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