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