mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/mman.h>
  11 #include <linux/slab.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/swap.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/namei.h>
  17 #include <linux/shm.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/random.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/mutex.h>
  29 #include <linux/capability.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/memcontrol.h>
  32
  33 #include <asm/pgtable.h>
  34 #include <asm/tlbflush.h>
  35 #include <linux/swapops.h>
  36 #include <linux/page_cgroup.h>
  37
  38 static DEFINE_SPINLOCK(swap_lock);
  39 static unsigned int nr_swapfiles;
  40 long nr_swap_pages;
  41 long total_swap_pages;
  42 static int swap_overflow;
  43 static int least_priority;
  44
  45 static const char Bad_file[] = "Bad swap file entry ";
  46 static const char Unused_file[] = "Unused swap file entry ";
  47 static const char Bad_offset[] = "Bad swap offset entry ";
  48 static const char Unused_offset[] = "Unused swap offset entry ";
  49
  50 static struct swap_list_t swap_list = {-1, -1};
  51
  52 static struct swap_info_struct *swap_info[MAX_SWAPFILES];
  53
  54 static DEFINE_MUTEX(swapon_mutex);
  55
  56 static inline int swap_count(unsigned short ent)
  57 {
  58         return ent & ~SWAP_HAS_CACHE;
  59 }
  60
  61 /* returns 1 if swap entry is freed */
  62 static int
  63 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
  64 {
  65         swp_entry_t entry = swp_entry(si->type, offset);
  66         struct page *page;
  67         int ret = 0;
  68
  69         page = find_get_page(&swapper_space, entry.val);
  70         if (!page)
  71                 return 0;
  72         /*
  73          * This function is called from scan_swap_map() and it's called
  74          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
  75          * We have to use trylock for avoiding deadlock. This is a special
  76          * case and you should use try_to_free_swap() with explicit lock_page()
  77          * in usual operations.
  78          */
  79         if (trylock_page(page)) {
  80                 ret = try_to_free_swap(page);
  81                 unlock_page(page);
  82         }
  83         page_cache_release(page);
  84         return ret;
  85 }
  86
  87 /*
  88  * We need this because the bdev->unplug_fn can sleep and we cannot
  89  * hold swap_lock while calling the unplug_fn. And swap_lock
  90  * cannot be turned into a mutex.
  91  */
  92 static DECLARE_RWSEM(swap_unplug_sem);
  93
  94 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  95 {
  96         swp_entry_t entry;
  97
  98         down_read(&swap_unplug_sem);
  99         entry.val = page_private(page);
 100         if (PageSwapCache(page)) {
 101                 struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
 102                 struct backing_dev_info *bdi;
 103
 104                 /*
 105                  * If the page is removed from swapcache from under us (with a
 106                  * racy try_to_unuse/swapoff) we need an additional reference
 107                  * count to avoid reading garbage from page_private(page) above.
 108                  * If the WARN_ON triggers during a swapoff it maybe the race
 109                  * condition and it's harmless. However if it triggers without
 110                  * swapoff it signals a problem.
 111                  */
 112                 WARN_ON(page_count(page) <= 1);
 113
 114                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
 115                 blk_run_backing_dev(bdi, page);
 116         }
 117         up_read(&swap_unplug_sem);
 118 }
 119
 120 /*
 121  * swapon tell device that all the old swap contents can be discarded,
 122  * to allow the swap device to optimize its wear-levelling.
 123  */
 124 static int discard_swap(struct swap_info_struct *si)
 125 {
 126         struct swap_extent *se;
 127         sector_t start_block;
 128         sector_t nr_blocks;
 129         int err = 0;
 130
 131         /* Do not discard the swap header page! */
 132         se = &si->first_swap_extent;
 133         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
 134         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
 135         if (nr_blocks) {
 136                 err = blkdev_issue_discard(si->bdev, start_block,
 137                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
 138                 if (err)
 139                         return err;
 140                 cond_resched();
 141         }
 142
 143         list_for_each_entry(se, &si->first_swap_extent.list, list) {
 144                 start_block = se->start_block << (PAGE_SHIFT - 9);
 145                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
 146
 147                 err = blkdev_issue_discard(si->bdev, start_block,
 148                                 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
 149                 if (err)
 150                         break;
 151
 152                 cond_resched();
 153         }
 154         return err;             /* That will often be -EOPNOTSUPP */
 155 }
 156
 157 /*
 158  * swap allocation tell device that a cluster of swap can now be discarded,
 159  * to allow the swap device to optimize its wear-levelling.
 160  */
 161 static void discard_swap_cluster(struct swap_info_struct *si,
 162                                  pgoff_t start_page, pgoff_t nr_pages)
 163 {
 164         struct swap_extent *se = si->curr_swap_extent;
 165         int found_extent = 0;
 166
 167         while (nr_pages) {
 168                 struct list_head *lh;
 169
 170                 if (se->start_page <= start_page &&
 171                     start_page < se->start_page + se->nr_pages) {
 172                         pgoff_t offset = start_page - se->start_page;
 173                         sector_t start_block = se->start_block + offset;
 174                         sector_t nr_blocks = se->nr_pages - offset;
 175
 176                         if (nr_blocks > nr_pages)
 177                                 nr_blocks = nr_pages;
 178                         start_page += nr_blocks;
 179                         nr_pages -= nr_blocks;
 180
 181                         if (!found_extent++)
 182                                 si->curr_swap_extent = se;
 183
 184                         start_block <<= PAGE_SHIFT - 9;
 185                         nr_blocks <<= PAGE_SHIFT - 9;
 186                         if (blkdev_issue_discard(si->bdev, start_block,
 187                                     nr_blocks, GFP_NOIO, DISCARD_FL_BARRIER))
 188                                 break;
 189                 }
 190
 191                 lh = se->list.next;
 192                 se = list_entry(lh, struct swap_extent, list);
 193         }
 194 }
 195
 196 static int wait_for_discard(void *word)
 197 {
 198         schedule();
 199         return 0;
 200 }
 201
 202 #define SWAPFILE_CLUSTER        256
 203 #define LATENCY_LIMIT           256
 204
 205 static inline unsigned long scan_swap_map(struct swap_info_struct *si,
 206                                           unsigned short usage)
 207 {
 208         unsigned long offset;
 209         unsigned long scan_base;
 210         unsigned long last_in_cluster = 0;
 211         int latency_ration = LATENCY_LIMIT;
 212         int found_free_cluster = 0;
 213
 214         /*
 215          * We try to cluster swap pages by allocating them sequentially
 216          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
 217          * way, however, we resort to first-free allocation, starting
 218          * a new cluster.  This prevents us from scattering swap pages
 219          * all over the entire swap partition, so that we reduce
 220          * overall disk seek times between swap pages.  -- sct
 221          * But we do now try to find an empty cluster.  -Andrea
 222          * And we let swap pages go all over an SSD partition.  Hugh
 223          */
 224
 225         si->flags += SWP_SCANNING;
 226         scan_base = offset = si->cluster_next;
 227
 228         if (unlikely(!si->cluster_nr--)) {
 229                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
 230                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
 231                         goto checks;
 232                 }
 233                 if (si->flags & SWP_DISCARDABLE) {
 234                         /*
 235                          * Start range check on racing allocations, in case
 236                          * they overlap the cluster we eventually decide on
 237                          * (we scan without swap_lock to allow preemption).
 238                          * It's hardly conceivable that cluster_nr could be
 239                          * wrapped during our scan, but don't depend on it.
 240                          */
 241                         if (si->lowest_alloc)
 242                                 goto checks;
 243                         si->lowest_alloc = si->max;
 244                         si->highest_alloc = 0;
 245                 }
 246                 spin_unlock(&swap_lock);
 247
 248                 /*
 249                  * If seek is expensive, start searching for new cluster from
 250                  * start of partition, to minimize the span of allocated swap.
 251                  * But if seek is cheap, search from our current position, so
 252                  * that swap is allocated from all over the partition: if the
 253                  * Flash Translation Layer only remaps within limited zones,
 254                  * we don't want to wear out the first zone too quickly.
 255                  */
 256                 if (!(si->flags & SWP_SOLIDSTATE))
 257                         scan_base = offset = si->lowest_bit;
 258                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 259
 260                 /* Locate the first empty (unaligned) cluster */
 261                 for (; last_in_cluster <= si->highest_bit; offset++) {
 262                         if (si->swap_map[offset])
 263                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 264                         else if (offset == last_in_cluster) {
 265                                 spin_lock(&swap_lock);
 266                                 offset -= SWAPFILE_CLUSTER - 1;
 267                                 si->cluster_next = offset;
 268                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 269                                 found_free_cluster = 1;
 270                                 goto checks;
 271                         }
 272                         if (unlikely(--latency_ration < 0)) {
 273                                 cond_resched();
 274                                 latency_ration = LATENCY_LIMIT;
 275                         }
 276                 }
 277
 278                 offset = si->lowest_bit;
 279                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 280
 281                 /* Locate the first empty (unaligned) cluster */
 282                 for (; last_in_cluster < scan_base; offset++) {
 283                         if (si->swap_map[offset])
 284                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 285                         else if (offset == last_in_cluster) {
 286                                 spin_lock(&swap_lock);
 287                                 offset -= SWAPFILE_CLUSTER - 1;
 288                                 si->cluster_next = offset;
 289                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 290                                 found_free_cluster = 1;
 291                                 goto checks;
 292                         }
 293                         if (unlikely(--latency_ration < 0)) {
 294                                 cond_resched();
 295                                 latency_ration = LATENCY_LIMIT;
 296                         }
 297                 }
 298
 299                 offset = scan_base;
 300                 spin_lock(&swap_lock);
 301                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 302                 si->lowest_alloc = 0;
 303         }
 304
 305 checks:
 306         if (!(si->flags & SWP_WRITEOK))
 307                 goto no_page;
 308         if (!si->highest_bit)
 309                 goto no_page;
 310         if (offset > si->highest_bit)
 311                 scan_base = offset = si->lowest_bit;
 312
 313         /* reuse swap entry of cache-only swap if not busy. */
 314         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 315                 int swap_was_freed;
 316                 spin_unlock(&swap_lock);
 317                 swap_was_freed = __try_to_reclaim_swap(si, offset);
 318                 spin_lock(&swap_lock);
 319                 /* entry was freed successfully, try to use this again */
 320                 if (swap_was_freed)
 321                         goto checks;
 322                 goto scan; /* check next one */
 323         }
 324
 325         if (si->swap_map[offset])
 326                 goto scan;
 327
 328         if (offset == si->lowest_bit)
 329                 si->lowest_bit++;
 330         if (offset == si->highest_bit)
 331                 si->highest_bit--;
 332         si->inuse_pages++;
 333         if (si->inuse_pages == si->pages) {
 334                 si->lowest_bit = si->max;
 335                 si->highest_bit = 0;
 336         }
 337         si->swap_map[offset] = usage;
 338         si->cluster_next = offset + 1;
 339         si->flags -= SWP_SCANNING;
 340
 341         if (si->lowest_alloc) {
 342                 /*
 343                  * Only set when SWP_DISCARDABLE, and there's a scan
 344                  * for a free cluster in progress or just completed.
 345                  */
 346                 if (found_free_cluster) {
 347                         /*
 348                          * To optimize wear-levelling, discard the
 349                          * old data of the cluster, taking care not to
 350                          * discard any of its pages that have already
 351                          * been allocated by racing tasks (offset has
 352                          * already stepped over any at the beginning).
 353                          */
 354                         if (offset < si->highest_alloc &&
 355                             si->lowest_alloc <= last_in_cluster)
 356                                 last_in_cluster = si->lowest_alloc - 1;
 357                         si->flags |= SWP_DISCARDING;
 358                         spin_unlock(&swap_lock);
 359
 360                         if (offset < last_in_cluster)
 361                                 discard_swap_cluster(si, offset,
 362                                         last_in_cluster - offset + 1);
 363
 364                         spin_lock(&swap_lock);
 365                         si->lowest_alloc = 0;
 366                         si->flags &= ~SWP_DISCARDING;
 367
 368                         smp_mb();       /* wake_up_bit advises this */
 369                         wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
 370
 371                 } else if (si->flags & SWP_DISCARDING) {
 372                         /*
 373                          * Delay using pages allocated by racing tasks
 374                          * until the whole discard has been issued. We
 375                          * could defer that delay until swap_writepage,
 376                          * but it's easier to keep this self-contained.
 377                          */
 378                         spin_unlock(&swap_lock);
 379                         wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
 380                                 wait_for_discard, TASK_UNINTERRUPTIBLE);
 381                         spin_lock(&swap_lock);
 382                 } else {
 383                         /*
 384                          * Note pages allocated by racing tasks while
 385                          * scan for a free cluster is in progress, so
 386                          * that its final discard can exclude them.
 387                          */
 388                         if (offset < si->lowest_alloc)
 389                                 si->lowest_alloc = offset;
 390                         if (offset > si->highest_alloc)
 391                                 si->highest_alloc = offset;
 392                 }
 393         }
 394         return offset;
 395
 396 scan:
 397         spin_unlock(&swap_lock);
 398         while (++offset <= si->highest_bit) {
 399                 if (!si->swap_map[offset]) {
 400                         spin_lock(&swap_lock);
 401                         goto checks;
 402                 }
 403                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 404                         spin_lock(&swap_lock);
 405                         goto checks;
 406                 }
 407                 if (unlikely(--latency_ration < 0)) {
 408                         cond_resched();
 409                         latency_ration = LATENCY_LIMIT;
 410                 }
 411         }
 412         offset = si->lowest_bit;
 413         while (++offset < scan_base) {
 414                 if (!si->swap_map[offset]) {
 415                         spin_lock(&swap_lock);
 416                         goto checks;
 417                 }
 418                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
 419                         spin_lock(&swap_lock);
 420                         goto checks;
 421                 }
 422                 if (unlikely(--latency_ration < 0)) {
 423                         cond_resched();
 424                         latency_ration = LATENCY_LIMIT;
 425                 }
 426         }
 427         spin_lock(&swap_lock);
 428
 429 no_page:
 430         si->flags -= SWP_SCANNING;
 431         return 0;
 432 }
 433
 434 swp_entry_t get_swap_page(void)
 435 {
 436         struct swap_info_struct *si;
 437         pgoff_t offset;
 438         int type, next;
 439         int wrapped = 0;
 440
 441         spin_lock(&swap_lock);
 442         if (nr_swap_pages <= 0)
 443                 goto noswap;
 444         nr_swap_pages--;
 445
 446         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 447                 si = swap_info[type];
 448                 next = si->next;
 449                 if (next < 0 ||
 450                     (!wrapped && si->prio != swap_info[next]->prio)) {
 451                         next = swap_list.head;
 452                         wrapped++;
 453                 }
 454
 455                 if (!si->highest_bit)
 456                         continue;
 457                 if (!(si->flags & SWP_WRITEOK))
 458                         continue;
 459
 460                 swap_list.next = next;
 461                 /* This is called for allocating swap entry for cache */
 462                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
 463                 if (offset) {
 464                         spin_unlock(&swap_lock);
 465                         return swp_entry(type, offset);
 466                 }
 467                 next = swap_list.next;
 468         }
 469
 470         nr_swap_pages++;
 471 noswap:
 472         spin_unlock(&swap_lock);
 473         return (swp_entry_t) {0};
 474 }
 475
 476 /* The only caller of this function is now susupend routine */
 477 swp_entry_t get_swap_page_of_type(int type)
 478 {
 479         struct swap_info_struct *si;
 480         pgoff_t offset;
 481
 482         spin_lock(&swap_lock);
 483         si = swap_info[type];
 484         if (si && (si->flags & SWP_WRITEOK)) {
 485                 nr_swap_pages--;
 486                 /* This is called for allocating swap entry, not cache */
 487                 offset = scan_swap_map(si, 1);
 488                 if (offset) {
 489                         spin_unlock(&swap_lock);
 490                         return swp_entry(type, offset);
 491                 }
 492                 nr_swap_pages++;
 493         }
 494         spin_unlock(&swap_lock);
 495         return (swp_entry_t) {0};
 496 }
 497
 498 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
 499 {
 500         struct swap_info_struct *p;
 501         unsigned long offset, type;
 502
 503         if (!entry.val)
 504                 goto out;
 505         type = swp_type(entry);
 506         if (type >= nr_swapfiles)
 507                 goto bad_nofile;
 508         p = swap_info[type];
 509         if (!(p->flags & SWP_USED))
 510                 goto bad_device;
 511         offset = swp_offset(entry);
 512         if (offset >= p->max)
 513                 goto bad_offset;
 514         if (!p->swap_map[offset])
 515                 goto bad_free;
 516         spin_lock(&swap_lock);
 517         return p;
 518
 519 bad_free:
 520         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 521         goto out;
 522 bad_offset:
 523         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 524         goto out;
 525 bad_device:
 526         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 527         goto out;
 528 bad_nofile:
 529         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 530 out:
 531         return NULL;
 532 }
 533
 534 static unsigned short swap_entry_free(struct swap_info_struct *p,
 535                            swp_entry_t entry, unsigned short usage)
 536 {
 537         unsigned long offset = swp_offset(entry);
 538         unsigned short count;
 539         unsigned short has_cache;
 540
 541         count = p->swap_map[offset];
 542         has_cache = count & SWAP_HAS_CACHE;
 543         count &= ~SWAP_HAS_CACHE;
 544
 545         if (usage == SWAP_HAS_CACHE) {
 546                 VM_BUG_ON(!has_cache);
 547                 has_cache = 0;
 548         } else if (count < SWAP_MAP_MAX)
 549                 count--;
 550
 551         if (!count)
 552                 mem_cgroup_uncharge_swap(entry);
 553
 554         usage = count | has_cache;
 555         p->swap_map[offset] = usage;
 556
 557         /* free if no reference */
 558         if (!usage) {
 559                 if (offset < p->lowest_bit)
 560                         p->lowest_bit = offset;
 561                 if (offset > p->highest_bit)
 562                         p->highest_bit = offset;
 563                 if (swap_list.next >= 0 &&
 564                     p->prio > swap_info[swap_list.next]->prio)
 565                         swap_list.next = p->type;
 566                 nr_swap_pages++;
 567                 p->inuse_pages--;
 568         }
 569
 570         return usage;
 571 }
 572
 573 /*
 574  * Caller has made sure that the swapdevice corresponding to entry
 575  * is still around or has not been recycled.
 576  */
 577 void swap_free(swp_entry_t entry)
 578 {
 579         struct swap_info_struct *p;
 580
 581         p = swap_info_get(entry);
 582         if (p) {
 583                 swap_entry_free(p, entry, 1);
 584                 spin_unlock(&swap_lock);
 585         }
 586 }
 587
 588 /*
 589  * Called after dropping swapcache to decrease refcnt to swap entries.
 590  */
 591 void swapcache_free(swp_entry_t entry, struct page *page)
 592 {
 593         struct swap_info_struct *p;
 594         unsigned short count;
 595
 596         p = swap_info_get(entry);
 597         if (p) {
 598                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
 599                 if (page)
 600                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
 601                 spin_unlock(&swap_lock);
 602         }
 603 }
 604
 605 /*
 606  * How many references to page are currently swapped out?
 607  */
 608 static inline int page_swapcount(struct page *page)
 609 {
 610         int count = 0;
 611         struct swap_info_struct *p;
 612         swp_entry_t entry;
 613
 614         entry.val = page_private(page);
 615         p = swap_info_get(entry);
 616         if (p) {
 617                 count = swap_count(p->swap_map[swp_offset(entry)]);
 618                 spin_unlock(&swap_lock);
 619         }
 620         return count;
 621 }
 622
 623 /*
 624  * We can write to an anon page without COW if there are no other references
 625  * to it.  And as a side-effect, free up its swap: because the old content
 626  * on disk will never be read, and seeking back there to write new content
 627  * later would only waste time away from clustering.
 628  */
 629 int reuse_swap_page(struct page *page)
 630 {
 631         int count;
 632
 633         VM_BUG_ON(!PageLocked(page));
 634         count = page_mapcount(page);
 635         if (count <= 1 && PageSwapCache(page)) {
 636                 count += page_swapcount(page);
 637                 if (count == 1 && !PageWriteback(page)) {
 638                         delete_from_swap_cache(page);
 639                         SetPageDirty(page);
 640                 }
 641         }
 642         return count == 1;
 643 }
 644
 645 /*
 646  * If swap is getting full, or if there are no more mappings of this page,
 647  * then try_to_free_swap is called to free its swap space.
 648  */
 649 int try_to_free_swap(struct page *page)
 650 {
 651         VM_BUG_ON(!PageLocked(page));
 652
 653         if (!PageSwapCache(page))
 654                 return 0;
 655         if (PageWriteback(page))
 656                 return 0;
 657         if (page_swapcount(page))
 658                 return 0;
 659
 660         delete_from_swap_cache(page);
 661         SetPageDirty(page);
 662         return 1;
 663 }
 664
 665 /*
 666  * Free the swap entry like above, but also try to
 667  * free the page cache entry if it is the last user.
 668  */
 669 int free_swap_and_cache(swp_entry_t entry)
 670 {
 671         struct swap_info_struct *p;
 672         struct page *page = NULL;
 673
 674         if (non_swap_entry(entry))
 675                 return 1;
 676
 677         p = swap_info_get(entry);
 678         if (p) {
 679                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
 680                         page = find_get_page(&swapper_space, entry.val);
 681                         if (page && !trylock_page(page)) {
 682                                 page_cache_release(page);
 683                                 page = NULL;
 684                         }
 685                 }
 686                 spin_unlock(&swap_lock);
 687         }
 688         if (page) {
 689                 /*
 690                  * Not mapped elsewhere, or swap space full? Free it!
 691                  * Also recheck PageSwapCache now page is locked (above).
 692                  */
 693                 if (PageSwapCache(page) && !PageWriteback(page) &&
 694                                 (!page_mapped(page) || vm_swap_full())) {
 695                         delete_from_swap_cache(page);
 696                         SetPageDirty(page);
 697                 }
 698                 unlock_page(page);
 699                 page_cache_release(page);
 700         }
 701         return p != NULL;
 702 }
 703
 704 #ifdef CONFIG_HIBERNATION
 705 /*
 706  * Find the swap type that corresponds to given device (if any).
 707  *
 708  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 709  * from 0, in which the swap header is expected to be located.
 710  *
 711  * This is needed for the suspend to disk (aka swsusp).
 712  */
 713 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 714 {
 715         struct block_device *bdev = NULL;
 716         int type;
 717
 718         if (device)
 719                 bdev = bdget(device);
 720
 721         spin_lock(&swap_lock);
 722         for (type = 0; type < nr_swapfiles; type++) {
 723                 struct swap_info_struct *sis = swap_info[type];
 724
 725                 if (!(sis->flags & SWP_WRITEOK))
 726                         continue;
 727
 728                 if (!bdev) {
 729                         if (bdev_p)
 730                                 *bdev_p = bdgrab(sis->bdev);
 731
 732                         spin_unlock(&swap_lock);
 733                         return type;
 734                 }
 735                 if (bdev == sis->bdev) {
 736                         struct swap_extent *se = &sis->first_swap_extent;
 737
 738                         if (se->start_block == offset) {
 739                                 if (bdev_p)
 740                                         *bdev_p = bdgrab(sis->bdev);
 741
 742                                 spin_unlock(&swap_lock);
 743                                 bdput(bdev);
 744                                 return type;
 745                         }
 746                 }
 747         }
 748         spin_unlock(&swap_lock);
 749         if (bdev)
 750                 bdput(bdev);
 751
 752         return -ENODEV;
 753 }
 754
 755 /*
 756  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 757  * corresponding to given index in swap_info (swap type).
 758  */
 759 sector_t swapdev_block(int type, pgoff_t offset)
 760 {
 761         struct block_device *bdev;
 762
 763         if ((unsigned int)type >= nr_swapfiles)
 764                 return 0;
 765         if (!(swap_info[type]->flags & SWP_WRITEOK))
 766                 return 0;
 767         return map_swap_page(swp_entry(type, offset), &bdev);
 768 }
 769
 770 /*
 771  * Return either the total number of swap pages of given type, or the number
 772  * of free pages of that type (depending on @free)
 773  *
 774  * This is needed for software suspend
 775  */
 776 unsigned int count_swap_pages(int type, int free)
 777 {
 778         unsigned int n = 0;
 779
 780         spin_lock(&swap_lock);
 781         if ((unsigned int)type < nr_swapfiles) {
 782                 struct swap_info_struct *sis = swap_info[type];
 783
 784                 if (sis->flags & SWP_WRITEOK) {
 785                         n = sis->pages;
 786                         if (free)
 787                                 n -= sis->inuse_pages;
 788                 }
 789         }
 790         spin_unlock(&swap_lock);
 791         return n;
 792 }
 793 #endif /* CONFIG_HIBERNATION */
 794
 795 /*
 796  * No need to decide whether this PTE shares the swap entry with others,
 797  * just let do_wp_page work it out if a write is requested later - to
 798  * force COW, vm_page_prot omits write permission from any private vma.
 799  */
 800 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 801                 unsigned long addr, swp_entry_t entry, struct page *page)
 802 {
 803         struct mem_cgroup *ptr = NULL;
 804         spinlock_t *ptl;
 805         pte_t *pte;
 806         int ret = 1;
 807
 808         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
 809                 ret = -ENOMEM;
 810                 goto out_nolock;
 811         }
 812
 813         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 814         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 815                 if (ret > 0)
 816                         mem_cgroup_cancel_charge_swapin(ptr);
 817                 ret = 0;
 818                 goto out;
 819         }
 820
 821         inc_mm_counter(vma->vm_mm, anon_rss);
 822         get_page(page);
 823         set_pte_at(vma->vm_mm, addr, pte,
 824                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 825         page_add_anon_rmap(page, vma, addr);
 826         mem_cgroup_commit_charge_swapin(page, ptr);
 827         swap_free(entry);
 828         /*
 829          * Move the page to the active list so it is not
 830          * immediately swapped out again after swapon.
 831          */
 832         activate_page(page);
 833 out:
 834         pte_unmap_unlock(pte, ptl);
 835 out_nolock:
 836         return ret;
 837 }
 838
 839 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 840                                 unsigned long addr, unsigned long end,
 841                                 swp_entry_t entry, struct page *page)
 842 {
 843         pte_t swp_pte = swp_entry_to_pte(entry);
 844         pte_t *pte;
 845         int ret = 0;
 846
 847         /*
 848          * We don't actually need pte lock while scanning for swp_pte: since
 849          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 850          * page table while we're scanning; though it could get zapped, and on
 851          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 852          * of unmatched parts which look like swp_pte, so unuse_pte must
 853          * recheck under pte lock.  Scanning without pte lock lets it be
 854          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 855          */
 856         pte = pte_offset_map(pmd, addr);
 857         do {
 858                 /*
 859                  * swapoff spends a _lot_ of time in this loop!
 860                  * Test inline before going to call unuse_pte.
 861                  */
 862                 if (unlikely(pte_same(*pte, swp_pte))) {
 863                         pte_unmap(pte);
 864                         ret = unuse_pte(vma, pmd, addr, entry, page);
 865                         if (ret)
 866                                 goto out;
 867                         pte = pte_offset_map(pmd, addr);
 868                 }
 869         } while (pte++, addr += PAGE_SIZE, addr != end);
 870         pte_unmap(pte - 1);
 871 out:
 872         return ret;
 873 }
 874
 875 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 876                                 unsigned long addr, unsigned long end,
 877                                 swp_entry_t entry, struct page *page)
 878 {
 879         pmd_t *pmd;
 880         unsigned long next;
 881         int ret;
 882
 883         pmd = pmd_offset(pud, addr);
 884         do {
 885                 next = pmd_addr_end(addr, end);
 886                 if (pmd_none_or_clear_bad(pmd))
 887                         continue;
 888                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 889                 if (ret)
 890                         return ret;
 891         } while (pmd++, addr = next, addr != end);
 892         return 0;
 893 }
 894
 895 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 896                                 unsigned long addr, unsigned long end,
 897                                 swp_entry_t entry, struct page *page)
 898 {
 899         pud_t *pud;
 900         unsigned long next;
 901         int ret;
 902
 903         pud = pud_offset(pgd, addr);
 904         do {
 905                 next = pud_addr_end(addr, end);
 906                 if (pud_none_or_clear_bad(pud))
 907                         continue;
 908                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 909                 if (ret)
 910                         return ret;
 911         } while (pud++, addr = next, addr != end);
 912         return 0;
 913 }
 914
 915 static int unuse_vma(struct vm_area_struct *vma,
 916                                 swp_entry_t entry, struct page *page)
 917 {
 918         pgd_t *pgd;
 919         unsigned long addr, end, next;
 920         int ret;
 921
 922         if (page->mapping) {
 923                 addr = page_address_in_vma(page, vma);
 924                 if (addr == -EFAULT)
 925                         return 0;
 926                 else
 927                         end = addr + PAGE_SIZE;
 928         } else {
 929                 addr = vma->vm_start;
 930                 end = vma->vm_end;
 931         }
 932
 933         pgd = pgd_offset(vma->vm_mm, addr);
 934         do {
 935                 next = pgd_addr_end(addr, end);
 936                 if (pgd_none_or_clear_bad(pgd))
 937                         continue;
 938                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 939                 if (ret)
 940                         return ret;
 941         } while (pgd++, addr = next, addr != end);
 942         return 0;
 943 }
 944
 945 static int unuse_mm(struct mm_struct *mm,
 946                                 swp_entry_t entry, struct page *page)
 947 {
 948         struct vm_area_struct *vma;
 949         int ret = 0;
 950
 951         if (!down_read_trylock(&mm->mmap_sem)) {
 952                 /*
 953                  * Activate page so shrink_inactive_list is unlikely to unmap
 954                  * its ptes while lock is dropped, so swapoff can make progress.
 955                  */
 956                 activate_page(page);
 957                 unlock_page(page);
 958                 down_read(&mm->mmap_sem);
 959                 lock_page(page);
 960         }
 961         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 962                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
 963                         break;
 964         }
 965         up_read(&mm->mmap_sem);
 966         return (ret < 0)? ret: 0;
 967 }
 968
 969 /*
 970  * Scan swap_map from current position to next entry still in use.
 971  * Recycle to start on reaching the end, returning 0 when empty.
 972  */
 973 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 974                                         unsigned int prev)
 975 {
 976         unsigned int max = si->max;
 977         unsigned int i = prev;
 978         int count;
 979
 980         /*
 981          * No need for swap_lock here: we're just looking
 982          * for whether an entry is in use, not modifying it; false
 983          * hits are okay, and sys_swapoff() has already prevented new
 984          * allocations from this area (while holding swap_lock).
 985          */
 986         for (;;) {
 987                 if (++i >= max) {
 988                         if (!prev) {
 989                                 i = 0;
 990                                 break;
 991                         }
 992                         /*
 993                          * No entries in use at top of swap_map,
 994                          * loop back to start and recheck there.
 995                          */
 996                         max = prev + 1;
 997                         prev = 0;
 998                         i = 1;
 999                 }
1000                 count = si->swap_map[i];
1001                 if (count && swap_count(count) != SWAP_MAP_BAD)
1002                         break;
1003         }
1004         return i;
1005 }
1006
1007 /*
1008  * We completely avoid races by reading each swap page in advance,
1009  * and then search for the process using it.  All the necessary
1010  * page table adjustments can then be made atomically.
1011  */
1012 static int try_to_unuse(unsigned int type)
1013 {
1014         struct swap_info_struct *si = swap_info[type];
1015         struct mm_struct *start_mm;
1016         unsigned short *swap_map;
1017         unsigned short swcount;
1018         struct page *page;
1019         swp_entry_t entry;
1020         unsigned int i = 0;
1021         int retval = 0;
1022         int reset_overflow = 0;
1023         int shmem;
1024
1025         /*
1026          * When searching mms for an entry, a good strategy is to
1027          * start at the first mm we freed the previous entry from
1028          * (though actually we don't notice whether we or coincidence
1029          * freed the entry).  Initialize this start_mm with a hold.
1030          *
1031          * A simpler strategy would be to start at the last mm we
1032          * freed the previous entry from; but that would take less
1033          * advantage of mmlist ordering, which clusters forked mms
1034          * together, child after parent.  If we race with dup_mmap(), we
1035          * prefer to resolve parent before child, lest we miss entries
1036          * duplicated after we scanned child: using last mm would invert
1037          * that.  Though it's only a serious concern when an overflowed
1038          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
1039          */
1040         start_mm = &init_mm;
1041         atomic_inc(&init_mm.mm_users);
1042
1043         /*
1044          * Keep on scanning until all entries have gone.  Usually,
1045          * one pass through swap_map is enough, but not necessarily:
1046          * there are races when an instance of an entry might be missed.
1047          */
1048         while ((i = find_next_to_unuse(si, i)) != 0) {
1049                 if (signal_pending(current)) {
1050                         retval = -EINTR;
1051                         break;
1052                 }
1053
1054                 /*
1055                  * Get a page for the entry, using the existing swap
1056                  * cache page if there is one.  Otherwise, get a clean
1057                  * page and read the swap into it.
1058                  */
1059                 swap_map = &si->swap_map[i];
1060                 entry = swp_entry(type, i);
1061                 page = read_swap_cache_async(entry,
1062                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1063                 if (!page) {
1064                         /*
1065                          * Either swap_duplicate() failed because entry
1066                          * has been freed independently, and will not be
1067                          * reused since sys_swapoff() already disabled
1068                          * allocation from here, or alloc_page() failed.
1069                          */
1070                         if (!*swap_map)
1071                                 continue;
1072                         retval = -ENOMEM;
1073                         break;
1074                 }
1075
1076                 /*
1077                  * Don't hold on to start_mm if it looks like exiting.
1078                  */
1079                 if (atomic_read(&start_mm->mm_users) == 1) {
1080                         mmput(start_mm);
1081                         start_mm = &init_mm;
1082                         atomic_inc(&init_mm.mm_users);
1083                 }
1084
1085                 /*
1086                  * Wait for and lock page.  When do_swap_page races with
1087                  * try_to_unuse, do_swap_page can handle the fault much
1088                  * faster than try_to_unuse can locate the entry.  This
1089                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1090                  * defer to do_swap_page in such a case - in some tests,
1091                  * do_swap_page and try_to_unuse repeatedly compete.
1092                  */
1093                 wait_on_page_locked(page);
1094                 wait_on_page_writeback(page);
1095                 lock_page(page);
1096                 wait_on_page_writeback(page);
1097
1098                 /*
1099                  * Remove all references to entry.
1100                  * Whenever we reach init_mm, there's no address space
1101                  * to search, but use it as a reminder to search shmem.
1102                  */
1103                 shmem = 0;
1104                 swcount = *swap_map;
1105                 if (swap_count(swcount)) {
1106                         if (start_mm == &init_mm)
1107                                 shmem = shmem_unuse(entry, page);
1108                         else
1109                                 retval = unuse_mm(start_mm, entry, page);
1110                 }
1111                 if (swap_count(*swap_map)) {
1112                         int set_start_mm = (*swap_map >= swcount);
1113                         struct list_head *p = &start_mm->mmlist;
1114                         struct mm_struct *new_start_mm = start_mm;
1115                         struct mm_struct *prev_mm = start_mm;
1116                         struct mm_struct *mm;
1117
1118                         atomic_inc(&new_start_mm->mm_users);
1119                         atomic_inc(&prev_mm->mm_users);
1120                         spin_lock(&mmlist_lock);
1121                         while (swap_count(*swap_map) && !retval && !shmem &&
1122                                         (p = p->next) != &start_mm->mmlist) {
1123                                 mm = list_entry(p, struct mm_struct, mmlist);
1124                                 if (!atomic_inc_not_zero(&mm->mm_users))
1125                                         continue;
1126                                 spin_unlock(&mmlist_lock);
1127                                 mmput(prev_mm);
1128                                 prev_mm = mm;
1129
1130                                 cond_resched();
1131
1132                                 swcount = *swap_map;
1133                                 if (!swap_count(swcount)) /* any usage ? */
1134                                         ;
1135                                 else if (mm == &init_mm) {
1136                                         set_start_mm = 1;
1137                                         shmem = shmem_unuse(entry, page);
1138                                 } else
1139                                         retval = unuse_mm(mm, entry, page);
1140
1141                                 if (set_start_mm && *swap_map < swcount) {
1142                                         mmput(new_start_mm);
1143                                         atomic_inc(&mm->mm_users);
1144                                         new_start_mm = mm;
1145                                         set_start_mm = 0;
1146                                 }
1147                                 spin_lock(&mmlist_lock);
1148                         }
1149                         spin_unlock(&mmlist_lock);
1150                         mmput(prev_mm);
1151                         mmput(start_mm);
1152                         start_mm = new_start_mm;
1153                 }
1154                 if (shmem) {
1155                         /* page has already been unlocked and released */
1156                         if (shmem > 0)
1157                                 continue;
1158                         retval = shmem;
1159                         break;
1160                 }
1161                 if (retval) {
1162                         unlock_page(page);
1163                         page_cache_release(page);
1164                         break;
1165                 }
1166
1167                 /*
1168                  * How could swap count reach 0x7ffe ?
1169                  * There's no way to repeat a swap page within an mm
1170                  * (except in shmem, where it's the shared object which takes
1171                  * the reference count)?
1172                  * We believe SWAP_MAP_MAX cannot occur.(if occur, unsigned
1173                  * short is too small....)
1174                  * If that's wrong, then we should worry more about
1175                  * exit_mmap() and do_munmap() cases described above:
1176                  * we might be resetting SWAP_MAP_MAX too early here.
1177                  * We know "Undead"s can happen, they're okay, so don't
1178                  * report them; but do report if we reset SWAP_MAP_MAX.
1179                  */
1180                 /* We might release the lock_page() in unuse_mm(). */
1181                 if (!PageSwapCache(page) || page_private(page) != entry.val)
1182                         goto retry;
1183
1184                 if (swap_count(*swap_map) == SWAP_MAP_MAX) {
1185                         spin_lock(&swap_lock);
1186                         *swap_map = SWAP_HAS_CACHE;
1187                         spin_unlock(&swap_lock);
1188                         reset_overflow = 1;
1189                 }
1190
1191                 /*
1192                  * If a reference remains (rare), we would like to leave
1193                  * the page in the swap cache; but try_to_unmap could
1194                  * then re-duplicate the entry once we drop page lock,
1195                  * so we might loop indefinitely; also, that page could
1196                  * not be swapped out to other storage meanwhile.  So:
1197                  * delete from cache even if there's another reference,
1198                  * after ensuring that the data has been saved to disk -
1199                  * since if the reference remains (rarer), it will be
1200                  * read from disk into another page.  Splitting into two
1201                  * pages would be incorrect if swap supported "shared
1202                  * private" pages, but they are handled by tmpfs files.
1203                  */
1204                 if (swap_count(*swap_map) &&
1205                      PageDirty(page) && PageSwapCache(page)) {
1206                         struct writeback_control wbc = {
1207                                 .sync_mode = WB_SYNC_NONE,
1208                         };
1209
1210                         swap_writepage(page, &wbc);
1211                         lock_page(page);
1212                         wait_on_page_writeback(page);
1213                 }
1214
1215                 /*
1216                  * It is conceivable that a racing task removed this page from
1217                  * swap cache just before we acquired the page lock at the top,
1218                  * or while we dropped it in unuse_mm().  The page might even
1219                  * be back in swap cache on another swap area: that we must not
1220                  * delete, since it may not have been written out to swap yet.
1221                  */
1222                 if (PageSwapCache(page) &&
1223                     likely(page_private(page) == entry.val))
1224                         delete_from_swap_cache(page);
1225
1226                 /*
1227                  * So we could skip searching mms once swap count went
1228                  * to 1, we did not mark any present ptes as dirty: must
1229                  * mark page dirty so shrink_page_list will preserve it.
1230                  */
1231                 SetPageDirty(page);
1232 retry:
1233                 unlock_page(page);
1234                 page_cache_release(page);
1235
1236                 /*
1237                  * Make sure that we aren't completely killing
1238                  * interactive performance.
1239                  */
1240                 cond_resched();
1241         }
1242
1243         mmput(start_mm);
1244         if (reset_overflow) {
1245                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
1246                 swap_overflow = 0;
1247         }
1248         return retval;
1249 }
1250
1251 /*
1252  * After a successful try_to_unuse, if no swap is now in use, we know
1253  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1254  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1255  * added to the mmlist just after page_duplicate - before would be racy.
1256  */
1257 static void drain_mmlist(void)
1258 {
1259         struct list_head *p, *next;
1260         unsigned int type;
1261
1262         for (type = 0; type < nr_swapfiles; type++)
1263                 if (swap_info[type]->inuse_pages)
1264                         return;
1265         spin_lock(&mmlist_lock);
1266         list_for_each_safe(p, next, &init_mm.mmlist)
1267                 list_del_init(p);
1268         spin_unlock(&mmlist_lock);
1269 }
1270
1271 /*
1272  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1273  * corresponds to page offset `offset'.  Note that the type of this function
1274  * is sector_t, but it returns page offset into the bdev, not sector offset.
1275  */
1276 sector_t map_swap_page(swp_entry_t entry, struct block_device **bdev)
1277 {
1278         struct swap_info_struct *sis;
1279         struct swap_extent *start_se;
1280         struct swap_extent *se;
1281         pgoff_t offset;
1282
1283         sis = swap_info[swp_type(entry)];
1284         *bdev = sis->bdev;
1285
1286         offset = swp_offset(entry);
1287         start_se = sis->curr_swap_extent;
1288         se = start_se;
1289
1290         for ( ; ; ) {
1291                 struct list_head *lh;
1292
1293                 if (se->start_page <= offset &&
1294                                 offset < (se->start_page + se->nr_pages)) {
1295                         return se->start_block + (offset - se->start_page);
1296                 }
1297                 lh = se->list.next;
1298                 se = list_entry(lh, struct swap_extent, list);
1299                 sis->curr_swap_extent = se;
1300                 BUG_ON(se == start_se);         /* It *must* be present */
1301         }
1302 }
1303
1304 /*
1305  * Free all of a swapdev's extent information
1306  */
1307 static void destroy_swap_extents(struct swap_info_struct *sis)
1308 {
1309         while (!list_empty(&sis->first_swap_extent.list)) {
1310                 struct swap_extent *se;
1311
1312                 se = list_entry(sis->first_swap_extent.list.next,
1313                                 struct swap_extent, list);
1314                 list_del(&se->list);
1315                 kfree(se);
1316         }
1317 }
1318
1319 /*
1320  * Add a block range (and the corresponding page range) into this swapdev's
1321  * extent list.  The extent list is kept sorted in page order.
1322  *
1323  * This function rather assumes that it is called in ascending page order.
1324  */
1325 static int
1326 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1327                 unsigned long nr_pages, sector_t start_block)
1328 {
1329         struct swap_extent *se;
1330         struct swap_extent *new_se;
1331         struct list_head *lh;
1332
1333         if (start_page == 0) {
1334                 se = &sis->first_swap_extent;
1335                 sis->curr_swap_extent = se;
1336                 se->start_page = 0;
1337                 se->nr_pages = nr_pages;
1338                 se->start_block = start_block;
1339                 return 1;
1340         } else {
1341                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1342                 se = list_entry(lh, struct swap_extent, list);
1343                 BUG_ON(se->start_page + se->nr_pages != start_page);
1344                 if (se->start_block + se->nr_pages == start_block) {
1345                         /* Merge it */
1346                         se->nr_pages += nr_pages;
1347                         return 0;
1348                 }
1349         }
1350
1351         /*
1352          * No merge.  Insert a new extent, preserving ordering.
1353          */
1354         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1355         if (new_se == NULL)
1356                 return -ENOMEM;
1357         new_se->start_page = start_page;
1358         new_se->nr_pages = nr_pages;
1359         new_se->start_block = start_block;
1360
1361         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1362         return 1;
1363 }
1364
1365 /*
1366  * A `swap extent' is a simple thing which maps a contiguous range of pages
1367  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1368  * is built at swapon time and is then used at swap_writepage/swap_readpage
1369  * time for locating where on disk a page belongs.
1370  *
1371  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1372  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1373  * swap files identically.
1374  *
1375  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1376  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1377  * swapfiles are handled *identically* after swapon time.
1378  *
1379  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1380  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1381  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1382  * requirements, they are simply tossed out - we will never use those blocks
1383  * for swapping.
1384  *
1385  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1386  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1387  * which will scribble on the fs.
1388  *
1389  * The amount of disk space which a single swap extent represents varies.
1390  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1391  * extents in the list.  To avoid much list walking, we cache the previous
1392  * search location in `curr_swap_extent', and start new searches from there.
1393  * This is extremely effective.  The average number of iterations in
1394  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1395  */
1396 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1397 {
1398         struct inode *inode;
1399         unsigned blocks_per_page;
1400         unsigned long page_no;
1401         unsigned blkbits;
1402         sector_t probe_block;
1403         sector_t last_block;
1404         sector_t lowest_block = -1;
1405         sector_t highest_block = 0;
1406         int nr_extents = 0;
1407         int ret;
1408
1409         inode = sis->swap_file->f_mapping->host;
1410         if (S_ISBLK(inode->i_mode)) {
1411                 ret = add_swap_extent(sis, 0, sis->max, 0);
1412                 *span = sis->pages;
1413                 goto out;
1414         }
1415
1416         blkbits = inode->i_blkbits;
1417         blocks_per_page = PAGE_SIZE >> blkbits;
1418
1419         /*
1420          * Map all the blocks into the extent list.  This code doesn't try
1421          * to be very smart.
1422          */
1423         probe_block = 0;
1424         page_no = 0;
1425         last_block = i_size_read(inode) >> blkbits;
1426         while ((probe_block + blocks_per_page) <= last_block &&
1427                         page_no < sis->max) {
1428                 unsigned block_in_page;
1429                 sector_t first_block;
1430
1431                 first_block = bmap(inode, probe_block);
1432                 if (first_block == 0)
1433                         goto bad_bmap;
1434
1435                 /*
1436                  * It must be PAGE_SIZE aligned on-disk
1437                  */
1438                 if (first_block & (blocks_per_page - 1)) {
1439                         probe_block++;
1440                         goto reprobe;
1441                 }
1442
1443                 for (block_in_page = 1; block_in_page < blocks_per_page;
1444                                         block_in_page++) {
1445                         sector_t block;
1446
1447                         block = bmap(inode, probe_block + block_in_page);
1448                         if (block == 0)
1449                                 goto bad_bmap;
1450                         if (block != first_block + block_in_page) {
1451                                 /* Discontiguity */
1452                                 probe_block++;
1453                                 goto reprobe;
1454                         }
1455                 }
1456
1457                 first_block >>= (PAGE_SHIFT - blkbits);
1458                 if (page_no) {  /* exclude the header page */
1459                         if (first_block < lowest_block)
1460                                 lowest_block = first_block;
1461                         if (first_block > highest_block)
1462                                 highest_block = first_block;
1463                 }
1464
1465                 /*
1466                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1467                  */
1468                 ret = add_swap_extent(sis, page_no, 1, first_block);
1469                 if (ret < 0)
1470                         goto out;
1471                 nr_extents += ret;
1472                 page_no++;
1473                 probe_block += blocks_per_page;
1474 reprobe:
1475                 continue;
1476         }
1477         ret = nr_extents;
1478         *span = 1 + highest_block - lowest_block;
1479         if (page_no == 0)
1480                 page_no = 1;    /* force Empty message */
1481         sis->max = page_no;
1482         sis->pages = page_no - 1;
1483         sis->highest_bit = page_no - 1;
1484 out:
1485         return ret;
1486 bad_bmap:
1487         printk(KERN_ERR "swapon: swapfile has holes\n");
1488         ret = -EINVAL;
1489         goto out;
1490 }
1491
1492 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1493 {
1494         struct swap_info_struct *p = NULL;
1495         unsigned short *swap_map;
1496         struct file *swap_file, *victim;
1497         struct address_space *mapping;
1498         struct inode *inode;
1499         char *pathname;
1500         int i, type, prev;
1501         int err;
1502
1503         if (!capable(CAP_SYS_ADMIN))
1504                 return -EPERM;
1505
1506         pathname = getname(specialfile);
1507         err = PTR_ERR(pathname);
1508         if (IS_ERR(pathname))
1509                 goto out;
1510
1511         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1512         putname(pathname);
1513         err = PTR_ERR(victim);
1514         if (IS_ERR(victim))
1515                 goto out;
1516
1517         mapping = victim->f_mapping;
1518         prev = -1;
1519         spin_lock(&swap_lock);
1520         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1521                 p = swap_info[type];
1522                 if (p->flags & SWP_WRITEOK) {
1523                         if (p->swap_file->f_mapping == mapping)
1524                                 break;
1525                 }
1526                 prev = type;
1527         }
1528         if (type < 0) {
1529                 err = -EINVAL;
1530                 spin_unlock(&swap_lock);
1531                 goto out_dput;
1532         }
1533         if (!security_vm_enough_memory(p->pages))
1534                 vm_unacct_memory(p->pages);
1535         else {
1536                 err = -ENOMEM;
1537                 spin_unlock(&swap_lock);
1538                 goto out_dput;
1539         }
1540         if (prev < 0)
1541                 swap_list.head = p->next;
1542         else
1543                 swap_info[prev]->next = p->next;
1544         if (type == swap_list.next) {
1545                 /* just pick something that's safe... */
1546                 swap_list.next = swap_list.head;
1547         }
1548         if (p->prio < 0) {
1549                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1550                         swap_info[i]->prio = p->prio--;
1551                 least_priority++;
1552         }
1553         nr_swap_pages -= p->pages;
1554         total_swap_pages -= p->pages;
1555         p->flags &= ~SWP_WRITEOK;
1556         spin_unlock(&swap_lock);
1557
1558         current->flags |= PF_OOM_ORIGIN;
1559         err = try_to_unuse(type);
1560         current->flags &= ~PF_OOM_ORIGIN;
1561
1562         if (err) {
1563                 /* re-insert swap space back into swap_list */
1564                 spin_lock(&swap_lock);
1565                 if (p->prio < 0)
1566                         p->prio = --least_priority;
1567                 prev = -1;
1568                 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1569                         if (p->prio >= swap_info[i]->prio)
1570                                 break;
1571                         prev = i;
1572                 }
1573                 p->next = i;
1574                 if (prev < 0)
1575                         swap_list.head = swap_list.next = type;
1576                 else
1577                         swap_info[prev]->next = type;
1578                 nr_swap_pages += p->pages;
1579                 total_swap_pages += p->pages;
1580                 p->flags |= SWP_WRITEOK;
1581                 spin_unlock(&swap_lock);
1582                 goto out_dput;
1583         }
1584
1585         /* wait for any unplug function to finish */
1586         down_write(&swap_unplug_sem);
1587         up_write(&swap_unplug_sem);
1588
1589         destroy_swap_extents(p);
1590         mutex_lock(&swapon_mutex);
1591         spin_lock(&swap_lock);
1592         drain_mmlist();
1593
1594         /* wait for anyone still in scan_swap_map */
1595         p->highest_bit = 0;             /* cuts scans short */
1596         while (p->flags >= SWP_SCANNING) {
1597                 spin_unlock(&swap_lock);
1598                 schedule_timeout_uninterruptible(1);
1599                 spin_lock(&swap_lock);
1600         }
1601
1602         swap_file = p->swap_file;
1603         p->swap_file = NULL;
1604         p->max = 0;
1605         swap_map = p->swap_map;
1606         p->swap_map = NULL;
1607         p->flags = 0;
1608         spin_unlock(&swap_lock);
1609         mutex_unlock(&swapon_mutex);
1610         vfree(swap_map);
1611         /* Destroy swap account informatin */
1612         swap_cgroup_swapoff(type);
1613
1614         inode = mapping->host;
1615         if (S_ISBLK(inode->i_mode)) {
1616                 struct block_device *bdev = I_BDEV(inode);
1617                 set_blocksize(bdev, p->old_block_size);
1618                 bd_release(bdev);
1619         } else {
1620                 mutex_lock(&inode->i_mutex);
1621                 inode->i_flags &= ~S_SWAPFILE;
1622                 mutex_unlock(&inode->i_mutex);
1623         }
1624         filp_close(swap_file, NULL);
1625         err = 0;
1626
1627 out_dput:
1628         filp_close(victim, NULL);
1629 out:
1630         return err;
1631 }
1632
1633 #ifdef CONFIG_PROC_FS
1634 /* iterator */
1635 static void *swap_start(struct seq_file *swap, loff_t *pos)
1636 {
1637         struct swap_info_struct *si;
1638         int type;
1639         loff_t l = *pos;
1640
1641         mutex_lock(&swapon_mutex);
1642
1643         if (!l)
1644                 return SEQ_START_TOKEN;
1645
1646         for (type = 0; type < nr_swapfiles; type++) {
1647                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1648                 si = swap_info[type];
1649                 if (!(si->flags & SWP_USED) || !si->swap_map)
1650                         continue;
1651                 if (!--l)
1652                         return si;
1653         }
1654
1655         return NULL;
1656 }
1657
1658 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1659 {
1660         struct swap_info_struct *si = v;
1661         int type;
1662
1663         if (v == SEQ_START_TOKEN)
1664                 type = 0;
1665         else
1666                 type = si->type + 1;
1667
1668         for (; type < nr_swapfiles; type++) {
1669                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1670                 si = swap_info[type];
1671                 if (!(si->flags & SWP_USED) || !si->swap_map)
1672                         continue;
1673                 ++*pos;
1674                 return si;
1675         }
1676
1677         return NULL;
1678 }
1679
1680 static void swap_stop(struct seq_file *swap, void *v)
1681 {
1682         mutex_unlock(&swapon_mutex);
1683 }
1684
1685 static int swap_show(struct seq_file *swap, void *v)
1686 {
1687         struct swap_info_struct *si = v;
1688         struct file *file;
1689         int len;
1690
1691         if (si == SEQ_START_TOKEN) {
1692                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1693                 return 0;
1694         }
1695
1696         file = si->swap_file;
1697         len = seq_path(swap, &file->f_path, " \t\n\\");
1698         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1699                         len < 40 ? 40 - len : 1, " ",
1700                         S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1701                                 "partition" : "file\t",
1702                         si->pages << (PAGE_SHIFT - 10),
1703                         si->inuse_pages << (PAGE_SHIFT - 10),
1704                         si->prio);
1705         return 0;
1706 }
1707
1708 static const struct seq_operations swaps_op = {
1709         .start =        swap_start,
1710         .next =         swap_next,
1711         .stop =         swap_stop,
1712         .show =         swap_show
1713 };
1714
1715 static int swaps_open(struct inode *inode, struct file *file)
1716 {
1717         return seq_open(file, &swaps_op);
1718 }
1719
1720 static const struct file_operations proc_swaps_operations = {
1721         .open           = swaps_open,
1722         .read           = seq_read,
1723         .llseek         = seq_lseek,
1724         .release        = seq_release,
1725 };
1726
1727 static int __init procswaps_init(void)
1728 {
1729         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1730         return 0;
1731 }
1732 __initcall(procswaps_init);
1733 #endif /* CONFIG_PROC_FS */
1734
1735 #ifdef MAX_SWAPFILES_CHECK
1736 static int __init max_swapfiles_check(void)
1737 {
1738         MAX_SWAPFILES_CHECK();
1739         return 0;
1740 }
1741 late_initcall(max_swapfiles_check);
1742 #endif
1743
1744 /*
1745  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1746  *
1747  * The swapon system call
1748  */
1749 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
1750 {
1751         struct swap_info_struct *p;
1752         char *name = NULL;
1753         struct block_device *bdev = NULL;
1754         struct file *swap_file = NULL;
1755         struct address_space *mapping;
1756         unsigned int type;
1757         int i, prev;
1758         int error;
1759         union swap_header *swap_header = NULL;
1760         unsigned int nr_good_pages = 0;
1761         int nr_extents = 0;
1762         sector_t span;
1763         unsigned long maxpages = 1;
1764         unsigned long swapfilepages;
1765         unsigned short *swap_map = NULL;
1766         struct page *page = NULL;
1767         struct inode *inode = NULL;
1768         int did_down = 0;
1769
1770         if (!capable(CAP_SYS_ADMIN))
1771                 return -EPERM;
1772
1773         p = kzalloc(sizeof(*p), GFP_KERNEL);
1774         if (!p)
1775                 return -ENOMEM;
1776
1777         spin_lock(&swap_lock);
1778         for (type = 0; type < nr_swapfiles; type++) {
1779                 if (!(swap_info[type]->flags & SWP_USED))
1780                         break;
1781         }
1782         error = -EPERM;
1783         if (type >= MAX_SWAPFILES) {
1784                 spin_unlock(&swap_lock);
1785                 kfree(p);
1786                 goto out;
1787         }
1788         if (type >= nr_swapfiles) {
1789                 p->type = type;
1790                 swap_info[type] = p;
1791                 /*
1792                  * Write swap_info[type] before nr_swapfiles, in case a
1793                  * racing procfs swap_start() or swap_next() is reading them.
1794                  * (We never shrink nr_swapfiles, we never free this entry.)
1795                  */
1796                 smp_wmb();
1797                 nr_swapfiles++;
1798         } else {
1799                 kfree(p);
1800                 p = swap_info[type];
1801                 /*
1802                  * Do not memset this entry: a racing procfs swap_next()
1803                  * would be relying on p->type to remain valid.
1804                  */
1805         }
1806         INIT_LIST_HEAD(&p->first_swap_extent.list);
1807         p->flags = SWP_USED;
1808         p->next = -1;
1809         spin_unlock(&swap_lock);
1810
1811         name = getname(specialfile);
1812         error = PTR_ERR(name);
1813         if (IS_ERR(name)) {
1814                 name = NULL;
1815                 goto bad_swap_2;
1816         }
1817         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1818         error = PTR_ERR(swap_file);
1819         if (IS_ERR(swap_file)) {
1820                 swap_file = NULL;
1821                 goto bad_swap_2;
1822         }
1823
1824         p->swap_file = swap_file;
1825         mapping = swap_file->f_mapping;
1826         inode = mapping->host;
1827
1828         error = -EBUSY;
1829         for (i = 0; i < nr_swapfiles; i++) {
1830                 struct swap_info_struct *q = swap_info[i];
1831
1832                 if (i == type || !q->swap_file)
1833                         continue;
1834                 if (mapping == q->swap_file->f_mapping)
1835                         goto bad_swap;
1836         }
1837
1838         error = -EINVAL;
1839         if (S_ISBLK(inode->i_mode)) {
1840                 bdev = I_BDEV(inode);
1841                 error = bd_claim(bdev, sys_swapon);
1842                 if (error < 0) {
1843                         bdev = NULL;
1844                         error = -EINVAL;
1845                         goto bad_swap;
1846                 }
1847                 p->old_block_size = block_size(bdev);
1848                 error = set_blocksize(bdev, PAGE_SIZE);
1849                 if (error < 0)
1850                         goto bad_swap;
1851                 p->bdev = bdev;
1852         } else if (S_ISREG(inode->i_mode)) {
1853                 p->bdev = inode->i_sb->s_bdev;
1854                 mutex_lock(&inode->i_mutex);
1855                 did_down = 1;
1856                 if (IS_SWAPFILE(inode)) {
1857                         error = -EBUSY;
1858                         goto bad_swap;
1859                 }
1860         } else {
1861                 goto bad_swap;
1862         }
1863
1864         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1865
1866         /*
1867          * Read the swap header.
1868          */
1869         if (!mapping->a_ops->readpage) {
1870                 error = -EINVAL;
1871                 goto bad_swap;
1872         }
1873         page = read_mapping_page(mapping, 0, swap_file);
1874         if (IS_ERR(page)) {
1875                 error = PTR_ERR(page);
1876                 goto bad_swap;
1877         }
1878         swap_header = kmap(page);
1879
1880         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1881                 printk(KERN_ERR "Unable to find swap-space signature\n");
1882                 error = -EINVAL;
1883                 goto bad_swap;
1884         }
1885
1886         /* swap partition endianess hack... */
1887         if (swab32(swap_header->info.version) == 1) {
1888                 swab32s(&swap_header->info.version);
1889                 swab32s(&swap_header->info.last_page);
1890                 swab32s(&swap_header->info.nr_badpages);
1891                 for (i = 0; i < swap_header->info.nr_badpages; i++)
1892                         swab32s(&swap_header->info.badpages[i]);
1893         }
1894         /* Check the swap header's sub-version */
1895         if (swap_header->info.version != 1) {
1896                 printk(KERN_WARNING
1897                        "Unable to handle swap header version %d\n",
1898                        swap_header->info.version);
1899                 error = -EINVAL;
1900                 goto bad_swap;
1901         }
1902
1903         p->lowest_bit  = 1;
1904         p->cluster_next = 1;
1905         p->cluster_nr = 0;
1906
1907         /*
1908          * Find out how many pages are allowed for a single swap
1909          * device. There are two limiting factors: 1) the number of
1910          * bits for the swap offset in the swp_entry_t type and
1911          * 2) the number of bits in the a swap pte as defined by
1912          * the different architectures. In order to find the
1913          * largest possible bit mask a swap entry with swap type 0
1914          * and swap offset ~0UL is created, encoded to a swap pte,
1915          * decoded to a swp_entry_t again and finally the swap
1916          * offset is extracted. This will mask all the bits from
1917          * the initial ~0UL mask that can't be encoded in either
1918          * the swp_entry_t or the architecture definition of a
1919          * swap pte.
1920          */
1921         maxpages = swp_offset(pte_to_swp_entry(
1922                         swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1;
1923         if (maxpages > swap_header->info.last_page)
1924                 maxpages = swap_header->info.last_page;
1925         p->highest_bit = maxpages - 1;
1926
1927         error = -EINVAL;
1928         if (!maxpages)
1929                 goto bad_swap;
1930         if (swapfilepages && maxpages > swapfilepages) {
1931                 printk(KERN_WARNING
1932                        "Swap area shorter than signature indicates\n");
1933                 goto bad_swap;
1934         }
1935         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1936                 goto bad_swap;
1937         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1938                 goto bad_swap;
1939
1940         /* OK, set up the swap map and apply the bad block list */
1941         swap_map = vmalloc(maxpages * sizeof(short));
1942         if (!swap_map) {
1943                 error = -ENOMEM;
1944                 goto bad_swap;
1945         }
1946
1947         memset(swap_map, 0, maxpages * sizeof(short));
1948         for (i = 0; i < swap_header->info.nr_badpages; i++) {
1949                 int page_nr = swap_header->info.badpages[i];
1950                 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) {
1951                         error = -EINVAL;
1952                         goto bad_swap;
1953                 }
1954                 swap_map[page_nr] = SWAP_MAP_BAD;
1955         }
1956
1957         error = swap_cgroup_swapon(type, maxpages);
1958         if (error)
1959                 goto bad_swap;
1960
1961         nr_good_pages = swap_header->info.last_page -
1962                         swap_header->info.nr_badpages -
1963                         1 /* header page */;
1964
1965         if (nr_good_pages) {
1966                 swap_map[0] = SWAP_MAP_BAD;
1967                 p->max = maxpages;
1968                 p->pages = nr_good_pages;
1969                 nr_extents = setup_swap_extents(p, &span);
1970                 if (nr_extents < 0) {
1971                         error = nr_extents;
1972                         goto bad_swap;
1973                 }
1974                 nr_good_pages = p->pages;
1975         }
1976         if (!nr_good_pages) {
1977                 printk(KERN_WARNING "Empty swap-file\n");
1978                 error = -EINVAL;
1979                 goto bad_swap;
1980         }
1981
1982         if (p->bdev) {
1983                 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
1984                         p->flags |= SWP_SOLIDSTATE;
1985                         p->cluster_next = 1 + (random32() % p->highest_bit);
1986                 }
1987                 if (discard_swap(p) == 0)
1988                         p->flags |= SWP_DISCARDABLE;
1989         }
1990
1991         mutex_lock(&swapon_mutex);
1992         spin_lock(&swap_lock);
1993         if (swap_flags & SWAP_FLAG_PREFER)
1994                 p->prio =
1995                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1996         else
1997                 p->prio = --least_priority;
1998         p->swap_map = swap_map;
1999         p->flags |= SWP_WRITEOK;
2000         nr_swap_pages += nr_good_pages;
2001         total_swap_pages += nr_good_pages;
2002
2003         printk(KERN_INFO "Adding %uk swap on %s.  "
2004                         "Priority:%d extents:%d across:%lluk %s%s\n",
2005                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
2006                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2007                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2008                 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2009
2010         /* insert swap space into swap_list: */
2011         prev = -1;
2012         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
2013                 if (p->prio >= swap_info[i]->prio)
2014                         break;
2015                 prev = i;
2016         }
2017         p->next = i;
2018         if (prev < 0)
2019                 swap_list.head = swap_list.next = type;
2020         else
2021                 swap_info[prev]->next = type;
2022         spin_unlock(&swap_lock);
2023         mutex_unlock(&swapon_mutex);
2024         error = 0;
2025         goto out;
2026 bad_swap:
2027         if (bdev) {
2028                 set_blocksize(bdev, p->old_block_size);
2029                 bd_release(bdev);
2030         }
2031         destroy_swap_extents(p);
2032         swap_cgroup_swapoff(type);
2033 bad_swap_2:
2034         spin_lock(&swap_lock);
2035         p->swap_file = NULL;
2036         p->flags = 0;
2037         spin_unlock(&swap_lock);
2038         vfree(swap_map);
2039         if (swap_file)
2040                 filp_close(swap_file, NULL);
2041 out:
2042         if (page && !IS_ERR(page)) {
2043                 kunmap(page);
2044                 page_cache_release(page);
2045         }
2046         if (name)
2047                 putname(name);
2048         if (did_down) {
2049                 if (!error)
2050                         inode->i_flags |= S_SWAPFILE;
2051                 mutex_unlock(&inode->i_mutex);
2052         }
2053         return error;
2054 }
2055
2056 void si_swapinfo(struct sysinfo *val)
2057 {
2058         unsigned int type;
2059         unsigned long nr_to_be_unused = 0;
2060
2061         spin_lock(&swap_lock);
2062         for (type = 0; type < nr_swapfiles; type++) {
2063                 struct swap_info_struct *si = swap_info[type];
2064
2065                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2066                         nr_to_be_unused += si->inuse_pages;
2067         }
2068         val->freeswap = nr_swap_pages + nr_to_be_unused;
2069         val->totalswap = total_swap_pages + nr_to_be_unused;
2070         spin_unlock(&swap_lock);
2071 }
2072
2073 /*
2074  * Verify that a swap entry is valid and increment its swap map count.
2075  *
2076  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
2077  * "permanent", but will be reclaimed by the next swapoff.
2078  * Returns error code in following case.
2079  * - success -> 0
2080  * - swp_entry is invalid -> EINVAL
2081  * - swp_entry is migration entry -> EINVAL
2082  * - swap-cache reference is requested but there is already one. -> EEXIST
2083  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2084  */
2085 static int __swap_duplicate(swp_entry_t entry, unsigned short usage)
2086 {
2087         struct swap_info_struct *p;
2088         unsigned long offset, type;
2089         unsigned short count;
2090         unsigned short has_cache;
2091         int err = -EINVAL;
2092
2093         if (non_swap_entry(entry))
2094                 goto out;
2095
2096         type = swp_type(entry);
2097         if (type >= nr_swapfiles)
2098                 goto bad_file;
2099         p = swap_info[type];
2100         offset = swp_offset(entry);
2101
2102         spin_lock(&swap_lock);
2103         if (unlikely(offset >= p->max))
2104                 goto unlock_out;
2105
2106         count = p->swap_map[offset];
2107         has_cache = count & SWAP_HAS_CACHE;
2108         count &= ~SWAP_HAS_CACHE;
2109         err = 0;
2110
2111         if (usage == SWAP_HAS_CACHE) {
2112
2113                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2114                 if (!has_cache && count)
2115                         has_cache = SWAP_HAS_CACHE;
2116                 else if (has_cache)             /* someone else added cache */
2117                         err = -EEXIST;
2118                 else                            /* no users remaining */
2119                         err = -ENOENT;
2120
2121         } else if (count || has_cache) {
2122
2123                 if (count < SWAP_MAP_MAX - 1)
2124                         count++;
2125                 else if (count <= SWAP_MAP_MAX) {
2126                         if (swap_overflow++ < 5)
2127                                 printk(KERN_WARNING
2128                                        "swap_dup: swap entry overflow\n");
2129                         count = SWAP_MAP_MAX;
2130                 } else
2131                         err = -EINVAL;
2132         } else
2133                 err = -ENOENT;                  /* unused swap entry */
2134
2135         p->swap_map[offset] = count | has_cache;
2136
2137 unlock_out:
2138         spin_unlock(&swap_lock);
2139 out:
2140         return err;
2141
2142 bad_file:
2143         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2144         goto out;
2145 }
2146
2147 /*
2148  * increase reference count of swap entry by 1.
2149  */
2150 void swap_duplicate(swp_entry_t entry)
2151 {
2152         __swap_duplicate(entry, 1);
2153 }
2154
2155 /*
2156  * @entry: swap entry for which we allocate swap cache.
2157  *
2158  * Called when allocating swap cache for existing swap entry,
2159  * This can return error codes. Returns 0 at success.
2160  * -EBUSY means there is a swap cache.
2161  * Note: return code is different from swap_duplicate().
2162  */
2163 int swapcache_prepare(swp_entry_t entry)
2164 {
2165         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2166 }
2167
2168 /*
2169  * swap_lock prevents swap_map being freed. Don't grab an extra
2170  * reference on the swaphandle, it doesn't matter if it becomes unused.
2171  */
2172 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2173 {
2174         struct swap_info_struct *si;
2175         int our_page_cluster = page_cluster;
2176         pgoff_t target, toff;
2177         pgoff_t base, end;
2178         int nr_pages = 0;
2179
2180         if (!our_page_cluster)  /* no readahead */
2181                 return 0;
2182
2183         si = swap_info[swp_type(entry)];
2184         target = swp_offset(entry);
2185         base = (target >> our_page_cluster) << our_page_cluster;
2186         end = base + (1 << our_page_cluster);
2187         if (!base)              /* first page is swap header */
2188                 base++;
2189
2190         spin_lock(&swap_lock);
2191         if (end > si->max)      /* don't go beyond end of map */
2192                 end = si->max;
2193
2194         /* Count contiguous allocated slots above our target */
2195         for (toff = target; ++toff < end; nr_pages++) {
2196                 /* Don't read in free or bad pages */
2197                 if (!si->swap_map[toff])
2198                         break;
2199                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2200                         break;
2201         }
2202         /* Count contiguous allocated slots below our target */
2203         for (toff = target; --toff >= base; nr_pages++) {
2204                 /* Don't read in free or bad pages */
2205                 if (!si->swap_map[toff])
2206                         break;
2207                 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2208                         break;
2209         }
2210         spin_unlock(&swap_lock);
2211
2212         /*
2213          * Indicate starting offset, and return number of pages to get:
2214          * if only 1, say 0, since there's then no readahead to be done.
2215          */
2216         *offset = ++toff;
2217         return nr_pages? ++nr_pages: 0;
2218 }