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
  29 #include <asm/pgtable.h>
  30 #include <asm/tlbflush.h>
  31 #include <linux/swapops.h>
  32
  33 spinlock_t swaplock = SPIN_LOCK_UNLOCKED;
  34 unsigned int nr_swapfiles;
  35 long total_swap_pages;
  36 static int swap_overflow;
  37
  38 EXPORT_SYMBOL(total_swap_pages);
  39
  40 static const char Bad_file[] = "Bad swap file entry ";
  41 static const char Unused_file[] = "Unused swap file entry ";
  42 static const char Bad_offset[] = "Bad swap offset entry ";
  43 static const char Unused_offset[] = "Unused swap offset entry ";
  44
  45 struct swap_list_t swap_list = {-1, -1};
  46
  47 struct swap_info_struct swap_info[MAX_SWAPFILES];
  48
  49 static DECLARE_MUTEX(swapon_sem);
  50
  51 /*
  52  * We need this because the bdev->unplug_fn can sleep and we cannot
  53  * hold swap_list_lock while calling the unplug_fn. And swap_list_lock
  54  * cannot be turned into a semaphore.
  55  */
  56 static DECLARE_RWSEM(swap_unplug_sem);
  57
  58 #define SWAPFILE_CLUSTER 256
  59
  60 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  61 {
  62         swp_entry_t entry;
  63
  64         down_read(&swap_unplug_sem);
  65         entry.val = page->private;
  66         if (PageSwapCache(page)) {
  67                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  68                 struct backing_dev_info *bdi;
  69
  70                 /*
  71                  * If the page is removed from swapcache from under us (with a
  72                  * racy try_to_unuse/swapoff) we need an additional reference
  73                  * count to avoid reading garbage from page->private above. If
  74                  * the WARN_ON triggers during a swapoff it maybe the race
  75                  * condition and it's harmless. However if it triggers without
  76                  * swapoff it signals a problem.
  77                  */
  78                 WARN_ON(page_count(page) <= 1);
  79
  80                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  81                 bdi->unplug_io_fn(bdi, page);
  82         }
  83         up_read(&swap_unplug_sem);
  84 }
  85
  86 static inline int scan_swap_map(struct swap_info_struct *si)
  87 {
  88         unsigned long offset;
  89         /*
  90          * We try to cluster swap pages by allocating them
  91          * sequentially in swap.  Once we've allocated
  92          * SWAPFILE_CLUSTER pages this way, however, we resort to
  93          * first-free allocation, starting a new cluster.  This
  94          * prevents us from scattering swap pages all over the entire
  95          * swap partition, so that we reduce overall disk seek times
  96          * between swap pages.  -- sct */
  97         if (si->cluster_nr) {
  98                 while (si->cluster_next <= si->highest_bit) {
  99                         offset = si->cluster_next++;
 100                         if (si->swap_map[offset])
 101                                 continue;
 102                         si->cluster_nr--;
 103                         goto got_page;
 104                 }
 105         }
 106         si->cluster_nr = SWAPFILE_CLUSTER;
 107
 108         /* try to find an empty (even not aligned) cluster. */
 109         offset = si->lowest_bit;
 110  check_next_cluster:
 111         if (offset+SWAPFILE_CLUSTER-1 <= si->highest_bit)
 112         {
 113                 int nr;
 114                 for (nr = offset; nr < offset+SWAPFILE_CLUSTER; nr++)
 115                         if (si->swap_map[nr])
 116                         {
 117                                 offset = nr+1;
 118                                 goto check_next_cluster;
 119                         }
 120                 /* We found a completly empty cluster, so start
 121                  * using it.
 122                  */
 123                 goto got_page;
 124         }
 125         /* No luck, so now go finegrined as usual. -Andrea */
 126         for (offset = si->lowest_bit; offset <= si->highest_bit ; offset++) {
 127                 if (si->swap_map[offset])
 128                         continue;
 129                 si->lowest_bit = offset+1;
 130         got_page:
 131                 if (offset == si->lowest_bit)
 132                         si->lowest_bit++;
 133                 if (offset == si->highest_bit)
 134                         si->highest_bit--;
 135                 if (si->lowest_bit > si->highest_bit) {
 136                         si->lowest_bit = si->max;
 137                         si->highest_bit = 0;
 138                 }
 139                 si->swap_map[offset] = 1;
 140                 si->inuse_pages++;
 141                 nr_swap_pages--;
 142                 si->cluster_next = offset+1;
 143                 return offset;
 144         }
 145         si->lowest_bit = si->max;
 146         si->highest_bit = 0;
 147         return 0;
 148 }
 149
 150 swp_entry_t get_swap_page(void)
 151 {
 152         struct swap_info_struct * p;
 153         unsigned long offset;
 154         swp_entry_t entry;
 155         int type, wrapped = 0;
 156
 157         entry.val = 0;  /* Out of memory */
 158         swap_list_lock();
 159         type = swap_list.next;
 160         if (type < 0)
 161                 goto out;
 162         if (nr_swap_pages <= 0)
 163                 goto out;
 164
 165         while (1) {
 166                 p = &swap_info[type];
 167                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
 168                         swap_device_lock(p);
 169                         offset = scan_swap_map(p);
 170                         swap_device_unlock(p);
 171                         if (offset) {
 172                                 entry = swp_entry(type,offset);
 173                                 type = swap_info[type].next;
 174                                 if (type < 0 ||
 175                                         p->prio != swap_info[type].prio) {
 176                                                 swap_list.next = swap_list.head;
 177                                 } else {
 178                                         swap_list.next = type;
 179                                 }
 180                                 goto out;
 181                         }
 182                 }
 183                 type = p->next;
 184                 if (!wrapped) {
 185                         if (type < 0 || p->prio != swap_info[type].prio) {
 186                                 type = swap_list.head;
 187                                 wrapped = 1;
 188                         }
 189                 } else
 190                         if (type < 0)
 191                                 goto out;       /* out of swap space */
 192         }
 193 out:
 194         swap_list_unlock();
 195         return entry;
 196 }
 197
 198 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 199 {
 200         struct swap_info_struct * p;
 201         unsigned long offset, type;
 202
 203         if (!entry.val)
 204                 goto out;
 205         type = swp_type(entry);
 206         if (type >= nr_swapfiles)
 207                 goto bad_nofile;
 208         p = & swap_info[type];
 209         if (!(p->flags & SWP_USED))
 210                 goto bad_device;
 211         offset = swp_offset(entry);
 212         if (offset >= p->max)
 213                 goto bad_offset;
 214         if (!p->swap_map[offset])
 215                 goto bad_free;
 216         swap_list_lock();
 217         if (p->prio > swap_info[swap_list.next].prio)
 218                 swap_list.next = type;
 219         swap_device_lock(p);
 220         return p;
 221
 222 bad_free:
 223         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 224         goto out;
 225 bad_offset:
 226         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 227         goto out;
 228 bad_device:
 229         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 230         goto out;
 231 bad_nofile:
 232         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 233 out:
 234         return NULL;
 235 }
 236
 237 static void swap_info_put(struct swap_info_struct * p)
 238 {
 239         swap_device_unlock(p);
 240         swap_list_unlock();
 241 }
 242
 243 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 244 {
 245         int count = p->swap_map[offset];
 246
 247         if (count < SWAP_MAP_MAX) {
 248                 count--;
 249                 p->swap_map[offset] = count;
 250                 if (!count) {
 251                         if (offset < p->lowest_bit)
 252                                 p->lowest_bit = offset;
 253                         if (offset > p->highest_bit)
 254                                 p->highest_bit = offset;
 255                         nr_swap_pages++;
 256                         p->inuse_pages--;
 257                 }
 258         }
 259         return count;
 260 }
 261
 262 /*
 263  * Caller has made sure that the swapdevice corresponding to entry
 264  * is still around or has not been recycled.
 265  */
 266 void swap_free(swp_entry_t entry)
 267 {
 268         struct swap_info_struct * p;
 269
 270         p = swap_info_get(entry);
 271         if (p) {
 272                 swap_entry_free(p, swp_offset(entry));
 273                 swap_info_put(p);
 274         }
 275 }
 276
 277 /*
 278  * Check if we're the only user of a swap page,
 279  * when the page is locked.
 280  */
 281 static int exclusive_swap_page(struct page *page)
 282 {
 283         int retval = 0;
 284         struct swap_info_struct * p;
 285         swp_entry_t entry;
 286
 287         entry.val = page->private;
 288         p = swap_info_get(entry);
 289         if (p) {
 290                 /* Is the only swap cache user the cache itself? */
 291                 if (p->swap_map[swp_offset(entry)] == 1) {
 292                         /* Recheck the page count with the swapcache lock held.. */
 293                         spin_lock_irq(&swapper_space.tree_lock);
 294                         if (page_count(page) == 2)
 295                                 retval = 1;
 296                         spin_unlock_irq(&swapper_space.tree_lock);
 297                 }
 298                 swap_info_put(p);
 299         }
 300         return retval;
 301 }
 302
 303 /*
 304  * We can use this swap cache entry directly
 305  * if there are no other references to it.
 306  *
 307  * Here "exclusive_swap_page()" does the real
 308  * work, but we opportunistically check whether
 309  * we need to get all the locks first..
 310  */
 311 int can_share_swap_page(struct page *page)
 312 {
 313         int retval = 0;
 314
 315         if (!PageLocked(page))
 316                 BUG();
 317         switch (page_count(page)) {
 318         case 3:
 319                 if (!PagePrivate(page))
 320                         break;
 321                 /* Fallthrough */
 322         case 2:
 323                 if (!PageSwapCache(page))
 324                         break;
 325                 retval = exclusive_swap_page(page);
 326                 break;
 327         case 1:
 328                 if (PageReserved(page))
 329                         break;
 330                 retval = 1;
 331         }
 332         return retval;
 333 }
 334
 335 /*
 336  * Work out if there are any other processes sharing this
 337  * swap cache page. Free it if you can. Return success.
 338  */
 339 int remove_exclusive_swap_page(struct page *page)
 340 {
 341         int retval;
 342         struct swap_info_struct * p;
 343         swp_entry_t entry;
 344
 345         BUG_ON(PagePrivate(page));
 346         BUG_ON(!PageLocked(page));
 347
 348         if (!PageSwapCache(page))
 349                 return 0;
 350         if (PageWriteback(page))
 351                 return 0;
 352         if (page_count(page) != 2) /* 2: us + cache */
 353                 return 0;
 354
 355         entry.val = page->private;
 356         p = swap_info_get(entry);
 357         if (!p)
 358                 return 0;
 359
 360         /* Is the only swap cache user the cache itself? */
 361         retval = 0;
 362         if (p->swap_map[swp_offset(entry)] == 1) {
 363                 /* Recheck the page count with the swapcache lock held.. */
 364                 spin_lock_irq(&swapper_space.tree_lock);
 365                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 366                         __delete_from_swap_cache(page);
 367                         SetPageDirty(page);
 368                         retval = 1;
 369                 }
 370                 spin_unlock_irq(&swapper_space.tree_lock);
 371         }
 372         swap_info_put(p);
 373
 374         if (retval) {
 375                 swap_free(entry);
 376                 page_cache_release(page);
 377         }
 378
 379         return retval;
 380 }
 381
 382 /*
 383  * Free the swap entry like above, but also try to
 384  * free the page cache entry if it is the last user.
 385  */
 386 void free_swap_and_cache(swp_entry_t entry)
 387 {
 388         struct swap_info_struct * p;
 389         struct page *page = NULL;
 390
 391         p = swap_info_get(entry);
 392         if (p) {
 393                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 394                         spin_lock_irq(&swapper_space.tree_lock);
 395                         page = radix_tree_lookup(&swapper_space.page_tree,
 396                                 entry.val);
 397                         if (page && TestSetPageLocked(page))
 398                                 page = NULL;
 399                         spin_unlock_irq(&swapper_space.tree_lock);
 400                 }
 401                 swap_info_put(p);
 402         }
 403         if (page) {
 404                 int one_user;
 405
 406                 BUG_ON(PagePrivate(page));
 407                 page_cache_get(page);
 408                 one_user = (page_count(page) == 2);
 409                 /* Only cache user (+us), or swap space full? Free it! */
 410                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
 411                         delete_from_swap_cache(page);
 412                         SetPageDirty(page);
 413                 }
 414                 unlock_page(page);
 415                 page_cache_release(page);
 416         }
 417 }
 418
 419 /*
 420  * The swap entry has been read in advance, and we return 1 to indicate
 421  * that the page has been used or is no longer needed.
 422  *
 423  * Always set the resulting pte to be nowrite (the same as COW pages
 424  * after one process has exited).  We don't know just how many PTEs will
 425  * share this swap entry, so be cautious and let do_wp_page work out
 426  * what to do if a write is requested later.
 427  */
 428 /* vma->vm_mm->page_table_lock is held */
 429 static void
 430 unuse_pte(struct vm_area_struct *vma, unsigned long address, pte_t *dir,
 431         swp_entry_t entry, struct page *page)
 432 {
 433         vma->vm_mm->rss++;
 434         get_page(page);
 435         set_pte(dir, pte_mkold(mk_pte(page, vma->vm_page_prot)));
 436         page_add_anon_rmap(page, vma, address);
 437         swap_free(entry);
 438 }
 439
 440 /* vma->vm_mm->page_table_lock is held */
 441 static unsigned long unuse_pmd(struct vm_area_struct * vma, pmd_t *dir,
 442         unsigned long address, unsigned long size, unsigned long offset,
 443         swp_entry_t entry, struct page *page)
 444 {
 445         pte_t * pte;
 446         unsigned long end;
 447         pte_t swp_pte = swp_entry_to_pte(entry);
 448
 449         if (pmd_none(*dir))
 450                 return 0;
 451         if (pmd_bad(*dir)) {
 452                 pmd_ERROR(*dir);
 453                 pmd_clear(dir);
 454                 return 0;
 455         }
 456         pte = pte_offset_map(dir, address);
 457         offset += address & PMD_MASK;
 458         address &= ~PMD_MASK;
 459         end = address + size;
 460         if (end > PMD_SIZE)
 461                 end = PMD_SIZE;
 462         do {
 463                 /*
 464                  * swapoff spends a _lot_ of time in this loop!
 465                  * Test inline before going to call unuse_pte.
 466                  */
 467                 if (unlikely(pte_same(*pte, swp_pte))) {
 468                         unuse_pte(vma, offset + address, pte, entry, page);
 469                         pte_unmap(pte);
 470
 471                         /*
 472                          * Move the page to the active list so it is not
 473                          * immediately swapped out again after swapon.
 474                          */
 475                         activate_page(page);
 476
 477                         /* add 1 since address may be 0 */
 478                         return 1 + offset + address;
 479                 }
 480                 address += PAGE_SIZE;
 481                 pte++;
 482         } while (address && (address < end));
 483         pte_unmap(pte - 1);
 484         return 0;
 485 }
 486
 487 /* vma->vm_mm->page_table_lock is held */
 488 static unsigned long unuse_pgd(struct vm_area_struct * vma, pgd_t *dir,
 489         unsigned long address, unsigned long size,
 490         swp_entry_t entry, struct page *page)
 491 {
 492         pmd_t * pmd;
 493         unsigned long offset, end;
 494         unsigned long foundaddr;
 495
 496         if (pgd_none(*dir))
 497                 return 0;
 498         if (pgd_bad(*dir)) {
 499                 pgd_ERROR(*dir);
 500                 pgd_clear(dir);
 501                 return 0;
 502         }
 503         pmd = pmd_offset(dir, address);
 504         offset = address & PGDIR_MASK;
 505         address &= ~PGDIR_MASK;
 506         end = address + size;
 507         if (end > PGDIR_SIZE)
 508                 end = PGDIR_SIZE;
 509         if (address >= end)
 510                 BUG();
 511         do {
 512                 foundaddr = unuse_pmd(vma, pmd, address, end - address,
 513                                                 offset, entry, page);
 514                 if (foundaddr)
 515                         return foundaddr;
 516                 address = (address + PMD_SIZE) & PMD_MASK;
 517                 pmd++;
 518         } while (address && (address < end));
 519         return 0;
 520 }
 521
 522 /* vma->vm_mm->page_table_lock is held */
 523 static unsigned long unuse_vma(struct vm_area_struct * vma, pgd_t *pgdir,
 524         swp_entry_t entry, struct page *page)
 525 {
 526         unsigned long start = vma->vm_start, end = vma->vm_end;
 527         unsigned long foundaddr;
 528
 529         if (start >= end)
 530                 BUG();
 531         do {
 532                 foundaddr = unuse_pgd(vma, pgdir, start, end - start,
 533                                                 entry, page);
 534                 if (foundaddr)
 535                         return foundaddr;
 536                 start = (start + PGDIR_SIZE) & PGDIR_MASK;
 537                 pgdir++;
 538         } while (start && (start < end));
 539         return 0;
 540 }
 541
 542 static int unuse_process(struct mm_struct * mm,
 543                         swp_entry_t entry, struct page* page)
 544 {
 545         struct vm_area_struct* vma;
 546         unsigned long foundaddr = 0;
 547
 548         /*
 549          * Go through process' page directory.
 550          */
 551         down_read(&mm->mmap_sem);
 552         spin_lock(&mm->page_table_lock);
 553         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 554                 if (!is_vm_hugetlb_page(vma)) {
 555                         pgd_t * pgd = pgd_offset(mm, vma->vm_start);
 556                         foundaddr = unuse_vma(vma, pgd, entry, page);
 557                         if (foundaddr)
 558                                 break;
 559                 }
 560         }
 561         spin_unlock(&mm->page_table_lock);
 562         up_read(&mm->mmap_sem);
 563         /*
 564          * Currently unuse_process cannot fail, but leave error handling
 565          * at call sites for now, since we change it from time to time.
 566          */
 567         return 0;
 568 }
 569
 570 /*
 571  * Scan swap_map from current position to next entry still in use.
 572  * Recycle to start on reaching the end, returning 0 when empty.
 573  */
 574 static int find_next_to_unuse(struct swap_info_struct *si, int prev)
 575 {
 576         int max = si->max;
 577         int i = prev;
 578         int count;
 579
 580         /*
 581          * No need for swap_device_lock(si) here: we're just looking
 582          * for whether an entry is in use, not modifying it; false
 583          * hits are okay, and sys_swapoff() has already prevented new
 584          * allocations from this area (while holding swap_list_lock()).
 585          */
 586         for (;;) {
 587                 if (++i >= max) {
 588                         if (!prev) {
 589                                 i = 0;
 590                                 break;
 591                         }
 592                         /*
 593                          * No entries in use at top of swap_map,
 594                          * loop back to start and recheck there.
 595                          */
 596                         max = prev + 1;
 597                         prev = 0;
 598                         i = 1;
 599                 }
 600                 count = si->swap_map[i];
 601                 if (count && count != SWAP_MAP_BAD)
 602                         break;
 603         }
 604         return i;
 605 }
 606
 607 /*
 608  * We completely avoid races by reading each swap page in advance,
 609  * and then search for the process using it.  All the necessary
 610  * page table adjustments can then be made atomically.
 611  */
 612 static int try_to_unuse(unsigned int type)
 613 {
 614         struct swap_info_struct * si = &swap_info[type];
 615         struct mm_struct *start_mm;
 616         unsigned short *swap_map;
 617         unsigned short swcount;
 618         struct page *page;
 619         swp_entry_t entry;
 620         int i = 0;
 621         int retval = 0;
 622         int reset_overflow = 0;
 623         int shmem;
 624
 625         /*
 626          * When searching mms for an entry, a good strategy is to
 627          * start at the first mm we freed the previous entry from
 628          * (though actually we don't notice whether we or coincidence
 629          * freed the entry).  Initialize this start_mm with a hold.
 630          *
 631          * A simpler strategy would be to start at the last mm we
 632          * freed the previous entry from; but that would take less
 633          * advantage of mmlist ordering (now preserved by swap_out()),
 634          * which clusters forked address spaces together, most recent
 635          * child immediately after parent.  If we race with dup_mmap(),
 636          * we very much want to resolve parent before child, otherwise
 637          * we may miss some entries: using last mm would invert that.
 638          */
 639         start_mm = &init_mm;
 640         atomic_inc(&init_mm.mm_users);
 641
 642         /*
 643          * Keep on scanning until all entries have gone.  Usually,
 644          * one pass through swap_map is enough, but not necessarily:
 645          * mmput() removes mm from mmlist before exit_mmap() and its
 646          * zap_page_range().  That's not too bad, those entries are
 647          * on their way out, and handled faster there than here.
 648          * do_munmap() behaves similarly, taking the range out of mm's
 649          * vma list before zap_page_range().  But unfortunately, when
 650          * unmapping a part of a vma, it takes the whole out first,
 651          * then reinserts what's left after (might even reschedule if
 652          * open() method called) - so swap entries may be invisible
 653          * to swapoff for a while, then reappear - but that is rare.
 654          */
 655         while ((i = find_next_to_unuse(si, i)) != 0) {
 656                 if (signal_pending(current)) {
 657                         retval = -EINTR;
 658                         break;
 659                 }
 660
 661                 /*
 662                  * Get a page for the entry, using the existing swap
 663                  * cache page if there is one.  Otherwise, get a clean
 664                  * page and read the swap into it.
 665                  */
 666                 swap_map = &si->swap_map[i];
 667                 entry = swp_entry(type, i);
 668                 page = read_swap_cache_async(entry, NULL, 0);
 669                 if (!page) {
 670                         /*
 671                          * Either swap_duplicate() failed because entry
 672                          * has been freed independently, and will not be
 673                          * reused since sys_swapoff() already disabled
 674                          * allocation from here, or alloc_page() failed.
 675                          */
 676                         if (!*swap_map)
 677                                 continue;
 678                         retval = -ENOMEM;
 679                         break;
 680                 }
 681
 682                 /*
 683                  * Don't hold on to start_mm if it looks like exiting.
 684                  */
 685                 if (atomic_read(&start_mm->mm_users) == 1) {
 686                         mmput(start_mm);
 687                         start_mm = &init_mm;
 688                         atomic_inc(&init_mm.mm_users);
 689                 }
 690
 691                 /*
 692                  * Wait for and lock page.  When do_swap_page races with
 693                  * try_to_unuse, do_swap_page can handle the fault much
 694                  * faster than try_to_unuse can locate the entry.  This
 695                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 696                  * defer to do_swap_page in such a case - in some tests,
 697                  * do_swap_page and try_to_unuse repeatedly compete.
 698                  */
 699                 wait_on_page_locked(page);
 700                 wait_on_page_writeback(page);
 701                 lock_page(page);
 702                 wait_on_page_writeback(page);
 703
 704                 /*
 705                  * Remove all references to entry, without blocking.
 706                  * Whenever we reach init_mm, there's no address space
 707                  * to search, but use it as a reminder to search shmem.
 708                  */
 709                 shmem = 0;
 710                 swcount = *swap_map;
 711                 if (swcount > 1) {
 712                         if (start_mm == &init_mm)
 713                                 shmem = shmem_unuse(entry, page);
 714                         else
 715                                 retval = unuse_process(start_mm, entry, page);
 716                 }
 717                 if (*swap_map > 1) {
 718                         int set_start_mm = (*swap_map >= swcount);
 719                         struct list_head *p = &start_mm->mmlist;
 720                         struct mm_struct *new_start_mm = start_mm;
 721                         struct mm_struct *prev_mm = start_mm;
 722                         struct mm_struct *mm;
 723
 724                         atomic_inc(&new_start_mm->mm_users);
 725                         atomic_inc(&prev_mm->mm_users);
 726                         spin_lock(&mmlist_lock);
 727                         while (*swap_map > 1 && !retval &&
 728                                         (p = p->next) != &start_mm->mmlist) {
 729                                 mm = list_entry(p, struct mm_struct, mmlist);
 730                                 atomic_inc(&mm->mm_users);
 731                                 spin_unlock(&mmlist_lock);
 732                                 mmput(prev_mm);
 733                                 prev_mm = mm;
 734
 735                                 cond_resched();
 736
 737                                 swcount = *swap_map;
 738                                 if (swcount <= 1)
 739                                         ;
 740                                 else if (mm == &init_mm) {
 741                                         set_start_mm = 1;
 742                                         shmem = shmem_unuse(entry, page);
 743                                 } else
 744                                         retval = unuse_process(mm, entry, page);
 745                                 if (set_start_mm && *swap_map < swcount) {
 746                                         mmput(new_start_mm);
 747                                         atomic_inc(&mm->mm_users);
 748                                         new_start_mm = mm;
 749                                         set_start_mm = 0;
 750                                 }
 751                                 spin_lock(&mmlist_lock);
 752                         }
 753                         spin_unlock(&mmlist_lock);
 754                         mmput(prev_mm);
 755                         mmput(start_mm);
 756                         start_mm = new_start_mm;
 757                 }
 758                 if (retval) {
 759                         unlock_page(page);
 760                         page_cache_release(page);
 761                         break;
 762                 }
 763
 764                 /*
 765                  * How could swap count reach 0x7fff when the maximum
 766                  * pid is 0x7fff, and there's no way to repeat a swap
 767                  * page within an mm (except in shmem, where it's the
 768                  * shared object which takes the reference count)?
 769                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 770                  *
 771                  * If that's wrong, then we should worry more about
 772                  * exit_mmap() and do_munmap() cases described above:
 773                  * we might be resetting SWAP_MAP_MAX too early here.
 774                  * We know "Undead"s can happen, they're okay, so don't
 775                  * report them; but do report if we reset SWAP_MAP_MAX.
 776                  */
 777                 if (*swap_map == SWAP_MAP_MAX) {
 778                         swap_device_lock(si);
 779                         *swap_map = 1;
 780                         swap_device_unlock(si);
 781                         reset_overflow = 1;
 782                 }
 783
 784                 /*
 785                  * If a reference remains (rare), we would like to leave
 786                  * the page in the swap cache; but try_to_unmap could
 787                  * then re-duplicate the entry once we drop page lock,
 788                  * so we might loop indefinitely; also, that page could
 789                  * not be swapped out to other storage meanwhile.  So:
 790                  * delete from cache even if there's another reference,
 791                  * after ensuring that the data has been saved to disk -
 792                  * since if the reference remains (rarer), it will be
 793                  * read from disk into another page.  Splitting into two
 794                  * pages would be incorrect if swap supported "shared
 795                  * private" pages, but they are handled by tmpfs files.
 796                  *
 797                  * Note shmem_unuse already deleted a swappage from
 798                  * the swap cache, unless the move to filepage failed:
 799                  * in which case it left swappage in cache, lowered its
 800                  * swap count to pass quickly through the loops above,
 801                  * and now we must reincrement count to try again later.
 802                  */
 803                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 804                         struct writeback_control wbc = {
 805                                 .sync_mode = WB_SYNC_NONE,
 806                         };
 807
 808                         swap_writepage(page, &wbc);
 809                         lock_page(page);
 810                         wait_on_page_writeback(page);
 811                 }
 812                 if (PageSwapCache(page)) {
 813                         if (shmem)
 814                                 swap_duplicate(entry);
 815                         else
 816                                 delete_from_swap_cache(page);
 817                 }
 818
 819                 /*
 820                  * So we could skip searching mms once swap count went
 821                  * to 1, we did not mark any present ptes as dirty: must
 822                  * mark page dirty so shrink_list will preserve it.
 823                  */
 824                 SetPageDirty(page);
 825                 unlock_page(page);
 826                 page_cache_release(page);
 827
 828                 /*
 829                  * Make sure that we aren't completely killing
 830                  * interactive performance.
 831                  */
 832                 cond_resched();
 833         }
 834
 835         mmput(start_mm);
 836         if (reset_overflow) {
 837                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 838                 swap_overflow = 0;
 839         }
 840         return retval;
 841 }
 842
 843 /*
 844  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 845  * corresponds to page offset `offset'.
 846  */
 847 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 848 {
 849         struct swap_extent *se = sis->curr_swap_extent;
 850         struct swap_extent *start_se = se;
 851
 852         for ( ; ; ) {
 853                 struct list_head *lh;
 854
 855                 if (se->start_page <= offset &&
 856                                 offset < (se->start_page + se->nr_pages)) {
 857                         return se->start_block + (offset - se->start_page);
 858                 }
 859                 lh = se->list.prev;
 860                 if (lh == &sis->extent_list)
 861                         lh = lh->prev;
 862                 se = list_entry(lh, struct swap_extent, list);
 863                 sis->curr_swap_extent = se;
 864                 BUG_ON(se == start_se);         /* It *must* be present */
 865         }
 866 }
 867
 868 /*
 869  * Free all of a swapdev's extent information
 870  */
 871 static void destroy_swap_extents(struct swap_info_struct *sis)
 872 {
 873         while (!list_empty(&sis->extent_list)) {
 874                 struct swap_extent *se;
 875
 876                 se = list_entry(sis->extent_list.next,
 877                                 struct swap_extent, list);
 878                 list_del(&se->list);
 879                 kfree(se);
 880         }
 881         sis->nr_extents = 0;
 882 }
 883
 884 /*
 885  * Add a block range (and the corresponding page range) into this swapdev's
 886  * extent list.  The extent list is kept sorted in block order.
 887  *
 888  * This function rather assumes that it is called in ascending sector_t order.
 889  * It doesn't look for extent coalescing opportunities.
 890  */
 891 static int
 892 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 893                 unsigned long nr_pages, sector_t start_block)
 894 {
 895         struct swap_extent *se;
 896         struct swap_extent *new_se;
 897         struct list_head *lh;
 898
 899         lh = sis->extent_list.next;     /* The highest-addressed block */
 900         while (lh != &sis->extent_list) {
 901                 se = list_entry(lh, struct swap_extent, list);
 902                 if (se->start_block + se->nr_pages == start_block &&
 903                     se->start_page  + se->nr_pages == start_page) {
 904                         /* Merge it */
 905                         se->nr_pages += nr_pages;
 906                         return 0;
 907                 }
 908                 lh = lh->next;
 909         }
 910
 911         /*
 912          * No merge.  Insert a new extent, preserving ordering.
 913          */
 914         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 915         if (new_se == NULL)
 916                 return -ENOMEM;
 917         new_se->start_page = start_page;
 918         new_se->nr_pages = nr_pages;
 919         new_se->start_block = start_block;
 920
 921         lh = sis->extent_list.prev;     /* The lowest block */
 922         while (lh != &sis->extent_list) {
 923                 se = list_entry(lh, struct swap_extent, list);
 924                 if (se->start_block > start_block)
 925                         break;
 926                 lh = lh->prev;
 927         }
 928         list_add_tail(&new_se->list, lh);
 929         sis->nr_extents++;
 930         return 0;
 931 }
 932
 933 /*
 934  * A `swap extent' is a simple thing which maps a contiguous range of pages
 935  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 936  * is built at swapon time and is then used at swap_writepage/swap_readpage
 937  * time for locating where on disk a page belongs.
 938  *
 939  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 940  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 941  * swap files identically.
 942  *
 943  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
 944  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
 945  * swapfiles are handled *identically* after swapon time.
 946  *
 947  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
 948  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
 949  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
 950  * requirements, they are simply tossed out - we will never use those blocks
 951  * for swapping.
 952  *
 953  * For S_ISREG swapfiles we hold i_sem across the life of the swapon.  This
 954  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
 955  * which will scribble on the fs.
 956  *
 957  * The amount of disk space which a single swap extent represents varies.
 958  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
 959  * extents in the list.  To avoid much list walking, we cache the previous
 960  * search location in `curr_swap_extent', and start new searches from there.
 961  * This is extremely effective.  The average number of iterations in
 962  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
 963  */
 964 static int setup_swap_extents(struct swap_info_struct *sis)
 965 {
 966         struct inode *inode;
 967         unsigned blocks_per_page;
 968         unsigned long page_no;
 969         unsigned blkbits;
 970         sector_t probe_block;
 971         sector_t last_block;
 972         int ret;
 973
 974         inode = sis->swap_file->f_mapping->host;
 975         if (S_ISBLK(inode->i_mode)) {
 976                 ret = add_swap_extent(sis, 0, sis->max, 0);
 977                 goto done;
 978         }
 979
 980         blkbits = inode->i_blkbits;
 981         blocks_per_page = PAGE_SIZE >> blkbits;
 982
 983         /*
 984          * Map all the blocks into the extent list.  This code doesn't try
 985          * to be very smart.
 986          */
 987         probe_block = 0;
 988         page_no = 0;
 989         last_block = i_size_read(inode) >> blkbits;
 990         while ((probe_block + blocks_per_page) <= last_block &&
 991                         page_no < sis->max) {
 992                 unsigned block_in_page;
 993                 sector_t first_block;
 994
 995                 first_block = bmap(inode, probe_block);
 996                 if (first_block == 0)
 997                         goto bad_bmap;
 998
 999                 /*
1000                  * It must be PAGE_SIZE aligned on-disk
1001                  */
1002                 if (first_block & (blocks_per_page - 1)) {
1003                         probe_block++;
1004                         goto reprobe;
1005                 }
1006
1007                 for (block_in_page = 1; block_in_page < blocks_per_page;
1008                                         block_in_page++) {
1009                         sector_t block;
1010
1011                         block = bmap(inode, probe_block + block_in_page);
1012                         if (block == 0)
1013                                 goto bad_bmap;
1014                         if (block != first_block + block_in_page) {
1015                                 /* Discontiguity */
1016                                 probe_block++;
1017                                 goto reprobe;
1018                         }
1019                 }
1020
1021                 /*
1022                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1023                  */
1024                 ret = add_swap_extent(sis, page_no, 1,
1025                                 first_block >> (PAGE_SHIFT - blkbits));
1026                 if (ret)
1027                         goto out;
1028                 page_no++;
1029                 probe_block += blocks_per_page;
1030 reprobe:
1031                 continue;
1032         }
1033         ret = 0;
1034         if (page_no == 0)
1035                 ret = -EINVAL;
1036         sis->max = page_no;
1037         sis->highest_bit = page_no - 1;
1038 done:
1039         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1040                                         struct swap_extent, list);
1041         goto out;
1042 bad_bmap:
1043         printk(KERN_ERR "swapon: swapfile has holes\n");
1044         ret = -EINVAL;
1045 out:
1046         return ret;
1047 }
1048
1049 #if 0   /* We don't need this yet */
1050 #include <linux/backing-dev.h>
1051 int page_queue_congested(struct page *page)
1052 {
1053         struct backing_dev_info *bdi;
1054
1055         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1056
1057         if (PageSwapCache(page)) {
1058                 swp_entry_t entry = { .val = page->private };
1059                 struct swap_info_struct *sis;
1060
1061                 sis = get_swap_info_struct(swp_type(entry));
1062                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1063         } else
1064                 bdi = page->mapping->backing_dev_info;
1065         return bdi_write_congested(bdi);
1066 }
1067 #endif
1068
1069 asmlinkage long sys_swapoff(const char __user * specialfile)
1070 {
1071         struct swap_info_struct * p = NULL;
1072         unsigned short *swap_map;
1073         struct file *swap_file, *victim;
1074         struct address_space *mapping;
1075         struct inode *inode;
1076         char * pathname;
1077         int i, type, prev;
1078         int err;
1079
1080         if (!capable(CAP_SYS_ADMIN))
1081                 return -EPERM;
1082
1083         pathname = getname(specialfile);
1084         err = PTR_ERR(pathname);
1085         if (IS_ERR(pathname))
1086                 goto out;
1087
1088         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1089         putname(pathname);
1090         err = PTR_ERR(victim);
1091         if (IS_ERR(victim))
1092                 goto out;
1093
1094         mapping = victim->f_mapping;
1095         prev = -1;
1096         swap_list_lock();
1097         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1098                 p = swap_info + type;
1099                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1100                         if (p->swap_file->f_mapping == mapping)
1101                                 break;
1102                 }
1103                 prev = type;
1104         }
1105         if (type < 0) {
1106                 err = -EINVAL;
1107                 swap_list_unlock();
1108                 goto out_dput;
1109         }
1110         if (!security_vm_enough_memory(p->pages))
1111                 vm_unacct_memory(p->pages);
1112         else {
1113                 err = -ENOMEM;
1114                 swap_list_unlock();
1115                 goto out_dput;
1116         }
1117         if (prev < 0) {
1118                 swap_list.head = p->next;
1119         } else {
1120                 swap_info[prev].next = p->next;
1121         }
1122         if (type == swap_list.next) {
1123                 /* just pick something that's safe... */
1124                 swap_list.next = swap_list.head;
1125         }
1126         nr_swap_pages -= p->pages;
1127         total_swap_pages -= p->pages;
1128         p->flags &= ~SWP_WRITEOK;
1129         swap_list_unlock();
1130         current->flags |= PF_SWAPOFF;
1131         err = try_to_unuse(type);
1132         current->flags &= ~PF_SWAPOFF;
1133
1134         /* wait for any unplug function to finish */
1135         down_write(&swap_unplug_sem);
1136         up_write(&swap_unplug_sem);
1137
1138         if (err) {
1139                 /* re-insert swap space back into swap_list */
1140                 swap_list_lock();
1141                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1142                         if (p->prio >= swap_info[i].prio)
1143                                 break;
1144                 p->next = i;
1145                 if (prev < 0)
1146                         swap_list.head = swap_list.next = p - swap_info;
1147                 else
1148                         swap_info[prev].next = p - swap_info;
1149                 nr_swap_pages += p->pages;
1150                 total_swap_pages += p->pages;
1151                 p->flags |= SWP_WRITEOK;
1152                 swap_list_unlock();
1153                 goto out_dput;
1154         }
1155         down(&swapon_sem);
1156         swap_list_lock();
1157         swap_device_lock(p);
1158         swap_file = p->swap_file;
1159         p->swap_file = NULL;
1160         p->max = 0;
1161         swap_map = p->swap_map;
1162         p->swap_map = NULL;
1163         p->flags = 0;
1164         destroy_swap_extents(p);
1165         swap_device_unlock(p);
1166         swap_list_unlock();
1167         up(&swapon_sem);
1168         vfree(swap_map);
1169         inode = mapping->host;
1170         if (S_ISBLK(inode->i_mode)) {
1171                 struct block_device *bdev = I_BDEV(inode);
1172                 set_blocksize(bdev, p->old_block_size);
1173                 bd_release(bdev);
1174         } else {
1175                 down(&inode->i_sem);
1176                 inode->i_flags &= ~S_SWAPFILE;
1177                 up(&inode->i_sem);
1178         }
1179         filp_close(swap_file, NULL);
1180         err = 0;
1181
1182 out_dput:
1183         filp_close(victim, NULL);
1184 out:
1185         return err;
1186 }
1187
1188 #ifdef CONFIG_PROC_FS
1189 /* iterator */
1190 static void *swap_start(struct seq_file *swap, loff_t *pos)
1191 {
1192         struct swap_info_struct *ptr = swap_info;
1193         int i;
1194         loff_t l = *pos;
1195
1196         down(&swapon_sem);
1197
1198         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1199                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1200                         continue;
1201                 if (!l--)
1202                         return ptr;
1203         }
1204
1205         return NULL;
1206 }
1207
1208 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1209 {
1210         struct swap_info_struct *ptr = v;
1211         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1212
1213         for (++ptr; ptr < endptr; ptr++) {
1214                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1215                         continue;
1216                 ++*pos;
1217                 return ptr;
1218         }
1219
1220         return NULL;
1221 }
1222
1223 static void swap_stop(struct seq_file *swap, void *v)
1224 {
1225         up(&swapon_sem);
1226 }
1227
1228 static int swap_show(struct seq_file *swap, void *v)
1229 {
1230         struct swap_info_struct *ptr = v;
1231         struct file *file;
1232         int len;
1233
1234         if (v == swap_info)
1235                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1236
1237         file = ptr->swap_file;
1238         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1239         seq_printf(swap, "%*s%s\t%d\t%ld\t%d\n",
1240                        len < 40 ? 40 - len : 1, " ",
1241                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1242                                 "partition" : "file\t",
1243                        ptr->pages << (PAGE_SHIFT - 10),
1244                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1245                        ptr->prio);
1246         return 0;
1247 }
1248
1249 static struct seq_operations swaps_op = {
1250         .start =        swap_start,
1251         .next =         swap_next,
1252         .stop =         swap_stop,
1253         .show =         swap_show
1254 };
1255
1256 static int swaps_open(struct inode *inode, struct file *file)
1257 {
1258         return seq_open(file, &swaps_op);
1259 }
1260
1261 static struct file_operations proc_swaps_operations = {
1262         .open           = swaps_open,
1263         .read           = seq_read,
1264         .llseek         = seq_lseek,
1265         .release        = seq_release,
1266 };
1267
1268 static int __init procswaps_init(void)
1269 {
1270         struct proc_dir_entry *entry;
1271
1272         entry = create_proc_entry("swaps", 0, NULL);
1273         if (entry)
1274                 entry->proc_fops = &proc_swaps_operations;
1275         return 0;
1276 }
1277 __initcall(procswaps_init);
1278 #endif /* CONFIG_PROC_FS */
1279
1280 /*
1281  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1282  *
1283  * The swapon system call
1284  */
1285 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1286 {
1287         struct swap_info_struct * p;
1288         char *name = NULL;
1289         struct block_device *bdev = NULL;
1290         struct file *swap_file = NULL;
1291         struct address_space *mapping;
1292         unsigned int type;
1293         int i, prev;
1294         int error;
1295         static int least_priority;
1296         union swap_header *swap_header = NULL;
1297         int swap_header_version;
1298         int nr_good_pages = 0;
1299         unsigned long maxpages = 1;
1300         int swapfilesize;
1301         unsigned short *swap_map;
1302         struct page *page = NULL;
1303         struct inode *inode = NULL;
1304         int did_down = 0;
1305
1306         if (!capable(CAP_SYS_ADMIN))
1307                 return -EPERM;
1308         swap_list_lock();
1309         p = swap_info;
1310         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1311                 if (!(p->flags & SWP_USED))
1312                         break;
1313         error = -EPERM;
1314         /*
1315          * Test if adding another swap device is possible. There are
1316          * two limiting factors: 1) the number of bits for the swap
1317          * type swp_entry_t definition and 2) the number of bits for
1318          * the swap type in the swap ptes as defined by the different
1319          * architectures. To honor both limitations a swap entry
1320          * with swap offset 0 and swap type ~0UL is created, encoded
1321          * to a swap pte, decoded to a swp_entry_t again and finally
1322          * the swap type part is extracted. This will mask all bits
1323          * from the initial ~0UL that can't be encoded in either the
1324          * swp_entry_t or the architecture definition of a swap pte.
1325          */
1326         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1327                 swap_list_unlock();
1328                 goto out;
1329         }
1330         if (type >= nr_swapfiles)
1331                 nr_swapfiles = type+1;
1332         INIT_LIST_HEAD(&p->extent_list);
1333         p->flags = SWP_USED;
1334         p->nr_extents = 0;
1335         p->swap_file = NULL;
1336         p->old_block_size = 0;
1337         p->swap_map = NULL;
1338         p->lowest_bit = 0;
1339         p->highest_bit = 0;
1340         p->cluster_nr = 0;
1341         p->inuse_pages = 0;
1342         p->sdev_lock = SPIN_LOCK_UNLOCKED;
1343         p->next = -1;
1344         if (swap_flags & SWAP_FLAG_PREFER) {
1345                 p->prio =
1346                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1347         } else {
1348                 p->prio = --least_priority;
1349         }
1350         swap_list_unlock();
1351         name = getname(specialfile);
1352         error = PTR_ERR(name);
1353         if (IS_ERR(name)) {
1354                 name = NULL;
1355                 goto bad_swap_2;
1356         }
1357         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1358         error = PTR_ERR(swap_file);
1359         if (IS_ERR(swap_file)) {
1360                 swap_file = NULL;
1361                 goto bad_swap_2;
1362         }
1363
1364         p->swap_file = swap_file;
1365         mapping = swap_file->f_mapping;
1366         inode = mapping->host;
1367
1368         error = -EBUSY;
1369         for (i = 0; i < nr_swapfiles; i++) {
1370                 struct swap_info_struct *q = &swap_info[i];
1371
1372                 if (i == type || !q->swap_file)
1373                         continue;
1374                 if (mapping == q->swap_file->f_mapping)
1375                         goto bad_swap;
1376         }
1377
1378         error = -EINVAL;
1379         if (S_ISBLK(inode->i_mode)) {
1380                 bdev = I_BDEV(inode);
1381                 error = bd_claim(bdev, sys_swapon);
1382                 if (error < 0) {
1383                         bdev = NULL;
1384                         goto bad_swap;
1385                 }
1386                 p->old_block_size = block_size(bdev);
1387                 error = set_blocksize(bdev, PAGE_SIZE);
1388                 if (error < 0)
1389                         goto bad_swap;
1390                 p->bdev = bdev;
1391         } else if (S_ISREG(inode->i_mode)) {
1392                 p->bdev = inode->i_sb->s_bdev;
1393                 down(&inode->i_sem);
1394                 did_down = 1;
1395                 if (IS_SWAPFILE(inode)) {
1396                         error = -EBUSY;
1397                         goto bad_swap;
1398                 }
1399         } else {
1400                 goto bad_swap;
1401         }
1402
1403         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1404
1405         /*
1406          * Read the swap header.
1407          */
1408         if (!mapping->a_ops->readpage) {
1409                 error = -EINVAL;
1410                 goto bad_swap;
1411         }
1412         page = read_cache_page(mapping, 0,
1413                         (filler_t *)mapping->a_ops->readpage, swap_file);
1414         if (IS_ERR(page)) {
1415                 error = PTR_ERR(page);
1416                 goto bad_swap;
1417         }
1418         wait_on_page_locked(page);
1419         if (!PageUptodate(page))
1420                 goto bad_swap;
1421         kmap(page);
1422         swap_header = page_address(page);
1423
1424         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1425                 swap_header_version = 1;
1426         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1427                 swap_header_version = 2;
1428         else {
1429                 printk("Unable to find swap-space signature\n");
1430                 error = -EINVAL;
1431                 goto bad_swap;
1432         }
1433
1434         switch (swap_header_version) {
1435         case 1:
1436                 printk(KERN_ERR "version 0 swap is no longer supported. "
1437                         "Use mkswap -v1 %s\n", name);
1438                 error = -EINVAL;
1439                 goto bad_swap;
1440         case 2:
1441                 /* Check the swap header's sub-version and the size of
1442                    the swap file and bad block lists */
1443                 if (swap_header->info.version != 1) {
1444                         printk(KERN_WARNING
1445                                "Unable to handle swap header version %d\n",
1446                                swap_header->info.version);
1447                         error = -EINVAL;
1448                         goto bad_swap;
1449                 }
1450
1451                 p->lowest_bit  = 1;
1452                 /*
1453                  * Find out how many pages are allowed for a single swap
1454                  * device. There are two limiting factors: 1) the number of
1455                  * bits for the swap offset in the swp_entry_t type and
1456                  * 2) the number of bits in the a swap pte as defined by
1457                  * the different architectures. In order to find the
1458                  * largest possible bit mask a swap entry with swap type 0
1459                  * and swap offset ~0UL is created, encoded to a swap pte,
1460                  * decoded to a swp_entry_t again and finally the swap
1461                  * offset is extracted. This will mask all the bits from
1462                  * the initial ~0UL mask that can't be encoded in either
1463                  * the swp_entry_t or the architecture definition of a
1464                  * swap pte.
1465                  */
1466                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1467                 if (maxpages > swap_header->info.last_page)
1468                         maxpages = swap_header->info.last_page;
1469                 p->highest_bit = maxpages - 1;
1470
1471                 error = -EINVAL;
1472                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1473                         goto bad_swap;
1474
1475                 /* OK, set up the swap map and apply the bad block list */
1476                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1477                         error = -ENOMEM;
1478                         goto bad_swap;
1479                 }
1480
1481                 error = 0;
1482                 memset(p->swap_map, 0, maxpages * sizeof(short));
1483                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1484                         int page = swap_header->info.badpages[i];
1485                         if (page <= 0 || page >= swap_header->info.last_page)
1486                                 error = -EINVAL;
1487                         else
1488                                 p->swap_map[page] = SWAP_MAP_BAD;
1489                 }
1490                 nr_good_pages = swap_header->info.last_page -
1491                                 swap_header->info.nr_badpages -
1492                                 1 /* header page */;
1493                 if (error)
1494                         goto bad_swap;
1495         }
1496
1497         if (swapfilesize && maxpages > swapfilesize) {
1498                 printk(KERN_WARNING
1499                        "Swap area shorter than signature indicates\n");
1500                 error = -EINVAL;
1501                 goto bad_swap;
1502         }
1503         if (!nr_good_pages) {
1504                 printk(KERN_WARNING "Empty swap-file\n");
1505                 error = -EINVAL;
1506                 goto bad_swap;
1507         }
1508         p->swap_map[0] = SWAP_MAP_BAD;
1509         p->max = maxpages;
1510         p->pages = nr_good_pages;
1511
1512         error = setup_swap_extents(p);
1513         if (error)
1514                 goto bad_swap;
1515
1516         down(&swapon_sem);
1517         swap_list_lock();
1518         swap_device_lock(p);
1519         p->flags = SWP_ACTIVE;
1520         nr_swap_pages += nr_good_pages;
1521         total_swap_pages += nr_good_pages;
1522         printk(KERN_INFO "Adding %dk swap on %s.  Priority:%d extents:%d\n",
1523                 nr_good_pages<<(PAGE_SHIFT-10), name,
1524                 p->prio, p->nr_extents);
1525
1526         /* insert swap space into swap_list: */
1527         prev = -1;
1528         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1529                 if (p->prio >= swap_info[i].prio) {
1530                         break;
1531                 }
1532                 prev = i;
1533         }
1534         p->next = i;
1535         if (prev < 0) {
1536                 swap_list.head = swap_list.next = p - swap_info;
1537         } else {
1538                 swap_info[prev].next = p - swap_info;
1539         }
1540         swap_device_unlock(p);
1541         swap_list_unlock();
1542         up(&swapon_sem);
1543         error = 0;
1544         goto out;
1545 bad_swap:
1546         if (bdev) {
1547                 set_blocksize(bdev, p->old_block_size);
1548                 bd_release(bdev);
1549         }
1550 bad_swap_2:
1551         swap_list_lock();
1552         swap_map = p->swap_map;
1553         p->swap_file = NULL;
1554         p->swap_map = NULL;
1555         p->flags = 0;
1556         if (!(swap_flags & SWAP_FLAG_PREFER))
1557                 ++least_priority;
1558         swap_list_unlock();
1559         destroy_swap_extents(p);
1560         if (swap_map)
1561                 vfree(swap_map);
1562         if (swap_file)
1563                 filp_close(swap_file, NULL);
1564 out:
1565         if (page && !IS_ERR(page)) {
1566                 kunmap(page);
1567                 page_cache_release(page);
1568         }
1569         if (name)
1570                 putname(name);
1571         if (did_down) {
1572                 if (!error)
1573                         inode->i_flags |= S_SWAPFILE;
1574                 up(&inode->i_sem);
1575         }
1576         return error;
1577 }
1578
1579 void si_swapinfo(struct sysinfo *val)
1580 {
1581         unsigned int i;
1582         unsigned long nr_to_be_unused = 0;
1583
1584         swap_list_lock();
1585         for (i = 0; i < nr_swapfiles; i++) {
1586                 if (!(swap_info[i].flags & SWP_USED) ||
1587                      (swap_info[i].flags & SWP_WRITEOK))
1588                         continue;
1589                 nr_to_be_unused += swap_info[i].inuse_pages;
1590         }
1591         val->freeswap = nr_swap_pages + nr_to_be_unused;
1592         val->totalswap = total_swap_pages + nr_to_be_unused;
1593         swap_list_unlock();
1594 }
1595
1596 /*
1597  * Verify that a swap entry is valid and increment its swap map count.
1598  *
1599  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1600  * "permanent", but will be reclaimed by the next swapoff.
1601  */
1602 int swap_duplicate(swp_entry_t entry)
1603 {
1604         struct swap_info_struct * p;
1605         unsigned long offset, type;
1606         int result = 0;
1607
1608         type = swp_type(entry);
1609         if (type >= nr_swapfiles)
1610                 goto bad_file;
1611         p = type + swap_info;
1612         offset = swp_offset(entry);
1613
1614         swap_device_lock(p);
1615         if (offset < p->max && p->swap_map[offset]) {
1616                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1617                         p->swap_map[offset]++;
1618                         result = 1;
1619                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1620                         if (swap_overflow++ < 5)
1621                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1622                         p->swap_map[offset] = SWAP_MAP_MAX;
1623                         result = 1;
1624                 }
1625         }
1626         swap_device_unlock(p);
1627 out:
1628         return result;
1629
1630 bad_file:
1631         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1632         goto out;
1633 }
1634
1635 struct swap_info_struct *
1636 get_swap_info_struct(unsigned type)
1637 {
1638         return &swap_info[type];
1639 }
1640
1641 /*
1642  * swap_device_lock prevents swap_map being freed. Don't grab an extra
1643  * reference on the swaphandle, it doesn't matter if it becomes unused.
1644  */
1645 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1646 {
1647         int ret = 0, i = 1 << page_cluster;
1648         unsigned long toff;
1649         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1650
1651         if (!page_cluster)      /* no readahead */
1652                 return 0;
1653         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1654         if (!toff)              /* first page is swap header */
1655                 toff++, i--;
1656         *offset = toff;
1657
1658         swap_device_lock(swapdev);
1659         do {
1660                 /* Don't read-ahead past the end of the swap area */
1661                 if (toff >= swapdev->max)
1662                         break;
1663                 /* Don't read in free or bad pages */
1664                 if (!swapdev->swap_map[toff])
1665                         break;
1666                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1667                         break;
1668                 toff++;
1669                 ret++;
1670         } while (--i);
1671         swap_device_unlock(swapdev);
1672         return ret;
1673 }