next = p->next;
1613         if (type == swap_list.next) {
1614                 /* just pick something that's safe... */
1615                 swap_list.next = swap_list.head;
1616         }
1617         spin_lock(&p->lock);
1618         if (p->prio < 0) {
1619                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1620                         swap_info[i]->prio = p->prio--;
1621                 least_priority++;
1622         }
1623         atomic_long_sub(p->pages, &nr_swap_pages);
1624         total_swap_pages -= p->pages;
1625         p->flags &= ~SWP_WRITEOK;
1626         spin_unlock(&p->lock);
1627         spin_unlock(&swap_lock);
1628
1629         set_current_oom_origin();
1630         err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1631         clear_current_oom_origin();
1632
1633         if (err) {
1634                 /* re-insert swap space back into swap_list */
1635                 reinsert_swap_info(p);
1636                 goto out_dput;
1637         }
1638
1639         destroy_swap_extents(p);
1640         if (p->flags & SWP_CONTINUED)
1641                 free_swap_count_continuations(p);
1642
1643         mutex_lock(&swapon_mutex);
1644         spin_lock(&swap_lock);
1645         spin_lock(&p->lock);
1646         drain_mmlist();
1647
1648         /* wait for anyone still in scan_swap_map */
1649         p->highest_bit = 0;             /* cuts scans short */
1650         while (p->flags >= SWP_SCANNING) {
1651                 spin_unlock(&p->lock);
1652                 spin_unlock(&swap_lock);
1653                 schedule_timeout_uninterruptible(1);
1654                 spin_lock(&swap_lock);
1655                 spin_lock(&p->lock);
1656         }
1657
1658         swap_file = p->swap_file;
1659         p->swap_file = NULL;
1660         p->max = 0;
1661         swap_map = p->swap_map;
1662         p->swap_map = NULL;
1663         p->flags = 0;
1664         frontswap_map = frontswap_map_get(p);
1665         frontswap_map_set(p, NULL);
1666         spin_unlock(&p->lock);
1667         spin_unlock(&swap_lock);
1668         frontswap_invalidate_area(type);
1669         mutex_unlock(&swapon_mutex);
1670         vfree(swap_map);
1671         vfree(frontswap_map);
1672         /* Destroy swap account informatin */
1673         swap_cgroup_swapoff(type);
1674
1675         inode = mapping->host;
1676         if (S_ISBLK(inode->i_mode)) {
1677                 struct block_device *bdev = I_BDEV(inode);
1678                 set_blocksize(bdev, p->old_block_size);
1679                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1680         } else {
1681                 mutex_lock(&inode->i_mutex);
1682                 inode->i_flags &= ~S_SWAPFILE;
1683                 mutex_unlock(&inode->i_mutex);
1684         }
1685         filp_close(swap_file, NULL);
1686         err = 0;
1687         atomic_inc(&proc_poll_event);
1688         wake_up_interruptible(&proc_poll_wait);
1689
1690 out_dput:
1691         filp_close(victim, NULL);
1692 out:
1693         putname(pathname);
1694         return err;
1695 }
1696
1697 #ifdef CONFIG_PROC_FS
1698 static unsigned swaps_poll(struct file *file, poll_table *wait)
1699 {
1700         struct seq_file *seq = file->private_data;
1701
1702         poll_wait(file, &proc_poll_wait, wait);
1703
1704         if (seq->poll_event != atomic_read(&proc_poll_event)) {
1705                 seq->poll_event = atomic_read(&proc_poll_event);
1706                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1707         }
1708
1709         return POLLIN | POLLRDNORM;
1710 }
1711
1712 /* iterator */
1713 static void *swap_start(struct seq_file *swap, loff_t *pos)
1714 {
1715         struct swap_info_struct *si;
1716         int type;
1717         loff_t l = *pos;
1718
1719         mutex_lock(&swapon_mutex);
1720
1721         if (!l)
1722                 return SEQ_START_TOKEN;
1723
1724         for (type = 0; 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                 if (!--l)
1730                         return si;
1731         }
1732
1733         return NULL;
1734 }
1735
1736 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1737 {
1738         struct swap_info_struct *si = v;
1739         int type;
1740
1741         if (v == SEQ_START_TOKEN)
1742                 type = 0;
1743         else
1744                 type = si->type + 1;
1745
1746         for (; type < nr_swapfiles; type++) {
1747                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1748                 si = swap_info[type];
1749                 if (!(si->flags & SWP_USED) || !si->swap_map)
1750                         continue;
1751                 ++*pos;
1752                 return si;
1753         }
1754
1755         return NULL;
1756 }
1757
1758 static void swap_stop(struct seq_file *swap, void *v)
1759 {
1760         mutex_unlock(&swapon_mutex);
1761 }
1762
1763 static int swap_show(struct seq_file *swap, void *v)
1764 {
1765         struct swap_info_struct *si = v;
1766         struct file *file;
1767         int len;
1768
1769         if (si == SEQ_START_TOKEN) {
1770                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1771                 return 0;
1772         }
1773
1774         file = si->swap_file;
1775         len = seq_path(swap, &file->f_path, " \t\n\\");
1776         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1777                         len < 40 ? 40 - len : 1, " ",
1778                         S_ISBLK(file_inode(file)->i_mode) ?
1779                                 "partition" : "file\t",
1780                         si->pages << (PAGE_SHIFT - 10),
1781                         si->inuse_pages << (PAGE_SHIFT - 10),
1782                         si->prio);
1783         return 0;
1784 }
1785
1786 static const struct seq_operations swaps_op = {
1787         .start =        swap_start,
1788         .next =         swap_next,
1789         .stop =         swap_stop,
1790         .show =         swap_show
1791 };
1792
1793 static int swaps_open(struct inode *inode, struct file *file)
1794 {
1795         struct seq_file *seq;
1796         int ret;
1797
1798         ret = seq_open(file, &swaps_op);
1799         if (ret)
1800                 return ret;
1801
1802         seq = file->private_data;
1803         seq->poll_event = atomic_read(&proc_poll_event);
1804         return 0;
1805 }
1806
1807 static const struct file_operations proc_swaps_operations = {
1808         .open           = swaps_open,
1809         .read           = seq_read,
1810         .llseek         = seq_lseek,
1811         .release        = seq_release,
1812         .poll           = swaps_poll,
1813 };
1814
1815 static int __init procswaps_init(void)
1816 {
1817         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1818         return 0;
1819 }
1820 __initcall(procswaps_init);
1821 #endif /* CONFIG_PROC_FS */
1822
1823 #ifdef MAX_SWAPFILES_CHECK
1824 static int __init max_swapfiles_check(void)
1825 {
1826         MAX_SWAPFILES_CHECK();
1827         return 0;
1828 }
1829 late_initcall(max_swapfiles_check);
1830 #endif
1831
1832 static struct swap_info_struct *alloc_swap_info(void)
1833 {
1834         struct swap_info_struct *p;
1835         unsigned int type;
1836
1837         p = kzalloc(sizeof(*p), GFP_KERNEL);
1838         if (!p)
1839                 return ERR_PTR(-ENOMEM);
1840
1841         spin_lock(&swap_lock);
1842         for (type = 0; type < nr_swapfiles; type++) {
1843                 if (!(swap_info[type]->flags & SWP_USED))
1844                         break;
1845         }
1846         if (type >= MAX_SWAPFILES) {
1847                 spin_unlock(&swap_lock);
1848                 kfree(p);
1849                 return ERR_PTR(-EPERM);
1850         }
1851         if (type >= nr_swapfiles) {
1852                 p->type = type;
1853                 swap_info[type] = p;
1854                 /*
1855                  * Write swap_info[type] before nr_swapfiles, in case a
1856                  * racing procfs swap_start() or swap_next() is reading them.
1857                  * (We never shrink nr_swapfiles, we never free this entry.)
1858                  */
1859                 smp_wmb();
1860                 nr_swapfiles++;
1861         } else {
1862                 kfree(p);
1863                 p = swap_info[type];
1864                 /*
1865                  * Do not memset this entry: a racing procfs swap_next()
1866                  * would be relying on p->type to remain valid.
1867                  */
1868         }
1869         INIT_LIST_HEAD(&p->first_swap_extent.list);
1870         p->flags = SWP_USED;
1871         p->next = -1;
1872         spin_unlock(&swap_lock);
1873         spin_lock_init(&p->lock);
1874
1875         return p;
1876 }
1877
1878 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1879 {
1880         int error;
1881
1882         if (S_ISBLK(inode->i_mode)) {
1883                 p->bdev = bdgrab(I_BDEV(inode));
1884                 error = blkdev_get(p->bdev,
1885                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1886                                    sys_swapon);
1887                 if (error < 0) {
1888                         p->bdev = NULL;
1889                         return -EINVAL;
1890                 }
1891                 p->old_block_size = block_size(p->bdev);
1892                 error = set_blocksize(p->bdev, PAGE_SIZE);
1893                 if (error < 0)
1894                         return error;
1895                 p->flags |= SWP_BLKDEV;
1896         } else if (S_ISREG(inode->i_mode)) {
1897                 p->bdev = inode->i_sb->s_bdev;
1898                 mutex_lock(&inode->i_mutex);
1899                 if (IS_SWAPFILE(inode))
1900                         return -EBUSY;
1901         } else
1902                 return -EINVAL;
1903
1904         return 0;
1905 }
1906
1907 static unsigned long read_swap_header(struct swap_info_struct *p,
1908                                         union swap_header *swap_header,
1909                                         struct inode *inode)
1910 {
1911         int i;
1912         unsigned long maxpages;
1913         unsigned long swapfilepages;
1914
1915         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1916                 printk(KERN_ERR "Unable to find swap-space signature\n");
1917                 return 0;
1918         }
1919
1920         /* swap partition endianess hack... */
1921         if (swab32(swap_header->info.version) == 1) {
1922                 swab32s(&swap_header->info.version);
1923                 swab32s(&swap_header->info.last_page);
1924                 swab32s(&swap_header->info.nr_badpages);
1925                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1926                         swab32s(&swap_header->info.badpages[i]);
1927         }
1928         /* Check the swap header's sub-version */
1929         if (swap_header->info.version != 1) {
1930                 printk(KERN_WARNING
1931                        "Unable to handle swap header version %d\n",
1932                        swap_header->info.version);
1933                 return 0;
1934         }
1935
1936         p->lowest_bit  = 1;
1937         p->cluster_next = 1;
1938         p->cluster_nr = 0;
1939
1940         /*
1941          * Find out how many pages are allowed for a single swap
1942          * device. There are two limiting factors: 1) the number
1943          * of bits for the swap offset in the swp_entry_t type, and
1944          * 2) the number of bits in the swap pte as defined by the
1945          * different architectures. In order to find the
1946          * largest possible bit mask, a swap entry with swap type 0
1947          * and swap offset ~0UL is created, encoded to a swap pte,
1948          * decoded to a swp_entry_t again, and finally the swap
1949          * offset is extracted. This will mask all the bits from
1950          * the initial ~0UL mask that can't be encoded in either
1951          * the swp_entry_t or the architecture definition of a
1952          * swap pte.
1953          */
1954         maxpages = swp_offset(pte_to_swp_entry(
1955                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1956         if (maxpages > swap_header->info.last_page) {
1957                 maxpages = swap_header->info.last_page + 1;
1958                 /* p->max is an unsigned int: don't overflow it */
1959                 if ((unsigned int)maxpages == 0)
1960                         maxpages = UINT_MAX;
1961         }
1962         p->highest_bit = maxpages - 1;
1963
1964         if (!maxpages)
1965                 return 0;
1966         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1967         if (swapfilepages && maxpages > swapfilepages) {
1968                 printk(KERN_WARNING
1969                        "Swap area shorter than signature indicates\n");
1970                 return 0;
1971         }
1972         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1973                 return 0;
1974         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1975                 return 0;
1976
1977         return maxpages;
1978 }
1979
1980 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1981                                         union swap_header *swap_header,
1982                                         unsigned char *swap_map,
1983                                         unsigned long maxpages,
1984                                         sector_t *span)
1985 {
1986         int i;
1987         unsigned int nr_good_pages;
1988         int nr_extents;
1989
1990         nr_good_pages = maxpages - 1;   /* omit header page */
1991
1992         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1993                 unsigned int page_nr = swap_header->info.badpages[i];
1994                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1995                         return -EINVAL;
1996                 if (page_nr < maxpages) {
1997                         swap_map[page_nr] = SWAP_MAP_BAD;
1998                         nr_good_pages--;
1999                 }
2000         }
2001
2002         if (nr_good_pages) {
2003                 swap_map[0] = SWAP_MAP_BAD;
2004                 p->max = maxpages;
2005                 p->pages = nr_good_pages;
2006                 nr_extents = setup_swap_extents(p, span);
2007                 if (nr_extents < 0)
2008                         return nr_extents;
2009                 nr_good_pages = p->pages;
2010         }
2011         if (!nr_good_pages) {
2012                 printk(KERN_WARNING "Empty swap-file\n");
2013                 return -EINVAL;
2014         }
2015
2016         return nr_extents;
2017 }
2018
2019 /*
2020  * Helper to sys_swapon determining if a given swap
2021  * backing device queue supports DISCARD operations.
2022  */
2023 static bool swap_discardable(struct swap_info_struct *si)
2024 {
2025         struct request_queue *q = bdev_get_queue(si->bdev);
2026
2027         if (!q || !blk_queue_discard(q))
2028                 return false;
2029
2030         return true;
2031 }
2032
2033 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2034 {
2035         struct swap_info_struct *p;
2036         struct filename *name;
2037         struct file *swap_file = NULL;
2038         struct address_space *mapping;
2039         int i;
2040         int prio;
2041         int error;
2042         union swap_header *swap_header;
2043         int nr_extents;
2044         sector_t span;
2045         unsigned long maxpages;
2046         unsigned char *swap_map = NULL;
2047         unsigned long *frontswap_map = NULL;
2048         struct page *page = NULL;
2049         struct inode *inode = NULL;
2050
2051         if (swap_flags & ~SWAP_FLAGS_VALID)
2052                 return -EINVAL;
2053
2054         if (!capable(CAP_SYS_ADMIN))
2055                 return -EPERM;
2056
2057         p = alloc_swap_info();
2058         if (IS_ERR(p))
2059                 return PTR_ERR(p);
2060
2061         name = getname(specialfile);
2062         if (IS_ERR(name)) {
2063                 error = PTR_ERR(name);
2064                 name = NULL;
2065                 goto bad_swap;
2066         }
2067         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2068         if (IS_ERR(swap_file)) {
2069                 error = PTR_ERR(swap_file);
2070                 swap_file = NULL;
2071                 goto bad_swap;
2072         }
2073
2074         p->swap_file = swap_file;
2075         mapping = swap_file->f_mapping;
2076
2077         for (i = 0; i < nr_swapfiles; i++) {
2078                 struct swap_info_struct *q = swap_info[i];
2079
2080                 if (q == p || !q->swap_file)
2081                         continue;
2082                 if (mapping == q->swap_file->f_mapping) {
2083                         error = -EBUSY;
2084                         goto bad_swap;
2085                 }
2086         }
2087
2088         inode = mapping->host;
2089         /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2090         error = claim_swapfile(p, inode);
2091         if (unlikely(error))
2092                 goto bad_swap;
2093
2094         /*
2095          * Read the swap header.
2096          */
2097         if (!mapping->a_ops->readpage) {
2098                 error = -EINVAL;
2099                 goto bad_swap;
2100         }
2101         page = read_mapping_page(mapping, 0, swap_file);
2102         if (IS_ERR(page)) {
2103                 error = PTR_ERR(page);
2104                 goto bad_swap;
2105         }
2106         swap_header = kmap(page);
2107
2108         maxpages = read_swap_header(p, swap_header, inode);
2109         if (unlikely(!maxpages)) {
2110                 error = -EINVAL;
2111                 goto bad_swap;
2112         }
2113
2114         /* OK, set up the swap map and apply the bad block list */
2115         swap_map = vzalloc(maxpages);
2116         if (!swap_map) {
2117                 error = -ENOMEM;
2118                 goto bad_swap;
2119         }
2120
2121         error = swap_cgroup_swapon(p->type, maxpages);
2122         if (error)
2123                 goto bad_swap;
2124
2125         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2126                 maxpages, &span);
2127         if (unlikely(nr_extents < 0)) {
2128                 error = nr_extents;
2129                 goto bad_swap;
2130         }
2131         /* frontswap enabled? set up bit-per-page map for frontswap */
2132         if (frontswap_enabled)
2133                 frontswap_map = vzalloc(BITS_TO_LONGS(maxpages) * sizeof(long));
2134
2135         if (p->bdev) {
2136                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2137                         p->flags |= SWP_SOLIDSTATE;
2138                         p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2139                 }
2140
2141                 if ((swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2142                         /*
2143                          * When discard is enabled for swap with no particular
2144                          * policy flagged, we set all swap discard flags here in
2145                          * order to sustain backward compatibility with older
2146                          * swapon(8) releases.
2147                          */
2148                         p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2149                                      SWP_PAGE_DISCARD);
2150
2151                         /*
2152                          * By flagging sys_swapon, a sysadmin can tell us to
2153                          * either do single-time area discards only, or to just
2154                          * perform discards for released swap page-clusters.
2155                          * Now it's time to adjust the p->flags accordingly.
2156                          */
2157                         if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2158                                 p->flags &= ~SWP_PAGE_DISCARD;
2159                         else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2160                                 p->flags &= ~SWP_AREA_DISCARD;
2161
2162                         /* issue a swapon-time discard if it's still required */
2163                         if (p->flags & SWP_AREA_DISCARD) {
2164                                 int err = discard_swap(p);
2165                                 if (unlikely(err))
2166                                         printk(KERN_ERR
2167                                                "swapon: discard_swap(%p): %d\n",
2168                                                 p, err);
2169                         }
2170                 }
2171         }
2172
2173         mutex_lock(&swapon_mutex);
2174         prio = -1;
2175         if (swap_flags & SWAP_FLAG_PREFER)
2176                 prio =
2177                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2178         enable_swap_info(p, prio, swap_map, frontswap_map);
2179
2180         printk(KERN_INFO "Adding %uk swap on %s.  "
2181                         "Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
2182                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2183                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2184                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2185                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2186                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
2187                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
2188                 (frontswap_map) ? "FS" : "");
2189
2190         mutex_unlock(&swapon_mutex);
2191         atomic_inc(&proc_poll_event);
2192         wake_up_interruptible(&proc_poll_wait);
2193
2194         if (S_ISREG(inode->i_mode))
2195                 inode->i_flags |= S_SWAPFILE;
2196         error = 0;
2197         goto out;
2198 bad_swap:
2199         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2200                 set_blocksize(p->bdev, p->old_block_size);
2201                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2202         }
2203         destroy_swap_extents(p);
2204         swap_cgroup_swapoff(p->type);
2205         spin_lock(&swap_lock);
2206         p->swap_file = NULL;
2207         p->flags = 0;
2208         spin_unlock(&swap_lock);
2209         vfree(swap_map);
2210         if (swap_file) {
2211                 if (inode && S_ISREG(inode->i_mode)) {
2212                         mutex_unlock(&inode->i_mutex);
2213                         inode = NULL;
2214                 }
2215                 filp_close(swap_file, NULL);
2216         }
2217 out:
2218         if (page && !IS_ERR(page)) {
2219                 kunmap(page);
2220                 page_cache_release(page);
2221         }
2222         if (name)
2223                 putname(name);
2224         if (inode && S_ISREG(inode->i_mode))
2225                 mutex_unlock(&inode->i_mutex);
2226         return error;
2227 }
2228
2229 void si_swapinfo(struct sysinfo *val)
2230 {
2231         unsigned int type;
2232         unsigned long nr_to_be_unused = 0;
2233
2234         spin_lock(&swap_lock);
2235         for (type = 0; type < nr_swapfiles; type++) {
2236                 struct swap_info_struct *si = swap_info[type];
2237
2238                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2239                         nr_to_be_unused += si->inuse_pages;
2240         }
2241         val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
2242         val->totalswap = total_swap_pages + nr_to_be_unused;
2243         spin_unlock(&swap_lock);
2244 }
2245
2246 /*
2247  * Verify that a swap entry is valid and increment its swap map count.
2248  *
2249  * Returns error code in following case.
2250  * - success -> 0
2251  * - swp_entry is invalid -> EINVAL
2252  * - swp_entry is migration entry -> EINVAL
2253  * - swap-cache reference is requested but there is already one. -> EEXIST
2254  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2255  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2256  */
2257 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2258 {
2259         struct swap_info_struct *p;
2260         unsigned long offset, type;
2261         unsigned char count;
2262         unsigned char has_cache;
2263         int err = -EINVAL;
2264
2265         if (non_swap_entry(entry))
2266                 goto out;
2267
2268         type = swp_type(entry);
2269         if (type >= nr_swapfiles)
2270                 goto bad_file;
2271         p = swap_info[type];
2272         offset = swp_offset(entry);
2273
2274         spin_lock(&p->lock);
2275         if (unlikely(offset >= p->max))
2276                 goto unlock_out;
2277
2278         count = p->swap_map[offset];
2279         has_cache = count & SWAP_HAS_CACHE;
2280         count &= ~SWAP_HAS_CACHE;
2281         err = 0;
2282
2283         if (usage == SWAP_HAS_CACHE) {
2284
2285                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2286                 if (!has_cache && count)
2287                         has_cache = SWAP_HAS_CACHE;
2288                 else if (has_cache)             /* someone else added cache */
2289                         err = -EEXIST;
2290                 else                            /* no users remaining */
2291                         err = -ENOENT;
2292
2293         } else if (count || has_cache) {
2294
2295                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2296                         count += usage;
2297                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2298                         err = -EINVAL;
2299                 else if (swap_count_continued(p, offset, count))
2300                         count = COUNT_CONTINUED;
2301                 else
2302                         err = -ENOMEM;
2303         } else
2304                 err = -ENOENT;                  /* unused swap entry */
2305
2306         p->swap_map[offset] = count | has_cache;
2307
2308 unlock_out:
2309         spin_unlock(&p->lock);
2310 out:
2311         return err;
2312
2313 bad_file:
2314         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2315         goto out;
2316 }
2317
2318 /*
2319  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2320  * (in which case its reference count is never incremented).
2321  */
2322 void swap_shmem_alloc(swp_entry_t entry)
2323 {
2324         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2325 }
2326
2327 /*
2328  * Increase reference count of swap entry by 1.
2329  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2330  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2331  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2332  * might occur if a page table entry has got corrupted.
2333  */
2334 int swap_duplicate(swp_entry_t entry)
2335 {
2336         int err = 0;
2337
2338         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2339                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2340         return err;
2341 }
2342
2343 /*
2344  * @entry: swap entry for which we allocate swap cache.
2345  *
2346  * Called when allocating swap cache for existing swap entry,
2347  * This can return error codes. Returns 0 at success.
2348  * -EBUSY means there is a swap cache.
2349  * Note: return code is different from swap_duplicate().
2350  */
2351 int swapcache_prepare(swp_entry_t entry)
2352 {
2353         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2354 }
2355
2356 struct swap_info_struct *page_swap_info(struct page *page)
2357 {
2358         swp_entry_t swap = { .val = page_private(page) };
2359         BUG_ON(!PageSwapCache(page));
2360         return swap_info[swp_type(swap)];
2361 }
2362
2363 /*
2364  * out-of-line __page_file_ methods to avoid include hell.
2365  */
2366 struct address_space *__page_file_mapping(struct page *page)
2367 {
2368         VM_BUG_ON(!PageSwapCache(page));
2369         return page_swap_info(page)->swap_file->f_mapping;
2370 }
2371 EXPORT_SYMBOL_GPL(__page_file_mapping);
2372
2373 pgoff_t __page_file_index(struct page *page)
2374 {
2375         swp_entry_t swap = { .val = page_private(page) };
2376         VM_BUG_ON(!PageSwapCache(page));
2377         return swp_offset(swap);
2378 }
2379 EXPORT_SYMBOL_GPL(__page_file_index);
2380
2381 /*
2382  * add_swap_count_continuation - called when a swap count is duplicated
2383  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2384  * page of the original vmalloc'ed swap_map, to hold the continuation count
2385  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2386  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2387  *
2388  * These continuation pages are seldom referenced: the common paths all work
2389  * on the original swap_map, only referring to a continuation page when the
2390  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2391  *
2392  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2393  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2394  * can be called after dropping locks.
2395  */
2396 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2397 {
2398         struct swap_info_struct *si;
2399         struct page *head;
2400         struct page *page;
2401         struct page *list_page;
2402         pgoff_t offset;
2403         unsigned char count;
2404
2405         /*
2406          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2407          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2408          */
2409         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2410
2411         si = swap_info_get(entry);
2412         if (!si) {
2413                 /*
2414                  * An acceptable race has occurred since the failing
2415                  * __swap_duplicate(): the swap entry has been freed,
2416                  * perhaps even the whole swap_map cleared for swapoff.
2417                  */
2418                 goto outer;
2419         }
2420
2421         offset = swp_offset(entry);
2422         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2423
2424         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2425                 /*
2426                  * The higher the swap count, the more likely it is that tasks
2427                  * will race to add swap count continuation: we need to avoid
2428                  * over-provisioning.
2429                  */
2430                 goto out;
2431         }
2432
2433         if (!page) {
2434                 spin_unlock(&si->lock);
2435                 return -ENOMEM;
2436         }
2437
2438         /*
2439          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2440          * no architecture is using highmem pages for kernel pagetables: so it
2441          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2442          */
2443         head = vmalloc_to_page(si->swap_map + offset);
2444         offset &= ~PAGE_MASK;
2445
2446         /*
2447          * Page allocation does not initialize the page's lru field,
2448          * but it does always reset its private field.
2449          */
2450         if (!page_private(head)) {
2451                 BUG_ON(count & COUNT_CONTINUED);
2452                 INIT_LIST_HEAD(&head->lru);
2453                 set_page_private(head, SWP_CONTINUED);
2454                 si->flags |= SWP_CONTINUED;
2455         }
2456
2457         list_for_each_entry(list_page, &head->lru, lru) {
2458                 unsigned char *map;
2459
2460                 /*
2461                  * If the previous map said no continuation, but we've found
2462                  * a continuation page, free our allocation and use this one.
2463                  */
2464                 if (!(count & COUNT_CONTINUED))
2465                         goto out;
2466
2467                 map = kmap_atomic(list_page) + offset;
2468                 count = *map;
2469                 kunmap_atomic(map);
2470
2471                 /*
2472                  * If this continuation count now has some space in it,
2473                  * free our allocation and use this one.
2474                  */
2475                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2476                         goto out;
2477         }
2478
2479         list_add_tail(&page->lru, &head->lru);
2480         page = NULL;                    /* now it's attached, don't free it */
2481 out:
2482         spin_unlock(&si->lock);
2483 outer:
2484         if (page)
2485                 __free_page(page);
2486         return 0;
2487 }
2488
2489 /*
2490  * swap_count_continued - when the original swap_map count is incremented
2491  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2492  * into, carry if so, or else fail until a new continuation page is allocated;
2493  * when the original swap_map count is decremented from 0 with continuation,
2494  * borrow from the continuation and report whether it still holds more.
2495  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2496  */
2497 static bool swap_count_continued(struct swap_info_struct *si,
2498                                  pgoff_t offset, unsigned char count)
2499 {
2500         struct page *head;
2501         struct page *page;
2502         unsigned char *map;
2503
2504         head = vmalloc_to_page(si->swap_map + offset);
2505         if (page_private(head) != SWP_CONTINUED) {
2506                 BUG_ON(count & COUNT_CONTINUED);
2507                 return false;           /* need to add count continuation */
2508         }
2509
2510         offset &= ~PAGE_MASK;
2511         page = list_entry(head->lru.next, struct page, lru);
2512         map = kmap_atomic(page) + offset;
2513
2514         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2515                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2516
2517         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2518                 /*
2519                  * Think of how you add 1 to 999
2520                  */
2521                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2522                         kunmap_atomic(map);
2523                         page = list_entry(page->lru.next, struct page, lru);
2524                         BUG_ON(page == head);
2525                         map = kmap_atomic(page) + offset;
2526                 }
2527                 if (*map == SWAP_CONT_MAX) {
2528                         kunmap_atomic(map);
2529                         page = list_entry(page->lru.next, struct page, lru);
2530                         if (page == head)
2531                                 return false;   /* add count continuation */
2532                         map = kmap_atomic(page) + offset;
2533 init_map:               *map = 0;               /* we didn't zero the page */
2534                 }
2535                 *map += 1;
2536                 kunmap_atomic(map);
2537                 page = list_entry(page->lru.prev, struct page, lru);
2538                 while (page != head) {
2539                         map = kmap_atomic(page) + offset;
2540                         *map = COUNT_CONTINUED;
2541                         kunmap_atomic(map);
2542                         page = list_entry(page->lru.prev, struct page, lru);
2543                 }
2544                 return true;                    /* incremented */
2545
2546         } else {                                /* decrementing */
2547                 /*
2548                  * Think of how you subtract 1 from 1000
2549                  */
2550                 BUG_ON(count != COUNT_CONTINUED);
2551                 while (*map == COUNT_CONTINUED) {
2552                         kunmap_atomic(map);
2553                         page = list_entry(page->lru.next, struct page, lru);
2554                         BUG_ON(page == head);
2555                         map = kmap_atomic(page) + offset;
2556                 }
2557                 BUG_ON(*map == 0);
2558                 *map -= 1;
2559                 if (*map == 0)
2560                         count = 0;
2561                 kunmap_atomic(map);
2562                 page = list_entry(page->lru.prev, struct page, lru);
2563                 while (page != head) {
2564                         map = kmap_atomic(page) + offset;
2565                         *map = SWAP_CONT_MAX | count;
2566                         count = COUNT_CONTINUED;
2567                         kunmap_atomic(map);
2568                         page = list_entry(page->lru.prev, struct page, lru);
2569                 }
2570                 return count == COUNT_CONTINUED;
2571         }
2572 }
2573
2574 /*
2575  * free_swap_count_continuations - swapoff free all the continuation pages
2576  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2577  */
2578 static void free_swap_count_continuations(struct swap_info_struct *si)
2579 {
2580         pgoff_t offset;
2581
2582         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2583                 struct page *head;
2584                 head = vmalloc_to_page(si->swap_map + offset);
2585                 if (page_private(head)) {
2586                         struct list_head *this, *next;
2587                         list_for_each_safe(this, next, &head->lru) {
2588                                 struct page *page;
2589                                 page = list_entry(this, struct page, lru);
2590                                 list_del(this);
2591                                 __free_page(page);
2592                         }
2593                 }
2594         }
2595 }