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