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/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
  28 #include <asm/pgtable.h>
  29 #include <asm/tlbflush.h>
  30 #include <linux/swapops.h>
  31
  32 spinlock_t swaplock = SPIN_LOCK_UNLOCKED;
  33 unsigned int nr_swapfiles;
  34 int total_swap_pages;
  35 static int swap_overflow;
  36
  37 EXPORT_SYMBOL(total_swap_pages);
  38
  39 static const char Bad_file[] = "Bad swap file entry ";
  40 static const char Unused_file[] = "Unused swap file entry ";
  41 static const char Bad_offset[] = "Bad swap offset entry ";
  42 static const char Unused_offset[] = "Unused swap offset entry ";
  43
  44 struct swap_list_t swap_list = {-1, -1};
  45
  46 struct swap_info_struct swap_info[MAX_SWAPFILES];
  47
  48 static DECLARE_MUTEX(swapon_sem);
  49
  50 /*
  51  * Array of backing blockdevs, for swap_unplug_fn.  We need this because the
  52  * bdev->unplug_fn can sleep and we cannot hold swap_list_lock while calling
  53  * the unplug_fn.  And swap_list_lock cannot be turned into a semaphore.
  54  */
  55 static DECLARE_MUTEX(swap_bdevs_sem);
  56 static struct block_device *swap_bdevs[MAX_SWAPFILES];
  57
  58 #define SWAPFILE_CLUSTER 256
  59
  60 /*
  61  * Caller holds swap_bdevs_sem
  62  */
  63 static void install_swap_bdev(struct block_device *bdev)
  64 {
  65         int i;
  66
  67         for (i = 0; i < MAX_SWAPFILES; i++) {
  68                 if (swap_bdevs[i] == NULL) {
  69                         swap_bdevs[i] = bdev;
  70                         return;
  71                 }
  72         }
  73         BUG();
  74 }
  75
  76 static void remove_swap_bdev(struct block_device *bdev)
  77 {
  78         int i;
  79
  80         for (i = 0; i < MAX_SWAPFILES; i++) {
  81                 if (swap_bdevs[i] == bdev) {
  82                         memcpy(&swap_bdevs[i], &swap_bdevs[i + 1],
  83                                 (MAX_SWAPFILES - i - 1) * sizeof(*swap_bdevs));
  84                         swap_bdevs[MAX_SWAPFILES - 1] = NULL;
  85                         return;
  86                 }
  87         }
  88         BUG();
  89 }
  90
  91 /*
  92  * Unlike a standard unplug_io_fn, swap_unplug_io_fn is never called
  93  * through swap's backing_dev_info (which is only used by shrink_list),
  94  * but directly from sync_page when PageSwapCache: and takes the page
  95  * as argument, so that it can find the right device from swp_entry_t.
  96  */
  97 void swap_unplug_io_fn(struct page *page)
  98 {
  99         swp_entry_t entry;
 100
 101         down(&swap_bdevs_sem);
 102         entry.val = page->private;
 103         if (PageSwapCache(page)) {
 104                 struct block_device *bdev = swap_bdevs[swp_type(entry)];
 105                 struct backing_dev_info *bdi;
 106
 107                 if (bdev) {
 108                         bdi = bdev->bd_inode->i_mapping->backing_dev_info;
 109                         (*bdi->unplug_io_fn)(bdi);
 110                 }
 111         }
 112         up(&swap_bdevs_sem);
 113 }
 114
 115 static inline int scan_swap_map(struct swap_info_struct *si)
 116 {
 117         unsigned long offset;
 118         /*
 119          * We try to cluster swap pages by allocating them
 120          * sequentially in swap.  Once we've allocated
 121          * SWAPFILE_CLUSTER pages this way, however, we resort to
 122          * first-free allocation, starting a new cluster.  This
 123          * prevents us from scattering swap pages all over the entire
 124          * swap partition, so that we reduce overall disk seek times
 125          * between swap pages.  -- sct */
 126         if (si->cluster_nr) {
 127                 while (si->cluster_next <= si->highest_bit) {
 128                         offset = si->cluster_next++;
 129                         if (si->swap_map[offset])
 130                                 continue;
 131                         si->cluster_nr--;
 132                         goto got_page;
 133                 }
 134         }
 135         si->cluster_nr = SWAPFILE_CLUSTER;
 136
 137         /* try to find an empty (even not aligned) cluster. */
 138         offset = si->lowest_bit;
 139  check_next_cluster:
 140         if (offset+SWAPFILE_CLUSTER-1 <= si->highest_bit)
 141         {
 142                 int nr;
 143                 for (nr = offset; nr < offset+SWAPFILE_CLUSTER; nr++)
 144                         if (si->swap_map[nr])
 145                         {
 146                                 offset = nr+1;
 147                                 goto check_next_cluster;
 148                         }
 149                 /* We found a completly empty cluster, so start
 150                  * using it.
 151                  */
 152                 goto got_page;
 153         }
 154         /* No luck, so now go finegrined as usual. -Andrea */
 155         for (offset = si->lowest_bit; offset <= si->highest_bit ; offset++) {
 156                 if (si->swap_map[offset])
 157                         continue;
 158                 si->lowest_bit = offset+1;
 159         got_page:
 160                 if (offset == si->lowest_bit)
 161                         si->lowest_bit++;
 162                 if (offset == si->highest_bit)
 163                         si->highest_bit--;
 164                 if (si->lowest_bit > si->highest_bit) {
 165                         si->lowest_bit = si->max;
 166                         si->highest_bit = 0;
 167                 }
 168                 si->swap_map[offset] = 1;
 169                 si->inuse_pages++;
 170                 nr_swap_pages--;
 171                 si->cluster_next = offset+1;
 172                 return offset;
 173         }
 174         si->lowest_bit = si->max;
 175         si->highest_bit = 0;
 176         return 0;
 177 }
 178
 179 swp_entry_t get_swap_page(void)
 180 {
 181         struct swap_info_struct * p;
 182         unsigned long offset;
 183         swp_entry_t entry;
 184         int type, wrapped = 0;
 185
 186         entry.val = 0;  /* Out of memory */
 187         swap_list_lock();
 188         type = swap_list.next;
 189         if (type < 0)
 190                 goto out;
 191         if (nr_swap_pages <= 0)
 192                 goto out;
 193
 194         while (1) {
 195                 p = &swap_info[type];
 196                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
 197                         swap_device_lock(p);
 198                         offset = scan_swap_map(p);
 199                         swap_device_unlock(p);
 200                         if (offset) {
 201                                 entry = swp_entry(type,offset);
 202                                 type = swap_info[type].next;
 203                                 if (type < 0 ||
 204                                         p->prio != swap_info[type].prio) {
 205                                                 swap_list.next = swap_list.head;
 206                                 } else {
 207                                         swap_list.next = type;
 208                                 }
 209                                 goto out;
 210                         }
 211                 }
 212                 type = p->next;
 213                 if (!wrapped) {
 214                         if (type < 0 || p->prio != swap_info[type].prio) {
 215                                 type = swap_list.head;
 216                                 wrapped = 1;
 217                         }
 218                 } else
 219                         if (type < 0)
 220                                 goto out;       /* out of swap space */
 221         }
 222 out:
 223         swap_list_unlock();
 224         return entry;
 225 }
 226
 227 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 228 {
 229         struct swap_info_struct * p;
 230         unsigned long offset, type;
 231
 232         if (!entry.val)
 233                 goto out;
 234         type = swp_type(entry);
 235         if (type >= nr_swapfiles)
 236                 goto bad_nofile;
 237         p = & swap_info[type];
 238         if (!(p->flags & SWP_USED))
 239                 goto bad_device;
 240         offset = swp_offset(entry);
 241         if (offset >= p->max)
 242                 goto bad_offset;
 243         if (!p->swap_map[offset])
 244                 goto bad_free;
 245         swap_list_lock();
 246         if (p->prio > swap_info[swap_list.next].prio)
 247                 swap_list.next = type;
 248         swap_device_lock(p);
 249         return p;
 250
 251 bad_free:
 252         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 253         goto out;
 254 bad_offset:
 255         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 256         goto out;
 257 bad_device:
 258         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 259         goto out;
 260 bad_nofile:
 261         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 262 out:
 263         return NULL;
 264 }
 265
 266 static void swap_info_put(struct swap_info_struct * p)
 267 {
 268         swap_device_unlock(p);
 269         swap_list_unlock();
 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                         nr_swap_pages++;
 285                         p->inuse_pages--;
 286                 }
 287         }
 288         return count;
 289 }
 290
 291 /*
 292  * Caller has made sure that the swapdevice corresponding to entry
 293  * is still around or has not been recycled.
 294  */
 295 void swap_free(swp_entry_t entry)
 296 {
 297         struct swap_info_struct * p;
 298
 299         p = swap_info_get(entry);
 300         if (p) {
 301                 swap_entry_free(p, swp_offset(entry));
 302                 swap_info_put(p);
 303         }
 304 }
 305
 306 /*
 307  * Check if we're the only user of a swap page,
 308  * when the page is locked.
 309  */
 310 static int exclusive_swap_page(struct page *page)
 311 {
 312         int retval = 0;
 313         struct swap_info_struct * p;
 314         swp_entry_t entry;
 315
 316         entry.val = page->private;
 317         p = swap_info_get(entry);
 318         if (p) {
 319                 /* Is the only swap cache user the cache itself? */
 320                 if (p->swap_map[swp_offset(entry)] == 1) {
 321                         /* Recheck the page count with the swapcache lock held.. */
 322                         spin_lock(&swapper_space.tree_lock);
 323                         if (page_count(page) == 2)
 324                                 retval = 1;
 325                         spin_unlock(&swapper_space.tree_lock);
 326                 }
 327                 swap_info_put(p);
 328         }
 329         return retval;
 330 }
 331
 332 /*
 333  * We can use this swap cache entry directly
 334  * if there are no other references to it.
 335  *
 336  * Here "exclusive_swap_page()" does the real
 337  * work, but we opportunistically check whether
 338  * we need to get all the locks first..
 339  */
 340 int can_share_swap_page(struct page *page)
 341 {
 342         int retval = 0;
 343
 344         if (!PageLocked(page))
 345                 BUG();
 346         switch (page_count(page)) {
 347         case 3:
 348                 if (!PagePrivate(page))
 349                         break;
 350                 /* Fallthrough */
 351         case 2:
 352                 if (!PageSwapCache(page))
 353                         break;
 354                 retval = exclusive_swap_page(page);
 355                 break;
 356         case 1:
 357                 if (PageReserved(page))
 358                         break;
 359                 retval = 1;
 360         }
 361         return retval;
 362 }
 363
 364 /*
 365  * Work out if there are any other processes sharing this
 366  * swap cache page. Free it if you can. Return success.
 367  */
 368 int remove_exclusive_swap_page(struct page *page)
 369 {
 370         int retval;
 371         struct swap_info_struct * p;
 372         swp_entry_t entry;
 373
 374         BUG_ON(PagePrivate(page));
 375         BUG_ON(!PageLocked(page));
 376
 377         if (!PageSwapCache(page))
 378                 return 0;
 379         if (PageWriteback(page))
 380                 return 0;
 381         if (page_count(page) != 2) /* 2: us + cache */
 382                 return 0;
 383
 384         entry.val = page->private;
 385         p = swap_info_get(entry);
 386         if (!p)
 387                 return 0;
 388
 389         /* Is the only swap cache user the cache itself? */
 390         retval = 0;
 391         if (p->swap_map[swp_offset(entry)] == 1) {
 392                 /* Recheck the page count with the swapcache lock held.. */
 393                 spin_lock(&swapper_space.tree_lock);
 394                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 395                         __delete_from_swap_cache(page);
 396                         SetPageDirty(page);
 397                         retval = 1;
 398                 }
 399                 spin_unlock(&swapper_space.tree_lock);
 400         }
 401         swap_info_put(p);
 402
 403         if (retval) {
 404                 swap_free(entry);
 405                 page_cache_release(page);
 406         }
 407
 408         return retval;
 409 }
 410
 411 /*
 412  * Free the swap entry like above, but also try to
 413  * free the page cache entry if it is the last user.
 414  */
 415 void free_swap_and_cache(swp_entry_t entry)
 416 {
 417         struct swap_info_struct * p;
 418         struct page *page = NULL;
 419
 420         p = swap_info_get(entry);
 421         if (p) {
 422                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 423                         spin_lock(&swapper_space.tree_lock);
 424                         page = radix_tree_lookup(&swapper_space.page_tree,
 425                                 entry.val);
 426                         if (page && TestSetPageLocked(page))
 427                                 page = NULL;
 428                         spin_unlock(&swapper_space.tree_lock);
 429                 }
 430                 swap_info_put(p);
 431         }
 432         if (page) {
 433                 int one_user;
 434
 435                 BUG_ON(PagePrivate(page));
 436                 page_cache_get(page);
 437                 one_user = (page_count(page) == 2);
 438                 /* Only cache user (+us), or swap space full? Free it! */
 439                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
 440                         delete_from_swap_cache(page);
 441                         SetPageDirty(page);
 442                 }
 443                 unlock_page(page);
 444                 page_cache_release(page);
 445         }
 446 }
 447
 448 /*
 449  * The swap entry has been read in advance, and we return 1 to indicate
 450  * that the page has been used or is no longer needed.
 451  *
 452  * Always set the resulting pte to be nowrite (the same as COW pages
 453  * after one process has exited).  We don't know just how many PTEs will
 454  * share this swap entry, so be cautious and let do_wp_page work out
 455  * what to do if a write is requested later.
 456  */
 457 /* vma->vm_mm->page_table_lock is held */
 458 static void
 459 unuse_pte(struct vm_area_struct *vma, unsigned long address, pte_t *dir,
 460         swp_entry_t entry, struct page *page, struct pte_chain **pte_chainp)
 461 {
 462         vma->vm_mm->rss++;
 463         get_page(page);
 464         set_pte(dir, pte_mkold(mk_pte(page, vma->vm_page_prot)));
 465         *pte_chainp = page_add_rmap(page, dir, *pte_chainp);
 466         swap_free(entry);
 467 }
 468
 469 /* vma->vm_mm->page_table_lock is held */
 470 static int unuse_pmd(struct vm_area_struct * vma, pmd_t *dir,
 471         unsigned long address, unsigned long size, unsigned long offset,
 472         swp_entry_t entry, struct page *page, struct pte_chain **pte_chainp)
 473 {
 474         pte_t * pte;
 475         unsigned long end;
 476         pte_t swp_pte = swp_entry_to_pte(entry);
 477
 478         if (pmd_none(*dir))
 479                 return 0;
 480         if (pmd_bad(*dir)) {
 481                 pmd_ERROR(*dir);
 482                 pmd_clear(dir);
 483                 return 0;
 484         }
 485         pte = pte_offset_map(dir, address);
 486         offset += address & PMD_MASK;
 487         address &= ~PMD_MASK;
 488         end = address + size;
 489         if (end > PMD_SIZE)
 490                 end = PMD_SIZE;
 491         do {
 492                 /*
 493                  * swapoff spends a _lot_ of time in this loop!
 494                  * Test inline before going to call unuse_pte.
 495                  */
 496                 if (unlikely(pte_same(*pte, swp_pte))) {
 497                         unuse_pte(vma, offset + address, pte,
 498                                         entry, page, pte_chainp);
 499                         pte_unmap(pte);
 500                         return 1;
 501                 }
 502                 address += PAGE_SIZE;
 503                 pte++;
 504         } while (address && (address < end));
 505         pte_unmap(pte - 1);
 506         return 0;
 507 }
 508
 509 /* vma->vm_mm->page_table_lock is held */
 510 static int unuse_pgd(struct vm_area_struct * vma, pgd_t *dir,
 511         unsigned long address, unsigned long size,
 512         swp_entry_t entry, struct page *page, struct pte_chain **pte_chainp)
 513 {
 514         pmd_t * pmd;
 515         unsigned long offset, end;
 516
 517         if (pgd_none(*dir))
 518                 return 0;
 519         if (pgd_bad(*dir)) {
 520                 pgd_ERROR(*dir);
 521                 pgd_clear(dir);
 522                 return 0;
 523         }
 524         pmd = pmd_offset(dir, address);
 525         offset = address & PGDIR_MASK;
 526         address &= ~PGDIR_MASK;
 527         end = address + size;
 528         if (end > PGDIR_SIZE)
 529                 end = PGDIR_SIZE;
 530         if (address >= end)
 531                 BUG();
 532         do {
 533                 if (unuse_pmd(vma, pmd, address, end - address,
 534                                 offset, entry, page, pte_chainp))
 535                         return 1;
 536                 address = (address + PMD_SIZE) & PMD_MASK;
 537                 pmd++;
 538         } while (address && (address < end));
 539         return 0;
 540 }
 541
 542 /* vma->vm_mm->page_table_lock is held */
 543 static int unuse_vma(struct vm_area_struct * vma, pgd_t *pgdir,
 544         swp_entry_t entry, struct page *page, struct pte_chain **pte_chainp)
 545 {
 546         unsigned long start = vma->vm_start, end = vma->vm_end;
 547
 548         if (start >= end)
 549                 BUG();
 550         do {
 551                 if (unuse_pgd(vma, pgdir, start, end - start,
 552                                 entry, page, pte_chainp))
 553                         return 1;
 554                 start = (start + PGDIR_SIZE) & PGDIR_MASK;
 555                 pgdir++;
 556         } while (start && (start < end));
 557         return 0;
 558 }
 559
 560 static int unuse_process(struct mm_struct * mm,
 561                         swp_entry_t entry, struct page* page)
 562 {
 563         struct vm_area_struct* vma;
 564         struct pte_chain *pte_chain;
 565
 566         pte_chain = pte_chain_alloc(GFP_KERNEL);
 567         if (!pte_chain)
 568                 return -ENOMEM;
 569
 570         /*
 571          * Go through process' page directory.
 572          */
 573         spin_lock(&mm->page_table_lock);
 574         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 575                 pgd_t * pgd = pgd_offset(mm, vma->vm_start);
 576                 if (unuse_vma(vma, pgd, entry, page, &pte_chain))
 577                         break;
 578         }
 579         spin_unlock(&mm->page_table_lock);
 580         pte_chain_free(pte_chain);
 581         return 0;
 582 }
 583
 584 /*
 585  * Scan swap_map from current position to next entry still in use.
 586  * Recycle to start on reaching the end, returning 0 when empty.
 587  */
 588 static int find_next_to_unuse(struct swap_info_struct *si, int prev)
 589 {
 590         int max = si->max;
 591         int i = prev;
 592         int count;
 593
 594         /*
 595          * No need for swap_device_lock(si) here: we're just looking
 596          * for whether an entry is in use, not modifying it; false
 597          * hits are okay, and sys_swapoff() has already prevented new
 598          * allocations from this area (while holding swap_list_lock()).
 599          */
 600         for (;;) {
 601                 if (++i >= max) {
 602                         if (!prev) {
 603                                 i = 0;
 604                                 break;
 605                         }
 606                         /*
 607                          * No entries in use at top of swap_map,
 608                          * loop back to start and recheck there.
 609                          */
 610                         max = prev + 1;
 611                         prev = 0;
 612                         i = 1;
 613                 }
 614                 count = si->swap_map[i];
 615                 if (count && count != SWAP_MAP_BAD)
 616                         break;
 617         }
 618         return i;
 619 }
 620
 621 /*
 622  * We completely avoid races by reading each swap page in advance,
 623  * and then search for the process using it.  All the necessary
 624  * page table adjustments can then be made atomically.
 625  */
 626 static int try_to_unuse(unsigned int type)
 627 {
 628         struct swap_info_struct * si = &swap_info[type];
 629         struct mm_struct *start_mm;
 630         unsigned short *swap_map;
 631         unsigned short swcount;
 632         struct page *page;
 633         swp_entry_t entry;
 634         int i = 0;
 635         int retval = 0;
 636         int reset_overflow = 0;
 637         int shmem;
 638
 639         /*
 640          * When searching mms for an entry, a good strategy is to
 641          * start at the first mm we freed the previous entry from
 642          * (though actually we don't notice whether we or coincidence
 643          * freed the entry).  Initialize this start_mm with a hold.
 644          *
 645          * A simpler strategy would be to start at the last mm we
 646          * freed the previous entry from; but that would take less
 647          * advantage of mmlist ordering (now preserved by swap_out()),
 648          * which clusters forked address spaces together, most recent
 649          * child immediately after parent.  If we race with dup_mmap(),
 650          * we very much want to resolve parent before child, otherwise
 651          * we may miss some entries: using last mm would invert that.
 652          */
 653         start_mm = &init_mm;
 654         atomic_inc(&init_mm.mm_users);
 655
 656         /*
 657          * Keep on scanning until all entries have gone.  Usually,
 658          * one pass through swap_map is enough, but not necessarily:
 659          * mmput() removes mm from mmlist before exit_mmap() and its
 660          * zap_page_range().  That's not too bad, those entries are
 661          * on their way out, and handled faster there than here.
 662          * do_munmap() behaves similarly, taking the range out of mm's
 663          * vma list before zap_page_range().  But unfortunately, when
 664          * unmapping a part of a vma, it takes the whole out first,
 665          * then reinserts what's left after (might even reschedule if
 666          * open() method called) - so swap entries may be invisible
 667          * to swapoff for a while, then reappear - but that is rare.
 668          */
 669         while ((i = find_next_to_unuse(si, i))) {
 670                 if (signal_pending(current)) {
 671                         retval = -EINTR;
 672                         break;
 673                 }
 674
 675                 /*
 676                  * Get a page for the entry, using the existing swap
 677                  * cache page if there is one.  Otherwise, get a clean
 678                  * page and read the swap into it.
 679                  */
 680                 swap_map = &si->swap_map[i];
 681                 entry = swp_entry(type, i);
 682                 page = read_swap_cache_async(entry);
 683                 if (!page) {
 684                         /*
 685                          * Either swap_duplicate() failed because entry
 686                          * has been freed independently, and will not be
 687                          * reused since sys_swapoff() already disabled
 688                          * allocation from here, or alloc_page() failed.
 689                          */
 690                         if (!*swap_map)
 691                                 continue;
 692                         retval = -ENOMEM;
 693                         break;
 694                 }
 695
 696                 /*
 697                  * Don't hold on to start_mm if it looks like exiting.
 698                  */
 699                 if (atomic_read(&start_mm->mm_users) == 1) {
 700                         mmput(start_mm);
 701                         start_mm = &init_mm;
 702                         atomic_inc(&init_mm.mm_users);
 703                 }
 704
 705                 /*
 706                  * Wait for and lock page.  When do_swap_page races with
 707                  * try_to_unuse, do_swap_page can handle the fault much
 708                  * faster than try_to_unuse can locate the entry.  This
 709                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 710                  * defer to do_swap_page in such a case - in some tests,
 711                  * do_swap_page and try_to_unuse repeatedly compete.
 712                  */
 713                 wait_on_page_locked(page);
 714                 wait_on_page_writeback(page);
 715                 lock_page(page);
 716                 wait_on_page_writeback(page);
 717
 718                 /*
 719                  * Remove all references to entry, without blocking.
 720                  * Whenever we reach init_mm, there's no address space
 721                  * to search, but use it as a reminder to search shmem.
 722                  */
 723                 shmem = 0;
 724                 swcount = *swap_map;
 725                 if (swcount > 1) {
 726                         if (start_mm == &init_mm)
 727                                 shmem = shmem_unuse(entry, page);
 728                         else
 729                                 retval = unuse_process(start_mm, entry, page);
 730                 }
 731                 if (*swap_map > 1) {
 732                         int set_start_mm = (*swap_map >= swcount);
 733                         struct list_head *p = &start_mm->mmlist;
 734                         struct mm_struct *new_start_mm = start_mm;
 735                         struct mm_struct *prev_mm = start_mm;
 736                         struct mm_struct *mm;
 737
 738                         atomic_inc(&new_start_mm->mm_users);
 739                         atomic_inc(&prev_mm->mm_users);
 740                         spin_lock(&mmlist_lock);
 741                         while (*swap_map > 1 && !retval &&
 742                                         (p = p->next) != &start_mm->mmlist) {
 743                                 mm = list_entry(p, struct mm_struct, mmlist);
 744                                 atomic_inc(&mm->mm_users);
 745                                 spin_unlock(&mmlist_lock);
 746                                 mmput(prev_mm);
 747                                 prev_mm = mm;
 748
 749                                 cond_resched();
 750
 751                                 swcount = *swap_map;
 752                                 if (swcount <= 1)
 753                                         ;
 754                                 else if (mm == &init_mm) {
 755                                         set_start_mm = 1;
 756                                         shmem = shmem_unuse(entry, page);
 757                                 } else
 758                                         retval = unuse_process(mm, entry, page);
 759                                 if (set_start_mm && *swap_map < swcount) {
 760                                         mmput(new_start_mm);
 761                                         atomic_inc(&mm->mm_users);
 762                                         new_start_mm = mm;
 763                                         set_start_mm = 0;
 764                                 }
 765                                 spin_lock(&mmlist_lock);
 766                         }
 767                         spin_unlock(&mmlist_lock);
 768                         mmput(prev_mm);
 769                         mmput(start_mm);
 770                         start_mm = new_start_mm;
 771                 }
 772                 if (retval) {
 773                         unlock_page(page);
 774                         page_cache_release(page);
 775                         break;
 776                 }
 777
 778                 /*
 779                  * How could swap count reach 0x7fff when the maximum
 780                  * pid is 0x7fff, and there's no way to repeat a swap
 781                  * page within an mm (except in shmem, where it's the
 782                  * shared object which takes the reference count)?
 783                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 784                  *
 785                  * If that's wrong, then we should worry more about
 786                  * exit_mmap() and do_munmap() cases described above:
 787                  * we might be resetting SWAP_MAP_MAX too early here.
 788                  * We know "Undead"s can happen, they're okay, so don't
 789                  * report them; but do report if we reset SWAP_MAP_MAX.
 790                  */
 791                 if (*swap_map == SWAP_MAP_MAX) {
 792                         swap_device_lock(si);
 793                         *swap_map = 1;
 794                         swap_device_unlock(si);
 795                         reset_overflow = 1;
 796                 }
 797
 798                 /*
 799                  * If a reference remains (rare), we would like to leave
 800                  * the page in the swap cache; but try_to_unmap could
 801                  * then re-duplicate the entry once we drop page lock,
 802                  * so we might loop indefinitely; also, that page could
 803                  * not be swapped out to other storage meanwhile.  So:
 804                  * delete from cache even if there's another reference,
 805                  * after ensuring that the data has been saved to disk -
 806                  * since if the reference remains (rarer), it will be
 807                  * read from disk into another page.  Splitting into two
 808                  * pages would be incorrect if swap supported "shared
 809                  * private" pages, but they are handled by tmpfs files.
 810                  *
 811                  * Note shmem_unuse already deleted a swappage from
 812                  * the swap cache, unless the move to filepage failed:
 813                  * in which case it left swappage in cache, lowered its
 814                  * swap count to pass quickly through the loops above,
 815                  * and now we must reincrement count to try again later.
 816                  */
 817                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 818                         struct writeback_control wbc = {
 819                                 .sync_mode = WB_SYNC_NONE,
 820                         };
 821
 822                         swap_writepage(page, &wbc);
 823                         lock_page(page);
 824                         wait_on_page_writeback(page);
 825                 }
 826                 if (PageSwapCache(page)) {
 827                         if (shmem)
 828                                 swap_duplicate(entry);
 829                         else
 830                                 delete_from_swap_cache(page);
 831                 }
 832
 833                 /*
 834                  * So we could skip searching mms once swap count went
 835                  * to 1, we did not mark any present ptes as dirty: must
 836                  * mark page dirty so shrink_list will preserve it.
 837                  */
 838                 SetPageDirty(page);
 839                 unlock_page(page);
 840                 page_cache_release(page);
 841
 842                 /*
 843                  * Make sure that we aren't completely killing
 844                  * interactive performance.
 845                  */
 846                 cond_resched();
 847         }
 848
 849         mmput(start_mm);
 850         if (reset_overflow) {
 851                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 852                 swap_overflow = 0;
 853         }
 854         return retval;
 855 }
 856
 857 /*
 858  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 859  * corresponds to page offset `offset'.
 860  */
 861 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 862 {
 863         struct swap_extent *se = sis->curr_swap_extent;
 864         struct swap_extent *start_se = se;
 865
 866         for ( ; ; ) {
 867                 struct list_head *lh;
 868
 869                 if (se->start_page <= offset &&
 870                                 offset < (se->start_page + se->nr_pages)) {
 871                         return se->start_block + (offset - se->start_page);
 872                 }
 873                 lh = se->list.prev;
 874                 if (lh == &sis->extent_list)
 875                         lh = lh->prev;
 876                 se = list_entry(lh, struct swap_extent, list);
 877                 sis->curr_swap_extent = se;
 878                 BUG_ON(se == start_se);         /* It *must* be present */
 879         }
 880 }
 881
 882 /*
 883  * Free all of a swapdev's extent information
 884  */
 885 static void destroy_swap_extents(struct swap_info_struct *sis)
 886 {
 887         while (!list_empty(&sis->extent_list)) {
 888                 struct swap_extent *se;
 889
 890                 se = list_entry(sis->extent_list.next,
 891                                 struct swap_extent, list);
 892                 list_del(&se->list);
 893                 kfree(se);
 894         }
 895         sis->nr_extents = 0;
 896 }
 897
 898 /*
 899  * Add a block range (and the corresponding page range) into this swapdev's
 900  * extent list.  The extent list is kept sorted in block order.
 901  *
 902  * This function rather assumes that it is called in ascending sector_t order.
 903  * It doesn't look for extent coalescing opportunities.
 904  */
 905 static int
 906 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 907                 unsigned long nr_pages, sector_t start_block)
 908 {
 909         struct swap_extent *se;
 910         struct swap_extent *new_se;
 911         struct list_head *lh;
 912
 913         lh = sis->extent_list.next;     /* The highest-addressed block */
 914         while (lh != &sis->extent_list) {
 915                 se = list_entry(lh, struct swap_extent, list);
 916                 if (se->start_block + se->nr_pages == start_block &&
 917                     se->start_page  + se->nr_pages == start_page) {
 918                         /* Merge it */
 919                         se->nr_pages += nr_pages;
 920                         return 0;
 921                 }
 922                 lh = lh->next;
 923         }
 924
 925         /*
 926          * No merge.  Insert a new extent, preserving ordering.
 927          */
 928         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 929         if (new_se == NULL)
 930                 return -ENOMEM;
 931         new_se->start_page = start_page;
 932         new_se->nr_pages = nr_pages;
 933         new_se->start_block = start_block;
 934
 935         lh = sis->extent_list.prev;     /* The lowest block */
 936         while (lh != &sis->extent_list) {
 937                 se = list_entry(lh, struct swap_extent, list);
 938                 if (se->start_block > start_block)
 939                         break;
 940                 lh = lh->prev;
 941         }
 942         list_add_tail(&new_se->list, lh);
 943         sis->nr_extents++;
 944         return 0;
 945 }
 946
 947 /*
 948  * A `swap extent' is a simple thing which maps a contiguous range of pages
 949  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 950  * is built at swapon time and is then used at swap_writepage/swap_readpage
 951  * time for locating where on disk a page belongs.
 952  *
 953  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 954  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 955  * swap files identically.
 956  *
 957  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
 958  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
 959  * swapfiles are handled *identically* after swapon time.
 960  *
 961  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
 962  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
 963  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
 964  * requirements, they are simply tossed out - we will never use those blocks
 965  * for swapping.
 966  *
 967  * For S_ISREG swapfiles we hold i_sem across the life of the swapon.  This
 968  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
 969  * which will scribble on the fs.
 970  *
 971  * The amount of disk space which a single swap extent represents varies.
 972  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
 973  * extents in the list.  To avoid much list walking, we cache the previous
 974  * search location in `curr_swap_extent', and start new searches from there.
 975  * This is extremely effective.  The average number of iterations in
 976  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
 977  */
 978 static int setup_swap_extents(struct swap_info_struct *sis)
 979 {
 980         struct inode *inode;
 981         unsigned blocks_per_page;
 982         unsigned long page_no;
 983         unsigned blkbits;
 984         sector_t probe_block;
 985         sector_t last_block;
 986         int ret;
 987
 988         inode = sis->swap_file->f_mapping->host;
 989         if (S_ISBLK(inode->i_mode)) {
 990                 ret = add_swap_extent(sis, 0, sis->max, 0);
 991                 goto done;
 992         }
 993
 994         blkbits = inode->i_blkbits;
 995         blocks_per_page = PAGE_SIZE >> blkbits;
 996
 997         /*
 998          * Map all the blocks into the extent list.  This code doesn't try
 999          * to be very smart.
1000          */
1001         probe_block = 0;
1002         page_no = 0;
1003         last_block = i_size_read(inode) >> blkbits;
1004         while ((probe_block + blocks_per_page) <= last_block &&
1005                         page_no < sis->max) {
1006                 unsigned block_in_page;
1007                 sector_t first_block;
1008
1009                 first_block = bmap(inode, probe_block);
1010                 if (first_block == 0)
1011                         goto bad_bmap;
1012
1013                 /*
1014                  * It must be PAGE_SIZE aligned on-disk
1015                  */
1016                 if (first_block & (blocks_per_page - 1)) {
1017                         probe_block++;
1018                         goto reprobe;
1019                 }
1020
1021                 for (block_in_page = 1; block_in_page < blocks_per_page;
1022                                         block_in_page++) {
1023                         sector_t block;
1024
1025                         block = bmap(inode, probe_block + block_in_page);
1026                         if (block == 0)
1027                                 goto bad_bmap;
1028                         if (block != first_block + block_in_page) {
1029                                 /* Discontiguity */
1030                                 probe_block++;
1031                                 goto reprobe;
1032                         }
1033                 }
1034
1035                 /*
1036                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1037                  */
1038                 ret = add_swap_extent(sis, page_no, 1,
1039                                 first_block >> (PAGE_SHIFT - blkbits));
1040                 if (ret)
1041                         goto out;
1042                 page_no++;
1043                 probe_block += blocks_per_page;
1044 reprobe:
1045                 continue;
1046         }
1047         ret = 0;
1048         if (page_no == 0)
1049                 ret = -EINVAL;
1050         sis->max = page_no;
1051         sis->highest_bit = page_no - 1;
1052 done:
1053         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1054                                         struct swap_extent, list);
1055         goto out;
1056 bad_bmap:
1057         printk(KERN_ERR "swapon: swapfile has holes\n");
1058         ret = -EINVAL;
1059 out:
1060         return ret;
1061 }
1062
1063 #if 0   /* We don't need this yet */
1064 #include <linux/backing-dev.h>
1065 int page_queue_congested(struct page *page)
1066 {
1067         struct backing_dev_info *bdi;
1068
1069         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1070
1071         if (PageSwapCache(page)) {
1072                 swp_entry_t entry = { .val = page->private };
1073                 struct swap_info_struct *sis;
1074
1075                 sis = get_swap_info_struct(swp_type(entry));
1076                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1077         } else
1078                 bdi = page->mapping->backing_dev_info;
1079         return bdi_write_congested(bdi);
1080 }
1081 #endif
1082
1083 asmlinkage long sys_swapoff(const char __user * specialfile)
1084 {
1085         struct swap_info_struct * p = NULL;
1086         unsigned short *swap_map;
1087         struct file *swap_file, *victim;
1088         struct address_space *mapping;
1089         char * pathname;
1090         int i, type, prev;
1091         int err;
1092
1093         if (!capable(CAP_SYS_ADMIN))
1094                 return -EPERM;
1095
1096         pathname = getname(specialfile);
1097         err = PTR_ERR(pathname);
1098         if (IS_ERR(pathname))
1099                 goto out;
1100
1101         victim = filp_open(pathname, O_RDWR, 0);
1102         putname(pathname);
1103         err = PTR_ERR(victim);
1104         if (IS_ERR(victim))
1105                 goto out;
1106
1107         mapping = victim->f_mapping;
1108         prev = -1;
1109         swap_list_lock();
1110         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1111                 p = swap_info + type;
1112                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1113                         if (p->swap_file->f_mapping == mapping)
1114                                 break;
1115                 }
1116                 prev = type;
1117         }
1118         if (type < 0) {
1119                 err = -EINVAL;
1120                 swap_list_unlock();
1121                 goto out_dput;
1122         }
1123         if (!security_vm_enough_memory(p->pages))
1124                 vm_unacct_memory(p->pages);
1125         else {
1126                 err = -ENOMEM;
1127                 swap_list_unlock();
1128                 goto out_dput;
1129         }
1130         if (prev < 0) {
1131                 swap_list.head = p->next;
1132         } else {
1133                 swap_info[prev].next = p->next;
1134         }
1135         if (type == swap_list.next) {
1136                 /* just pick something that's safe... */
1137                 swap_list.next = swap_list.head;
1138         }
1139         nr_swap_pages -= p->pages;
1140         total_swap_pages -= p->pages;
1141         p->flags &= ~SWP_WRITEOK;
1142         swap_list_unlock();
1143         current->flags |= PF_SWAPOFF;
1144         err = try_to_unuse(type);
1145         current->flags &= ~PF_SWAPOFF;
1146         if (err) {
1147                 /* re-insert swap space back into swap_list */
1148                 swap_list_lock();
1149                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1150                         if (p->prio >= swap_info[i].prio)
1151                                 break;
1152                 p->next = i;
1153                 if (prev < 0)
1154                         swap_list.head = swap_list.next = p - swap_info;
1155                 else
1156                         swap_info[prev].next = p - swap_info;
1157                 nr_swap_pages += p->pages;
1158                 total_swap_pages += p->pages;
1159                 p->flags |= SWP_WRITEOK;
1160                 swap_list_unlock();
1161                 goto out_dput;
1162         }
1163         down(&swapon_sem);
1164         down(&swap_bdevs_sem);
1165         swap_list_lock();
1166         swap_device_lock(p);
1167         swap_file = p->swap_file;
1168         p->swap_file = NULL;
1169         p->max = 0;
1170         swap_map = p->swap_map;
1171         p->swap_map = NULL;
1172         p->flags = 0;
1173         destroy_swap_extents(p);
1174         swap_device_unlock(p);
1175         swap_list_unlock();
1176         remove_swap_bdev(p->bdev);
1177         up(&swap_bdevs_sem);
1178         up(&swapon_sem);
1179         vfree(swap_map);
1180         if (S_ISBLK(mapping->host->i_mode)) {
1181                 struct block_device *bdev = I_BDEV(mapping->host);
1182                 set_blocksize(bdev, p->old_block_size);
1183                 bd_release(bdev);
1184         } else {
1185                 up(&mapping->host->i_sem);
1186         }
1187         filp_close(swap_file, NULL);
1188         err = 0;
1189
1190 out_dput:
1191         filp_close(victim, NULL);
1192 out:
1193         return err;
1194 }
1195
1196 #ifdef CONFIG_PROC_FS
1197 /* iterator */
1198 static void *swap_start(struct seq_file *swap, loff_t *pos)
1199 {
1200         struct swap_info_struct *ptr = swap_info;
1201         int i;
1202         loff_t l = *pos;
1203
1204         down(&swapon_sem);
1205
1206         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1207                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1208                         continue;
1209                 if (!l--)
1210                         return ptr;
1211         }
1212
1213         return NULL;
1214 }
1215
1216 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1217 {
1218         struct swap_info_struct *ptr = v;
1219         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1220
1221         for (++ptr; ptr < endptr; ptr++) {
1222                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1223                         continue;
1224                 ++*pos;
1225                 return ptr;
1226         }
1227
1228         return NULL;
1229 }
1230
1231 static void swap_stop(struct seq_file *swap, void *v)
1232 {
1233         up(&swapon_sem);
1234 }
1235
1236 static int swap_show(struct seq_file *swap, void *v)
1237 {
1238         struct swap_info_struct *ptr = v;
1239         struct file *file;
1240         int len;
1241
1242         if (v == swap_info)
1243                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1244
1245         file = ptr->swap_file;
1246         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1247         seq_printf(swap, "%*s%s\t%d\t%ld\t%d\n",
1248                        len < 40 ? 40 - len : 1, " ",
1249                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1250                                 "partition" : "file\t",
1251                        ptr->pages << (PAGE_SHIFT - 10),
1252                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1253                        ptr->prio);
1254         return 0;
1255 }
1256
1257 static struct seq_operations swaps_op = {
1258         .start =        swap_start,
1259         .next =         swap_next,
1260         .stop =         swap_stop,
1261         .show =         swap_show
1262 };
1263
1264 static int swaps_open(struct inode *inode, struct file *file)
1265 {
1266         return seq_open(file, &swaps_op);
1267 }
1268
1269 static struct file_operations proc_swaps_operations = {
1270         .open           = swaps_open,
1271         .read           = seq_read,
1272         .llseek         = seq_lseek,
1273         .release        = seq_release,
1274 };
1275
1276 static int __init procswaps_init(void)
1277 {
1278         struct proc_dir_entry *entry;
1279
1280         entry = create_proc_entry("swaps", 0, NULL);
1281         if (entry)
1282                 entry->proc_fops = &proc_swaps_operations;
1283         return 0;
1284 }
1285 __initcall(procswaps_init);
1286 #endif /* CONFIG_PROC_FS */
1287
1288 /*
1289  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1290  *
1291  * The swapon system call
1292  */
1293 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1294 {
1295         struct swap_info_struct * p;
1296         char *name = NULL;
1297         struct block_device *bdev = NULL;
1298         struct file *swap_file = NULL;
1299         struct address_space *mapping;
1300         unsigned int type;
1301         int i, prev;
1302         int error;
1303         static int least_priority;
1304         union swap_header *swap_header = 0;
1305         int swap_header_version;
1306         int nr_good_pages = 0;
1307         unsigned long maxpages = 1;
1308         int swapfilesize;
1309         unsigned short *swap_map;
1310         struct page *page = NULL;
1311         struct inode *inode = NULL;
1312         int did_down = 0;
1313
1314         if (!capable(CAP_SYS_ADMIN))
1315                 return -EPERM;
1316         swap_list_lock();
1317         p = swap_info;
1318         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1319                 if (!(p->flags & SWP_USED))
1320                         break;
1321         error = -EPERM;
1322         /*
1323          * Test if adding another swap device is possible. There are
1324          * two limiting factors: 1) the number of bits for the swap
1325          * type swp_entry_t definition and 2) the number of bits for
1326          * the swap type in the swap ptes as defined by the different
1327          * architectures. To honor both limitations a swap entry
1328          * with swap offset 0 and swap type ~0UL is created, encoded
1329          * to a swap pte, decoded to a swp_entry_t again and finally
1330          * the swap type part is extracted. This will mask all bits
1331          * from the initial ~0UL that can't be encoded in either the
1332          * swp_entry_t or the architecture definition of a swap pte.
1333          */
1334         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1335                 swap_list_unlock();
1336                 goto out;
1337         }
1338         if (type >= nr_swapfiles)
1339                 nr_swapfiles = type+1;
1340         INIT_LIST_HEAD(&p->extent_list);
1341         p->flags = SWP_USED;
1342         p->nr_extents = 0;
1343         p->swap_file = NULL;
1344         p->old_block_size = 0;
1345         p->swap_map = NULL;
1346         p->lowest_bit = 0;
1347         p->highest_bit = 0;
1348         p->cluster_nr = 0;
1349         p->inuse_pages = 0;
1350         p->sdev_lock = SPIN_LOCK_UNLOCKED;
1351         p->next = -1;
1352         if (swap_flags & SWAP_FLAG_PREFER) {
1353                 p->prio =
1354                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1355         } else {
1356                 p->prio = --least_priority;
1357         }
1358         swap_list_unlock();
1359         name = getname(specialfile);
1360         error = PTR_ERR(name);
1361         if (IS_ERR(name)) {
1362                 name = NULL;
1363                 goto bad_swap_2;
1364         }
1365         swap_file = filp_open(name, O_RDWR, 0);
1366         error = PTR_ERR(swap_file);
1367         if (IS_ERR(swap_file)) {
1368                 swap_file = NULL;
1369                 goto bad_swap_2;
1370         }
1371
1372         p->swap_file = swap_file;
1373         mapping = swap_file->f_mapping;
1374         inode = mapping->host;
1375
1376         error = -EBUSY;
1377         for (i = 0; i < nr_swapfiles; i++) {
1378                 struct swap_info_struct *q = &swap_info[i];
1379
1380                 if (i == type || !q->swap_file)
1381                         continue;
1382                 if (mapping == q->swap_file->f_mapping)
1383                         goto bad_swap;
1384         }
1385
1386         error = -EINVAL;
1387         if (S_ISBLK(inode->i_mode)) {
1388                 bdev = I_BDEV(inode);
1389                 error = bd_claim(bdev, sys_swapon);
1390                 if (error < 0) {
1391                         bdev = NULL;
1392                         goto bad_swap;
1393                 }
1394                 p->old_block_size = block_size(bdev);
1395                 error = set_blocksize(bdev, PAGE_SIZE);
1396                 if (error < 0)
1397                         goto bad_swap;
1398                 p->bdev = bdev;
1399         } else if (S_ISREG(inode->i_mode)) {
1400                 p->bdev = inode->i_sb->s_bdev;
1401                 down(&inode->i_sem);
1402                 did_down = 1;
1403         } else {
1404                 goto bad_swap;
1405         }
1406
1407         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1408
1409         /*
1410          * Read the swap header.
1411          */
1412         if (!mapping->a_ops->readpage) {
1413                 error = -EINVAL;
1414                 goto bad_swap;
1415         }
1416         page = read_cache_page(mapping, 0,
1417                         (filler_t *)mapping->a_ops->readpage, swap_file);
1418         if (IS_ERR(page)) {
1419                 error = PTR_ERR(page);
1420                 goto bad_swap;
1421         }
1422         wait_on_page_locked(page);
1423         if (!PageUptodate(page))
1424                 goto bad_swap;
1425         kmap(page);
1426         swap_header = page_address(page);
1427
1428         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1429                 swap_header_version = 1;
1430         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1431                 swap_header_version = 2;
1432         else {
1433                 printk("Unable to find swap-space signature\n");
1434                 error = -EINVAL;
1435                 goto bad_swap;
1436         }
1437
1438         switch (swap_header_version) {
1439         case 1:
1440                 printk(KERN_ERR "version 0 swap is no longer supported. "
1441                         "Use mkswap -v1 %s\n", name);
1442                 error = -EINVAL;
1443                 goto bad_swap;
1444         case 2:
1445                 /* Check the swap header's sub-version and the size of
1446                    the swap file and bad block lists */
1447                 if (swap_header->info.version != 1) {
1448                         printk(KERN_WARNING
1449                                "Unable to handle swap header version %d\n",
1450                                swap_header->info.version);
1451                         error = -EINVAL;
1452                         goto bad_swap;
1453                 }
1454
1455                 p->lowest_bit  = 1;
1456                 /*
1457                  * Find out how many pages are allowed for a single swap
1458                  * device. There are two limiting factors: 1) the number of
1459                  * bits for the swap offset in the swp_entry_t type and
1460                  * 2) the number of bits in the a swap pte as defined by
1461                  * the different architectures. In order to find the
1462                  * largest possible bit mask a swap entry with swap type 0
1463                  * and swap offset ~0UL is created, encoded to a swap pte,
1464                  * decoded to a swp_entry_t again and finally the swap
1465                  * offset is extracted. This will mask all the bits from
1466                  * the initial ~0UL mask that can't be encoded in either
1467                  * the swp_entry_t or the architecture definition of a
1468                  * swap pte.
1469                  */
1470                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1471                 if (maxpages > swap_header->info.last_page)
1472                         maxpages = swap_header->info.last_page;
1473                 p->highest_bit = maxpages - 1;
1474
1475                 error = -EINVAL;
1476                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1477                         goto bad_swap;
1478
1479                 /* OK, set up the swap map and apply the bad block list */
1480                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1481                         error = -ENOMEM;
1482                         goto bad_swap;
1483                 }
1484
1485                 error = 0;
1486                 memset(p->swap_map, 0, maxpages * sizeof(short));
1487                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1488                         int page = swap_header->info.badpages[i];
1489                         if (page <= 0 || page >= swap_header->info.last_page)
1490                                 error = -EINVAL;
1491                         else
1492                                 p->swap_map[page] = SWAP_MAP_BAD;
1493                 }
1494                 nr_good_pages = swap_header->info.last_page -
1495                                 swap_header->info.nr_badpages -
1496                                 1 /* header page */;
1497                 if (error)
1498                         goto bad_swap;
1499         }
1500
1501         if (swapfilesize && maxpages > swapfilesize) {
1502                 printk(KERN_WARNING
1503                        "Swap area shorter than signature indicates\n");
1504                 error = -EINVAL;
1505                 goto bad_swap;
1506         }
1507         if (!nr_good_pages) {
1508                 printk(KERN_WARNING "Empty swap-file\n");
1509                 error = -EINVAL;
1510                 goto bad_swap;
1511         }
1512         p->swap_map[0] = SWAP_MAP_BAD;
1513         p->max = maxpages;
1514         p->pages = nr_good_pages;
1515
1516         error = setup_swap_extents(p);
1517         if (error)
1518                 goto bad_swap;
1519
1520         down(&swapon_sem);
1521         down(&swap_bdevs_sem);
1522         swap_list_lock();
1523         swap_device_lock(p);
1524         p->flags = SWP_ACTIVE;
1525         nr_swap_pages += nr_good_pages;
1526         total_swap_pages += nr_good_pages;
1527         printk(KERN_INFO "Adding %dk swap on %s.  Priority:%d extents:%d\n",
1528                 nr_good_pages<<(PAGE_SHIFT-10), name,
1529                 p->prio, p->nr_extents);
1530
1531         /* insert swap space into swap_list: */
1532         prev = -1;
1533         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1534                 if (p->prio >= swap_info[i].prio) {
1535                         break;
1536                 }
1537                 prev = i;
1538         }
1539         p->next = i;
1540         if (prev < 0) {
1541                 swap_list.head = swap_list.next = p - swap_info;
1542         } else {
1543                 swap_info[prev].next = p - swap_info;
1544         }
1545         swap_device_unlock(p);
1546         swap_list_unlock();
1547         install_swap_bdev(p->bdev);
1548         up(&swap_bdevs_sem);
1549         up(&swapon_sem);
1550         error = 0;
1551         goto out;
1552 bad_swap:
1553         if (bdev) {
1554                 set_blocksize(bdev, p->old_block_size);
1555                 bd_release(bdev);
1556         }
1557 bad_swap_2:
1558         swap_list_lock();
1559         swap_map = p->swap_map;
1560         p->swap_file = NULL;
1561         p->swap_map = NULL;
1562         p->flags = 0;
1563         if (!(swap_flags & SWAP_FLAG_PREFER))
1564                 ++least_priority;
1565         swap_list_unlock();
1566         destroy_swap_extents(p);
1567         if (swap_map)
1568                 vfree(swap_map);
1569         if (swap_file)
1570                 filp_close(swap_file, NULL);
1571 out:
1572         if (page && !IS_ERR(page)) {
1573                 kunmap(page);
1574                 page_cache_release(page);
1575         }
1576         if (name)
1577                 putname(name);
1578         if (error && did_down)
1579                 up(&inode->i_sem);
1580         return error;
1581 }
1582
1583 void si_swapinfo(struct sysinfo *val)
1584 {
1585         unsigned int i;
1586         unsigned long nr_to_be_unused = 0;
1587
1588         swap_list_lock();
1589         for (i = 0; i < nr_swapfiles; i++) {
1590                 if (!(swap_info[i].flags & SWP_USED) ||
1591                      (swap_info[i].flags & SWP_WRITEOK))
1592                         continue;
1593                 nr_to_be_unused += swap_info[i].inuse_pages;
1594         }
1595         val->freeswap = nr_swap_pages + nr_to_be_unused;
1596         val->totalswap = total_swap_pages + nr_to_be_unused;
1597         swap_list_unlock();
1598 }
1599
1600 /*
1601  * Verify that a swap entry is valid and increment its swap map count.
1602  *
1603  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1604  * "permanent", but will be reclaimed by the next swapoff.
1605  */
1606 int swap_duplicate(swp_entry_t entry)
1607 {
1608         struct swap_info_struct * p;
1609         unsigned long offset, type;
1610         int result = 0;
1611
1612         type = swp_type(entry);
1613         if (type >= nr_swapfiles)
1614                 goto bad_file;
1615         p = type + swap_info;
1616         offset = swp_offset(entry);
1617
1618         swap_device_lock(p);
1619         if (offset < p->max && p->swap_map[offset]) {
1620                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1621                         p->swap_map[offset]++;
1622                         result = 1;
1623                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1624                         if (swap_overflow++ < 5)
1625                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1626                         p->swap_map[offset] = SWAP_MAP_MAX;
1627                         result = 1;
1628                 }
1629         }
1630         swap_device_unlock(p);
1631 out:
1632         return result;
1633
1634 bad_file:
1635         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1636         goto out;
1637 }
1638
1639 struct swap_info_struct *
1640 get_swap_info_struct(unsigned type)
1641 {
1642         return &swap_info[type];
1643 }
1644
1645 /*
1646  * swap_device_lock prevents swap_map being freed. Don't grab an extra
1647  * reference on the swaphandle, it doesn't matter if it becomes unused.
1648  */
1649 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1650 {
1651         int ret = 0, i = 1 << page_cluster;
1652         unsigned long toff;
1653         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1654
1655         if (!page_cluster)      /* no readahead */
1656                 return 0;
1657         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1658         if (!toff)              /* first page is swap header */
1659                 toff++, i--;
1660         *offset = toff;
1661
1662         swap_device_lock(swapdev);
1663         do {
1664                 /* Don't read-ahead past the end of the swap area */
1665                 if (toff >= swapdev->max)
1666                         break;
1667                 /* Don't read in free or bad pages */
1668                 if (!swapdev->swap_map[toff])
1669                         break;
1670                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1671                         break;
1672                 toff++;
1673                 ret++;
1674         } while (--i);
1675         swap_device_unlock(swapdev);
1676         return ret;
1677 }