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 static inline int maybe_same_pte(pte_t pte, pte_t swp_pte)
 870 {
 871 #ifdef CONFIG_MEM_SOFT_DIRTY
 872         /*
 873          * When pte keeps soft dirty bit the pte generated
 874          * from swap entry does not has it, still it's same
 875          * pte from logical point of view.
 876          */
 877         pte_t swp_pte_dirty = pte_swp_mksoft_dirty(swp_pte);
 878         return pte_same(pte, swp_pte) || pte_same(pte, swp_pte_dirty);
 879 #else
 880         return pte_same(pte, swp_pte);
 881 #endif
 882 }
 883
 884 /*
 885  * No need to decide whether this PTE shares the swap entry with others,
 886  * just let do_wp_page work it out if a write is requested later - to
 887  * force COW, vm_page_prot omits write permission from any private vma.
 888  */
 889 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 890                 unsigned long addr, swp_entry_t entry, struct page *page)
 891 {
 892         struct page *swapcache;
 893         struct mem_cgroup *memcg;
 894         spinlock_t *ptl;
 895         pte_t *pte;
 896         int ret = 1;
 897
 898         swapcache = page;
 899         page = ksm_might_need_to_copy(page, vma, addr);
 900         if (unlikely(!page))
 901                 return -ENOMEM;
 902
 903         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
 904                                          GFP_KERNEL, &memcg)) {
 905                 ret = -ENOMEM;
 906                 goto out_nolock;
 907         }
 908
 909         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 910         if (unlikely(!maybe_same_pte(*pte, swp_entry_to_pte(entry)))) {
 911                 mem_cgroup_cancel_charge_swapin(memcg);
 912                 ret = 0;
 913                 goto out;
 914         }
 915
 916         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
 917         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
 918         get_page(page);
 919         set_pte_at(vma->vm_mm, addr, pte,
 920                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 921         if (page == swapcache)
 922                 page_add_anon_rmap(page, vma, addr);
 923         else /* ksm created a completely new copy */
 924                 page_add_new_anon_rmap(page, vma, addr);
 925         mem_cgroup_commit_charge_swapin(page, memcg);
 926         swap_free(entry);
 927         /*
 928          * Move the page to the active list so it is not
 929          * immediately swapped out again after swapon.
 930          */
 931         activate_page(page);
 932 out:
 933         pte_unmap_unlock(pte, ptl);
 934 out_nolock:
 935         if (page != swapcache) {
 936                 unlock_page(page);
 937                 put_page(page);
 938         }
 939         return ret;
 940 }
 941
 942 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 943                                 unsigned long addr, unsigned long end,
 944                                 swp_entry_t entry, struct page *page)
 945 {
 946         pte_t swp_pte = swp_entry_to_pte(entry);
 947         pte_t *pte;
 948         int ret = 0;
 949
 950         /*
 951          * We don't actually need pte lock while scanning for swp_pte: since
 952          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 953          * page table while we're scanning; though it could get zapped, and on
 954          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 955          * of unmatched parts which look like swp_pte, so unuse_pte must
 956          * recheck under pte lock.  Scanning without pte lock lets it be
 957          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 958          */
 959         pte = pte_offset_map(pmd, addr);
 960         do {
 961                 /*
 962                  * swapoff spends a _lot_ of time in this loop!
 963                  * Test inline before going to call unuse_pte.
 964                  */
 965                 if (unlikely(maybe_same_pte(*pte, swp_pte))) {
 966                         pte_unmap(pte);
 967                         ret = unuse_pte(vma, pmd, addr, entry, page);
 968                         if (ret)
 969                                 goto out;
 970                         pte = pte_offset_map(pmd, addr);
 971                 }
 972         } while (pte++, addr += PAGE_SIZE, addr != end);
 973         pte_unmap(pte - 1);
 974 out:
 975         return ret;
 976 }
 977
 978 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 979                                 unsigned long addr, unsigned long end,
 980                                 swp_entry_t entry, struct page *page)
 981 {
 982         pmd_t *pmd;
 983         unsigned long next;
 984         int ret;
 985
 986         pmd = pmd_offset(pud, addr);
 987         do {
 988                 next = pmd_addr_end(addr, end);
 989                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 990                         continue;
 991                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 992                 if (ret)
 993                         return ret;
 994         } while (pmd++, addr = next, addr != end);
 995         return 0;
 996 }
 997
 998 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 999                                 unsigned long addr, unsigned long end,
1000                                 swp_entry_t entry, struct page *page)
1001 {
1002         pud_t *pud;
1003         unsigned long next;
1004         int ret;
1005
1006         pud = pud_offset(pgd, addr);
1007         do {
1008                 next = pud_addr_end(addr, end);
1009                 if (pud_none_or_clear_bad(pud))
1010                         continue;
1011                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1012                 if (ret)
1013                         return ret;
1014         } while (pud++, addr = next, addr != end);
1015         return 0;
1016 }
1017
1018 static int unuse_vma(struct vm_area_struct *vma,
1019                                 swp_entry_t entry, struct page *page)
1020 {
1021         pgd_t *pgd;
1022         unsigned long addr, end, next;
1023         int ret;
1024
1025         if (page_anon_vma(page)) {
1026                 addr = page_address_in_vma(page, vma);
1027                 if (addr == -EFAULT)
1028                         return 0;
1029                 else
1030                         end = addr + PAGE_SIZE;
1031         } else {
1032                 addr = vma->vm_start;
1033                 end = vma->vm_end;
1034         }
1035
1036         pgd = pgd_offset(vma->vm_mm, addr);
1037         do {
1038                 next = pgd_addr_end(addr, end);
1039                 if (pgd_none_or_clear_bad(pgd))
1040                         continue;
1041                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1042                 if (ret)
1043                         return ret;
1044         } while (pgd++, addr = next, addr != end);
1045         return 0;
1046 }
1047
1048 static int unuse_mm(struct mm_struct *mm,
1049                                 swp_entry_t entry, struct page *page)
1050 {
1051         struct vm_area_struct *vma;
1052         int ret = 0;
1053
1054         if (!down_read_trylock(&mm->mmap_sem)) {
1055                 /*
1056                  * Activate page so shrink_inactive_list is unlikely to unmap
1057                  * its ptes while lock is dropped, so swapoff can make progress.
1058                  */
1059                 activate_page(page);
1060                 unlock_page(page);
1061                 down_read(&mm->mmap_sem);
1062                 lock_page(page);
1063         }
1064         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1065                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1066                         break;
1067         }
1068         up_read(&mm->mmap_sem);
1069         return (ret < 0)? ret: 0;
1070 }
1071
1072 /*
1073  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1074  * from current position to next entry still in use.
1075  * Recycle to start on reaching the end, returning 0 when empty.
1076  */
1077 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1078                                         unsigned int prev, bool frontswap)
1079 {
1080         unsigned int max = si->max;
1081         unsigned int i = prev;
1082         unsigned char count;
1083
1084         /*
1085          * No need for swap_lock here: we're just looking
1086          * for whether an entry is in use, not modifying it; false
1087          * hits are okay, and sys_swapoff() has already prevented new
1088          * allocations from this area (while holding swap_lock).
1089          */
1090         for (;;) {
1091                 if (++i >= max) {
1092                         if (!prev) {
1093                                 i = 0;
1094                                 break;
1095                         }
1096                         /*
1097                          * No entries in use at top of swap_map,
1098                          * loop back to start and recheck there.
1099                          */
1100                         max = prev + 1;
1101                         prev = 0;
1102                         i = 1;
1103                 }
1104                 if (frontswap) {
1105                         if (frontswap_test(si, i))
1106                                 break;
1107                         else
1108                                 continue;
1109                 }
1110                 count = si->swap_map[i];
1111                 if (count && swap_count(count) != SWAP_MAP_BAD)
1112                         break;
1113         }
1114         return i;
1115 }
1116
1117 /*
1118  * We completely avoid races by reading each swap page in advance,
1119  * and then search for the process using it.  All the necessary
1120  * page table adjustments can then be made atomically.
1121  *
1122  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1123  * pages_to_unuse==0 means all pages; ignored if frontswap is false
1124  */
1125 int try_to_unuse(unsigned int type, bool frontswap,
1126                  unsigned long pages_to_unuse)
1127 {
1128         struct swap_info_struct *si = swap_info[type];
1129         struct mm_struct *start_mm;
1130         unsigned char *swap_map;
1131         unsigned char swcount;
1132         struct page *page;
1133         swp_entry_t entry;
1134         unsigned int i = 0;
1135         int retval = 0;
1136
1137         /*
1138          * When searching mms for an entry, a good strategy is to
1139          * start at the first mm we freed the previous entry from
1140          * (though actually we don't notice whether we or coincidence
1141          * freed the entry).  Initialize this start_mm with a hold.
1142          *
1143          * A simpler strategy would be to start at the last mm we
1144          * freed the previous entry from; but that would take less
1145          * advantage of mmlist ordering, which clusters forked mms
1146          * together, child after parent.  If we race with dup_mmap(), we
1147          * prefer to resolve parent before child, lest we miss entries
1148          * duplicated after we scanned child: using last mm would invert
1149          * that.
1150          */
1151         start_mm = &init_mm;
1152         atomic_inc(&init_mm.mm_users);
1153
1154         /*
1155          * Keep on scanning until all entries have gone.  Usually,
1156          * one pass through swap_map is enough, but not necessarily:
1157          * there are races when an instance of an entry might be missed.
1158          */
1159         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1160                 if (signal_pending(current)) {
1161                         retval = -EINTR;
1162                         break;
1163                 }
1164
1165                 /*
1166                  * Get a page for the entry, using the existing swap
1167                  * cache page if there is one.  Otherwise, get a clean
1168                  * page and read the swap into it.
1169                  */
1170                 swap_map = &si->swap_map[i];
1171                 entry = swp_entry(type, i);
1172                 page = read_swap_cache_async(entry,
1173                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1174                 if (!page) {
1175                         /*
1176                          * Either swap_duplicate() failed because entry
1177                          * has been freed independently, and will not be
1178                          * reused since sys_swapoff() already disabled
1179                          * allocation from here, or alloc_page() failed.
1180                          */
1181                         if (!*swap_map)
1182                                 continue;
1183                         retval = -ENOMEM;
1184                         break;
1185                 }
1186
1187                 /*
1188                  * Don't hold on to start_mm if it looks like exiting.
1189                  */
1190                 if (atomic_read(&start_mm->mm_users) == 1) {
1191                         mmput(start_mm);
1192                         start_mm = &init_mm;
1193                         atomic_inc(&init_mm.mm_users);
1194                 }
1195
1196                 /*
1197                  * Wait for and lock page.  When do_swap_page races with
1198                  * try_to_unuse, do_swap_page can handle the fault much
1199                  * faster than try_to_unuse can locate the entry.  This
1200                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1201                  * defer to do_swap_page in such a case - in some tests,
1202                  * do_swap_page and try_to_unuse repeatedly compete.
1203                  */
1204                 wait_on_page_locked(page);
1205                 wait_on_page_writeback(page);
1206                 lock_page(page);
1207                 wait_on_page_writeback(page);
1208
1209                 /*
1210                  * Remove all references to entry.
1211                  */
1212                 swcount = *swap_map;
1213                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1214                         retval = shmem_unuse(entry, page);
1215                         /* page has already been unlocked and released */
1216                         if (retval < 0)
1217                                 break;
1218                         continue;
1219                 }
1220                 if (swap_count(swcount) && start_mm != &init_mm)
1221                         retval = unuse_mm(start_mm, entry, page);
1222
1223                 if (swap_count(*swap_map)) {
1224                         int set_start_mm = (*swap_map >= swcount);
1225                         struct list_head *p = &start_mm->mmlist;
1226                         struct mm_struct *new_start_mm = start_mm;
1227                         struct mm_struct *prev_mm = start_mm;
1228                         struct mm_struct *mm;
1229
1230                         atomic_inc(&new_start_mm->mm_users);
1231                         atomic_inc(&prev_mm->mm_users);
1232                         spin_lock(&mmlist_lock);
1233                         while (swap_count(*swap_map) && !retval &&
1234                                         (p = p->next) != &start_mm->mmlist) {
1235                                 mm = list_entry(p, struct mm_struct, mmlist);
1236                                 if (!atomic_inc_not_zero(&mm->mm_users))
1237                                         continue;
1238                                 spin_unlock(&mmlist_lock);
1239                                 mmput(prev_mm);
1240                                 prev_mm = mm;
1241
1242                                 cond_resched();
1243
1244                                 swcount = *swap_map;
1245                                 if (!swap_count(swcount)) /* any usage ? */
1246                                         ;
1247                                 else if (mm == &init_mm)
1248                                         set_start_mm = 1;
1249                                 else
1250                                         retval = unuse_mm(mm, entry, page);
1251
1252                                 if (set_start_mm && *swap_map < swcount) {
1253                                         mmput(new_start_mm);
1254                                         atomic_inc(&mm->mm_users);
1255                                         new_start_mm = mm;
1256                                         set_start_mm = 0;
1257                                 }
1258                                 spin_lock(&mmlist_lock);
1259                         }
1260                         spin_unlock(&mmlist_lock);
1261                         mmput(prev_mm);
1262                         mmput(start_mm);
1263                         start_mm = new_start_mm;
1264                 }
1265                 if (retval) {
1266                         unlock_page(page);
1267                         page_cache_release(page);
1268                         break;
1269                 }
1270
1271                 /*
1272                  * If a reference remains (rare), we would like to leave
1273                  * the page in the swap cache; but try_to_unmap could
1274                  * then re-duplicate the entry once we drop page lock,
1275                  * so we might loop indefinitely; also, that page could
1276                  * not be swapped out to other storage meanwhile.  So:
1277                  * delete from cache even if there's another reference,
1278                  * after ensuring that the data has been saved to disk -
1279                  * since if the reference remains (rarer), it will be
1280                  * read from disk into another page.  Splitting into two
1281                  * pages would be incorrect if swap supported "shared
1282                  * private" pages, but they are handled by tmpfs files.
1283                  *
1284                  * Given how unuse_vma() targets one particular offset
1285                  * in an anon_vma, once the anon_vma has been determined,
1286                  * this splitting happens to be just what is needed to
1287                  * handle where KSM pages have been swapped out: re-reading
1288                  * is unnecessarily slow, but we can fix that later on.
1289                  */
1290                 if (swap_count(*swap_map) &&
1291                      PageDirty(page) && PageSwapCache(page)) {
1292                         struct writeback_control wbc = {
1293                                 .sync_mode = WB_SYNC_NONE,
1294                         };
1295
1296                         swap_writepage(page, &wbc);
1297                         lock_page(page);
1298                         wait_on_page_writeback(page);
1299                 }
1300
1301                 /*
1302                  * It is conceivable that a racing task removed this page from
1303                  * swap cache just before we acquired the page lock at the top,
1304                  * or while we dropped it in unuse_mm().  The page might even
1305                  * be back in swap cache on another swap area: that we must not
1306                  * delete, since it may not have been written out to swap yet.
1307                  */
1308                 if (PageSwapCache(page) &&
1309                     likely(page_private(page) == entry.val))
1310                         delete_from_swap_cache(page);
1311
1312                 /*
1313                  * So we could skip searching mms once swap count went
1314                  * to 1, we did not mark any present ptes as dirty: must
1315                  * mark page dirty so shrink_page_list will preserve it.
1316                  */
1317                 SetPageDirty(page);
1318                 unlock_page(page);
1319                 page_cache_release(page);
1320
1321                 /*
1322                  * Make sure that we aren't completely killing
1323                  * interactive performance.
1324                  */
1325                 cond_resched();
1326                 if (frontswap && pages_to_unuse > 0) {
1327                         if (!--pages_to_unuse)
1328                                 break;
1329                 }
1330         }
1331
1332         mmput(start_mm);
1333         return retval;
1334 }
1335
1336 /*
1337  * After a successful try_to_unuse, if no swap is now in use, we know
1338  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1339  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1340  * added to the mmlist just after page_duplicate - before would be racy.
1341  */
1342 static void drain_mmlist(void)
1343 {
1344         struct list_head *p, *next;
1345         unsigned int type;
1346
1347         for (type = 0; type < nr_swapfiles; type++)
1348                 if (swap_info[type]->inuse_pages)
1349                         return;
1350         spin_lock(&mmlist_lock);
1351         list_for_each_safe(p, next, &init_mm.mmlist)
1352                 list_del_init(p);
1353         spin_unlock(&mmlist_lock);
1354 }
1355
1356 /*
1357  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1358  * corresponds to page offset for the specified swap entry.
1359  * Note that the type of this function is sector_t, but it returns page offset
1360  * into the bdev, not sector offset.
1361  */
1362 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1363 {
1364         struct swap_info_struct *sis;
1365         struct swap_extent *start_se;
1366         struct swap_extent *se;
1367         pgoff_t offset;
1368
1369         sis = swap_info[swp_type(entry)];
1370         *bdev = sis->bdev;
1371
1372         offset = swp_offset(entry);
1373         start_se = sis->curr_swap_extent;
1374         se = start_se;
1375
1376         for ( ; ; ) {
1377                 struct list_head *lh;
1378
1379                 if (se->start_page <= offset &&
1380                                 offset < (se->start_page + se->nr_pages)) {
1381                         return se->start_block + (offset - se->start_page);
1382                 }
1383                 lh = se->list.next;
1384                 se = list_entry(lh, struct swap_extent, list);
1385                 sis->curr_swap_extent = se;
1386                 BUG_ON(se == start_se);         /* It *must* be present */
1387         }
1388 }
1389
1390 /*
1391  * Returns the page offset into bdev for the specified page's swap entry.
1392  */
1393 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1394 {
1395         swp_entry_t entry;
1396         entry.val = page_private(page);
1397         return map_swap_entry(entry, bdev);
1398 }
1399
1400 /*
1401  * Free all of a swapdev's extent information
1402  */
1403 static void destroy_swap_extents(struct swap_info_struct *sis)
1404 {
1405         while (!list_empty(&sis->first_swap_extent.list)) {
1406                 struct swap_extent *se;
1407
1408                 se = list_entry(sis->first_swap_extent.list.next,
1409                                 struct swap_extent, list);
1410                 list_del(&se->list);
1411                 kfree(se);
1412         }
1413
1414         if (sis->flags & SWP_FILE) {
1415                 struct file *swap_file = sis->swap_file;
1416                 struct address_space *mapping = swap_file->f_mapping;
1417
1418                 sis->flags &= ~SWP_FILE;
1419                 mapping->a_ops->swap_deactivate(swap_file);
1420         }
1421 }
1422
1423 /*
1424  * Add a block range (and the corresponding page range) into this swapdev's
1425  * extent list.  The extent list is kept sorted in page order.
1426  *
1427  * This function rather assumes that it is called in ascending page order.
1428  */
1429 int
1430 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1431                 unsigned long nr_pages, sector_t start_block)
1432 {
1433         struct swap_extent *se;
1434         struct swap_extent *new_se;
1435         struct list_head *lh;
1436
1437         if (start_page == 0) {
1438                 se = &sis->first_swap_extent;
1439                 sis->curr_swap_extent = se;
1440                 se->start_page = 0;
1441                 se->nr_pages = nr_pages;
1442                 se->start_block = start_block;
1443                 return 1;
1444         } else {
1445                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1446                 se = list_entry(lh, struct swap_extent, list);
1447                 BUG_ON(se->start_page + se->nr_pages != start_page);
1448                 if (se->start_block + se->nr_pages == start_block) {
1449                         /* Merge it */
1450                         se->nr_pages += nr_pages;
1451                         return 0;
1452                 }
1453         }
1454
1455         /*
1456          * No merge.  Insert a new extent, preserving ordering.
1457          */
1458         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1459         if (new_se == NULL)
1460                 return -ENOMEM;
1461         new_se->start_page = start_page;
1462         new_se->nr_pages = nr_pages;
1463         new_se->start_block = start_block;
1464
1465         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1466         return 1;
1467 }
1468
1469 /*
1470  * A `swap extent' is a simple thing which maps a contiguous range of pages
1471  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1472  * is built at swapon time and is then used at swap_writepage/swap_readpage
1473  * time for locating where on disk a page belongs.
1474  *
1475  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1476  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1477  * swap files identically.
1478  *
1479  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1480  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1481  * swapfiles are handled *identically* after swapon time.
1482  *
1483  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1484  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1485  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1486  * requirements, they are simply tossed out - we will never use those blocks
1487  * for swapping.
1488  *
1489  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1490  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1491  * which will scribble on the fs.
1492  *
1493  * The amount of disk space which a single swap extent represents varies.
1494  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1495  * extents in the list.  To avoid much list walking, we cache the previous
1496  * search location in `curr_swap_extent', and start new searches from there.
1497  * This is extremely effective.  The average number of iterations in
1498  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1499  */
1500 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1501 {
1502         struct file *swap_file = sis->swap_file;
1503         struct address_space *mapping = swap_file->f_mapping;
1504         struct inode *inode = mapping->host;
1505         int ret;
1506
1507         if (S_ISBLK(inode->i_mode)) {
1508                 ret = add_swap_extent(sis, 0, sis->max, 0);
1509                 *span = sis->pages;
1510                 return ret;
1511         }
1512
1513         if (mapping->a_ops->swap_activate) {
1514                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1515                 if (!ret) {
1516                         sis->flags |= SWP_FILE;
1517                         ret = add_swap_extent(sis, 0, sis->max, 0);
1518                         *span = sis->pages;
1519                 }
1520                 return ret;
1521         }
1522
1523         return generic_swapfile_activate(sis, swap_file, span);
1524 }
1525
1526 static void _enable_swap_info(struct swap_info_struct *p, int prio,
1527                                 unsigned char *swap_map)
1528 {
1529         int i, prev;
1530
1531         if (prio >= 0)
1532                 p->prio = prio;
1533         else
1534                 p->prio = --least_priority;
1535         p->swap_map = swap_map;
1536         p->flags |= SWP_WRITEOK;
1537         atomic_long_add(p->pages, &nr_swap_pages);
1538         total_swap_pages += p->pages;
1539
1540         /* insert swap space into swap_list: */
1541         prev = -1;
1542         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1543                 if (p->prio >= swap_info[i]->prio)
1544                         break;
1545                 prev = i;
1546         }
1547         p->next = i;
1548         if (prev < 0)
1549                 swap_list.head = swap_list.next = p->type;
1550         else
1551                 swap_info[prev]->next = p->type;
1552 }
1553
1554 static void enable_swap_info(struct swap_info_struct *p, int prio,
1555                                 unsigned char *swap_map,
1556                                 unsigned long *frontswap_map)
1557 {
1558         frontswap_init(p->type, frontswap_map);
1559         spin_lock(&swap_lock);
1560         spin_lock(&p->lock);
1561          _enable_swap_info(p, prio, swap_map);
1562         spin_unlock(&p->lock);
1563         spin_unlock(&swap_lock);
1564 }
1565
1566 static void reinsert_swap_info(struct swap_info_struct *p)
1567 {
1568         spin_lock(&swap_lock);
1569         spin_lock(&p->lock);
1570         _enable_swap_info(p, p->prio, p->swap_map);
1571         spin_unlock(&p->lock);
1572         spin_unlock(&swap_lock);
1573 }
1574
1575 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1576 {
1577         struct swap_info_struct *p = NULL;
1578         unsigned char *swap_map;
1579         unsigned long *frontswap_map;
1580         struct file *swap_file, *victim;
1581         struct address_space *mapping;
1582         struct inode *inode;
1583         struct filename *pathname;
1584         int i, type, prev;
1585         int err;
1586
1587         if (!capable(CAP_SYS_ADMIN))
1588                 return -EPERM;
1589
1590         BUG_ON(!current->mm);
1591
1592         pathname = getname(specialfile);
1593         if (IS_ERR(pathname))
1594                 return PTR_ERR(pathname);
1595
1596         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
1597         err = PTR_ERR(victim);
1598         if (IS_ERR(victim))
1599                 goto out;
1600
1601         mapping = victim->f_mapping;
1602         prev = -1;
1603         spin_lock(&swap_lock);
1604         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1605                 p = swap_info[type];
1606                 if (p->flags & SWP_WRITEOK) {
1607                         if (p->swap_file->f_mapping == mapping)
1608                                 break;
1609                 }
1610                 prev = type;
1611         }
1612         if (type < 0) {
1613                 err = -EINVAL;
1614                 spin_unlock(&swap_lock);
1615                 goto out_dput;
1616         }
1617         if (!security_vm_enough_memory_mm(current->mm, p->pages))
1618                 vm_unacct_memory(p->pages);
1619         else {
1620                 err = -ENOMEM;
1621                 spin_unlock(&swap_lock);
1622                 goto out_dput;
1623         }
1624         if (prev < 0)
1625                 swap_list.head = p->next;
1626         else
1627                 swap_info[prev]->next = p->next;
1628         if (type == swap_list.next) {
1629                 /* just pick something that's safe... */
1630                 swap_list.next = swap_list.head;
1631         }
1632         spin_lock(&p->lock);
1633         if (p->prio < 0) {
1634                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1635                         swap_info[i]->prio = p->prio--;
1636                 least_priority++;
1637         }
1638         atomic_long_sub(p->pages, &nr_swap_pages);
1639         total_swap_pages -= p->pages;
1640         p->flags &= ~SWP_WRITEOK;
1641         spin_unlock(&p->lock);
1642         spin_unlock(&swap_lock);
1643
1644         set_current_oom_origin();
1645         err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1646         clear_current_oom_origin();
1647
1648         if (err) {
1649                 /* re-insert swap space back into swap_list */
1650                 reinsert_swap_info(p);
1651                 goto out_dput;
1652         }
1653
1654         destroy_swap_extents(p);
1655         if (p->flags & SWP_CONTINUED)
1656                 free_swap_count_continuations(p);
1657
1658         mutex_lock(&swapon_mutex);
1659         spin_lock(&swap_lock);
1660         spin_lock(&p->lock);
1661         drain_mmlist();
1662
1663         /* wait for anyone still in scan_swap_map */
1664         p->highest_bit = 0;             /* cuts scans short */
1665         while (p->flags >= SWP_SCANNING) {
1666                 spin_unlock(&p->lock);
1667                 spin_unlock(&swap_lock);
1668                 schedule_timeout_uninterruptible(1);
1669                 spin_lock(&swap_lock);
1670                 spin_lock(&p->lock);
1671         }
1672
1673         swap_file = p->swap_file;
1674         p->swap_file = NULL;
1675         p->max = 0;
1676         swap_map = p->swap_map;
1677         p->swap_map = NULL;
1678         p->flags = 0;
1679         frontswap_map = frontswap_map_get(p);
1680         frontswap_map_set(p, NULL);
1681         spin_unlock(&p->lock);
1682         spin_unlock(&swap_lock);
1683         frontswap_invalidate_area(type);
1684         mutex_unlock(&swapon_mutex);
1685         vfree(swap_map);
1686         vfree(frontswap_map);
1687         /* Destroy swap account informatin */
1688         swap_cgroup_swapoff(type);
1689
1690         inode = mapping->host;
1691         if (S_ISBLK(inode->i_mode)) {
1692                 struct block_device *bdev = I_BDEV(inode);
1693                 set_blocksize(bdev, p->old_block_size);
1694                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1695         } else {
1696                 mutex_lock(&inode->i_mutex);
1697                 inode->i_flags &= ~S_SWAPFILE;
1698                 mutex_unlock(&inode->i_mutex);
1699         }
1700         filp_close(swap_file, NULL);
1701         err = 0;
1702         atomic_inc(&proc_poll_event);
1703         wake_up_interruptible(&proc_poll_wait);
1704
1705 out_dput:
1706         filp_close(victim, NULL);
1707 out:
1708         putname(pathname);
1709         return err;
1710 }
1711
1712 #ifdef CONFIG_PROC_FS
1713 static unsigned swaps_poll(struct file *file, poll_table *wait)
1714 {
1715         struct seq_file *seq = file->private_data;
1716
1717         poll_wait(file, &proc_poll_wait, wait);
1718
1719         if (seq->poll_event != atomic_read(&proc_poll_event)) {
1720                 seq->poll_event = atomic_read(&proc_poll_event);
1721                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1722         }
1723
1724         return POLLIN | POLLRDNORM;
1725 }
1726
1727 /* iterator */
1728 static void *swap_start(struct seq_file *swap, loff_t *pos)
1729 {
1730         struct swap_info_struct *si;
1731         int type;
1732         loff_t l = *pos;
1733
1734         mutex_lock(&swapon_mutex);
1735
1736         if (!l)
1737                 return SEQ_START_TOKEN;
1738
1739         for (type = 0; type < nr_swapfiles; type++) {
1740                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1741                 si = swap_info[type];
1742                 if (!(si->flags & SWP_USED) || !si->swap_map)
1743                         continue;
1744                 if (!--l)
1745                         return si;
1746         }
1747
1748         return NULL;
1749 }
1750
1751 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1752 {
1753         struct swap_info_struct *si = v;
1754         int type;
1755
1756         if (v == SEQ_START_TOKEN)
1757                 type = 0;
1758         else
1759                 type = si->type + 1;
1760
1761         for (; type < nr_swapfiles; type++) {
1762                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1763                 si = swap_info[type];
1764                 if (!(si->flags & SWP_USED) || !si->swap_map)
1765                         continue;
1766                 ++*pos;
1767                 return si;
1768         }
1769
1770         return NULL;
1771 }
1772
1773 static void swap_stop(struct seq_file *swap, void *v)
1774 {
1775         mutex_unlock(&swapon_mutex);
1776 }
1777
1778 static int swap_show(struct seq_file *swap, void *v)
1779 {
1780         struct swap_info_struct *si = v;
1781         struct file *file;
1782         int len;
1783
1784         if (si == SEQ_START_TOKEN) {
1785                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1786                 return 0;
1787         }
1788
1789         file = si->swap_file;
1790         len = seq_path(swap, &file->f_path, " \t\n\\");
1791         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1792                         len < 40 ? 40 - len : 1, " ",
1793                         S_ISBLK(file_inode(file)->i_mode) ?
1794                                 "partition" : "file\t",
1795                         si->pages << (PAGE_SHIFT - 10),
1796                         si->inuse_pages << (PAGE_SHIFT - 10),
1797                         si->prio);
1798         return 0;
1799 }
1800
1801 static const struct seq_operations swaps_op = {
1802         .start =        swap_start,
1803         .next =         swap_next,
1804         .stop =         swap_stop,
1805         .show =         swap_show
1806 };
1807
1808 static int swaps_open(struct inode *inode, struct file *file)
1809 {
1810         struct seq_file *seq;
1811         int ret;
1812
1813         ret = seq_open(file, &swaps_op);
1814         if (ret)
1815                 return ret;
1816
1817         seq = file->private_data;
1818         seq->poll_event = atomic_read(&proc_poll_event);
1819         return 0;
1820 }
1821
1822 static const struct file_operations proc_swaps_operations = {
1823         .open           = swaps_open,
1824         .read           = seq_read,
1825         .llseek         = seq_lseek,
1826         .release        = seq_release,
1827         .poll           = swaps_poll,
1828 };
1829
1830 static int __init procswaps_init(void)
1831 {
1832         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1833         return 0;
1834 }
1835 __initcall(procswaps_init);
1836 #endif /* CONFIG_PROC_FS */
1837
1838 #ifdef MAX_SWAPFILES_CHECK
1839 static int __init max_swapfiles_check(void)
1840 {
1841         MAX_SWAPFILES_CHECK();
1842         return 0;
1843 }
1844 late_initcall(max_swapfiles_check);
1845 #endif
1846
1847 static struct swap_info_struct *alloc_swap_info(void)
1848 {
1849         struct swap_info_struct *p;
1850         unsigned int type;
1851
1852         p = kzalloc(sizeof(*p), GFP_KERNEL);
1853         if (!p)
1854                 return ERR_PTR(-ENOMEM);
1855
1856         spin_lock(&swap_lock);
1857         for (type = 0; type < nr_swapfiles; type++) {
1858                 if (!(swap_info[type]->flags & SWP_USED))
1859                         break;
1860         }
1861         if (type >= MAX_SWAPFILES) {
1862                 spin_unlock(&swap_lock);
1863                 kfree(p);
1864                 return ERR_PTR(-EPERM);
1865         }
1866         if (type >= nr_swapfiles) {
1867                 p->type = type;
1868                 swap_info[type] = p;
1869                 /*
1870                  * Write swap_info[type] before nr_swapfiles, in case a
1871                  * racing procfs swap_start() or swap_next() is reading them.
1872                  * (We never shrink nr_swapfiles, we never free this entry.)
1873                  */
1874                 smp_wmb();
1875                 nr_swapfiles++;
1876         } else {
1877                 kfree(p);
1878                 p = swap_info[type];
1879                 /*
1880                  * Do not memset this entry: a racing procfs swap_next()
1881                  * would be relying on p->type to remain valid.
1882                  */
1883         }
1884         INIT_LIST_HEAD(&p->first_swap_extent.list);
1885         p->flags = SWP_USED;
1886         p->next = -1;
1887         spin_unlock(&swap_lock);
1888         spin_lock_init(&p->lock);
1889
1890         return p;
1891 }
1892
1893 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1894 {
1895         int error;
1896
1897         if (S_ISBLK(inode->i_mode)) {
1898                 p->bdev = bdgrab(I_BDEV(inode));
1899                 error = blkdev_get(p->bdev,
1900                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1901                                    sys_swapon);
1902                 if (error < 0) {
1903                         p->bdev = NULL;
1904                         return -EINVAL;
1905                 }
1906                 p->old_block_size = block_size(p->bdev);
1907                 error = set_blocksize(p->bdev, PAGE_SIZE);
1908                 if (error < 0)
1909                         return error;
1910                 p->flags |= SWP_BLKDEV;
1911         } else if (S_ISREG(inode->i_mode)) {
1912                 p->bdev = inode->i_sb->s_bdev;
1913                 mutex_lock(&inode->i_mutex);
1914                 if (IS_SWAPFILE(inode))
1915                         return -EBUSY;
1916         } else
1917                 return -EINVAL;
1918
1919         return 0;
1920 }
1921
1922 static unsigned long read_swap_header(struct swap_info_struct *p,
1923                                         union swap_header *swap_header,
1924                                         struct inode *inode)
1925 {
1926         int i;
1927         unsigned long maxpages;
1928         unsigned long swapfilepages;
1929
1930         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1931                 printk(KERN_ERR "Unable to find swap-space signature\n");
1932                 return 0;
1933         }
1934
1935         /* swap partition endianess hack... */
1936         if (swab32(swap_header->info.version) == 1) {
1937                 swab32s(&swap_header->info.version);
1938                 swab32s(&swap_header->info.last_page);
1939                 swab32s(&swap_header->info.nr_badpages);
1940                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1941                         swab32s(&swap_header->info.badpages[i]);
1942         }
1943         /* Check the swap header's sub-version */
1944         if (swap_header->info.version != 1) {
1945                 printk(KERN_WARNING
1946                        "Unable to handle swap header version %d\n",
1947                        swap_header->info.version);
1948                 return 0;
1949         }
1950
1951         p->lowest_bit  = 1;
1952         p->cluster_next = 1;
1953         p->cluster_nr = 0;
1954
1955         /*
1956          * Find out how many pages are allowed for a single swap
1957          * device. There are two limiting factors: 1) the number
1958          * of bits for the swap offset in the swp_entry_t type, and
1959          * 2) the number of bits in the swap pte as defined by the
1960          * different architectures. In order to find the
1961          * largest possible bit mask, a swap entry with swap type 0
1962          * and swap offset ~0UL is created, encoded to a swap pte,
1963          * decoded to a swp_entry_t again, and finally the swap
1964          * offset is extracted. This will mask all the bits from
1965          * the initial ~0UL mask that can't be encoded in either
1966          * the swp_entry_t or the architecture definition of a
1967          * swap pte.
1968          */
1969         maxpages = swp_offset(pte_to_swp_entry(
1970                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
1971         if (maxpages > swap_header->info.last_page) {
1972                 maxpages = swap_header->info.last_page + 1;
1973                 /* p->max is an unsigned int: don't overflow it */
1974                 if ((unsigned int)maxpages == 0)
1975                         maxpages = UINT_MAX;
1976         }
1977         p->highest_bit = maxpages - 1;
1978
1979         if (!maxpages)
1980                 return 0;
1981         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1982         if (swapfilepages && maxpages > swapfilepages) {
1983                 printk(KERN_WARNING
1984                        "Swap area shorter than signature indicates\n");
1985                 return 0;
1986         }
1987         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1988                 return 0;
1989         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1990                 return 0;
1991
1992         return maxpages;
1993 }
1994
1995 static int setup_swap_map_and_extents(struct swap_info_struct *p,
1996                                         union swap_header *swap_header,
1997                                         unsigned char *swap_map,
1998                                         unsigned long maxpages,
1999                                         sector_t *span)
2000 {
2001         int i;
2002         unsigned int nr_good_pages;
2003         int nr_extents;
2004
2005         nr_good_pages = maxpages - 1;   /* omit header page */
2006
2007         for (i = 0; i < swap_header->info.nr_badpages; i++) {
2008                 unsigned int page_nr = swap_header->info.badpages[i];
2009                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2010                         return -EINVAL;
2011                 if (page_nr < maxpages) {
2012                         swap_map[page_nr] = SWAP_MAP_BAD;
2013                         nr_good_pages--;
2014                 }
2015         }
2016
2017         if (nr_good_pages) {
2018                 swap_map[0] = SWAP_MAP_BAD;
2019                 p->max = maxpages;
2020                 p->pages = nr_good_pages;
2021                 nr_extents = setup_swap_extents(p, span);
2022                 if (nr_extents < 0)
2023                         return nr_extents;
2024                 nr_good_pages = p->pages;
2025         }
2026         if (!nr_good_pages) {
2027                 printk(KERN_WARNING "Empty swap-file\n");
2028                 return -EINVAL;
2029         }
2030
2031         return nr_extents;
2032 }
2033
2034 /*
2035  * Helper to sys_swapon determining if a given swap
2036  * backing device queue supports DISCARD operations.
2037  */
2038 static bool swap_discardable(struct swap_info_struct *si)
2039 {
2040         struct request_queue *q = bdev_get_queue(si->bdev);
2041
2042         if (!q || !blk_queue_discard(q))
2043                 return false;
2044
2045         return true;
2046 }
2047
2048 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2049 {
2050         struct swap_info_struct *p;
2051         struct filename *name;
2052         struct file *swap_file = NULL;
2053         struct address_space *mapping;
2054         int i;
2055         int prio;
2056         int error;
2057         union swap_header *swap_header;
2058         int nr_extents;
2059         sector_t span;
2060         unsigned long maxpages;
2061         unsigned char *swap_map = NULL;
2062         unsigned long *frontswap_map = NULL;
2063         struct page *page = NULL;
2064         struct inode *inode = NULL;
2065
2066         if (swap_flags & ~SWAP_FLAGS_VALID)
2067                 return -EINVAL;
2068
2069         if (!capable(CAP_SYS_ADMIN))
2070                 return -EPERM;
2071
2072         p = alloc_swap_info();
2073         if (IS_ERR(p))
2074                 return PTR_ERR(p);
2075
2076         name = getname(specialfile);
2077         if (IS_ERR(name)) {
2078                 error = PTR_ERR(name);
2079                 name = NULL;
2080                 goto bad_swap;
2081         }
2082         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2083         if (IS_ERR(swap_file)) {
2084                 error = PTR_ERR(swap_file);
2085                 swap_file = NULL;
2086                 goto bad_swap;
2087         }
2088
2089         p->swap_file = swap_file;
2090         mapping = swap_file->f_mapping;
2091
2092         for (i = 0; i < nr_swapfiles; i++) {
2093                 struct swap_info_struct *q = swap_info[i];
2094
2095                 if (q == p || !q->swap_file)
2096                         continue;
2097                 if (mapping == q->swap_file->f_mapping) {
2098                         error = -EBUSY;
2099                         goto bad_swap;
2100                 }
2101         }
2102
2103         inode = mapping->host;
2104         /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2105         error = claim_swapfile(p, inode);
2106         if (unlikely(error))
2107                 goto bad_swap;
2108
2109         /*
2110          * Read the swap header.
2111          */
2112         if (!mapping->a_ops->readpage) {
2113                 error = -EINVAL;
2114                 goto bad_swap;
2115         }
2116         page = read_mapping_page(mapping, 0, swap_file);
2117         if (IS_ERR(page)) {
2118                 error = PTR_ERR(page);
2119                 goto bad_swap;
2120         }
2121         swap_header = kmap(page);
2122
2123         maxpages = read_swap_header(p, swap_header, inode);
2124         if (unlikely(!maxpages)) {
2125                 error = -EINVAL;
2126                 goto bad_swap;
2127         }
2128
2129         /* OK, set up the swap map and apply the bad block list */
2130         swap_map = vzalloc(maxpages);
2131         if (!swap_map) {
2132                 error = -ENOMEM;
2133                 goto bad_swap;
2134         }
2135
2136         error = swap_cgroup_swapon(p->type, maxpages);
2137         if (error)
2138                 goto bad_swap;
2139
2140         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2141                 maxpages, &span);
2142         if (unlikely(nr_extents < 0)) {
2143                 error = nr_extents;
2144                 goto bad_swap;
2145         }
2146         /* frontswap enabled? set up bit-per-page map for frontswap */
2147         if (frontswap_enabled)
2148                 frontswap_map = vzalloc(BITS_TO_LONGS(maxpages) * sizeof(long));
2149
2150         if (p->bdev) {
2151                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2152                         p->flags |= SWP_SOLIDSTATE;
2153                         p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2154                 }
2155
2156                 if ((swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2157                         /*
2158                          * When discard is enabled for swap with no particular
2159                          * policy flagged, we set all swap discard flags here in
2160                          * order to sustain backward compatibility with older
2161                          * swapon(8) releases.
2162                          */
2163                         p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2164                                      SWP_PAGE_DISCARD);
2165
2166                         /*
2167                          * By flagging sys_swapon, a sysadmin can tell us to
2168                          * either do single-time area discards only, or to just
2169                          * perform discards for released swap page-clusters.
2170                          * Now it's time to adjust the p->flags accordingly.
2171                          */
2172                         if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2173                                 p->flags &= ~SWP_PAGE_DISCARD;
2174                         else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2175                                 p->flags &= ~SWP_AREA_DISCARD;
2176
2177                         /* issue a swapon-time discard if it's still required */
2178                         if (p->flags & SWP_AREA_DISCARD) {
2179                                 int err = discard_swap(p);
2180                                 if (unlikely(err))
2181                                         printk(KERN_ERR
2182                                                "swapon: discard_swap(%p): %d\n",
2183                                                 p, err);
2184                         }
2185                 }
2186         }
2187
2188         mutex_lock(&swapon_mutex);
2189         prio = -1;
2190         if (swap_flags & SWAP_FLAG_PREFER)
2191                 prio =
2192                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2193         enable_swap_info(p, prio, swap_map, frontswap_map);
2194
2195         printk(KERN_INFO "Adding %uk swap on %s.  "
2196                         "Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
2197                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2198                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2199                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2200                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2201                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
2202                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
2203                 (frontswap_map) ? "FS" : "");
2204
2205         mutex_unlock(&swapon_mutex);
2206         atomic_inc(&proc_poll_event);
2207         wake_up_interruptible(&proc_poll_wait);
2208
2209         if (S_ISREG(inode->i_mode))
2210                 inode->i_flags |= S_SWAPFILE;
2211         error = 0;
2212         goto out;
2213 bad_swap:
2214         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2215                 set_blocksize(p->bdev, p->old_block_size);
2216                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2217         }
2218         destroy_swap_extents(p);
2219         swap_cgroup_swapoff(p->type);
2220         spin_lock(&swap_lock);
2221         p->swap_file = NULL;
2222         p->flags = 0;
2223         spin_unlock(&swap_lock);
2224         vfree(swap_map);
2225         if (swap_file) {
2226                 if (inode && S_ISREG(inode->i_mode)) {
2227                         mutex_unlock(&inode->i_mutex);
2228                         inode = NULL;
2229                 }
2230                 filp_close(swap_file, NULL);
2231         }
2232 out:
2233         if (page && !IS_ERR(page)) {
2234                 kunmap(page);
2235                 page_cache_release(page);
2236         }
2237         if (name)
2238                 putname(name);
2239         if (inode && S_ISREG(inode->i_mode))
2240                 mutex_unlock(&inode->i_mutex);
2241         return error;
2242 }
2243
2244 void si_swapinfo(struct sysinfo *val)
2245 {
2246         unsigned int type;
2247         unsigned long nr_to_be_unused = 0;
2248
2249         spin_lock(&swap_lock);
2250         for (type = 0; type < nr_swapfiles; type++) {
2251                 struct swap_info_struct *si = swap_info[type];
2252
2253                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2254                         nr_to_be_unused += si->inuse_pages;
2255         }
2256         val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
2257         val->totalswap = total_swap_pages + nr_to_be_unused;
2258         spin_unlock(&swap_lock);
2259 }
2260
2261 /*
2262  * Verify that a swap entry is valid and increment its swap map count.
2263  *
2264  * Returns error code in following case.
2265  * - success -> 0
2266  * - swp_entry is invalid -> EINVAL
2267  * - swp_entry is migration entry -> EINVAL
2268  * - swap-cache reference is requested but there is already one. -> EEXIST
2269  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2270  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2271  */
2272 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2273 {
2274         struct swap_info_struct *p;
2275         unsigned long offset, type;
2276         unsigned char count;
2277         unsigned char has_cache;
2278         int err = -EINVAL;
2279
2280         if (non_swap_entry(entry))
2281                 goto out;
2282
2283         type = swp_type(entry);
2284         if (type >= nr_swapfiles)
2285                 goto bad_file;
2286         p = swap_info[type];
2287         offset = swp_offset(entry);
2288
2289         spin_lock(&p->lock);
2290         if (unlikely(offset >= p->max))
2291                 goto unlock_out;
2292
2293         count = p->swap_map[offset];
2294         has_cache = count & SWAP_HAS_CACHE;
2295         count &= ~SWAP_HAS_CACHE;
2296         err = 0;
2297
2298         if (usage == SWAP_HAS_CACHE) {
2299
2300                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2301                 if (!has_cache && count)
2302                         has_cache = SWAP_HAS_CACHE;
2303                 else if (has_cache)             /* someone else added cache */
2304                         err = -EEXIST;
2305                 else                            /* no users remaining */
2306                         err = -ENOENT;
2307
2308         } else if (count || has_cache) {
2309
2310                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2311                         count += usage;
2312                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2313                         err = -EINVAL;
2314                 else if (swap_count_continued(p, offset, count))
2315                         count = COUNT_CONTINUED;
2316                 else
2317                         err = -ENOMEM;
2318         } else
2319                 err = -ENOENT;                  /* unused swap entry */
2320
2321         p->swap_map[offset] = count | has_cache;
2322
2323 unlock_out:
2324         spin_unlock(&p->lock);
2325 out:
2326         return err;
2327
2328 bad_file:
2329         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2330         goto out;
2331 }
2332
2333 /*
2334  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2335  * (in which case its reference count is never incremented).
2336  */
2337 void swap_shmem_alloc(swp_entry_t entry)
2338 {
2339         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2340 }
2341
2342 /*
2343  * Increase reference count of swap entry by 1.
2344  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2345  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2346  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2347  * might occur if a page table entry has got corrupted.
2348  */
2349 int swap_duplicate(swp_entry_t entry)
2350 {
2351         int err = 0;
2352
2353         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2354                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2355         return err;
2356 }
2357
2358 /*
2359  * @entry: swap entry for which we allocate swap cache.
2360  *
2361  * Called when allocating swap cache for existing swap entry,
2362  * This can return error codes. Returns 0 at success.
2363  * -EBUSY means there is a swap cache.
2364  * Note: return code is different from swap_duplicate().
2365  */
2366 int swapcache_prepare(swp_entry_t entry)
2367 {
2368         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2369 }
2370
2371 struct swap_info_struct *page_swap_info(struct page *page)
2372 {
2373         swp_entry_t swap = { .val = page_private(page) };
2374         BUG_ON(!PageSwapCache(page));
2375         return swap_info[swp_type(swap)];
2376 }
2377
2378 /*
2379  * out-of-line __page_file_ methods to avoid include hell.
2380  */
2381 struct address_space *__page_file_mapping(struct page *page)
2382 {
2383         VM_BUG_ON(!PageSwapCache(page));
2384         return page_swap_info(page)->swap_file->f_mapping;
2385 }
2386 EXPORT_SYMBOL_GPL(__page_file_mapping);
2387
2388 pgoff_t __page_file_index(struct page *page)
2389 {
2390         swp_entry_t swap = { .val = page_private(page) };
2391         VM_BUG_ON(!PageSwapCache(page));
2392         return swp_offset(swap);
2393 }
2394 EXPORT_SYMBOL_GPL(__page_file_index);
2395
2396 /*
2397  * add_swap_count_continuation - called when a swap count is duplicated
2398  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2399  * page of the original vmalloc'ed swap_map, to hold the continuation count
2400  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2401  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2402  *
2403  * These continuation pages are seldom referenced: the common paths all work
2404  * on the original swap_map, only referring to a continuation page when the
2405  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2406  *
2407  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2408  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2409  * can be called after dropping locks.
2410  */
2411 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2412 {
2413         struct swap_info_struct *si;
2414         struct page *head;
2415         struct page *page;
2416         struct page *list_page;
2417         pgoff_t offset;
2418         unsigned char count;
2419
2420         /*
2421          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2422          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2423          */
2424         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2425
2426         si = swap_info_get(entry);
2427         if (!si) {
2428                 /*
2429                  * An acceptable race has occurred since the failing
2430                  * __swap_duplicate(): the swap entry has been freed,
2431                  * perhaps even the whole swap_map cleared for swapoff.
2432                  */
2433                 goto outer;
2434         }
2435
2436         offset = swp_offset(entry);
2437         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2438
2439         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2440                 /*
2441                  * The higher the swap count, the more likely it is that tasks
2442                  * will race to add swap count continuation: we need to avoid
2443                  * over-provisioning.
2444                  */
2445                 goto out;
2446         }
2447
2448         if (!page) {
2449                 spin_unlock(&si->lock);
2450                 return -ENOMEM;
2451         }
2452
2453         /*
2454          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2455          * no architecture is using highmem pages for kernel pagetables: so it
2456          * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2457          */
2458         head = vmalloc_to_page(si->swap_map + offset);
2459         offset &= ~PAGE_MASK;
2460
2461         /*
2462          * Page allocation does not initialize the page's lru field,
2463          * but it does always reset its private field.
2464          */
2465         if (!page_private(head)) {
2466                 BUG_ON(count & COUNT_CONTINUED);
2467                 INIT_LIST_HEAD(&head->lru);
2468                 set_page_private(head, SWP_CONTINUED);
2469                 si->flags |= SWP_CONTINUED;
2470         }
2471
2472         list_for_each_entry(list_page, &head->lru, lru) {
2473                 unsigned char *map;
2474
2475                 /*
2476                  * If the previous map said no continuation, but we've found
2477                  * a continuation page, free our allocation and use this one.
2478                  */
2479                 if (!(count & COUNT_CONTINUED))
2480                         goto out;
2481
2482                 map = kmap_atomic(list_page) + offset;
2483                 count = *map;
2484                 kunmap_atomic(map);
2485
2486                 /*
2487                  * If this continuation count now has some space in it,
2488                  * free our allocation and use this one.
2489                  */
2490                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2491                         goto out;
2492         }
2493
2494         list_add_tail(&page->lru, &head->lru);
2495         page = NULL;                    /* now it's attached, don't free it */
2496 out:
2497         spin_unlock(&si->lock);
2498 outer:
2499         if (page)
2500                 __free_page(page);
2501         return 0;
2502 }
2503
2504 /*
2505  * swap_count_continued - when the original swap_map count is incremented
2506  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2507  * into, carry if so, or else fail until a new continuation page is allocated;
2508  * when the original swap_map count is decremented from 0 with continuation,
2509  * borrow from the continuation and report whether it still holds more.
2510  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2511  */
2512 static bool swap_count_continued(struct swap_info_struct *si,
2513                                  pgoff_t offset, unsigned char count)
2514 {
2515         struct page *head;
2516         struct page *page;
2517         unsigned char *map;
2518
2519         head = vmalloc_to_page(si->swap_map + offset);
2520         if (page_private(head) != SWP_CONTINUED) {
2521                 BUG_ON(count & COUNT_CONTINUED);
2522                 return false;           /* need to add count continuation */
2523         }
2524
2525         offset &= ~PAGE_MASK;
2526         page = list_entry(head->lru.next, struct page, lru);
2527         map = kmap_atomic(page) + offset;
2528
2529         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2530                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2531
2532         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2533                 /*
2534                  * Think of how you add 1 to 999
2535                  */
2536                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2537                         kunmap_atomic(map);
2538                         page = list_entry(page->lru.next, struct page, lru);
2539                         BUG_ON(page == head);
2540                         map = kmap_atomic(page) + offset;
2541                 }
2542                 if (*map == SWAP_CONT_MAX) {
2543                         kunmap_atomic(map);
2544                         page = list_entry(page->lru.next, struct page, lru);
2545                         if (page == head)
2546                                 return false;   /* add count continuation */
2547                         map = kmap_atomic(page) + offset;
2548 init_map:               *map = 0;               /* we didn't zero the page */
2549                 }
2550                 *map += 1;
2551                 kunmap_atomic(map);
2552                 page = list_entry(page->lru.prev, struct page, lru);
2553                 while (page != head) {
2554                         map = kmap_atomic(page) + offset;
2555                         *map = COUNT_CONTINUED;
2556                         kunmap_atomic(map);
2557                         page = list_entry(page->lru.prev, struct page, lru);
2558                 }
2559                 return true;                    /* incremented */
2560
2561         } else {                                /* decrementing */
2562                 /*
2563                  * Think of how you subtract 1 from 1000
2564                  */
2565                 BUG_ON(count != COUNT_CONTINUED);
2566                 while (*map == COUNT_CONTINUED) {
2567                         kunmap_atomic(map);
2568                         page = list_entry(page->lru.next, struct page, lru);
2569                         BUG_ON(page == head);
2570                         map = kmap_atomic(page) + offset;
2571                 }
2572                 BUG_ON(*map == 0);
2573                 *map -= 1;
2574                 if (*map == 0)
2575                         count = 0;
2576                 kunmap_atomic(map);
2577                 page = list_entry(page->lru.prev, struct page, lru);
2578                 while (page != head) {
2579                         map = kmap_atomic(page) + offset;
2580                         *map = SWAP_CONT_MAX | count;
2581                         count = COUNT_CONTINUED;
2582                         kunmap_atomic(map);
2583                         page = list_entry(page->lru.prev, struct page, lru);
2584                 }
2585                 return count == COUNT_CONTINUED;
2586         }
2587 }
2588
2589 /*
2590  * free_swap_count_continuations - swapoff free all the continuation pages
2591  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2592  */
2593 static void free_swap_count_continuations(struct swap_info_struct *si)
2594 {
2595         pgoff_t offset;
2596
2597         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2598                 struct page *head;
2599                 head = vmalloc_to_page(si->swap_map + offset);
2600                 if (page_private(head)) {
2601                         struct list_head *this, *next;
2602                         list_for_each_safe(this, next, &head->lru) {
2603                                 struct page *page;
2604                                 page = list_entry(this, struct page, lru);
2605                                 list_del(this);
2606                                 __free_page(page);
2607                         }
2608                 }
2609         }
2610 }