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 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,
 734                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
 735                 if (!page) {
 736                         /*
 737                          * Either swap_duplicate() failed because entry
 738                          * has been freed independently, and will not be
 739                          * reused since sys_swapoff() already disabled
 740                          * allocation from here, or alloc_page() failed.
 741                          */
 742                         if (!*swap_map)
 743                                 continue;
 744                         retval = -ENOMEM;
 745                         break;
 746                 }
 747
 748                 /*
 749                  * Don't hold on to start_mm if it looks like exiting.
 750                  */
 751                 if (atomic_read(&start_mm->mm_users) == 1) {
 752                         mmput(start_mm);
 753                         start_mm = &init_mm;
 754                         atomic_inc(&init_mm.mm_users);
 755                 }
 756
 757                 /*
 758                  * Wait for and lock page.  When do_swap_page races with
 759                  * try_to_unuse, do_swap_page can handle the fault much
 760                  * faster than try_to_unuse can locate the entry.  This
 761                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 762                  * defer to do_swap_page in such a case - in some tests,
 763                  * do_swap_page and try_to_unuse repeatedly compete.
 764                  */
 765                 wait_on_page_locked(page);
 766                 wait_on_page_writeback(page);
 767                 lock_page(page);
 768                 wait_on_page_writeback(page);
 769
 770                 /*
 771                  * Remove all references to entry.
 772                  * Whenever we reach init_mm, there's no address space
 773                  * to search, but use it as a reminder to search shmem.
 774                  */
 775                 shmem = 0;
 776                 swcount = *swap_map;
 777                 if (swcount > 1) {
 778                         if (start_mm == &init_mm)
 779                                 shmem = shmem_unuse(entry, page);
 780                         else
 781                                 retval = unuse_mm(start_mm, entry, page);
 782                 }
 783                 if (*swap_map > 1) {
 784                         int set_start_mm = (*swap_map >= swcount);
 785                         struct list_head *p = &start_mm->mmlist;
 786                         struct mm_struct *new_start_mm = start_mm;
 787                         struct mm_struct *prev_mm = start_mm;
 788                         struct mm_struct *mm;
 789
 790                         atomic_inc(&new_start_mm->mm_users);
 791                         atomic_inc(&prev_mm->mm_users);
 792                         spin_lock(&mmlist_lock);
 793                         while (*swap_map > 1 && !retval && !shmem &&
 794                                         (p = p->next) != &start_mm->mmlist) {
 795                                 mm = list_entry(p, struct mm_struct, mmlist);
 796                                 if (!atomic_inc_not_zero(&mm->mm_users))
 797                                         continue;
 798                                 spin_unlock(&mmlist_lock);
 799                                 mmput(prev_mm);
 800                                 prev_mm = mm;
 801
 802                                 cond_resched();
 803
 804                                 swcount = *swap_map;
 805                                 if (swcount <= 1)
 806                                         ;
 807                                 else if (mm == &init_mm) {
 808                                         set_start_mm = 1;
 809                                         shmem = shmem_unuse(entry, page);
 810                                 } else
 811                                         retval = unuse_mm(mm, entry, page);
 812                                 if (set_start_mm && *swap_map < swcount) {
 813                                         mmput(new_start_mm);
 814                                         atomic_inc(&mm->mm_users);
 815                                         new_start_mm = mm;
 816                                         set_start_mm = 0;
 817                                 }
 818                                 spin_lock(&mmlist_lock);
 819                         }
 820                         spin_unlock(&mmlist_lock);
 821                         mmput(prev_mm);
 822                         mmput(start_mm);
 823                         start_mm = new_start_mm;
 824                 }
 825                 if (shmem) {
 826                         /* page has already been unlocked and released */
 827                         if (shmem > 0)
 828                                 continue;
 829                         retval = shmem;
 830                         break;
 831                 }
 832                 if (retval) {
 833                         unlock_page(page);
 834                         page_cache_release(page);
 835                         break;
 836                 }
 837
 838                 /*
 839                  * How could swap count reach 0x7fff when the maximum
 840                  * pid is 0x7fff, and there's no way to repeat a swap
 841                  * page within an mm (except in shmem, where it's the
 842                  * shared object which takes the reference count)?
 843                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 844                  *
 845                  * If that's wrong, then we should worry more about
 846                  * exit_mmap() and do_munmap() cases described above:
 847                  * we might be resetting SWAP_MAP_MAX too early here.
 848                  * We know "Undead"s can happen, they're okay, so don't
 849                  * report them; but do report if we reset SWAP_MAP_MAX.
 850                  */
 851                 if (*swap_map == SWAP_MAP_MAX) {
 852                         spin_lock(&swap_lock);
 853                         *swap_map = 1;
 854                         spin_unlock(&swap_lock);
 855                         reset_overflow = 1;
 856                 }
 857
 858                 /*
 859                  * If a reference remains (rare), we would like to leave
 860                  * the page in the swap cache; but try_to_unmap could
 861                  * then re-duplicate the entry once we drop page lock,
 862                  * so we might loop indefinitely; also, that page could
 863                  * not be swapped out to other storage meanwhile.  So:
 864                  * delete from cache even if there's another reference,
 865                  * after ensuring that the data has been saved to disk -
 866                  * since if the reference remains (rarer), it will be
 867                  * read from disk into another page.  Splitting into two
 868                  * pages would be incorrect if swap supported "shared
 869                  * private" pages, but they are handled by tmpfs files.
 870                  */
 871                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 872                         struct writeback_control wbc = {
 873                                 .sync_mode = WB_SYNC_NONE,
 874                         };
 875
 876                         swap_writepage(page, &wbc);
 877                         lock_page(page);
 878                         wait_on_page_writeback(page);
 879                 }
 880                 if (PageSwapCache(page))
 881                         delete_from_swap_cache(page);
 882
 883                 /*
 884                  * So we could skip searching mms once swap count went
 885                  * to 1, we did not mark any present ptes as dirty: must
 886                  * mark page dirty so shrink_page_list will preserve it.
 887                  */
 888                 SetPageDirty(page);
 889                 unlock_page(page);
 890                 page_cache_release(page);
 891
 892                 /*
 893                  * Make sure that we aren't completely killing
 894                  * interactive performance.
 895                  */
 896                 cond_resched();
 897         }
 898
 899         mmput(start_mm);
 900         if (reset_overflow) {
 901                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 902                 swap_overflow = 0;
 903         }
 904         return retval;
 905 }
 906
 907 /*
 908  * After a successful try_to_unuse, if no swap is now in use, we know
 909  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 910  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 911  * added to the mmlist just after page_duplicate - before would be racy.
 912  */
 913 static void drain_mmlist(void)
 914 {
 915         struct list_head *p, *next;
 916         unsigned int i;
 917
 918         for (i = 0; i < nr_swapfiles; i++)
 919                 if (swap_info[i].inuse_pages)
 920                         return;
 921         spin_lock(&mmlist_lock);
 922         list_for_each_safe(p, next, &init_mm.mmlist)
 923                 list_del_init(p);
 924         spin_unlock(&mmlist_lock);
 925 }
 926
 927 /*
 928  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 929  * corresponds to page offset `offset'.
 930  */
 931 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 932 {
 933         struct swap_extent *se = sis->curr_swap_extent;
 934         struct swap_extent *start_se = se;
 935
 936         for ( ; ; ) {
 937                 struct list_head *lh;
 938
 939                 if (se->start_page <= offset &&
 940                                 offset < (se->start_page + se->nr_pages)) {
 941                         return se->start_block + (offset - se->start_page);
 942                 }
 943                 lh = se->list.next;
 944                 if (lh == &sis->extent_list)
 945                         lh = lh->next;
 946                 se = list_entry(lh, struct swap_extent, list);
 947                 sis->curr_swap_extent = se;
 948                 BUG_ON(se == start_se);         /* It *must* be present */
 949         }
 950 }
 951
 952 #ifdef CONFIG_HIBERNATION
 953 /*
 954  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 955  * corresponding to given index in swap_info (swap type).
 956  */
 957 sector_t swapdev_block(int swap_type, pgoff_t offset)
 958 {
 959         struct swap_info_struct *sis;
 960
 961         if (swap_type >= nr_swapfiles)
 962                 return 0;
 963
 964         sis = swap_info + swap_type;
 965         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
 966 }
 967 #endif /* CONFIG_HIBERNATION */
 968
 969 /*
 970  * Free all of a swapdev's extent information
 971  */
 972 static void destroy_swap_extents(struct swap_info_struct *sis)
 973 {
 974         while (!list_empty(&sis->extent_list)) {
 975                 struct swap_extent *se;
 976
 977                 se = list_entry(sis->extent_list.next,
 978                                 struct swap_extent, list);
 979                 list_del(&se->list);
 980                 kfree(se);
 981         }
 982 }
 983
 984 /*
 985  * Add a block range (and the corresponding page range) into this swapdev's
 986  * extent list.  The extent list is kept sorted in page order.
 987  *
 988  * This function rather assumes that it is called in ascending page order.
 989  */
 990 static int
 991 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 992                 unsigned long nr_pages, sector_t start_block)
 993 {
 994         struct swap_extent *se;
 995         struct swap_extent *new_se;
 996         struct list_head *lh;
 997
 998         lh = sis->extent_list.prev;     /* The highest page extent */
 999         if (lh != &sis->extent_list) {
1000                 se = list_entry(lh, struct swap_extent, list);
1001                 BUG_ON(se->start_page + se->nr_pages != start_page);
1002                 if (se->start_block + se->nr_pages == start_block) {
1003                         /* Merge it */
1004                         se->nr_pages += nr_pages;
1005                         return 0;
1006                 }
1007         }
1008
1009         /*
1010          * No merge.  Insert a new extent, preserving ordering.
1011          */
1012         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1013         if (new_se == NULL)
1014                 return -ENOMEM;
1015         new_se->start_page = start_page;
1016         new_se->nr_pages = nr_pages;
1017         new_se->start_block = start_block;
1018
1019         list_add_tail(&new_se->list, &sis->extent_list);
1020         return 1;
1021 }
1022
1023 /*
1024  * A `swap extent' is a simple thing which maps a contiguous range of pages
1025  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1026  * is built at swapon time and is then used at swap_writepage/swap_readpage
1027  * time for locating where on disk a page belongs.
1028  *
1029  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1030  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1031  * swap files identically.
1032  *
1033  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1034  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1035  * swapfiles are handled *identically* after swapon time.
1036  *
1037  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1038  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1039  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1040  * requirements, they are simply tossed out - we will never use those blocks
1041  * for swapping.
1042  *
1043  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1044  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1045  * which will scribble on the fs.
1046  *
1047  * The amount of disk space which a single swap extent represents varies.
1048  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1049  * extents in the list.  To avoid much list walking, we cache the previous
1050  * search location in `curr_swap_extent', and start new searches from there.
1051  * This is extremely effective.  The average number of iterations in
1052  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1053  */
1054 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1055 {
1056         struct inode *inode;
1057         unsigned blocks_per_page;
1058         unsigned long page_no;
1059         unsigned blkbits;
1060         sector_t probe_block;
1061         sector_t last_block;
1062         sector_t lowest_block = -1;
1063         sector_t highest_block = 0;
1064         int nr_extents = 0;
1065         int ret;
1066
1067         inode = sis->swap_file->f_mapping->host;
1068         if (S_ISBLK(inode->i_mode)) {
1069                 ret = add_swap_extent(sis, 0, sis->max, 0);
1070                 *span = sis->pages;
1071                 goto done;
1072         }
1073
1074         blkbits = inode->i_blkbits;
1075         blocks_per_page = PAGE_SIZE >> blkbits;
1076
1077         /*
1078          * Map all the blocks into the extent list.  This code doesn't try
1079          * to be very smart.
1080          */
1081         probe_block = 0;
1082         page_no = 0;
1083         last_block = i_size_read(inode) >> blkbits;
1084         while ((probe_block + blocks_per_page) <= last_block &&
1085                         page_no < sis->max) {
1086                 unsigned block_in_page;
1087                 sector_t first_block;
1088
1089                 first_block = bmap(inode, probe_block);
1090                 if (first_block == 0)
1091                         goto bad_bmap;
1092
1093                 /*
1094                  * It must be PAGE_SIZE aligned on-disk
1095                  */
1096                 if (first_block & (blocks_per_page - 1)) {
1097                         probe_block++;
1098                         goto reprobe;
1099                 }
1100
1101                 for (block_in_page = 1; block_in_page < blocks_per_page;
1102                                         block_in_page++) {
1103                         sector_t block;
1104
1105                         block = bmap(inode, probe_block + block_in_page);
1106                         if (block == 0)
1107                                 goto bad_bmap;
1108                         if (block != first_block + block_in_page) {
1109                                 /* Discontiguity */
1110                                 probe_block++;
1111                                 goto reprobe;
1112                         }
1113                 }
1114
1115                 first_block >>= (PAGE_SHIFT - blkbits);
1116                 if (page_no) {  /* exclude the header page */
1117                         if (first_block < lowest_block)
1118                                 lowest_block = first_block;
1119                         if (first_block > highest_block)
1120                                 highest_block = first_block;
1121                 }
1122
1123                 /*
1124                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1125                  */
1126                 ret = add_swap_extent(sis, page_no, 1, first_block);
1127                 if (ret < 0)
1128                         goto out;
1129                 nr_extents += ret;
1130                 page_no++;
1131                 probe_block += blocks_per_page;
1132 reprobe:
1133                 continue;
1134         }
1135         ret = nr_extents;
1136         *span = 1 + highest_block - lowest_block;
1137         if (page_no == 0)
1138                 page_no = 1;    /* force Empty message */
1139         sis->max = page_no;
1140         sis->pages = page_no - 1;
1141         sis->highest_bit = page_no - 1;
1142 done:
1143         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1144                                         struct swap_extent, list);
1145         goto out;
1146 bad_bmap:
1147         printk(KERN_ERR "swapon: swapfile has holes\n");
1148         ret = -EINVAL;
1149 out:
1150         return ret;
1151 }
1152
1153 #if 0   /* We don't need this yet */
1154 #include <linux/backing-dev.h>
1155 int page_queue_congested(struct page *page)
1156 {
1157         struct backing_dev_info *bdi;
1158
1159         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1160
1161         if (PageSwapCache(page)) {
1162                 swp_entry_t entry = { .val = page_private(page) };
1163                 struct swap_info_struct *sis;
1164
1165                 sis = get_swap_info_struct(swp_type(entry));
1166                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1167         } else
1168                 bdi = page->mapping->backing_dev_info;
1169         return bdi_write_congested(bdi);
1170 }
1171 #endif
1172
1173 asmlinkage long sys_swapoff(const char __user * specialfile)
1174 {
1175         struct swap_info_struct * p = NULL;
1176         unsigned short *swap_map;
1177         struct file *swap_file, *victim;
1178         struct address_space *mapping;
1179         struct inode *inode;
1180         char * pathname;
1181         int i, type, prev;
1182         int err;
1183
1184         if (!capable(CAP_SYS_ADMIN))
1185                 return -EPERM;
1186
1187         pathname = getname(specialfile);
1188         err = PTR_ERR(pathname);
1189         if (IS_ERR(pathname))
1190                 goto out;
1191
1192         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1193         putname(pathname);
1194         err = PTR_ERR(victim);
1195         if (IS_ERR(victim))
1196                 goto out;
1197
1198         mapping = victim->f_mapping;
1199         prev = -1;
1200         spin_lock(&swap_lock);
1201         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1202                 p = swap_info + type;
1203                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1204                         if (p->swap_file->f_mapping == mapping)
1205                                 break;
1206                 }
1207                 prev = type;
1208         }
1209         if (type < 0) {
1210                 err = -EINVAL;
1211                 spin_unlock(&swap_lock);
1212                 goto out_dput;
1213         }
1214         if (!security_vm_enough_memory(p->pages))
1215                 vm_unacct_memory(p->pages);
1216         else {
1217                 err = -ENOMEM;
1218                 spin_unlock(&swap_lock);
1219                 goto out_dput;
1220         }
1221         if (prev < 0) {
1222                 swap_list.head = p->next;
1223         } else {
1224                 swap_info[prev].next = p->next;
1225         }
1226         if (type == swap_list.next) {
1227                 /* just pick something that's safe... */
1228                 swap_list.next = swap_list.head;
1229         }
1230         nr_swap_pages -= p->pages;
1231         total_swap_pages -= p->pages;
1232         p->flags &= ~SWP_WRITEOK;
1233         spin_unlock(&swap_lock);
1234
1235         current->flags |= PF_SWAPOFF;
1236         err = try_to_unuse(type);
1237         current->flags &= ~PF_SWAPOFF;
1238
1239         if (err) {
1240                 /* re-insert swap space back into swap_list */
1241                 spin_lock(&swap_lock);
1242                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1243                         if (p->prio >= swap_info[i].prio)
1244                                 break;
1245                 p->next = i;
1246                 if (prev < 0)
1247                         swap_list.head = swap_list.next = p - swap_info;
1248                 else
1249                         swap_info[prev].next = p - swap_info;
1250                 nr_swap_pages += p->pages;
1251                 total_swap_pages += p->pages;
1252                 p->flags |= SWP_WRITEOK;
1253                 spin_unlock(&swap_lock);
1254                 goto out_dput;
1255         }
1256
1257         /* wait for any unplug function to finish */
1258         down_write(&swap_unplug_sem);
1259         up_write(&swap_unplug_sem);
1260
1261         destroy_swap_extents(p);
1262         mutex_lock(&swapon_mutex);
1263         spin_lock(&swap_lock);
1264         drain_mmlist();
1265
1266         /* wait for anyone still in scan_swap_map */
1267         p->highest_bit = 0;             /* cuts scans short */
1268         while (p->flags >= SWP_SCANNING) {
1269                 spin_unlock(&swap_lock);
1270                 schedule_timeout_uninterruptible(1);
1271                 spin_lock(&swap_lock);
1272         }
1273
1274         swap_file = p->swap_file;
1275         p->swap_file = NULL;
1276         p->max = 0;
1277         swap_map = p->swap_map;
1278         p->swap_map = NULL;
1279         p->flags = 0;
1280         spin_unlock(&swap_lock);
1281         mutex_unlock(&swapon_mutex);
1282         vfree(swap_map);
1283         inode = mapping->host;
1284         if (S_ISBLK(inode->i_mode)) {
1285                 struct block_device *bdev = I_BDEV(inode);
1286                 set_blocksize(bdev, p->old_block_size);
1287                 bd_release(bdev);
1288         } else {
1289                 mutex_lock(&inode->i_mutex);
1290                 inode->i_flags &= ~S_SWAPFILE;
1291                 mutex_unlock(&inode->i_mutex);
1292         }
1293         filp_close(swap_file, NULL);
1294         err = 0;
1295
1296 out_dput:
1297         filp_close(victim, NULL);
1298 out:
1299         return err;
1300 }
1301
1302 #ifdef CONFIG_PROC_FS
1303 /* iterator */
1304 static void *swap_start(struct seq_file *swap, loff_t *pos)
1305 {
1306         struct swap_info_struct *ptr = swap_info;
1307         int i;
1308         loff_t l = *pos;
1309
1310         mutex_lock(&swapon_mutex);
1311
1312         if (!l)
1313                 return SEQ_START_TOKEN;
1314
1315         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1316                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1317                         continue;
1318                 if (!--l)
1319                         return ptr;
1320         }
1321
1322         return NULL;
1323 }
1324
1325 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1326 {
1327         struct swap_info_struct *ptr;
1328         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1329
1330         if (v == SEQ_START_TOKEN)
1331                 ptr = swap_info;
1332         else {
1333                 ptr = v;
1334                 ptr++;
1335         }
1336
1337         for (; ptr < endptr; ptr++) {
1338                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1339                         continue;
1340                 ++*pos;
1341                 return ptr;
1342         }
1343
1344         return NULL;
1345 }
1346
1347 static void swap_stop(struct seq_file *swap, void *v)
1348 {
1349         mutex_unlock(&swapon_mutex);
1350 }
1351
1352 static int swap_show(struct seq_file *swap, void *v)
1353 {
1354         struct swap_info_struct *ptr = v;
1355         struct file *file;
1356         int len;
1357
1358         if (ptr == SEQ_START_TOKEN) {
1359                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1360                 return 0;
1361         }
1362
1363         file = ptr->swap_file;
1364         len = seq_path(swap, file->f_path.mnt, file->f_path.dentry, " \t\n\\");
1365         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1366                        len < 40 ? 40 - len : 1, " ",
1367                        S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1368                                 "partition" : "file\t",
1369                        ptr->pages << (PAGE_SHIFT - 10),
1370                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1371                        ptr->prio);
1372         return 0;
1373 }
1374
1375 static const struct seq_operations swaps_op = {
1376         .start =        swap_start,
1377         .next =         swap_next,
1378         .stop =         swap_stop,
1379         .show =         swap_show
1380 };
1381
1382 static int swaps_open(struct inode *inode, struct file *file)
1383 {
1384         return seq_open(file, &swaps_op);
1385 }
1386
1387 static const struct file_operations proc_swaps_operations = {
1388         .open           = swaps_open,
1389         .read           = seq_read,
1390         .llseek         = seq_lseek,
1391         .release        = seq_release,
1392 };
1393
1394 static int __init procswaps_init(void)
1395 {
1396         struct proc_dir_entry *entry;
1397
1398         entry = create_proc_entry("swaps", 0, NULL);
1399         if (entry)
1400                 entry->proc_fops = &proc_swaps_operations;
1401         return 0;
1402 }
1403 __initcall(procswaps_init);
1404 #endif /* CONFIG_PROC_FS */
1405
1406 /*
1407  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1408  *
1409  * The swapon system call
1410  */
1411 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1412 {
1413         struct swap_info_struct * p;
1414         char *name = NULL;
1415         struct block_device *bdev = NULL;
1416         struct file *swap_file = NULL;
1417         struct address_space *mapping;
1418         unsigned int type;
1419         int i, prev;
1420         int error;
1421         static int least_priority;
1422         union swap_header *swap_header = NULL;
1423         int swap_header_version;
1424         unsigned int nr_good_pages = 0;
1425         int nr_extents = 0;
1426         sector_t span;
1427         unsigned long maxpages = 1;
1428         int swapfilesize;
1429         unsigned short *swap_map;
1430         struct page *page = NULL;
1431         struct inode *inode = NULL;
1432         int did_down = 0;
1433
1434         if (!capable(CAP_SYS_ADMIN))
1435                 return -EPERM;
1436         spin_lock(&swap_lock);
1437         p = swap_info;
1438         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1439                 if (!(p->flags & SWP_USED))
1440                         break;
1441         error = -EPERM;
1442         if (type >= MAX_SWAPFILES) {
1443                 spin_unlock(&swap_lock);
1444                 goto out;
1445         }
1446         if (type >= nr_swapfiles)
1447                 nr_swapfiles = type+1;
1448         INIT_LIST_HEAD(&p->extent_list);
1449         p->flags = SWP_USED;
1450         p->swap_file = NULL;
1451         p->old_block_size = 0;
1452         p->swap_map = NULL;
1453         p->lowest_bit = 0;
1454         p->highest_bit = 0;
1455         p->cluster_nr = 0;
1456         p->inuse_pages = 0;
1457         p->next = -1;
1458         if (swap_flags & SWAP_FLAG_PREFER) {
1459                 p->prio =
1460                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1461         } else {
1462                 p->prio = --least_priority;
1463         }
1464         spin_unlock(&swap_lock);
1465         name = getname(specialfile);
1466         error = PTR_ERR(name);
1467         if (IS_ERR(name)) {
1468                 name = NULL;
1469                 goto bad_swap_2;
1470         }
1471         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1472         error = PTR_ERR(swap_file);
1473         if (IS_ERR(swap_file)) {
1474                 swap_file = NULL;
1475                 goto bad_swap_2;
1476         }
1477
1478         p->swap_file = swap_file;
1479         mapping = swap_file->f_mapping;
1480         inode = mapping->host;
1481
1482         error = -EBUSY;
1483         for (i = 0; i < nr_swapfiles; i++) {
1484                 struct swap_info_struct *q = &swap_info[i];
1485
1486                 if (i == type || !q->swap_file)
1487                         continue;
1488                 if (mapping == q->swap_file->f_mapping)
1489                         goto bad_swap;
1490         }
1491
1492         error = -EINVAL;
1493         if (S_ISBLK(inode->i_mode)) {
1494                 bdev = I_BDEV(inode);
1495                 error = bd_claim(bdev, sys_swapon);
1496                 if (error < 0) {
1497                         bdev = NULL;
1498                         error = -EINVAL;
1499                         goto bad_swap;
1500                 }
1501                 p->old_block_size = block_size(bdev);
1502                 error = set_blocksize(bdev, PAGE_SIZE);
1503                 if (error < 0)
1504                         goto bad_swap;
1505                 p->bdev = bdev;
1506         } else if (S_ISREG(inode->i_mode)) {
1507                 p->bdev = inode->i_sb->s_bdev;
1508                 mutex_lock(&inode->i_mutex);
1509                 did_down = 1;
1510                 if (IS_SWAPFILE(inode)) {
1511                         error = -EBUSY;
1512                         goto bad_swap;
1513                 }
1514         } else {
1515                 goto bad_swap;
1516         }
1517
1518         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1519
1520         /*
1521          * Read the swap header.
1522          */
1523         if (!mapping->a_ops->readpage) {
1524                 error = -EINVAL;
1525                 goto bad_swap;
1526         }
1527         page = read_mapping_page(mapping, 0, swap_file);
1528         if (IS_ERR(page)) {
1529                 error = PTR_ERR(page);
1530                 goto bad_swap;
1531         }
1532         kmap(page);
1533         swap_header = page_address(page);
1534
1535         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1536                 swap_header_version = 1;
1537         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1538                 swap_header_version = 2;
1539         else {
1540                 printk(KERN_ERR "Unable to find swap-space signature\n");
1541                 error = -EINVAL;
1542                 goto bad_swap;
1543         }
1544
1545         switch (swap_header_version) {
1546         case 1:
1547                 printk(KERN_ERR "version 0 swap is no longer supported. "
1548                         "Use mkswap -v1 %s\n", name);
1549                 error = -EINVAL;
1550                 goto bad_swap;
1551         case 2:
1552                 /* Check the swap header's sub-version and the size of
1553                    the swap file and bad block lists */
1554                 if (swap_header->info.version != 1) {
1555                         printk(KERN_WARNING
1556                                "Unable to handle swap header version %d\n",
1557                                swap_header->info.version);
1558                         error = -EINVAL;
1559                         goto bad_swap;
1560                 }
1561
1562                 p->lowest_bit  = 1;
1563                 p->cluster_next = 1;
1564
1565                 /*
1566                  * Find out how many pages are allowed for a single swap
1567                  * device. There are two limiting factors: 1) the number of
1568                  * bits for the swap offset in the swp_entry_t type and
1569                  * 2) the number of bits in the a swap pte as defined by
1570                  * the different architectures. In order to find the
1571                  * largest possible bit mask a swap entry with swap type 0
1572                  * and swap offset ~0UL is created, encoded to a swap pte,
1573                  * decoded to a swp_entry_t again and finally the swap
1574                  * offset is extracted. This will mask all the bits from
1575                  * the initial ~0UL mask that can't be encoded in either
1576                  * the swp_entry_t or the architecture definition of a
1577                  * swap pte.
1578                  */
1579                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1580                 if (maxpages > swap_header->info.last_page)
1581                         maxpages = swap_header->info.last_page;
1582                 p->highest_bit = maxpages - 1;
1583
1584                 error = -EINVAL;
1585                 if (!maxpages)
1586                         goto bad_swap;
1587                 if (swapfilesize && maxpages > swapfilesize) {
1588                         printk(KERN_WARNING
1589                                "Swap area shorter than signature indicates\n");
1590                         goto bad_swap;
1591                 }
1592                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1593                         goto bad_swap;
1594                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1595                         goto bad_swap;
1596
1597                 /* OK, set up the swap map and apply the bad block list */
1598                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1599                         error = -ENOMEM;
1600                         goto bad_swap;
1601                 }
1602
1603                 error = 0;
1604                 memset(p->swap_map, 0, maxpages * sizeof(short));
1605                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1606                         int page_nr = swap_header->info.badpages[i];
1607                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1608                                 error = -EINVAL;
1609                         else
1610                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1611                 }
1612                 nr_good_pages = swap_header->info.last_page -
1613                                 swap_header->info.nr_badpages -
1614                                 1 /* header page */;
1615                 if (error)
1616                         goto bad_swap;
1617         }
1618
1619         if (nr_good_pages) {
1620                 p->swap_map[0] = SWAP_MAP_BAD;
1621                 p->max = maxpages;
1622                 p->pages = nr_good_pages;
1623                 nr_extents = setup_swap_extents(p, &span);
1624                 if (nr_extents < 0) {
1625                         error = nr_extents;
1626                         goto bad_swap;
1627                 }
1628                 nr_good_pages = p->pages;
1629         }
1630         if (!nr_good_pages) {
1631                 printk(KERN_WARNING "Empty swap-file\n");
1632                 error = -EINVAL;
1633                 goto bad_swap;
1634         }
1635
1636         mutex_lock(&swapon_mutex);
1637         spin_lock(&swap_lock);
1638         p->flags = SWP_ACTIVE;
1639         nr_swap_pages += nr_good_pages;
1640         total_swap_pages += nr_good_pages;
1641
1642         printk(KERN_INFO "Adding %uk swap on %s.  "
1643                         "Priority:%d extents:%d across:%lluk\n",
1644                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1645                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1646
1647         /* insert swap space into swap_list: */
1648         prev = -1;
1649         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1650                 if (p->prio >= swap_info[i].prio) {
1651                         break;
1652                 }
1653                 prev = i;
1654         }
1655         p->next = i;
1656         if (prev < 0) {
1657                 swap_list.head = swap_list.next = p - swap_info;
1658         } else {
1659                 swap_info[prev].next = p - swap_info;
1660         }
1661         spin_unlock(&swap_lock);
1662         mutex_unlock(&swapon_mutex);
1663         error = 0;
1664         goto out;
1665 bad_swap:
1666         if (bdev) {
1667                 set_blocksize(bdev, p->old_block_size);
1668                 bd_release(bdev);
1669         }
1670         destroy_swap_extents(p);
1671 bad_swap_2:
1672         spin_lock(&swap_lock);
1673         swap_map = p->swap_map;
1674         p->swap_file = NULL;
1675         p->swap_map = NULL;
1676         p->flags = 0;
1677         if (!(swap_flags & SWAP_FLAG_PREFER))
1678                 ++least_priority;
1679         spin_unlock(&swap_lock);
1680         vfree(swap_map);
1681         if (swap_file)
1682                 filp_close(swap_file, NULL);
1683 out:
1684         if (page && !IS_ERR(page)) {
1685                 kunmap(page);
1686                 page_cache_release(page);
1687         }
1688         if (name)
1689                 putname(name);
1690         if (did_down) {
1691                 if (!error)
1692                         inode->i_flags |= S_SWAPFILE;
1693                 mutex_unlock(&inode->i_mutex);
1694         }
1695         return error;
1696 }
1697
1698 void si_swapinfo(struct sysinfo *val)
1699 {
1700         unsigned int i;
1701         unsigned long nr_to_be_unused = 0;
1702
1703         spin_lock(&swap_lock);
1704         for (i = 0; i < nr_swapfiles; i++) {
1705                 if (!(swap_info[i].flags & SWP_USED) ||
1706                      (swap_info[i].flags & SWP_WRITEOK))
1707                         continue;
1708                 nr_to_be_unused += swap_info[i].inuse_pages;
1709         }
1710         val->freeswap = nr_swap_pages + nr_to_be_unused;
1711         val->totalswap = total_swap_pages + nr_to_be_unused;
1712         spin_unlock(&swap_lock);
1713 }
1714
1715 /*
1716  * Verify that a swap entry is valid and increment its swap map count.
1717  *
1718  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1719  * "permanent", but will be reclaimed by the next swapoff.
1720  */
1721 int swap_duplicate(swp_entry_t entry)
1722 {
1723         struct swap_info_struct * p;
1724         unsigned long offset, type;
1725         int result = 0;
1726
1727         if (is_migration_entry(entry))
1728                 return 1;
1729
1730         type = swp_type(entry);
1731         if (type >= nr_swapfiles)
1732                 goto bad_file;
1733         p = type + swap_info;
1734         offset = swp_offset(entry);
1735
1736         spin_lock(&swap_lock);
1737         if (offset < p->max && p->swap_map[offset]) {
1738                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1739                         p->swap_map[offset]++;
1740                         result = 1;
1741                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1742                         if (swap_overflow++ < 5)
1743                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1744                         p->swap_map[offset] = SWAP_MAP_MAX;
1745                         result = 1;
1746                 }
1747         }
1748         spin_unlock(&swap_lock);
1749 out:
1750         return result;
1751
1752 bad_file:
1753         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1754         goto out;
1755 }
1756
1757 struct swap_info_struct *
1758 get_swap_info_struct(unsigned type)
1759 {
1760         return &swap_info[type];
1761 }
1762
1763 /*
1764  * swap_lock prevents swap_map being freed. Don't grab an extra
1765  * reference on the swaphandle, it doesn't matter if it becomes unused.
1766  */
1767 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1768 {
1769         struct swap_info_struct *si;
1770         int our_page_cluster = page_cluster;
1771         pgoff_t target, toff;
1772         pgoff_t base, end;
1773         int nr_pages = 0;
1774
1775         if (!our_page_cluster)  /* no readahead */
1776                 return 0;
1777
1778         si = &swap_info[swp_type(entry)];
1779         target = swp_offset(entry);
1780         base = (target >> our_page_cluster) << our_page_cluster;
1781         end = base + (1 << our_page_cluster);
1782         if (!base)              /* first page is swap header */
1783                 base++;
1784
1785         spin_lock(&swap_lock);
1786         if (end > si->max)      /* don't go beyond end of map */
1787                 end = si->max;
1788
1789         /* Count contiguous allocated slots above our target */
1790         for (toff = target; ++toff < end; nr_pages++) {
1791                 /* Don't read in free or bad pages */
1792                 if (!si->swap_map[toff])
1793                         break;
1794                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1795                         break;
1796         }
1797         /* Count contiguous allocated slots below our target */
1798         for (toff = target; --toff >= base; nr_pages++) {
1799                 /* Don't read in free or bad pages */
1800                 if (!si->swap_map[toff])
1801                         break;
1802                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1803                         break;
1804         }
1805         spin_unlock(&swap_lock);
1806
1807         /*
1808          * Indicate starting offset, and return number of pages to get:
1809          * if only 1, say 0, since there's then no readahead to be done.
1810          */
1811         *offset = ++toff;
1812         return nr_pages? ++nr_pages: 0;
1813 }