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/writeback.h>
  20 #include <linux/proc_fs.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/init.h>
  23 #include <linux/module.h>
  24 #include <linux/rmap.h>
  25 #include <linux/security.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/mutex.h>
  28 #include <linux/capability.h>
  29 #include <linux/syscalls.h>
  30
  31 #include <asm/pgtable.h>
  32 #include <asm/tlbflush.h>
  33 #include <linux/swapops.h>
  34
  35 DEFINE_SPINLOCK(swap_lock);
  36 unsigned int nr_swapfiles;
  37 long total_swap_pages;
  38 static int swap_overflow;
  39
  40 static const char Bad_file[] = "Bad swap file entry ";
  41 static const char Unused_file[] = "Unused swap file entry ";
  42 static const char Bad_offset[] = "Bad swap offset entry ";
  43 static const char Unused_offset[] = "Unused swap offset entry ";
  44
  45 struct swap_list_t swap_list = {-1, -1};
  46
  47 static struct swap_info_struct swap_info[MAX_SWAPFILES];
  48
  49 static DEFINE_MUTEX(swapon_mutex);
  50
  51 /*
  52  * We need this because the bdev->unplug_fn can sleep and we cannot
  53  * hold swap_lock while calling the unplug_fn. And swap_lock
  54  * cannot be turned into a mutex.
  55  */
  56 static DECLARE_RWSEM(swap_unplug_sem);
  57
  58 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  59 {
  60         swp_entry_t entry;
  61
  62         down_read(&swap_unplug_sem);
  63         entry.val = page_private(page);
  64         if (PageSwapCache(page)) {
  65                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  66                 struct backing_dev_info *bdi;
  67
  68                 /*
  69                  * If the page is removed from swapcache from under us (with a
  70                  * racy try_to_unuse/swapoff) we need an additional reference
  71                  * count to avoid reading garbage from page_private(page) above.
  72                  * If the WARN_ON triggers during a swapoff it maybe the race
  73                  * condition and it's harmless. However if it triggers without
  74                  * swapoff it signals a problem.
  75                  */
  76                 WARN_ON(page_count(page) <= 1);
  77
  78                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  79                 blk_run_backing_dev(bdi, page);
  80         }
  81         up_read(&swap_unplug_sem);
  82 }
  83
  84 #define SWAPFILE_CLUSTER        256
  85 #define LATENCY_LIMIT           256
  86
  87 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
  88 {
  89         unsigned long offset, last_in_cluster;
  90         int latency_ration = LATENCY_LIMIT;
  91
  92         /*
  93          * We try to cluster swap pages by allocating them sequentially
  94          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
  95          * way, however, we resort to first-free allocation, starting
  96          * a new cluster.  This prevents us from scattering swap pages
  97          * all over the entire swap partition, so that we reduce
  98          * overall disk seek times between swap pages.  -- sct
  99          * But we do now try to find an empty cluster.  -Andrea
 100          */
 101
 102         si->flags += SWP_SCANNING;
 103         if (unlikely(!si->cluster_nr)) {
 104                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 105                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
 106                         goto lowest;
 107                 spin_unlock(&swap_lock);
 108
 109                 offset = si->lowest_bit;
 110                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 111
 112                 /* Locate the first empty (unaligned) cluster */
 113                 for (; last_in_cluster <= si->highest_bit; offset++) {
 114                         if (si->swap_map[offset])
 115                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 116                         else if (offset == last_in_cluster) {
 117                                 spin_lock(&swap_lock);
 118                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
 119                                 goto cluster;
 120                         }
 121                         if (unlikely(--latency_ration < 0)) {
 122                                 cond_resched();
 123                                 latency_ration = LATENCY_LIMIT;
 124                         }
 125                 }
 126                 spin_lock(&swap_lock);
 127                 goto lowest;
 128         }
 129
 130         si->cluster_nr--;
 131 cluster:
 132         offset = si->cluster_next;
 133         if (offset > si->highest_bit)
 134 lowest:         offset = si->lowest_bit;
 135 checks: if (!(si->flags & SWP_WRITEOK))
 136                 goto no_page;
 137         if (!si->highest_bit)
 138                 goto no_page;
 139         if (!si->swap_map[offset]) {
 140                 if (offset == si->lowest_bit)
 141                         si->lowest_bit++;
 142                 if (offset == si->highest_bit)
 143                         si->highest_bit--;
 144                 si->inuse_pages++;
 145                 if (si->inuse_pages == si->pages) {
 146                         si->lowest_bit = si->max;
 147                         si->highest_bit = 0;
 148                 }
 149                 si->swap_map[offset] = 1;
 150                 si->cluster_next = offset + 1;
 151                 si->flags -= SWP_SCANNING;
 152                 return offset;
 153         }
 154
 155         spin_unlock(&swap_lock);
 156         while (++offset <= si->highest_bit) {
 157                 if (!si->swap_map[offset]) {
 158                         spin_lock(&swap_lock);
 159                         goto checks;
 160                 }
 161                 if (unlikely(--latency_ration < 0)) {
 162                         cond_resched();
 163                         latency_ration = LATENCY_LIMIT;
 164                 }
 165         }
 166         spin_lock(&swap_lock);
 167         goto lowest;
 168
 169 no_page:
 170         si->flags -= SWP_SCANNING;
 171         return 0;
 172 }
 173
 174 swp_entry_t get_swap_page(void)
 175 {
 176         struct swap_info_struct *si;
 177         pgoff_t offset;
 178         int type, next;
 179         int wrapped = 0;
 180
 181         spin_lock(&swap_lock);
 182         if (nr_swap_pages <= 0)
 183                 goto noswap;
 184         nr_swap_pages--;
 185
 186         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 187                 si = swap_info + type;
 188                 next = si->next;
 189                 if (next < 0 ||
 190                     (!wrapped && si->prio != swap_info[next].prio)) {
 191                         next = swap_list.head;
 192                         wrapped++;
 193                 }
 194
 195                 if (!si->highest_bit)
 196                         continue;
 197                 if (!(si->flags & SWP_WRITEOK))
 198                         continue;
 199
 200                 swap_list.next = next;
 201                 offset = scan_swap_map(si);
 202                 if (offset) {
 203                         spin_unlock(&swap_lock);
 204                         return swp_entry(type, offset);
 205                 }
 206                 next = swap_list.next;
 207         }
 208
 209         nr_swap_pages++;
 210 noswap:
 211         spin_unlock(&swap_lock);
 212         return (swp_entry_t) {0};
 213 }
 214
 215 swp_entry_t get_swap_page_of_type(int type)
 216 {
 217         struct swap_info_struct *si;
 218         pgoff_t offset;
 219
 220         spin_lock(&swap_lock);
 221         si = swap_info + type;
 222         if (si->flags & SWP_WRITEOK) {
 223                 nr_swap_pages--;
 224                 offset = scan_swap_map(si);
 225                 if (offset) {
 226                         spin_unlock(&swap_lock);
 227                         return swp_entry(type, offset);
 228                 }
 229                 nr_swap_pages++;
 230         }
 231         spin_unlock(&swap_lock);
 232         return (swp_entry_t) {0};
 233 }
 234
 235 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 236 {
 237         struct swap_info_struct * p;
 238         unsigned long offset, type;
 239
 240         if (!entry.val)
 241                 goto out;
 242         type = swp_type(entry);
 243         if (type >= nr_swapfiles)
 244                 goto bad_nofile;
 245         p = & swap_info[type];
 246         if (!(p->flags & SWP_USED))
 247                 goto bad_device;
 248         offset = swp_offset(entry);
 249         if (offset >= p->max)
 250                 goto bad_offset;
 251         if (!p->swap_map[offset])
 252                 goto bad_free;
 253         spin_lock(&swap_lock);
 254         return p;
 255
 256 bad_free:
 257         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 258         goto out;
 259 bad_offset:
 260         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 261         goto out;
 262 bad_device:
 263         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 264         goto out;
 265 bad_nofile:
 266         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 267 out:
 268         return NULL;
 269 }
 270
 271 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 272 {
 273         int count = p->swap_map[offset];
 274
 275         if (count < SWAP_MAP_MAX) {
 276                 count--;
 277                 p->swap_map[offset] = count;
 278                 if (!count) {
 279                         if (offset < p->lowest_bit)
 280                                 p->lowest_bit = offset;
 281                         if (offset > p->highest_bit)
 282                                 p->highest_bit = offset;
 283                         if (p->prio > swap_info[swap_list.next].prio)
 284                                 swap_list.next = p - swap_info;
 285                         nr_swap_pages++;
 286                         p->inuse_pages--;
 287                 }
 288         }
 289         return count;
 290 }
 291
 292 /*
 293  * Caller has made sure that the swapdevice corresponding to entry
 294  * is still around or has not been recycled.
 295  */
 296 void swap_free(swp_entry_t entry)
 297 {
 298         struct swap_info_struct * p;
 299
 300         p = swap_info_get(entry);
 301         if (p) {
 302                 swap_entry_free(p, swp_offset(entry));
 303                 spin_unlock(&swap_lock);
 304         }
 305 }
 306
 307 /*
 308  * How many references to page are currently swapped out?
 309  */
 310 static inline int page_swapcount(struct page *page)
 311 {
 312         int count = 0;
 313         struct swap_info_struct *p;
 314         swp_entry_t entry;
 315
 316         entry.val = page_private(page);
 317         p = swap_info_get(entry);
 318         if (p) {
 319                 /* Subtract the 1 for the swap cache itself */
 320                 count = p->swap_map[swp_offset(entry)] - 1;
 321                 spin_unlock(&swap_lock);
 322         }
 323         return count;
 324 }
 325
 326 /*
 327  * We can use this swap cache entry directly
 328  * if there are no other references to it.
 329  */
 330 int can_share_swap_page(struct page *page)
 331 {
 332         int count;
 333
 334         BUG_ON(!PageLocked(page));
 335         count = page_mapcount(page);
 336         if (count <= 1 && PageSwapCache(page))
 337                 count += page_swapcount(page);
 338         return count == 1;
 339 }
 340
 341 /*
 342  * Work out if there are any other processes sharing this
 343  * swap cache page. Free it if you can. Return success.
 344  */
 345 int remove_exclusive_swap_page(struct page *page)
 346 {
 347         int retval;
 348         struct swap_info_struct * p;
 349         swp_entry_t entry;
 350
 351         BUG_ON(PagePrivate(page));
 352         BUG_ON(!PageLocked(page));
 353
 354         if (!PageSwapCache(page))
 355                 return 0;
 356         if (PageWriteback(page))
 357                 return 0;
 358         if (page_count(page) != 2) /* 2: us + cache */
 359                 return 0;
 360
 361         entry.val = page_private(page);
 362         p = swap_info_get(entry);
 363         if (!p)
 364                 return 0;
 365
 366         /* Is the only swap cache user the cache itself? */
 367         retval = 0;
 368         if (p->swap_map[swp_offset(entry)] == 1) {
 369                 /* Recheck the page count with the swapcache lock held.. */
 370                 write_lock_irq(&swapper_space.tree_lock);
 371                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 372                         __delete_from_swap_cache(page);
 373                         SetPageDirty(page);
 374                         retval = 1;
 375                 }
 376                 write_unlock_irq(&swapper_space.tree_lock);
 377         }
 378         spin_unlock(&swap_lock);
 379
 380         if (retval) {
 381                 swap_free(entry);
 382                 page_cache_release(page);
 383         }
 384
 385         return retval;
 386 }
 387
 388 /*
 389  * Free the swap entry like above, but also try to
 390  * free the page cache entry if it is the last user.
 391  */
 392 void free_swap_and_cache(swp_entry_t entry)
 393 {
 394         struct swap_info_struct * p;
 395         struct page *page = NULL;
 396
 397         if (is_migration_entry(entry))
 398                 return;
 399
 400         p = swap_info_get(entry);
 401         if (p) {
 402                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 403                         page = find_get_page(&swapper_space, entry.val);
 404                         if (page && unlikely(TestSetPageLocked(page))) {
 405                                 page_cache_release(page);
 406                                 page = NULL;
 407                         }
 408                 }
 409                 spin_unlock(&swap_lock);
 410         }
 411         if (page) {
 412                 int one_user;
 413
 414                 BUG_ON(PagePrivate(page));
 415                 one_user = (page_count(page) == 2);
 416                 /* Only cache user (+us), or swap space full? Free it! */
 417                 /* Also recheck PageSwapCache after page is locked (above) */
 418                 if (PageSwapCache(page) && !PageWriteback(page) &&
 419                                         (one_user || vm_swap_full())) {
 420                         delete_from_swap_cache(page);
 421                         SetPageDirty(page);
 422                 }
 423                 unlock_page(page);
 424                 page_cache_release(page);
 425         }
 426 }
 427
 428 #ifdef CONFIG_SOFTWARE_SUSPEND
 429 /*
 430  * Find the swap type that corresponds to given device (if any).
 431  *
 432  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 433  * from 0, in which the swap header is expected to be located.
 434  *
 435  * This is needed for the suspend to disk (aka swsusp).
 436  */
 437 int swap_type_of(dev_t device, sector_t offset)
 438 {
 439         struct block_device *bdev = NULL;
 440         int i;
 441
 442         if (device)
 443                 bdev = bdget(device);
 444
 445         spin_lock(&swap_lock);
 446         for (i = 0; i < nr_swapfiles; i++) {
 447                 struct swap_info_struct *sis = swap_info + i;
 448
 449                 if (!(sis->flags & SWP_WRITEOK))
 450                         continue;
 451
 452                 if (!bdev) {
 453                         spin_unlock(&swap_lock);
 454                         return i;
 455                 }
 456                 if (bdev == sis->bdev) {
 457                         struct swap_extent *se;
 458
 459                         se = list_entry(sis->extent_list.next,
 460                                         struct swap_extent, list);
 461                         if (se->start_block == offset) {
 462                                 spin_unlock(&swap_lock);
 463                                 bdput(bdev);
 464                                 return i;
 465                         }
 466                 }
 467         }
 468         spin_unlock(&swap_lock);
 469         if (bdev)
 470                 bdput(bdev);
 471
 472         return -ENODEV;
 473 }
 474
 475 /*
 476  * Return either the total number of swap pages of given type, or the number
 477  * of free pages of that type (depending on @free)
 478  *
 479  * This is needed for software suspend
 480  */
 481 unsigned int count_swap_pages(int type, int free)
 482 {
 483         unsigned int n = 0;
 484
 485         if (type < nr_swapfiles) {
 486                 spin_lock(&swap_lock);
 487                 if (swap_info[type].flags & SWP_WRITEOK) {
 488                         n = swap_info[type].pages;
 489                         if (free)
 490                                 n -= swap_info[type].inuse_pages;
 491                 }
 492                 spin_unlock(&swap_lock);
 493         }
 494         return n;
 495 }
 496 #endif
 497
 498 /*
 499  * No need to decide whether this PTE shares the swap entry with others,
 500  * just let do_wp_page work it out if a write is requested later - to
 501  * force COW, vm_page_prot omits write permission from any private vma.
 502  */
 503 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
 504                 unsigned long addr, swp_entry_t entry, struct page *page)
 505 {
 506         inc_mm_counter(vma->vm_mm, anon_rss);
 507         get_page(page);
 508         set_pte_at(vma->vm_mm, addr, pte,
 509                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 510         page_add_anon_rmap(page, vma, addr);
 511         swap_free(entry);
 512         /*
 513          * Move the page to the active list so it is not
 514          * immediately swapped out again after swapon.
 515          */
 516         activate_page(page);
 517 }
 518
 519 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 520                                 unsigned long addr, unsigned long end,
 521                                 swp_entry_t entry, struct page *page)
 522 {
 523         pte_t swp_pte = swp_entry_to_pte(entry);
 524         pte_t *pte;
 525         spinlock_t *ptl;
 526         int found = 0;
 527
 528         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 529         do {
 530                 /*
 531                  * swapoff spends a _lot_ of time in this loop!
 532                  * Test inline before going to call unuse_pte.
 533                  */
 534                 if (unlikely(pte_same(*pte, swp_pte))) {
 535                         unuse_pte(vma, pte++, addr, entry, page);
 536                         found = 1;
 537                         break;
 538                 }
 539         } while (pte++, addr += PAGE_SIZE, addr != end);
 540         pte_unmap_unlock(pte - 1, ptl);
 541         return found;
 542 }
 543
 544 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 545                                 unsigned long addr, unsigned long end,
 546                                 swp_entry_t entry, struct page *page)
 547 {
 548         pmd_t *pmd;
 549         unsigned long next;
 550
 551         pmd = pmd_offset(pud, addr);
 552         do {
 553                 next = pmd_addr_end(addr, end);
 554                 if (pmd_none_or_clear_bad(pmd))
 555                         continue;
 556                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 557                         return 1;
 558         } while (pmd++, addr = next, addr != end);
 559         return 0;
 560 }
 561
 562 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 563                                 unsigned long addr, unsigned long end,
 564                                 swp_entry_t entry, struct page *page)
 565 {
 566         pud_t *pud;
 567         unsigned long next;
 568
 569         pud = pud_offset(pgd, addr);
 570         do {
 571                 next = pud_addr_end(addr, end);
 572                 if (pud_none_or_clear_bad(pud))
 573                         continue;
 574                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 575                         return 1;
 576         } while (pud++, addr = next, addr != end);
 577         return 0;
 578 }
 579
 580 static int unuse_vma(struct vm_area_struct *vma,
 581                                 swp_entry_t entry, struct page *page)
 582 {
 583         pgd_t *pgd;
 584         unsigned long addr, end, next;
 585
 586         if (page->mapping) {
 587                 addr = page_address_in_vma(page, vma);
 588                 if (addr == -EFAULT)
 589                         return 0;
 590                 else
 591                         end = addr + PAGE_SIZE;
 592         } else {
 593                 addr = vma->vm_start;
 594                 end = vma->vm_end;
 595         }
 596
 597         pgd = pgd_offset(vma->vm_mm, addr);
 598         do {
 599                 next = pgd_addr_end(addr, end);
 600                 if (pgd_none_or_clear_bad(pgd))
 601                         continue;
 602                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 603                         return 1;
 604         } while (pgd++, addr = next, addr != end);
 605         return 0;
 606 }
 607
 608 static int unuse_mm(struct mm_struct *mm,
 609                                 swp_entry_t entry, struct page *page)
 610 {
 611         struct vm_area_struct *vma;
 612
 613         if (!down_read_trylock(&mm->mmap_sem)) {
 614                 /*
 615                  * Activate page so shrink_cache is unlikely to unmap its
 616                  * ptes while lock is dropped, so swapoff can make progress.
 617                  */
 618                 activate_page(page);
 619                 unlock_page(page);
 620                 down_read(&mm->mmap_sem);
 621                 lock_page(page);
 622         }
 623         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 624                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 625                         break;
 626         }
 627         up_read(&mm->mmap_sem);
 628         /*
 629          * Currently unuse_mm cannot fail, but leave error handling
 630          * at call sites for now, since we change it from time to time.
 631          */
 632         return 0;
 633 }
 634
 635 /*
 636  * Scan swap_map from current position to next entry still in use.
 637  * Recycle to start on reaching the end, returning 0 when empty.
 638  */
 639 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 640                                         unsigned int prev)
 641 {
 642         unsigned int max = si->max;
 643         unsigned int i = prev;
 644         int count;
 645
 646         /*
 647          * No need for swap_lock here: we're just looking
 648          * for whether an entry is in use, not modifying it; false
 649          * hits are okay, and sys_swapoff() has already prevented new
 650          * allocations from this area (while holding swap_lock).
 651          */
 652         for (;;) {
 653                 if (++i >= max) {
 654                         if (!prev) {
 655                                 i = 0;
 656                                 break;
 657                         }
 658                         /*
 659                          * No entries in use at top of swap_map,
 660                          * loop back to start and recheck there.
 661                          */
 662                         max = prev + 1;
 663                         prev = 0;
 664                         i = 1;
 665                 }
 666                 count = si->swap_map[i];
 667                 if (count && count != SWAP_MAP_BAD)
 668                         break;
 669         }
 670         return i;
 671 }
 672
 673 /*
 674  * We completely avoid races by reading each swap page in advance,
 675  * and then search for the process using it.  All the necessary
 676  * page table adjustments can then be made atomically.
 677  */
 678 static int try_to_unuse(unsigned int type)
 679 {
 680         struct swap_info_struct * si = &swap_info[type];
 681         struct mm_struct *start_mm;
 682         unsigned short *swap_map;
 683         unsigned short swcount;
 684         struct page *page;
 685         swp_entry_t entry;
 686         unsigned int i = 0;
 687         int retval = 0;
 688         int reset_overflow = 0;
 689         int shmem;
 690
 691         /*
 692          * When searching mms for an entry, a good strategy is to
 693          * start at the first mm we freed the previous entry from
 694          * (though actually we don't notice whether we or coincidence
 695          * freed the entry).  Initialize this start_mm with a hold.
 696          *
 697          * A simpler strategy would be to start at the last mm we
 698          * freed the previous entry from; but that would take less
 699          * advantage of mmlist ordering, which clusters forked mms
 700          * together, child after parent.  If we race with dup_mmap(), we
 701          * prefer to resolve parent before child, lest we miss entries
 702          * duplicated after we scanned child: using last mm would invert
 703          * that.  Though it's only a serious concern when an overflowed
 704          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 705          */
 706         start_mm = &init_mm;
 707         atomic_inc(&init_mm.mm_users);
 708
 709         /*
 710          * Keep on scanning until all entries have gone.  Usually,
 711          * one pass through swap_map is enough, but not necessarily:
 712          * there are races when an instance of an entry might be missed.
 713          */
 714         while ((i = find_next_to_unuse(si, i)) != 0) {
 715                 if (signal_pending(current)) {
 716                         retval = -EINTR;
 717                         break;
 718                 }
 719
 720                 /*
 721                  * Get a page for the entry, using the existing swap
 722                  * cache page if there is one.  Otherwise, get a clean
 723                  * page and read the swap into it.
 724                  */
 725                 swap_map = &si->swap_map[i];
 726                 entry = swp_entry(type, i);
 727                 page = read_swap_cache_async(entry, NULL, 0);
 728                 if (!page) {
 729                         /*
 730                          * Either swap_duplicate() failed because entry
 731                          * has been freed independently, and will not be
 732                          * reused since sys_swapoff() already disabled
 733                          * allocation from here, or alloc_page() failed.
 734                          */
 735                         if (!*swap_map)
 736                                 continue;
 737                         retval = -ENOMEM;
 738                         break;
 739                 }
 740
 741                 /*
 742                  * Don't hold on to start_mm if it looks like exiting.
 743                  */
 744                 if (atomic_read(&start_mm->mm_users) == 1) {
 745                         mmput(start_mm);
 746                         start_mm = &init_mm;
 747                         atomic_inc(&init_mm.mm_users);
 748                 }
 749
 750                 /*
 751                  * Wait for and lock page.  When do_swap_page races with
 752                  * try_to_unuse, do_swap_page can handle the fault much
 753                  * faster than try_to_unuse can locate the entry.  This
 754                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 755                  * defer to do_swap_page in such a case - in some tests,
 756                  * do_swap_page and try_to_unuse repeatedly compete.
 757                  */
 758                 wait_on_page_locked(page);
 759                 wait_on_page_writeback(page);
 760                 lock_page(page);
 761                 wait_on_page_writeback(page);
 762
 763                 /*
 764                  * Remove all references to entry.
 765                  * Whenever we reach init_mm, there's no address space
 766                  * to search, but use it as a reminder to search shmem.
 767                  */
 768                 shmem = 0;
 769                 swcount = *swap_map;
 770                 if (swcount > 1) {
 771                         if (start_mm == &init_mm)
 772                                 shmem = shmem_unuse(entry, page);
 773                         else
 774                                 retval = unuse_mm(start_mm, entry, page);
 775                 }
 776                 if (*swap_map > 1) {
 777                         int set_start_mm = (*swap_map >= swcount);
 778                         struct list_head *p = &start_mm->mmlist;
 779                         struct mm_struct *new_start_mm = start_mm;
 780                         struct mm_struct *prev_mm = start_mm;
 781                         struct mm_struct *mm;
 782
 783                         atomic_inc(&new_start_mm->mm_users);
 784                         atomic_inc(&prev_mm->mm_users);
 785                         spin_lock(&mmlist_lock);
 786                         while (*swap_map > 1 && !retval &&
 787                                         (p = p->next) != &start_mm->mmlist) {
 788                                 mm = list_entry(p, struct mm_struct, mmlist);
 789                                 if (!atomic_inc_not_zero(&mm->mm_users))
 790                                         continue;
 791                                 spin_unlock(&mmlist_lock);
 792                                 mmput(prev_mm);
 793                                 prev_mm = mm;
 794
 795                                 cond_resched();
 796
 797                                 swcount = *swap_map;
 798                                 if (swcount <= 1)
 799                                         ;
 800                                 else if (mm == &init_mm) {
 801                                         set_start_mm = 1;
 802                                         shmem = shmem_unuse(entry, page);
 803                                 } else
 804                                         retval = unuse_mm(mm, entry, page);
 805                                 if (set_start_mm && *swap_map < swcount) {
 806                                         mmput(new_start_mm);
 807                                         atomic_inc(&mm->mm_users);
 808                                         new_start_mm = mm;
 809                                         set_start_mm = 0;
 810                                 }
 811                                 spin_lock(&mmlist_lock);
 812                         }
 813                         spin_unlock(&mmlist_lock);
 814                         mmput(prev_mm);
 815                         mmput(start_mm);
 816                         start_mm = new_start_mm;
 817                 }
 818                 if (retval) {
 819                         unlock_page(page);
 820                         page_cache_release(page);
 821                         break;
 822                 }
 823
 824                 /*
 825                  * How could swap count reach 0x7fff when the maximum
 826                  * pid is 0x7fff, and there's no way to repeat a swap
 827                  * page within an mm (except in shmem, where it's the
 828                  * shared object which takes the reference count)?
 829                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 830                  *
 831                  * If that's wrong, then we should worry more about
 832                  * exit_mmap() and do_munmap() cases described above:
 833                  * we might be resetting SWAP_MAP_MAX too early here.
 834                  * We know "Undead"s can happen, they're okay, so don't
 835                  * report them; but do report if we reset SWAP_MAP_MAX.
 836                  */
 837                 if (*swap_map == SWAP_MAP_MAX) {
 838                         spin_lock(&swap_lock);
 839                         *swap_map = 1;
 840                         spin_unlock(&swap_lock);
 841                         reset_overflow = 1;
 842                 }
 843
 844                 /*
 845                  * If a reference remains (rare), we would like to leave
 846                  * the page in the swap cache; but try_to_unmap could
 847                  * then re-duplicate the entry once we drop page lock,
 848                  * so we might loop indefinitely; also, that page could
 849                  * not be swapped out to other storage meanwhile.  So:
 850                  * delete from cache even if there's another reference,
 851                  * after ensuring that the data has been saved to disk -
 852                  * since if the reference remains (rarer), it will be
 853                  * read from disk into another page.  Splitting into two
 854                  * pages would be incorrect if swap supported "shared
 855                  * private" pages, but they are handled by tmpfs files.
 856                  *
 857                  * Note shmem_unuse already deleted a swappage from
 858                  * the swap cache, unless the move to filepage failed:
 859                  * in which case it left swappage in cache, lowered its
 860                  * swap count to pass quickly through the loops above,
 861                  * and now we must reincrement count to try again later.
 862                  */
 863                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 864                         struct writeback_control wbc = {
 865                                 .sync_mode = WB_SYNC_NONE,
 866                         };
 867
 868                         swap_writepage(page, &wbc);
 869                         lock_page(page);
 870                         wait_on_page_writeback(page);
 871                 }
 872                 if (PageSwapCache(page)) {
 873                         if (shmem)
 874                                 swap_duplicate(entry);
 875                         else
 876                                 delete_from_swap_cache(page);
 877                 }
 878
 879                 /*
 880                  * So we could skip searching mms once swap count went
 881                  * to 1, we did not mark any present ptes as dirty: must
 882                  * mark page dirty so shrink_list will preserve it.
 883                  */
 884                 SetPageDirty(page);
 885                 unlock_page(page);
 886                 page_cache_release(page);
 887
 888                 /*
 889                  * Make sure that we aren't completely killing
 890                  * interactive performance.
 891                  */
 892                 cond_resched();
 893         }
 894
 895         mmput(start_mm);
 896         if (reset_overflow) {
 897                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 898                 swap_overflow = 0;
 899         }
 900         return retval;
 901 }
 902
 903 /*
 904  * After a successful try_to_unuse, if no swap is now in use, we know
 905  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 906  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 907  * added to the mmlist just after page_duplicate - before would be racy.
 908  */
 909 static void drain_mmlist(void)
 910 {
 911         struct list_head *p, *next;
 912         unsigned int i;
 913
 914         for (i = 0; i < nr_swapfiles; i++)
 915                 if (swap_info[i].inuse_pages)
 916                         return;
 917         spin_lock(&mmlist_lock);
 918         list_for_each_safe(p, next, &init_mm.mmlist)
 919                 list_del_init(p);
 920         spin_unlock(&mmlist_lock);
 921 }
 922
 923 /*
 924  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 925  * corresponds to page offset `offset'.
 926  */
 927 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 928 {
 929         struct swap_extent *se = sis->curr_swap_extent;
 930         struct swap_extent *start_se = se;
 931
 932         for ( ; ; ) {
 933                 struct list_head *lh;
 934
 935                 if (se->start_page <= offset &&
 936                                 offset < (se->start_page + se->nr_pages)) {
 937                         return se->start_block + (offset - se->start_page);
 938                 }
 939                 lh = se->list.next;
 940                 if (lh == &sis->extent_list)
 941                         lh = lh->next;
 942                 se = list_entry(lh, struct swap_extent, list);
 943                 sis->curr_swap_extent = se;
 944                 BUG_ON(se == start_se);         /* It *must* be present */
 945         }
 946 }
 947
 948 #ifdef CONFIG_SOFTWARE_SUSPEND
 949 /*
 950  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 951  * corresponding to given index in swap_info (swap type).
 952  */
 953 sector_t swapdev_block(int swap_type, pgoff_t offset)
 954 {
 955         struct swap_info_struct *sis;
 956
 957         if (swap_type >= nr_swapfiles)
 958                 return 0;
 959
 960         sis = swap_info + swap_type;
 961         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
 962 }
 963 #endif /* CONFIG_SOFTWARE_SUSPEND */
 964
 965 /*
 966  * Free all of a swapdev's extent information
 967  */
 968 static void destroy_swap_extents(struct swap_info_struct *sis)
 969 {
 970         while (!list_empty(&sis->extent_list)) {
 971                 struct swap_extent *se;
 972
 973                 se = list_entry(sis->extent_list.next,
 974                                 struct swap_extent, list);
 975                 list_del(&se->list);
 976                 kfree(se);
 977         }
 978 }
 979
 980 /*
 981  * Add a block range (and the corresponding page range) into this swapdev's
 982  * extent list.  The extent list is kept sorted in page order.
 983  *
 984  * This function rather assumes that it is called in ascending page order.
 985  */
 986 static int
 987 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 988                 unsigned long nr_pages, sector_t start_block)
 989 {
 990         struct swap_extent *se;
 991         struct swap_extent *new_se;
 992         struct list_head *lh;
 993
 994         lh = sis->extent_list.prev;     /* The highest page extent */
 995         if (lh != &sis->extent_list) {
 996                 se = list_entry(lh, struct swap_extent, list);
 997                 BUG_ON(se->start_page + se->nr_pages != start_page);
 998                 if (se->start_block + se->nr_pages == start_block) {
 999                         /* Merge it */
1000                         se->nr_pages += nr_pages;
1001                         return 0;
1002                 }
1003         }
1004
1005         /*
1006          * No merge.  Insert a new extent, preserving ordering.
1007          */
1008         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1009         if (new_se == NULL)
1010                 return -ENOMEM;
1011         new_se->start_page = start_page;
1012         new_se->nr_pages = nr_pages;
1013         new_se->start_block = start_block;
1014
1015         list_add_tail(&new_se->list, &sis->extent_list);
1016         return 1;
1017 }
1018
1019 /*
1020  * A `swap extent' is a simple thing which maps a contiguous range of pages
1021  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1022  * is built at swapon time and is then used at swap_writepage/swap_readpage
1023  * time for locating where on disk a page belongs.
1024  *
1025  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1026  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1027  * swap files identically.
1028  *
1029  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1030  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1031  * swapfiles are handled *identically* after swapon time.
1032  *
1033  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1034  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1035  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1036  * requirements, they are simply tossed out - we will never use those blocks
1037  * for swapping.
1038  *
1039  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1040  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1041  * which will scribble on the fs.
1042  *
1043  * The amount of disk space which a single swap extent represents varies.
1044  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1045  * extents in the list.  To avoid much list walking, we cache the previous
1046  * search location in `curr_swap_extent', and start new searches from there.
1047  * This is extremely effective.  The average number of iterations in
1048  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1049  */
1050 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1051 {
1052         struct inode *inode;
1053         unsigned blocks_per_page;
1054         unsigned long page_no;
1055         unsigned blkbits;
1056         sector_t probe_block;
1057         sector_t last_block;
1058         sector_t lowest_block = -1;
1059         sector_t highest_block = 0;
1060         int nr_extents = 0;
1061         int ret;
1062
1063         inode = sis->swap_file->f_mapping->host;
1064         if (S_ISBLK(inode->i_mode)) {
1065                 ret = add_swap_extent(sis, 0, sis->max, 0);
1066                 *span = sis->pages;
1067                 goto done;
1068         }
1069
1070         blkbits = inode->i_blkbits;
1071         blocks_per_page = PAGE_SIZE >> blkbits;
1072
1073         /*
1074          * Map all the blocks into the extent list.  This code doesn't try
1075          * to be very smart.
1076          */
1077         probe_block = 0;
1078         page_no = 0;
1079         last_block = i_size_read(inode) >> blkbits;
1080         while ((probe_block + blocks_per_page) <= last_block &&
1081                         page_no < sis->max) {
1082                 unsigned block_in_page;
1083                 sector_t first_block;
1084
1085                 first_block = bmap(inode, probe_block);
1086                 if (first_block == 0)
1087                         goto bad_bmap;
1088
1089                 /*
1090                  * It must be PAGE_SIZE aligned on-disk
1091                  */
1092                 if (first_block & (blocks_per_page - 1)) {
1093                         probe_block++;
1094                         goto reprobe;
1095                 }
1096
1097                 for (block_in_page = 1; block_in_page < blocks_per_page;
1098                                         block_in_page++) {
1099                         sector_t block;
1100
1101                         block = bmap(inode, probe_block + block_in_page);
1102                         if (block == 0)
1103                                 goto bad_bmap;
1104                         if (block != first_block + block_in_page) {
1105                                 /* Discontiguity */
1106                                 probe_block++;
1107                                 goto reprobe;
1108                         }
1109                 }
1110
1111                 first_block >>= (PAGE_SHIFT - blkbits);
1112                 if (page_no) {  /* exclude the header page */
1113                         if (first_block < lowest_block)
1114                                 lowest_block = first_block;
1115                         if (first_block > highest_block)
1116                                 highest_block = first_block;
1117                 }
1118
1119                 /*
1120                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1121                  */
1122                 ret = add_swap_extent(sis, page_no, 1, first_block);
1123                 if (ret < 0)
1124                         goto out;
1125                 nr_extents += ret;
1126                 page_no++;
1127                 probe_block += blocks_per_page;
1128 reprobe:
1129                 continue;
1130         }
1131         ret = nr_extents;
1132         *span = 1 + highest_block - lowest_block;
1133         if (page_no == 0)
1134                 page_no = 1;    /* force Empty message */
1135         sis->max = page_no;
1136         sis->pages = page_no - 1;
1137         sis->highest_bit = page_no - 1;
1138 done:
1139         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1140                                         struct swap_extent, list);
1141         goto out;
1142 bad_bmap:
1143         printk(KERN_ERR "swapon: swapfile has holes\n");
1144         ret = -EINVAL;
1145 out:
1146         return ret;
1147 }
1148
1149 #if 0   /* We don't need this yet */
1150 #include <linux/backing-dev.h>
1151 int page_queue_congested(struct page *page)
1152 {
1153         struct backing_dev_info *bdi;
1154
1155         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1156
1157         if (PageSwapCache(page)) {
1158                 swp_entry_t entry = { .val = page_private(page) };
1159                 struct swap_info_struct *sis;
1160
1161                 sis = get_swap_info_struct(swp_type(entry));
1162                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1163         } else
1164                 bdi = page->mapping->backing_dev_info;
1165         return bdi_write_congested(bdi);
1166 }
1167 #endif
1168
1169 asmlinkage long sys_swapoff(const char __user * specialfile)
1170 {
1171         struct swap_info_struct * p = NULL;
1172         unsigned short *swap_map;
1173         struct file *swap_file, *victim;
1174         struct address_space *mapping;
1175         struct inode *inode;
1176         char * pathname;
1177         int i, type, prev;
1178         int err;
1179
1180         if (!capable(CAP_SYS_ADMIN))
1181                 return -EPERM;
1182
1183         pathname = getname(specialfile);
1184         err = PTR_ERR(pathname);
1185         if (IS_ERR(pathname))
1186                 goto out;
1187
1188         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1189         putname(pathname);
1190         err = PTR_ERR(victim);
1191         if (IS_ERR(victim))
1192                 goto out;
1193
1194         mapping = victim->f_mapping;
1195         prev = -1;
1196         spin_lock(&swap_lock);
1197         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1198                 p = swap_info + type;
1199                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1200                         if (p->swap_file->f_mapping == mapping)
1201                                 break;
1202                 }
1203                 prev = type;
1204         }
1205         if (type < 0) {
1206                 err = -EINVAL;
1207                 spin_unlock(&swap_lock);
1208                 goto out_dput;
1209         }
1210         if (!security_vm_enough_memory(p->pages))
1211                 vm_unacct_memory(p->pages);
1212         else {
1213                 err = -ENOMEM;
1214                 spin_unlock(&swap_lock);
1215                 goto out_dput;
1216         }
1217         if (prev < 0) {
1218                 swap_list.head = p->next;
1219         } else {
1220                 swap_info[prev].next = p->next;
1221         }
1222         if (type == swap_list.next) {
1223                 /* just pick something that's safe... */
1224                 swap_list.next = swap_list.head;
1225         }
1226         nr_swap_pages -= p->pages;
1227         total_swap_pages -= p->pages;
1228         p->flags &= ~SWP_WRITEOK;
1229         spin_unlock(&swap_lock);
1230
1231         current->flags |= PF_SWAPOFF;
1232         err = try_to_unuse(type);
1233         current->flags &= ~PF_SWAPOFF;
1234
1235         if (err) {
1236                 /* re-insert swap space back into swap_list */
1237                 spin_lock(&swap_lock);
1238                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1239                         if (p->prio >= swap_info[i].prio)
1240                                 break;
1241                 p->next = i;
1242                 if (prev < 0)
1243                         swap_list.head = swap_list.next = p - swap_info;
1244                 else
1245                         swap_info[prev].next = p - swap_info;
1246                 nr_swap_pages += p->pages;
1247                 total_swap_pages += p->pages;
1248                 p->flags |= SWP_WRITEOK;
1249                 spin_unlock(&swap_lock);
1250                 goto out_dput;
1251         }
1252
1253         /* wait for any unplug function to finish */
1254         down_write(&swap_unplug_sem);
1255         up_write(&swap_unplug_sem);
1256
1257         destroy_swap_extents(p);
1258         mutex_lock(&swapon_mutex);
1259         spin_lock(&swap_lock);
1260         drain_mmlist();
1261
1262         /* wait for anyone still in scan_swap_map */
1263         p->highest_bit = 0;             /* cuts scans short */
1264         while (p->flags >= SWP_SCANNING) {
1265                 spin_unlock(&swap_lock);
1266                 schedule_timeout_uninterruptible(1);
1267                 spin_lock(&swap_lock);
1268         }
1269
1270         swap_file = p->swap_file;
1271         p->swap_file = NULL;
1272         p->max = 0;
1273         swap_map = p->swap_map;
1274         p->swap_map = NULL;
1275         p->flags = 0;
1276         spin_unlock(&swap_lock);
1277         mutex_unlock(&swapon_mutex);
1278         vfree(swap_map);
1279         inode = mapping->host;
1280         if (S_ISBLK(inode->i_mode)) {
1281                 struct block_device *bdev = I_BDEV(inode);
1282                 set_blocksize(bdev, p->old_block_size);
1283                 bd_release(bdev);
1284         } else {
1285                 mutex_lock(&inode->i_mutex);
1286                 inode->i_flags &= ~S_SWAPFILE;
1287                 mutex_unlock(&inode->i_mutex);
1288         }
1289         filp_close(swap_file, NULL);
1290         err = 0;
1291
1292 out_dput:
1293         filp_close(victim, NULL);
1294 out:
1295         return err;
1296 }
1297
1298 #ifdef CONFIG_PROC_FS
1299 /* iterator */
1300 static void *swap_start(struct seq_file *swap, loff_t *pos)
1301 {
1302         struct swap_info_struct *ptr = swap_info;
1303         int i;
1304         loff_t l = *pos;
1305
1306         mutex_lock(&swapon_mutex);
1307
1308         if (!l)
1309                 return SEQ_START_TOKEN;
1310
1311         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1312                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1313                         continue;
1314                 if (!--l)
1315                         return ptr;
1316         }
1317
1318         return NULL;
1319 }
1320
1321 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1322 {
1323         struct swap_info_struct *ptr;
1324         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1325
1326         if (v == SEQ_START_TOKEN)
1327                 ptr = swap_info;
1328         else {
1329                 ptr = v;
1330                 ptr++;
1331         }
1332
1333         for (; ptr < endptr; ptr++) {
1334                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1335                         continue;
1336                 ++*pos;
1337                 return ptr;
1338         }
1339
1340         return NULL;
1341 }
1342
1343 static void swap_stop(struct seq_file *swap, void *v)
1344 {
1345         mutex_unlock(&swapon_mutex);
1346 }
1347
1348 static int swap_show(struct seq_file *swap, void *v)
1349 {
1350         struct swap_info_struct *ptr = v;
1351         struct file *file;
1352         int len;
1353
1354         if (ptr == SEQ_START_TOKEN) {
1355                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1356                 return 0;
1357         }
1358
1359         file = ptr->swap_file;
1360         len = seq_path(swap, file->f_path.mnt, file->f_path.dentry, " \t\n\\");
1361         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1362                        len < 40 ? 40 - len : 1, " ",
1363                        S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1364                                 "partition" : "file\t",
1365                        ptr->pages << (PAGE_SHIFT - 10),
1366                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1367                        ptr->prio);
1368         return 0;
1369 }
1370
1371 static const struct seq_operations swaps_op = {
1372         .start =        swap_start,
1373         .next =         swap_next,
1374         .stop =         swap_stop,
1375         .show =         swap_show
1376 };
1377
1378 static int swaps_open(struct inode *inode, struct file *file)
1379 {
1380         return seq_open(file, &swaps_op);
1381 }
1382
1383 static const struct file_operations proc_swaps_operations = {
1384         .open           = swaps_open,
1385         .read           = seq_read,
1386         .llseek         = seq_lseek,
1387         .release        = seq_release,
1388 };
1389
1390 static int __init procswaps_init(void)
1391 {
1392         struct proc_dir_entry *entry;
1393
1394         entry = create_proc_entry("swaps", 0, NULL);
1395         if (entry)
1396                 entry->proc_fops = &proc_swaps_operations;
1397         return 0;
1398 }
1399 __initcall(procswaps_init);
1400 #endif /* CONFIG_PROC_FS */
1401
1402 /*
1403  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1404  *
1405  * The swapon system call
1406  */
1407 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1408 {
1409         struct swap_info_struct * p;
1410         char *name = NULL;
1411         struct block_device *bdev = NULL;
1412         struct file *swap_file = NULL;
1413         struct address_space *mapping;
1414         unsigned int type;
1415         int i, prev;
1416         int error;
1417         static int least_priority;
1418         union swap_header *swap_header = NULL;
1419         int swap_header_version;
1420         unsigned int nr_good_pages = 0;
1421         int nr_extents = 0;
1422         sector_t span;
1423         unsigned long maxpages = 1;
1424         int swapfilesize;
1425         unsigned short *swap_map;
1426         struct page *page = NULL;
1427         struct inode *inode = NULL;
1428         int did_down = 0;
1429
1430         if (!capable(CAP_SYS_ADMIN))
1431                 return -EPERM;
1432         spin_lock(&swap_lock);
1433         p = swap_info;
1434         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1435                 if (!(p->flags & SWP_USED))
1436                         break;
1437         error = -EPERM;
1438         if (type >= MAX_SWAPFILES) {
1439                 spin_unlock(&swap_lock);
1440                 goto out;
1441         }
1442         if (type >= nr_swapfiles)
1443                 nr_swapfiles = type+1;
1444         INIT_LIST_HEAD(&p->extent_list);
1445         p->flags = SWP_USED;
1446         p->swap_file = NULL;
1447         p->old_block_size = 0;
1448         p->swap_map = NULL;
1449         p->lowest_bit = 0;
1450         p->highest_bit = 0;
1451         p->cluster_nr = 0;
1452         p->inuse_pages = 0;
1453         p->next = -1;
1454         if (swap_flags & SWAP_FLAG_PREFER) {
1455                 p->prio =
1456                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1457         } else {
1458                 p->prio = --least_priority;
1459         }
1460         spin_unlock(&swap_lock);
1461         name = getname(specialfile);
1462         error = PTR_ERR(name);
1463         if (IS_ERR(name)) {
1464                 name = NULL;
1465                 goto bad_swap_2;
1466         }
1467         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1468         error = PTR_ERR(swap_file);
1469         if (IS_ERR(swap_file)) {
1470                 swap_file = NULL;
1471                 goto bad_swap_2;
1472         }
1473
1474         p->swap_file = swap_file;
1475         mapping = swap_file->f_mapping;
1476         inode = mapping->host;
1477
1478         error = -EBUSY;
1479         for (i = 0; i < nr_swapfiles; i++) {
1480                 struct swap_info_struct *q = &swap_info[i];
1481
1482                 if (i == type || !q->swap_file)
1483                         continue;
1484                 if (mapping == q->swap_file->f_mapping)
1485                         goto bad_swap;
1486         }
1487
1488         error = -EINVAL;
1489         if (S_ISBLK(inode->i_mode)) {
1490                 bdev = I_BDEV(inode);
1491                 error = bd_claim(bdev, sys_swapon);
1492                 if (error < 0) {
1493                         bdev = NULL;
1494                         error = -EINVAL;
1495                         goto bad_swap;
1496                 }
1497                 p->old_block_size = block_size(bdev);
1498                 error = set_blocksize(bdev, PAGE_SIZE);
1499                 if (error < 0)
1500                         goto bad_swap;
1501                 p->bdev = bdev;
1502         } else if (S_ISREG(inode->i_mode)) {
1503                 p->bdev = inode->i_sb->s_bdev;
1504                 mutex_lock(&inode->i_mutex);
1505                 did_down = 1;
1506                 if (IS_SWAPFILE(inode)) {
1507                         error = -EBUSY;
1508                         goto bad_swap;
1509                 }
1510         } else {
1511                 goto bad_swap;
1512         }
1513
1514         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1515
1516         /*
1517          * Read the swap header.
1518          */
1519         if (!mapping->a_ops->readpage) {
1520                 error = -EINVAL;
1521                 goto bad_swap;
1522         }
1523         page = read_mapping_page(mapping, 0, swap_file);
1524         if (IS_ERR(page)) {
1525                 error = PTR_ERR(page);
1526                 goto bad_swap;
1527         }
1528         wait_on_page_locked(page);
1529         if (!PageUptodate(page))
1530                 goto bad_swap;
1531         kmap(page);
1532         swap_header = page_address(page);
1533
1534         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1535                 swap_header_version = 1;
1536         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1537                 swap_header_version = 2;
1538         else {
1539                 printk(KERN_ERR "Unable to find swap-space signature\n");
1540                 error = -EINVAL;
1541                 goto bad_swap;
1542         }
1543
1544         switch (swap_header_version) {
1545         case 1:
1546                 printk(KERN_ERR "version 0 swap is no longer supported. "
1547                         "Use mkswap -v1 %s\n", name);
1548                 error = -EINVAL;
1549                 goto bad_swap;
1550         case 2:
1551                 /* Check the swap header's sub-version and the size of
1552                    the swap file and bad block lists */
1553                 if (swap_header->info.version != 1) {
1554                         printk(KERN_WARNING
1555                                "Unable to handle swap header version %d\n",
1556                                swap_header->info.version);
1557                         error = -EINVAL;
1558                         goto bad_swap;
1559                 }
1560
1561                 p->lowest_bit  = 1;
1562                 p->cluster_next = 1;
1563
1564                 /*
1565                  * Find out how many pages are allowed for a single swap
1566                  * device. There are two limiting factors: 1) the number of
1567                  * bits for the swap offset in the swp_entry_t type and
1568                  * 2) the number of bits in the a swap pte as defined by
1569                  * the different architectures. In order to find the
1570                  * largest possible bit mask a swap entry with swap type 0
1571                  * and swap offset ~0UL is created, encoded to a swap pte,
1572                  * decoded to a swp_entry_t again and finally the swap
1573                  * offset is extracted. This will mask all the bits from
1574                  * the initial ~0UL mask that can't be encoded in either
1575                  * the swp_entry_t or the architecture definition of a
1576                  * swap pte.
1577                  */
1578                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1579                 if (maxpages > swap_header->info.last_page)
1580                         maxpages = swap_header->info.last_page;
1581                 p->highest_bit = maxpages - 1;
1582
1583                 error = -EINVAL;
1584                 if (!maxpages)
1585                         goto bad_swap;
1586                 if (swapfilesize && maxpages > swapfilesize) {
1587                         printk(KERN_WARNING
1588                                "Swap area shorter than signature indicates\n");
1589                         goto bad_swap;
1590                 }
1591                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1592                         goto bad_swap;
1593                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1594                         goto bad_swap;
1595
1596                 /* OK, set up the swap map and apply the bad block list */
1597                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1598                         error = -ENOMEM;
1599                         goto bad_swap;
1600                 }
1601
1602                 error = 0;
1603                 memset(p->swap_map, 0, maxpages * sizeof(short));
1604                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1605                         int page_nr = swap_header->info.badpages[i];
1606                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1607                                 error = -EINVAL;
1608                         else
1609                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1610                 }
1611                 nr_good_pages = swap_header->info.last_page -
1612                                 swap_header->info.nr_badpages -
1613                                 1 /* header page */;
1614                 if (error)
1615                         goto bad_swap;
1616         }
1617
1618         if (nr_good_pages) {
1619                 p->swap_map[0] = SWAP_MAP_BAD;
1620                 p->max = maxpages;
1621                 p->pages = nr_good_pages;
1622                 nr_extents = setup_swap_extents(p, &span);
1623                 if (nr_extents < 0) {
1624                         error = nr_extents;
1625                         goto bad_swap;
1626                 }
1627                 nr_good_pages = p->pages;
1628         }
1629         if (!nr_good_pages) {
1630                 printk(KERN_WARNING "Empty swap-file\n");
1631                 error = -EINVAL;
1632                 goto bad_swap;
1633         }
1634
1635         mutex_lock(&swapon_mutex);
1636         spin_lock(&swap_lock);
1637         p->flags = SWP_ACTIVE;
1638         nr_swap_pages += nr_good_pages;
1639         total_swap_pages += nr_good_pages;
1640
1641         printk(KERN_INFO "Adding %uk swap on %s.  "
1642                         "Priority:%d extents:%d across:%lluk\n",
1643                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1644                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1645
1646         /* insert swap space into swap_list: */
1647         prev = -1;
1648         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1649                 if (p->prio >= swap_info[i].prio) {
1650                         break;
1651                 }
1652                 prev = i;
1653         }
1654         p->next = i;
1655         if (prev < 0) {
1656                 swap_list.head = swap_list.next = p - swap_info;
1657         } else {
1658                 swap_info[prev].next = p - swap_info;
1659         }
1660         spin_unlock(&swap_lock);
1661         mutex_unlock(&swapon_mutex);
1662         error = 0;
1663         goto out;
1664 bad_swap:
1665         if (bdev) {
1666                 set_blocksize(bdev, p->old_block_size);
1667                 bd_release(bdev);
1668         }
1669         destroy_swap_extents(p);
1670 bad_swap_2:
1671         spin_lock(&swap_lock);
1672         swap_map = p->swap_map;
1673         p->swap_file = NULL;
1674         p->swap_map = NULL;
1675         p->flags = 0;
1676         if (!(swap_flags & SWAP_FLAG_PREFER))
1677                 ++least_priority;
1678         spin_unlock(&swap_lock);
1679         vfree(swap_map);
1680         if (swap_file)
1681                 filp_close(swap_file, NULL);
1682 out:
1683         if (page && !IS_ERR(page)) {
1684                 kunmap(page);
1685                 page_cache_release(page);
1686         }
1687         if (name)
1688                 putname(name);
1689         if (did_down) {
1690                 if (!error)
1691                         inode->i_flags |= S_SWAPFILE;
1692                 mutex_unlock(&inode->i_mutex);
1693         }
1694         return error;
1695 }
1696
1697 void si_swapinfo(struct sysinfo *val)
1698 {
1699         unsigned int i;
1700         unsigned long nr_to_be_unused = 0;
1701
1702         spin_lock(&swap_lock);
1703         for (i = 0; i < nr_swapfiles; i++) {
1704                 if (!(swap_info[i].flags & SWP_USED) ||
1705                      (swap_info[i].flags & SWP_WRITEOK))
1706                         continue;
1707                 nr_to_be_unused += swap_info[i].inuse_pages;
1708         }
1709         val->freeswap = nr_swap_pages + nr_to_be_unused;
1710         val->totalswap = total_swap_pages + nr_to_be_unused;
1711         spin_unlock(&swap_lock);
1712 }
1713
1714 /*
1715  * Verify that a swap entry is valid and increment its swap map count.
1716  *
1717  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1718  * "permanent", but will be reclaimed by the next swapoff.
1719  */
1720 int swap_duplicate(swp_entry_t entry)
1721 {
1722         struct swap_info_struct * p;
1723         unsigned long offset, type;
1724         int result = 0;
1725
1726         if (is_migration_entry(entry))
1727                 return 1;
1728
1729         type = swp_type(entry);
1730         if (type >= nr_swapfiles)
1731                 goto bad_file;
1732         p = type + swap_info;
1733         offset = swp_offset(entry);
1734
1735         spin_lock(&swap_lock);
1736         if (offset < p->max && p->swap_map[offset]) {
1737                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1738                         p->swap_map[offset]++;
1739                         result = 1;
1740                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1741                         if (swap_overflow++ < 5)
1742                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1743                         p->swap_map[offset] = SWAP_MAP_MAX;
1744                         result = 1;
1745                 }
1746         }
1747         spin_unlock(&swap_lock);
1748 out:
1749         return result;
1750
1751 bad_file:
1752         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1753         goto out;
1754 }
1755
1756 struct swap_info_struct *
1757 get_swap_info_struct(unsigned type)
1758 {
1759         return &swap_info[type];
1760 }
1761
1762 /*
1763  * swap_lock prevents swap_map being freed. Don't grab an extra
1764  * reference on the swaphandle, it doesn't matter if it becomes unused.
1765  */
1766 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1767 {
1768         int our_page_cluster = page_cluster;
1769         int ret = 0, i = 1 << our_page_cluster;
1770         unsigned long toff;
1771         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1772
1773         if (!our_page_cluster)  /* no readahead */
1774                 return 0;
1775         toff = (swp_offset(entry) >> our_page_cluster) << our_page_cluster;
1776         if (!toff)              /* first page is swap header */
1777                 toff++, i--;
1778         *offset = toff;
1779
1780         spin_lock(&swap_lock);
1781         do {
1782                 /* Don't read-ahead past the end of the swap area */
1783                 if (toff >= swapdev->max)
1784                         break;
1785                 /* Don't read in free or bad pages */
1786                 if (!swapdev->swap_map[toff])
1787                         break;
1788                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1789                         break;
1790                 toff++;
1791                 ret++;
1792         } while (--i);
1793         spin_unlock(&swap_lock);
1794         return ret;
1795 }