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