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_HIBERNATION
 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, struct block_device **bdev_p)
 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                         if (bdev_p)
 454                                 *bdev_p = sis->bdev;
 455
 456                         spin_unlock(&swap_lock);
 457                         return i;
 458                 }
 459                 if (bdev == sis->bdev) {
 460                         struct swap_extent *se;
 461
 462                         se = list_entry(sis->extent_list.next,
 463                                         struct swap_extent, list);
 464                         if (se->start_block == offset) {
 465                                 if (bdev_p)
 466                                         *bdev_p = sis->bdev;
 467
 468                                 spin_unlock(&swap_lock);
 469                                 bdput(bdev);
 470                                 return i;
 471                         }
 472                 }
 473         }
 474         spin_unlock(&swap_lock);
 475         if (bdev)
 476                 bdput(bdev);
 477
 478         return -ENODEV;
 479 }
 480
 481 /*
 482  * Return either the total number of swap pages of given type, or the number
 483  * of free pages of that type (depending on @free)
 484  *
 485  * This is needed for software suspend
 486  */
 487 unsigned int count_swap_pages(int type, int free)
 488 {
 489         unsigned int n = 0;
 490
 491         if (type < nr_swapfiles) {
 492                 spin_lock(&swap_lock);
 493                 if (swap_info[type].flags & SWP_WRITEOK) {
 494                         n = swap_info[type].pages;
 495                         if (free)
 496                                 n -= swap_info[type].inuse_pages;
 497                 }
 498                 spin_unlock(&swap_lock);
 499         }
 500         return n;
 501 }
 502 #endif
 503
 504 /*
 505  * No need to decide whether this PTE shares the swap entry with others,
 506  * just let do_wp_page work it out if a write is requested later - to
 507  * force COW, vm_page_prot omits write permission from any private vma.
 508  */
 509 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 510                 unsigned long addr, swp_entry_t entry, struct page *page)
 511 {
 512         spinlock_t *ptl;
 513         pte_t *pte;
 514         int found = 1;
 515
 516         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 517         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 518                 found = 0;
 519                 goto out;
 520         }
 521
 522         inc_mm_counter(vma->vm_mm, anon_rss);
 523         get_page(page);
 524         set_pte_at(vma->vm_mm, addr, pte,
 525                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 526         page_add_anon_rmap(page, vma, addr);
 527         swap_free(entry);
 528         /*
 529          * Move the page to the active list so it is not
 530          * immediately swapped out again after swapon.
 531          */
 532         activate_page(page);
 533 out:
 534         pte_unmap_unlock(pte, ptl);
 535         return found;
 536 }
 537
 538 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 539                                 unsigned long addr, unsigned long end,
 540                                 swp_entry_t entry, struct page *page)
 541 {
 542         pte_t swp_pte = swp_entry_to_pte(entry);
 543         pte_t *pte;
 544         int found = 0;
 545
 546         /*
 547          * We don't actually need pte lock while scanning for swp_pte: since
 548          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 549          * page table while we're scanning; though it could get zapped, and on
 550          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 551          * of unmatched parts which look like swp_pte, so unuse_pte must
 552          * recheck under pte lock.  Scanning without pte lock lets it be
 553          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 554          */
 555         pte = pte_offset_map(pmd, addr);
 556         do {
 557                 /*
 558                  * swapoff spends a _lot_ of time in this loop!
 559                  * Test inline before going to call unuse_pte.
 560                  */
 561                 if (unlikely(pte_same(*pte, swp_pte))) {
 562                         pte_unmap(pte);
 563                         found = unuse_pte(vma, pmd, addr, entry, page);
 564                         if (found)
 565                                 goto out;
 566                         pte = pte_offset_map(pmd, addr);
 567                 }
 568         } while (pte++, addr += PAGE_SIZE, addr != end);
 569         pte_unmap(pte - 1);
 570 out:
 571         return found;
 572 }
 573
 574 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 575                                 unsigned long addr, unsigned long end,
 576                                 swp_entry_t entry, struct page *page)
 577 {
 578         pmd_t *pmd;
 579         unsigned long next;
 580
 581         pmd = pmd_offset(pud, addr);
 582         do {
 583                 next = pmd_addr_end(addr, end);
 584                 if (pmd_none_or_clear_bad(pmd))
 585                         continue;
 586                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 587                         return 1;
 588         } while (pmd++, addr = next, addr != end);
 589         return 0;
 590 }
 591
 592 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 593                                 unsigned long addr, unsigned long end,
 594                                 swp_entry_t entry, struct page *page)
 595 {
 596         pud_t *pud;
 597         unsigned long next;
 598
 599         pud = pud_offset(pgd, addr);
 600         do {
 601                 next = pud_addr_end(addr, end);
 602                 if (pud_none_or_clear_bad(pud))
 603                         continue;
 604                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 605                         return 1;
 606         } while (pud++, addr = next, addr != end);
 607         return 0;
 608 }
 609
 610 static int unuse_vma(struct vm_area_struct *vma,
 611                                 swp_entry_t entry, struct page *page)
 612 {
 613         pgd_t *pgd;
 614         unsigned long addr, end, next;
 615
 616         if (page->mapping) {
 617                 addr = page_address_in_vma(page, vma);
 618                 if (addr == -EFAULT)
 619                         return 0;
 620                 else
 621                         end = addr + PAGE_SIZE;
 622         } else {
 623                 addr = vma->vm_start;
 624                 end = vma->vm_end;
 625         }
 626
 627         pgd = pgd_offset(vma->vm_mm, addr);
 628         do {
 629                 next = pgd_addr_end(addr, end);
 630                 if (pgd_none_or_clear_bad(pgd))
 631                         continue;
 632                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 633                         return 1;
 634         } while (pgd++, addr = next, addr != end);
 635         return 0;
 636 }
 637
 638 static int unuse_mm(struct mm_struct *mm,
 639                                 swp_entry_t entry, struct page *page)
 640 {
 641         struct vm_area_struct *vma;
 642
 643         if (!down_read_trylock(&mm->mmap_sem)) {
 644                 /*
 645                  * Activate page so shrink_cache is unlikely to unmap its
 646                  * ptes while lock is dropped, so swapoff can make progress.
 647                  */
 648                 activate_page(page);
 649                 unlock_page(page);
 650                 down_read(&mm->mmap_sem);
 651                 lock_page(page);
 652         }
 653         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 654                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 655                         break;
 656         }
 657         up_read(&mm->mmap_sem);
 658         /*
 659          * Currently unuse_mm cannot fail, but leave error handling
 660          * at call sites for now, since we change it from time to time.
 661          */
 662         return 0;
 663 }
 664
 665 /*
 666  * Scan swap_map from current position to next entry still in use.
 667  * Recycle to start on reaching the end, returning 0 when empty.
 668  */
 669 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 670                                         unsigned int prev)
 671 {
 672         unsigned int max = si->max;
 673         unsigned int i = prev;
 674         int count;
 675
 676         /*
 677          * No need for swap_lock here: we're just looking
 678          * for whether an entry is in use, not modifying it; false
 679          * hits are okay, and sys_swapoff() has already prevented new
 680          * allocations from this area (while holding swap_lock).
 681          */
 682         for (;;) {
 683                 if (++i >= max) {
 684                         if (!prev) {
 685                                 i = 0;
 686                                 break;
 687                         }
 688                         /*
 689                          * No entries in use at top of swap_map,
 690                          * loop back to start and recheck there.
 691                          */
 692                         max = prev + 1;
 693                         prev = 0;
 694                         i = 1;
 695                 }
 696                 count = si->swap_map[i];
 697                 if (count && count != SWAP_MAP_BAD)
 698                         break;
 699         }
 700         return i;
 701 }
 702
 703 /*
 704  * We completely avoid races by reading each swap page in advance,
 705  * and then search for the process using it.  All the necessary
 706  * page table adjustments can then be made atomically.
 707  */
 708 static int try_to_unuse(unsigned int type)
 709 {
 710         struct swap_info_struct * si = &swap_info[type];
 711         struct mm_struct *start_mm;
 712         unsigned short *swap_map;
 713         unsigned short swcount;
 714         struct page *page;
 715         swp_entry_t entry;
 716         unsigned int i = 0;
 717         int retval = 0;
 718         int reset_overflow = 0;
 719         int shmem;
 720
 721         /*
 722          * When searching mms for an entry, a good strategy is to
 723          * start at the first mm we freed the previous entry from
 724          * (though actually we don't notice whether we or coincidence
 725          * freed the entry).  Initialize this start_mm with a hold.
 726          *
 727          * A simpler strategy would be to start at the last mm we
 728          * freed the previous entry from; but that would take less
 729          * advantage of mmlist ordering, which clusters forked mms
 730          * together, child after parent.  If we race with dup_mmap(), we
 731          * prefer to resolve parent before child, lest we miss entries
 732          * duplicated after we scanned child: using last mm would invert
 733          * that.  Though it's only a serious concern when an overflowed
 734          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 735          */
 736         start_mm = &init_mm;
 737         atomic_inc(&init_mm.mm_users);
 738
 739         /*
 740          * Keep on scanning until all entries have gone.  Usually,
 741          * one pass through swap_map is enough, but not necessarily:
 742          * there are races when an instance of an entry might be missed.
 743          */
 744         while ((i = find_next_to_unuse(si, i)) != 0) {
 745                 if (signal_pending(current)) {
 746                         retval = -EINTR;
 747                         break;
 748                 }
 749
 750                 /*
 751                  * Get a page for the entry, using the existing swap
 752                  * cache page if there is one.  Otherwise, get a clean
 753                  * page and read the swap into it.
 754                  */
 755                 swap_map = &si->swap_map[i];
 756                 entry = swp_entry(type, i);
 757                 page = read_swap_cache_async(entry,
 758                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
 759                 if (!page) {
 760                         /*
 761                          * Either swap_duplicate() failed because entry
 762                          * has been freed independently, and will not be
 763                          * reused since sys_swapoff() already disabled
 764                          * allocation from here, or alloc_page() failed.
 765                          */
 766                         if (!*swap_map)
 767                                 continue;
 768                         retval = -ENOMEM;
 769                         break;
 770                 }
 771
 772                 /*
 773                  * Don't hold on to start_mm if it looks like exiting.
 774                  */
 775                 if (atomic_read(&start_mm->mm_users) == 1) {
 776                         mmput(start_mm);
 777                         start_mm = &init_mm;
 778                         atomic_inc(&init_mm.mm_users);
 779                 }
 780
 781                 /*
 782                  * Wait for and lock page.  When do_swap_page races with
 783                  * try_to_unuse, do_swap_page can handle the fault much
 784                  * faster than try_to_unuse can locate the entry.  This
 785                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 786                  * defer to do_swap_page in such a case - in some tests,
 787                  * do_swap_page and try_to_unuse repeatedly compete.
 788                  */
 789                 wait_on_page_locked(page);
 790                 wait_on_page_writeback(page);
 791                 lock_page(page);
 792                 wait_on_page_writeback(page);
 793
 794                 /*
 795                  * Remove all references to entry.
 796                  * Whenever we reach init_mm, there's no address space
 797                  * to search, but use it as a reminder to search shmem.
 798                  */
 799                 shmem = 0;
 800                 swcount = *swap_map;
 801                 if (swcount > 1) {
 802                         if (start_mm == &init_mm)
 803                                 shmem = shmem_unuse(entry, page);
 804                         else
 805                                 retval = unuse_mm(start_mm, entry, page);
 806                 }
 807                 if (*swap_map > 1) {
 808                         int set_start_mm = (*swap_map >= swcount);
 809                         struct list_head *p = &start_mm->mmlist;
 810                         struct mm_struct *new_start_mm = start_mm;
 811                         struct mm_struct *prev_mm = start_mm;
 812                         struct mm_struct *mm;
 813
 814                         atomic_inc(&new_start_mm->mm_users);
 815                         atomic_inc(&prev_mm->mm_users);
 816                         spin_lock(&mmlist_lock);
 817                         while (*swap_map > 1 && !retval &&
 818                                         (p = p->next) != &start_mm->mmlist) {
 819                                 mm = list_entry(p, struct mm_struct, mmlist);
 820                                 if (!atomic_inc_not_zero(&mm->mm_users))
 821                                         continue;
 822                                 spin_unlock(&mmlist_lock);
 823                                 mmput(prev_mm);
 824                                 prev_mm = mm;
 825
 826                                 cond_resched();
 827
 828                                 swcount = *swap_map;
 829                                 if (swcount <= 1)
 830                                         ;
 831                                 else if (mm == &init_mm) {
 832                                         set_start_mm = 1;
 833                                         shmem = shmem_unuse(entry, page);
 834                                 } else
 835                                         retval = unuse_mm(mm, entry, page);
 836                                 if (set_start_mm && *swap_map < swcount) {
 837                                         mmput(new_start_mm);
 838                                         atomic_inc(&mm->mm_users);
 839                                         new_start_mm = mm;
 840                                         set_start_mm = 0;
 841                                 }
 842                                 spin_lock(&mmlist_lock);
 843                         }
 844                         spin_unlock(&mmlist_lock);
 845                         mmput(prev_mm);
 846                         mmput(start_mm);
 847                         start_mm = new_start_mm;
 848                 }
 849                 if (retval) {
 850                         unlock_page(page);
 851                         page_cache_release(page);
 852                         break;
 853                 }
 854
 855                 /*
 856                  * How could swap count reach 0x7fff when the maximum
 857                  * pid is 0x7fff, and there's no way to repeat a swap
 858                  * page within an mm (except in shmem, where it's the
 859                  * shared object which takes the reference count)?
 860                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 861                  *
 862                  * If that's wrong, then we should worry more about
 863                  * exit_mmap() and do_munmap() cases described above:
 864                  * we might be resetting SWAP_MAP_MAX too early here.
 865                  * We know "Undead"s can happen, they're okay, so don't
 866                  * report them; but do report if we reset SWAP_MAP_MAX.
 867                  */
 868                 if (*swap_map == SWAP_MAP_MAX) {
 869                         spin_lock(&swap_lock);
 870                         *swap_map = 1;
 871                         spin_unlock(&swap_lock);
 872                         reset_overflow = 1;
 873                 }
 874
 875                 /*
 876                  * If a reference remains (rare), we would like to leave
 877                  * the page in the swap cache; but try_to_unmap could
 878                  * then re-duplicate the entry once we drop page lock,
 879                  * so we might loop indefinitely; also, that page could
 880                  * not be swapped out to other storage meanwhile.  So:
 881                  * delete from cache even if there's another reference,
 882                  * after ensuring that the data has been saved to disk -
 883                  * since if the reference remains (rarer), it will be
 884                  * read from disk into another page.  Splitting into two
 885                  * pages would be incorrect if swap supported "shared
 886                  * private" pages, but they are handled by tmpfs files.
 887                  *
 888                  * Note shmem_unuse already deleted a swappage from
 889                  * the swap cache, unless the move to filepage failed:
 890                  * in which case it left swappage in cache, lowered its
 891                  * swap count to pass quickly through the loops above,
 892                  * and now we must reincrement count to try again later.
 893                  */
 894                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 895                         struct writeback_control wbc = {
 896                                 .sync_mode = WB_SYNC_NONE,
 897                         };
 898
 899                         swap_writepage(page, &wbc);
 900                         lock_page(page);
 901                         wait_on_page_writeback(page);
 902                 }
 903                 if (PageSwapCache(page)) {
 904                         if (shmem)
 905                                 swap_duplicate(entry);
 906                         else
 907                                 delete_from_swap_cache(page);
 908                 }
 909
 910                 /*
 911                  * So we could skip searching mms once swap count went
 912                  * to 1, we did not mark any present ptes as dirty: must
 913                  * mark page dirty so shrink_page_list will preserve it.
 914                  */
 915                 SetPageDirty(page);
 916                 unlock_page(page);
 917                 page_cache_release(page);
 918
 919                 /*
 920                  * Make sure that we aren't completely killing
 921                  * interactive performance.
 922                  */
 923                 cond_resched();
 924         }
 925
 926         mmput(start_mm);
 927         if (reset_overflow) {
 928                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 929                 swap_overflow = 0;
 930         }
 931         return retval;
 932 }
 933
 934 /*
 935  * After a successful try_to_unuse, if no swap is now in use, we know
 936  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 937  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 938  * added to the mmlist just after page_duplicate - before would be racy.
 939  */
 940 static void drain_mmlist(void)
 941 {
 942         struct list_head *p, *next;
 943         unsigned int i;
 944
 945         for (i = 0; i < nr_swapfiles; i++)
 946                 if (swap_info[i].inuse_pages)
 947                         return;
 948         spin_lock(&mmlist_lock);
 949         list_for_each_safe(p, next, &init_mm.mmlist)
 950                 list_del_init(p);
 951         spin_unlock(&mmlist_lock);
 952 }
 953
 954 /*
 955  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 956  * corresponds to page offset `offset'.
 957  */
 958 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 959 {
 960         struct swap_extent *se = sis->curr_swap_extent;
 961         struct swap_extent *start_se = se;
 962
 963         for ( ; ; ) {
 964                 struct list_head *lh;
 965
 966                 if (se->start_page <= offset &&
 967                                 offset < (se->start_page + se->nr_pages)) {
 968                         return se->start_block + (offset - se->start_page);
 969                 }
 970                 lh = se->list.next;
 971                 if (lh == &sis->extent_list)
 972                         lh = lh->next;
 973                 se = list_entry(lh, struct swap_extent, list);
 974                 sis->curr_swap_extent = se;
 975                 BUG_ON(se == start_se);         /* It *must* be present */
 976         }
 977 }
 978
 979 #ifdef CONFIG_HIBERNATION
 980 /*
 981  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 982  * corresponding to given index in swap_info (swap type).
 983  */
 984 sector_t swapdev_block(int swap_type, pgoff_t offset)
 985 {
 986         struct swap_info_struct *sis;
 987
 988         if (swap_type >= nr_swapfiles)
 989                 return 0;
 990
 991         sis = swap_info + swap_type;
 992         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
 993 }
 994 #endif /* CONFIG_HIBERNATION */
 995
 996 /*
 997  * Free all of a swapdev's extent information
 998  */
 999 static void destroy_swap_extents(struct swap_info_struct *sis)
1000 {
1001         while (!list_empty(&sis->extent_list)) {
1002                 struct swap_extent *se;
1003
1004                 se = list_entry(sis->extent_list.next,
1005                                 struct swap_extent, list);
1006                 list_del(&se->list);
1007                 kfree(se);
1008         }
1009 }
1010
1011 /*
1012  * Add a block range (and the corresponding page range) into this swapdev's
1013  * extent list.  The extent list is kept sorted in page order.
1014  *
1015  * This function rather assumes that it is called in ascending page order.
1016  */
1017 static int
1018 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1019                 unsigned long nr_pages, sector_t start_block)
1020 {
1021         struct swap_extent *se;
1022         struct swap_extent *new_se;
1023         struct list_head *lh;
1024
1025         lh = sis->extent_list.prev;     /* The highest page extent */
1026         if (lh != &sis->extent_list) {
1027                 se = list_entry(lh, struct swap_extent, list);
1028                 BUG_ON(se->start_page + se->nr_pages != start_page);
1029                 if (se->start_block + se->nr_pages == start_block) {
1030                         /* Merge it */
1031                         se->nr_pages += nr_pages;
1032                         return 0;
1033                 }
1034         }
1035
1036         /*
1037          * No merge.  Insert a new extent, preserving ordering.
1038          */
1039         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1040         if (new_se == NULL)
1041                 return -ENOMEM;
1042         new_se->start_page = start_page;
1043         new_se->nr_pages = nr_pages;
1044         new_se->start_block = start_block;
1045
1046         list_add_tail(&new_se->list, &sis->extent_list);
1047         return 1;
1048 }
1049
1050 /*
1051  * A `swap extent' is a simple thing which maps a contiguous range of pages
1052  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1053  * is built at swapon time and is then used at swap_writepage/swap_readpage
1054  * time for locating where on disk a page belongs.
1055  *
1056  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1057  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1058  * swap files identically.
1059  *
1060  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1061  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1062  * swapfiles are handled *identically* after swapon time.
1063  *
1064  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1065  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1066  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1067  * requirements, they are simply tossed out - we will never use those blocks
1068  * for swapping.
1069  *
1070  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1071  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1072  * which will scribble on the fs.
1073  *
1074  * The amount of disk space which a single swap extent represents varies.
1075  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1076  * extents in the list.  To avoid much list walking, we cache the previous
1077  * search location in `curr_swap_extent', and start new searches from there.
1078  * This is extremely effective.  The average number of iterations in
1079  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1080  */
1081 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1082 {
1083         struct inode *inode;
1084         unsigned blocks_per_page;
1085         unsigned long page_no;
1086         unsigned blkbits;
1087         sector_t probe_block;
1088         sector_t last_block;
1089         sector_t lowest_block = -1;
1090         sector_t highest_block = 0;
1091         int nr_extents = 0;
1092         int ret;
1093
1094         inode = sis->swap_file->f_mapping->host;
1095         if (S_ISBLK(inode->i_mode)) {
1096                 ret = add_swap_extent(sis, 0, sis->max, 0);
1097                 *span = sis->pages;
1098                 goto done;
1099         }
1100
1101         blkbits = inode->i_blkbits;
1102         blocks_per_page = PAGE_SIZE >> blkbits;
1103
1104         /*
1105          * Map all the blocks into the extent list.  This code doesn't try
1106          * to be very smart.
1107          */
1108         probe_block = 0;
1109         page_no = 0;
1110         last_block = i_size_read(inode) >> blkbits;
1111         while ((probe_block + blocks_per_page) <= last_block &&
1112                         page_no < sis->max) {
1113                 unsigned block_in_page;
1114                 sector_t first_block;
1115
1116                 first_block = bmap(inode, probe_block);
1117                 if (first_block == 0)
1118                         goto bad_bmap;
1119
1120                 /*
1121                  * It must be PAGE_SIZE aligned on-disk
1122                  */
1123                 if (first_block & (blocks_per_page - 1)) {
1124                         probe_block++;
1125                         goto reprobe;
1126                 }
1127
1128                 for (block_in_page = 1; block_in_page < blocks_per_page;
1129                                         block_in_page++) {
1130                         sector_t block;
1131
1132                         block = bmap(inode, probe_block + block_in_page);
1133                         if (block == 0)
1134                                 goto bad_bmap;
1135                         if (block != first_block + block_in_page) {
1136                                 /* Discontiguity */
1137                                 probe_block++;
1138                                 goto reprobe;
1139                         }
1140                 }
1141
1142                 first_block >>= (PAGE_SHIFT - blkbits);
1143                 if (page_no) {  /* exclude the header page */
1144                         if (first_block < lowest_block)
1145                                 lowest_block = first_block;
1146                         if (first_block > highest_block)
1147                                 highest_block = first_block;
1148                 }
1149
1150                 /*
1151                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1152                  */
1153                 ret = add_swap_extent(sis, page_no, 1, first_block);
1154                 if (ret < 0)
1155                         goto out;
1156                 nr_extents += ret;
1157                 page_no++;
1158                 probe_block += blocks_per_page;
1159 reprobe:
1160                 continue;
1161         }
1162         ret = nr_extents;
1163         *span = 1 + highest_block - lowest_block;
1164         if (page_no == 0)
1165                 page_no = 1;    /* force Empty message */
1166         sis->max = page_no;
1167         sis->pages = page_no - 1;
1168         sis->highest_bit = page_no - 1;
1169 done:
1170         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1171                                         struct swap_extent, list);
1172         goto out;
1173 bad_bmap:
1174         printk(KERN_ERR "swapon: swapfile has holes\n");
1175         ret = -EINVAL;
1176 out:
1177         return ret;
1178 }
1179
1180 #if 0   /* We don't need this yet */
1181 #include <linux/backing-dev.h>
1182 int page_queue_congested(struct page *page)
1183 {
1184         struct backing_dev_info *bdi;
1185
1186         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1187
1188         if (PageSwapCache(page)) {
1189                 swp_entry_t entry = { .val = page_private(page) };
1190                 struct swap_info_struct *sis;
1191
1192                 sis = get_swap_info_struct(swp_type(entry));
1193                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1194         } else
1195                 bdi = page->mapping->backing_dev_info;
1196         return bdi_write_congested(bdi);
1197 }
1198 #endif
1199
1200 asmlinkage long sys_swapoff(const char __user * specialfile)
1201 {
1202         struct swap_info_struct * p = NULL;
1203         unsigned short *swap_map;
1204         struct file *swap_file, *victim;
1205         struct address_space *mapping;
1206         struct inode *inode;
1207         char * pathname;
1208         int i, type, prev;
1209         int err;
1210
1211         if (!capable(CAP_SYS_ADMIN))
1212                 return -EPERM;
1213
1214         pathname = getname(specialfile);
1215         err = PTR_ERR(pathname);
1216         if (IS_ERR(pathname))
1217                 goto out;
1218
1219         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1220         putname(pathname);
1221         err = PTR_ERR(victim);
1222         if (IS_ERR(victim))
1223                 goto out;
1224
1225         mapping = victim->f_mapping;
1226         prev = -1;
1227         spin_lock(&swap_lock);
1228         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1229                 p = swap_info + type;
1230                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1231                         if (p->swap_file->f_mapping == mapping)
1232                                 break;
1233                 }
1234                 prev = type;
1235         }
1236         if (type < 0) {
1237                 err = -EINVAL;
1238                 spin_unlock(&swap_lock);
1239                 goto out_dput;
1240         }
1241         if (!security_vm_enough_memory(p->pages))
1242                 vm_unacct_memory(p->pages);
1243         else {
1244                 err = -ENOMEM;
1245                 spin_unlock(&swap_lock);
1246                 goto out_dput;
1247         }
1248         if (prev < 0) {
1249                 swap_list.head = p->next;
1250         } else {
1251                 swap_info[prev].next = p->next;
1252         }
1253         if (type == swap_list.next) {
1254                 /* just pick something that's safe... */
1255                 swap_list.next = swap_list.head;
1256         }
1257         nr_swap_pages -= p->pages;
1258         total_swap_pages -= p->pages;
1259         p->flags &= ~SWP_WRITEOK;
1260         spin_unlock(&swap_lock);
1261
1262         current->flags |= PF_SWAPOFF;
1263         err = try_to_unuse(type);
1264         current->flags &= ~PF_SWAPOFF;
1265
1266         if (err) {
1267                 /* re-insert swap space back into swap_list */
1268                 spin_lock(&swap_lock);
1269                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1270                         if (p->prio >= swap_info[i].prio)
1271                                 break;
1272                 p->next = i;
1273                 if (prev < 0)
1274                         swap_list.head = swap_list.next = p - swap_info;
1275                 else
1276                         swap_info[prev].next = p - swap_info;
1277                 nr_swap_pages += p->pages;
1278                 total_swap_pages += p->pages;
1279                 p->flags |= SWP_WRITEOK;
1280                 spin_unlock(&swap_lock);
1281                 goto out_dput;
1282         }
1283
1284         /* wait for any unplug function to finish */
1285         down_write(&swap_unplug_sem);
1286         up_write(&swap_unplug_sem);
1287
1288         destroy_swap_extents(p);
1289         mutex_lock(&swapon_mutex);
1290         spin_lock(&swap_lock);
1291         drain_mmlist();
1292
1293         /* wait for anyone still in scan_swap_map */
1294         p->highest_bit = 0;             /* cuts scans short */
1295         while (p->flags >= SWP_SCANNING) {
1296                 spin_unlock(&swap_lock);
1297                 schedule_timeout_uninterruptible(1);
1298                 spin_lock(&swap_lock);
1299         }
1300
1301         swap_file = p->swap_file;
1302         p->swap_file = NULL;
1303         p->max = 0;
1304         swap_map = p->swap_map;
1305         p->swap_map = NULL;
1306         p->flags = 0;
1307         spin_unlock(&swap_lock);
1308         mutex_unlock(&swapon_mutex);
1309         vfree(swap_map);
1310         inode = mapping->host;
1311         if (S_ISBLK(inode->i_mode)) {
1312                 struct block_device *bdev = I_BDEV(inode);
1313                 set_blocksize(bdev, p->old_block_size);
1314                 bd_release(bdev);
1315         } else {
1316                 mutex_lock(&inode->i_mutex);
1317                 inode->i_flags &= ~S_SWAPFILE;
1318                 mutex_unlock(&inode->i_mutex);
1319         }
1320         filp_close(swap_file, NULL);
1321         err = 0;
1322
1323 out_dput:
1324         filp_close(victim, NULL);
1325 out:
1326         return err;
1327 }
1328
1329 #ifdef CONFIG_PROC_FS
1330 /* iterator */
1331 static void *swap_start(struct seq_file *swap, loff_t *pos)
1332 {
1333         struct swap_info_struct *ptr = swap_info;
1334         int i;
1335         loff_t l = *pos;
1336
1337         mutex_lock(&swapon_mutex);
1338
1339         if (!l)
1340                 return SEQ_START_TOKEN;
1341
1342         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1343                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1344                         continue;
1345                 if (!--l)
1346                         return ptr;
1347         }
1348
1349         return NULL;
1350 }
1351
1352 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1353 {
1354         struct swap_info_struct *ptr;
1355         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1356
1357         if (v == SEQ_START_TOKEN)
1358                 ptr = swap_info;
1359         else {
1360                 ptr = v;
1361                 ptr++;
1362         }
1363
1364         for (; ptr < endptr; ptr++) {
1365                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1366                         continue;
1367                 ++*pos;
1368                 return ptr;
1369         }
1370
1371         return NULL;
1372 }
1373
1374 static void swap_stop(struct seq_file *swap, void *v)
1375 {
1376         mutex_unlock(&swapon_mutex);
1377 }
1378
1379 static int swap_show(struct seq_file *swap, void *v)
1380 {
1381         struct swap_info_struct *ptr = v;
1382         struct file *file;
1383         int len;
1384
1385         if (ptr == SEQ_START_TOKEN) {
1386                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1387                 return 0;
1388         }
1389
1390         file = ptr->swap_file;
1391         len = seq_path(swap, file->f_path.mnt, file->f_path.dentry, " \t\n\\");
1392         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1393                        len < 40 ? 40 - len : 1, " ",
1394                        S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1395                                 "partition" : "file\t",
1396                        ptr->pages << (PAGE_SHIFT - 10),
1397                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1398                        ptr->prio);
1399         return 0;
1400 }
1401
1402 static const struct seq_operations swaps_op = {
1403         .start =        swap_start,
1404         .next =         swap_next,
1405         .stop =         swap_stop,
1406         .show =         swap_show
1407 };
1408
1409 static int swaps_open(struct inode *inode, struct file *file)
1410 {
1411         return seq_open(file, &swaps_op);
1412 }
1413
1414 static const struct file_operations proc_swaps_operations = {
1415         .open           = swaps_open,
1416         .read           = seq_read,
1417         .llseek         = seq_lseek,
1418         .release        = seq_release,
1419 };
1420
1421 static int __init procswaps_init(void)
1422 {
1423         struct proc_dir_entry *entry;
1424
1425         entry = create_proc_entry("swaps", 0, NULL);
1426         if (entry)
1427                 entry->proc_fops = &proc_swaps_operations;
1428         return 0;
1429 }
1430 __initcall(procswaps_init);
1431 #endif /* CONFIG_PROC_FS */
1432
1433 /*
1434  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1435  *
1436  * The swapon system call
1437  */
1438 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1439 {
1440         struct swap_info_struct * p;
1441         char *name = NULL;
1442         struct block_device *bdev = NULL;
1443         struct file *swap_file = NULL;
1444         struct address_space *mapping;
1445         unsigned int type;
1446         int i, prev;
1447         int error;
1448         static int least_priority;
1449         union swap_header *swap_header = NULL;
1450         int swap_header_version;
1451         unsigned int nr_good_pages = 0;
1452         int nr_extents = 0;
1453         sector_t span;
1454         unsigned long maxpages = 1;
1455         int swapfilesize;
1456         unsigned short *swap_map;
1457         struct page *page = NULL;
1458         struct inode *inode = NULL;
1459         int did_down = 0;
1460
1461         if (!capable(CAP_SYS_ADMIN))
1462                 return -EPERM;
1463         spin_lock(&swap_lock);
1464         p = swap_info;
1465         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1466                 if (!(p->flags & SWP_USED))
1467                         break;
1468         error = -EPERM;
1469         if (type >= MAX_SWAPFILES) {
1470                 spin_unlock(&swap_lock);
1471                 goto out;
1472         }
1473         if (type >= nr_swapfiles)
1474                 nr_swapfiles = type+1;
1475         INIT_LIST_HEAD(&p->extent_list);
1476         p->flags = SWP_USED;
1477         p->swap_file = NULL;
1478         p->old_block_size = 0;
1479         p->swap_map = NULL;
1480         p->lowest_bit = 0;
1481         p->highest_bit = 0;
1482         p->cluster_nr = 0;
1483         p->inuse_pages = 0;
1484         p->next = -1;
1485         if (swap_flags & SWAP_FLAG_PREFER) {
1486                 p->prio =
1487                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1488         } else {
1489                 p->prio = --least_priority;
1490         }
1491         spin_unlock(&swap_lock);
1492         name = getname(specialfile);
1493         error = PTR_ERR(name);
1494         if (IS_ERR(name)) {
1495                 name = NULL;
1496                 goto bad_swap_2;
1497         }
1498         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1499         error = PTR_ERR(swap_file);
1500         if (IS_ERR(swap_file)) {
1501                 swap_file = NULL;
1502                 goto bad_swap_2;
1503         }
1504
1505         p->swap_file = swap_file;
1506         mapping = swap_file->f_mapping;
1507         inode = mapping->host;
1508
1509         error = -EBUSY;
1510         for (i = 0; i < nr_swapfiles; i++) {
1511                 struct swap_info_struct *q = &swap_info[i];
1512
1513                 if (i == type || !q->swap_file)
1514                         continue;
1515                 if (mapping == q->swap_file->f_mapping)
1516                         goto bad_swap;
1517         }
1518
1519         error = -EINVAL;
1520         if (S_ISBLK(inode->i_mode)) {
1521                 bdev = I_BDEV(inode);
1522                 error = bd_claim(bdev, sys_swapon);
1523                 if (error < 0) {
1524                         bdev = NULL;
1525                         error = -EINVAL;
1526                         goto bad_swap;
1527                 }
1528                 p->old_block_size = block_size(bdev);
1529                 error = set_blocksize(bdev, PAGE_SIZE);
1530                 if (error < 0)
1531                         goto bad_swap;
1532                 p->bdev = bdev;
1533         } else if (S_ISREG(inode->i_mode)) {
1534                 p->bdev = inode->i_sb->s_bdev;
1535                 mutex_lock(&inode->i_mutex);
1536                 did_down = 1;
1537                 if (IS_SWAPFILE(inode)) {
1538                         error = -EBUSY;
1539                         goto bad_swap;
1540                 }
1541         } else {
1542                 goto bad_swap;
1543         }
1544
1545         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1546
1547         /*
1548          * Read the swap header.
1549          */
1550         if (!mapping->a_ops->readpage) {
1551                 error = -EINVAL;
1552                 goto bad_swap;
1553         }
1554         page = read_mapping_page(mapping, 0, swap_file);
1555         if (IS_ERR(page)) {
1556                 error = PTR_ERR(page);
1557                 goto bad_swap;
1558         }
1559         kmap(page);
1560         swap_header = page_address(page);
1561
1562         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1563                 swap_header_version = 1;
1564         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1565                 swap_header_version = 2;
1566         else {
1567                 printk(KERN_ERR "Unable to find swap-space signature\n");
1568                 error = -EINVAL;
1569                 goto bad_swap;
1570         }
1571
1572         switch (swap_header_version) {
1573         case 1:
1574                 printk(KERN_ERR "version 0 swap is no longer supported. "
1575                         "Use mkswap -v1 %s\n", name);
1576                 error = -EINVAL;
1577                 goto bad_swap;
1578         case 2:
1579                 /* Check the swap header's sub-version and the size of
1580                    the swap file and bad block lists */
1581                 if (swap_header->info.version != 1) {
1582                         printk(KERN_WARNING
1583                                "Unable to handle swap header version %d\n",
1584                                swap_header->info.version);
1585                         error = -EINVAL;
1586                         goto bad_swap;
1587                 }
1588
1589                 p->lowest_bit  = 1;
1590                 p->cluster_next = 1;
1591
1592                 /*
1593                  * Find out how many pages are allowed for a single swap
1594                  * device. There are two limiting factors: 1) the number of
1595                  * bits for the swap offset in the swp_entry_t type and
1596                  * 2) the number of bits in the a swap pte as defined by
1597                  * the different architectures. In order to find the
1598                  * largest possible bit mask a swap entry with swap type 0
1599                  * and swap offset ~0UL is created, encoded to a swap pte,
1600                  * decoded to a swp_entry_t again and finally the swap
1601                  * offset is extracted. This will mask all the bits from
1602                  * the initial ~0UL mask that can't be encoded in either
1603                  * the swp_entry_t or the architecture definition of a
1604                  * swap pte.
1605                  */
1606                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1607                 if (maxpages > swap_header->info.last_page)
1608                         maxpages = swap_header->info.last_page;
1609                 p->highest_bit = maxpages - 1;
1610
1611                 error = -EINVAL;
1612                 if (!maxpages)
1613                         goto bad_swap;
1614                 if (swapfilesize && maxpages > swapfilesize) {
1615                         printk(KERN_WARNING
1616                                "Swap area shorter than signature indicates\n");
1617                         goto bad_swap;
1618                 }
1619                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1620                         goto bad_swap;
1621                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1622                         goto bad_swap;
1623
1624                 /* OK, set up the swap map and apply the bad block list */
1625                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1626                         error = -ENOMEM;
1627                         goto bad_swap;
1628                 }
1629
1630                 error = 0;
1631                 memset(p->swap_map, 0, maxpages * sizeof(short));
1632                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1633                         int page_nr = swap_header->info.badpages[i];
1634                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1635                                 error = -EINVAL;
1636                         else
1637                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1638                 }
1639                 nr_good_pages = swap_header->info.last_page -
1640                                 swap_header->info.nr_badpages -
1641                                 1 /* header page */;
1642                 if (error)
1643                         goto bad_swap;
1644         }
1645
1646         if (nr_good_pages) {
1647                 p->swap_map[0] = SWAP_MAP_BAD;
1648                 p->max = maxpages;
1649                 p->pages = nr_good_pages;
1650                 nr_extents = setup_swap_extents(p, &span);
1651                 if (nr_extents < 0) {
1652                         error = nr_extents;
1653                         goto bad_swap;
1654                 }
1655                 nr_good_pages = p->pages;
1656         }
1657         if (!nr_good_pages) {
1658                 printk(KERN_WARNING "Empty swap-file\n");
1659                 error = -EINVAL;
1660                 goto bad_swap;
1661         }
1662
1663         mutex_lock(&swapon_mutex);
1664         spin_lock(&swap_lock);
1665         p->flags = SWP_ACTIVE;
1666         nr_swap_pages += nr_good_pages;
1667         total_swap_pages += nr_good_pages;
1668
1669         printk(KERN_INFO "Adding %uk swap on %s.  "
1670                         "Priority:%d extents:%d across:%lluk\n",
1671                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1672                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1673
1674         /* insert swap space into swap_list: */
1675         prev = -1;
1676         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1677                 if (p->prio >= swap_info[i].prio) {
1678                         break;
1679                 }
1680                 prev = i;
1681         }
1682         p->next = i;
1683         if (prev < 0) {
1684                 swap_list.head = swap_list.next = p - swap_info;
1685         } else {
1686                 swap_info[prev].next = p - swap_info;
1687         }
1688         spin_unlock(&swap_lock);
1689         mutex_unlock(&swapon_mutex);
1690         error = 0;
1691         goto out;
1692 bad_swap:
1693         if (bdev) {
1694                 set_blocksize(bdev, p->old_block_size);
1695                 bd_release(bdev);
1696         }
1697         destroy_swap_extents(p);
1698 bad_swap_2:
1699         spin_lock(&swap_lock);
1700         swap_map = p->swap_map;
1701         p->swap_file = NULL;
1702         p->swap_map = NULL;
1703         p->flags = 0;
1704         if (!(swap_flags & SWAP_FLAG_PREFER))
1705                 ++least_priority;
1706         spin_unlock(&swap_lock);
1707         vfree(swap_map);
1708         if (swap_file)
1709                 filp_close(swap_file, NULL);
1710 out:
1711         if (page && !IS_ERR(page)) {
1712                 kunmap(page);
1713                 page_cache_release(page);
1714         }
1715         if (name)
1716                 putname(name);
1717         if (did_down) {
1718                 if (!error)
1719                         inode->i_flags |= S_SWAPFILE;
1720                 mutex_unlock(&inode->i_mutex);
1721         }
1722         return error;
1723 }
1724
1725 void si_swapinfo(struct sysinfo *val)
1726 {
1727         unsigned int i;
1728         unsigned long nr_to_be_unused = 0;
1729
1730         spin_lock(&swap_lock);
1731         for (i = 0; i < nr_swapfiles; i++) {
1732                 if (!(swap_info[i].flags & SWP_USED) ||
1733                      (swap_info[i].flags & SWP_WRITEOK))
1734                         continue;
1735                 nr_to_be_unused += swap_info[i].inuse_pages;
1736         }
1737         val->freeswap = nr_swap_pages + nr_to_be_unused;
1738         val->totalswap = total_swap_pages + nr_to_be_unused;
1739         spin_unlock(&swap_lock);
1740 }
1741
1742 /*
1743  * Verify that a swap entry is valid and increment its swap map count.
1744  *
1745  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1746  * "permanent", but will be reclaimed by the next swapoff.
1747  */
1748 int swap_duplicate(swp_entry_t entry)
1749 {
1750         struct swap_info_struct * p;
1751         unsigned long offset, type;
1752         int result = 0;
1753
1754         if (is_migration_entry(entry))
1755                 return 1;
1756
1757         type = swp_type(entry);
1758         if (type >= nr_swapfiles)
1759                 goto bad_file;
1760         p = type + swap_info;
1761         offset = swp_offset(entry);
1762
1763         spin_lock(&swap_lock);
1764         if (offset < p->max && p->swap_map[offset]) {
1765                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1766                         p->swap_map[offset]++;
1767                         result = 1;
1768                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1769                         if (swap_overflow++ < 5)
1770                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1771                         p->swap_map[offset] = SWAP_MAP_MAX;
1772                         result = 1;
1773                 }
1774         }
1775         spin_unlock(&swap_lock);
1776 out:
1777         return result;
1778
1779 bad_file:
1780         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1781         goto out;
1782 }
1783
1784 struct swap_info_struct *
1785 get_swap_info_struct(unsigned type)
1786 {
1787         return &swap_info[type];
1788 }
1789
1790 /*
1791  * swap_lock prevents swap_map being freed. Don't grab an extra
1792  * reference on the swaphandle, it doesn't matter if it becomes unused.
1793  */
1794 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1795 {
1796         struct swap_info_struct *si;
1797         int our_page_cluster = page_cluster;
1798         pgoff_t target, toff;
1799         pgoff_t base, end;
1800         int nr_pages = 0;
1801
1802         if (!our_page_cluster)  /* no readahead */
1803                 return 0;
1804
1805         si = &swap_info[swp_type(entry)];
1806         target = swp_offset(entry);
1807         base = (target >> our_page_cluster) << our_page_cluster;
1808         end = base + (1 << our_page_cluster);
1809         if (!base)              /* first page is swap header */
1810                 base++;
1811
1812         spin_lock(&swap_lock);
1813         if (end > si->max)      /* don't go beyond end of map */
1814                 end = si->max;
1815
1816         /* Count contiguous allocated slots above our target */
1817         for (toff = target; ++toff < end; nr_pages++) {
1818                 /* Don't read in free or bad pages */
1819                 if (!si->swap_map[toff])
1820                         break;
1821                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1822                         break;
1823         }
1824         /* Count contiguous allocated slots below our target */
1825         for (toff = target; --toff >= base; nr_pages++) {
1826                 /* Don't read in free or bad pages */
1827                 if (!si->swap_map[toff])
1828                         break;
1829                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1830                         break;
1831         }
1832         spin_unlock(&swap_lock);
1833
1834         /*
1835          * Indicate starting offset, and return number of pages to get:
1836          * if only 1, say 0, since there's then no readahead to be done.
1837          */
1838         *offset = ++toff;
1839         return nr_pages? ++nr_pages: 0;
1840 }