err.no Git - linux-2.6/blob - mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/config.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/mman.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/namei.h>
  18 #include <linux/shm.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/syscalls.h>
  29
  30 #include <asm/pgtable.h>
  31 #include <asm/tlbflush.h>
  32 #include <linux/swapops.h>
  33
  34 DEFINE_SPINLOCK(swaplock);
  35 unsigned int nr_swapfiles;
  36 long 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                 blk_run_backing_dev(bdi, page);
  83         }
  84         up_read(&swap_unplug_sem);
  85 }
  86
  87 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
  88 {
  89         unsigned long offset, last_in_cluster;
  90
  91         /*
  92          * We try to cluster swap pages by allocating them sequentially
  93          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
  94          * way, however, we resort to first-free allocation, starting
  95          * a new cluster.  This prevents us from scattering swap pages
  96          * all over the entire swap partition, so that we reduce
  97          * overall disk seek times between swap pages.  -- sct
  98          * But we do now try to find an empty cluster.  -Andrea
  99          */
 100
 101         si->flags += SWP_SCANNING;
 102         if (unlikely(!si->cluster_nr)) {
 103                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 104                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
 105                         goto lowest;
 106                 swap_device_unlock(si);
 107
 108                 offset = si->lowest_bit;
 109                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 110
 111                 /* Locate the first empty (unaligned) cluster */
 112                 for (; last_in_cluster <= si->highest_bit; offset++) {
 113                         if (si->swap_map[offset])
 114                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 115                         else if (offset == last_in_cluster) {
 116                                 swap_device_lock(si);
 117                                 si->cluster_next = offset-SWAPFILE_CLUSTER-1;
 118                                 goto cluster;
 119                         }
 120                 }
 121                 swap_device_lock(si);
 122                 goto lowest;
 123         }
 124
 125         si->cluster_nr--;
 126 cluster:
 127         offset = si->cluster_next;
 128         if (offset > si->highest_bit)
 129 lowest:         offset = si->lowest_bit;
 130 checks: if (!(si->flags & SWP_WRITEOK))
 131                 goto no_page;
 132         if (!si->highest_bit)
 133                 goto no_page;
 134         if (!si->swap_map[offset]) {
 135                 if (offset == si->lowest_bit)
 136                         si->lowest_bit++;
 137                 if (offset == si->highest_bit)
 138                         si->highest_bit--;
 139                 si->inuse_pages++;
 140                 if (si->inuse_pages == si->pages) {
 141                         si->lowest_bit = si->max;
 142                         si->highest_bit = 0;
 143                 }
 144                 si->swap_map[offset] = 1;
 145                 si->cluster_next = offset + 1;
 146                 si->flags -= SWP_SCANNING;
 147                 return offset;
 148         }
 149
 150         swap_device_unlock(si);
 151         while (++offset <= si->highest_bit) {
 152                 if (!si->swap_map[offset]) {
 153                         swap_device_lock(si);
 154                         goto checks;
 155                 }
 156         }
 157         swap_device_lock(si);
 158         goto lowest;
 159
 160 no_page:
 161         si->flags -= SWP_SCANNING;
 162         return 0;
 163 }
 164
 165 swp_entry_t get_swap_page(void)
 166 {
 167         struct swap_info_struct *si;
 168         pgoff_t offset;
 169         int type, next;
 170         int wrapped = 0;
 171
 172         swap_list_lock();
 173         if (nr_swap_pages <= 0)
 174                 goto noswap;
 175         nr_swap_pages--;
 176
 177         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 178                 si = swap_info + type;
 179                 next = si->next;
 180                 if (next < 0 ||
 181                     (!wrapped && si->prio != swap_info[next].prio)) {
 182                         next = swap_list.head;
 183                         wrapped++;
 184                 }
 185
 186                 if (!si->highest_bit)
 187                         continue;
 188                 if (!(si->flags & SWP_WRITEOK))
 189                         continue;
 190
 191                 swap_list.next = next;
 192                 swap_device_lock(si);
 193                 swap_list_unlock();
 194                 offset = scan_swap_map(si);
 195                 swap_device_unlock(si);
 196                 if (offset)
 197                         return swp_entry(type, offset);
 198                 swap_list_lock();
 199                 next = swap_list.next;
 200         }
 201
 202         nr_swap_pages++;
 203 noswap:
 204         swap_list_unlock();
 205         return (swp_entry_t) {0};
 206 }
 207
 208 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 209 {
 210         struct swap_info_struct * p;
 211         unsigned long offset, type;
 212
 213         if (!entry.val)
 214                 goto out;
 215         type = swp_type(entry);
 216         if (type >= nr_swapfiles)
 217                 goto bad_nofile;
 218         p = & swap_info[type];
 219         if (!(p->flags & SWP_USED))
 220                 goto bad_device;
 221         offset = swp_offset(entry);
 222         if (offset >= p->max)
 223                 goto bad_offset;
 224         if (!p->swap_map[offset])
 225                 goto bad_free;
 226         swap_list_lock();
 227         swap_device_lock(p);
 228         return p;
 229
 230 bad_free:
 231         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 232         goto out;
 233 bad_offset:
 234         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 235         goto out;
 236 bad_device:
 237         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 238         goto out;
 239 bad_nofile:
 240         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 241 out:
 242         return NULL;
 243 }
 244
 245 static void swap_info_put(struct swap_info_struct * p)
 246 {
 247         swap_device_unlock(p);
 248         swap_list_unlock();
 249 }
 250
 251 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 252 {
 253         int count = p->swap_map[offset];
 254
 255         if (count < SWAP_MAP_MAX) {
 256                 count--;
 257                 p->swap_map[offset] = count;
 258                 if (!count) {
 259                         if (offset < p->lowest_bit)
 260                                 p->lowest_bit = offset;
 261                         if (offset > p->highest_bit)
 262                                 p->highest_bit = offset;
 263                         if (p->prio > swap_info[swap_list.next].prio)
 264                                 swap_list.next = p - swap_info;
 265                         nr_swap_pages++;
 266                         p->inuse_pages--;
 267                 }
 268         }
 269         return count;
 270 }
 271
 272 /*
 273  * Caller has made sure that the swapdevice corresponding to entry
 274  * is still around or has not been recycled.
 275  */
 276 void swap_free(swp_entry_t entry)
 277 {
 278         struct swap_info_struct * p;
 279
 280         p = swap_info_get(entry);
 281         if (p) {
 282                 swap_entry_free(p, swp_offset(entry));
 283                 swap_info_put(p);
 284         }
 285 }
 286
 287 /*
 288  * How many references to page are currently swapped out?
 289  */
 290 static inline int page_swapcount(struct page *page)
 291 {
 292         int count = 0;
 293         struct swap_info_struct *p;
 294         swp_entry_t entry;
 295
 296         entry.val = page->private;
 297         p = swap_info_get(entry);
 298         if (p) {
 299                 /* Subtract the 1 for the swap cache itself */
 300                 count = p->swap_map[swp_offset(entry)] - 1;
 301                 swap_info_put(p);
 302         }
 303         return count;
 304 }
 305
 306 /*
 307  * We can use this swap cache entry directly
 308  * if there are no other references to it.
 309  */
 310 int can_share_swap_page(struct page *page)
 311 {
 312         int count;
 313
 314         BUG_ON(!PageLocked(page));
 315         count = page_mapcount(page);
 316         if (count <= 1 && PageSwapCache(page))
 317                 count += page_swapcount(page);
 318         return count == 1;
 319 }
 320
 321 /*
 322  * Work out if there are any other processes sharing this
 323  * swap cache page. Free it if you can. Return success.
 324  */
 325 int remove_exclusive_swap_page(struct page *page)
 326 {
 327         int retval;
 328         struct swap_info_struct * p;
 329         swp_entry_t entry;
 330
 331         BUG_ON(PagePrivate(page));
 332         BUG_ON(!PageLocked(page));
 333
 334         if (!PageSwapCache(page))
 335                 return 0;
 336         if (PageWriteback(page))
 337                 return 0;
 338         if (page_count(page) != 2) /* 2: us + cache */
 339                 return 0;
 340
 341         entry.val = page->private;
 342         p = swap_info_get(entry);
 343         if (!p)
 344                 return 0;
 345
 346         /* Is the only swap cache user the cache itself? */
 347         retval = 0;
 348         if (p->swap_map[swp_offset(entry)] == 1) {
 349                 /* Recheck the page count with the swapcache lock held.. */
 350                 write_lock_irq(&swapper_space.tree_lock);
 351                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 352                         __delete_from_swap_cache(page);
 353                         SetPageDirty(page);
 354                         retval = 1;
 355                 }
 356                 write_unlock_irq(&swapper_space.tree_lock);
 357         }
 358         swap_info_put(p);
 359
 360         if (retval) {
 361                 swap_free(entry);
 362                 page_cache_release(page);
 363         }
 364
 365         return retval;
 366 }
 367
 368 /*
 369  * Free the swap entry like above, but also try to
 370  * free the page cache entry if it is the last user.
 371  */
 372 void free_swap_and_cache(swp_entry_t entry)
 373 {
 374         struct swap_info_struct * p;
 375         struct page *page = NULL;
 376
 377         p = swap_info_get(entry);
 378         if (p) {
 379                 if (swap_entry_free(p, swp_offset(entry)) == 1)
 380                         page = find_trylock_page(&swapper_space, entry.val);
 381                 swap_info_put(p);
 382         }
 383         if (page) {
 384                 int one_user;
 385
 386                 BUG_ON(PagePrivate(page));
 387                 page_cache_get(page);
 388                 one_user = (page_count(page) == 2);
 389                 /* Only cache user (+us), or swap space full? Free it! */
 390                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
 391                         delete_from_swap_cache(page);
 392                         SetPageDirty(page);
 393                 }
 394                 unlock_page(page);
 395                 page_cache_release(page);
 396         }
 397 }
 398
 399 /*
 400  * Always set the resulting pte to be nowrite (the same as COW pages
 401  * after one process has exited).  We don't know just how many PTEs will
 402  * share this swap entry, so be cautious and let do_wp_page work out
 403  * what to do if a write is requested later.
 404  *
 405  * vma->vm_mm->page_table_lock is held.
 406  */
 407 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
 408                 unsigned long addr, swp_entry_t entry, struct page *page)
 409 {
 410         inc_mm_counter(vma->vm_mm, rss);
 411         get_page(page);
 412         set_pte_at(vma->vm_mm, addr, pte,
 413                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 414         page_add_anon_rmap(page, vma, addr);
 415         swap_free(entry);
 416         /*
 417          * Move the page to the active list so it is not
 418          * immediately swapped out again after swapon.
 419          */
 420         activate_page(page);
 421 }
 422
 423 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 424                                 unsigned long addr, unsigned long end,
 425                                 swp_entry_t entry, struct page *page)
 426 {
 427         pte_t *pte;
 428         pte_t swp_pte = swp_entry_to_pte(entry);
 429
 430         pte = pte_offset_map(pmd, addr);
 431         do {
 432                 /*
 433                  * swapoff spends a _lot_ of time in this loop!
 434                  * Test inline before going to call unuse_pte.
 435                  */
 436                 if (unlikely(pte_same(*pte, swp_pte))) {
 437                         unuse_pte(vma, pte, addr, entry, page);
 438                         pte_unmap(pte);
 439                         return 1;
 440                 }
 441         } while (pte++, addr += PAGE_SIZE, addr != end);
 442         pte_unmap(pte - 1);
 443         return 0;
 444 }
 445
 446 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 447                                 unsigned long addr, unsigned long end,
 448                                 swp_entry_t entry, struct page *page)
 449 {
 450         pmd_t *pmd;
 451         unsigned long next;
 452
 453         pmd = pmd_offset(pud, addr);
 454         do {
 455                 next = pmd_addr_end(addr, end);
 456                 if (pmd_none_or_clear_bad(pmd))
 457                         continue;
 458                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 459                         return 1;
 460         } while (pmd++, addr = next, addr != end);
 461         return 0;
 462 }
 463
 464 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 465                                 unsigned long addr, unsigned long end,
 466                                 swp_entry_t entry, struct page *page)
 467 {
 468         pud_t *pud;
 469         unsigned long next;
 470
 471         pud = pud_offset(pgd, addr);
 472         do {
 473                 next = pud_addr_end(addr, end);
 474                 if (pud_none_or_clear_bad(pud))
 475                         continue;
 476                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 477                         return 1;
 478         } while (pud++, addr = next, addr != end);
 479         return 0;
 480 }
 481
 482 static int unuse_vma(struct vm_area_struct *vma,
 483                                 swp_entry_t entry, struct page *page)
 484 {
 485         pgd_t *pgd;
 486         unsigned long addr, end, next;
 487
 488         if (page->mapping) {
 489                 addr = page_address_in_vma(page, vma);
 490                 if (addr == -EFAULT)
 491                         return 0;
 492                 else
 493                         end = addr + PAGE_SIZE;
 494         } else {
 495                 addr = vma->vm_start;
 496                 end = vma->vm_end;
 497         }
 498
 499         pgd = pgd_offset(vma->vm_mm, addr);
 500         do {
 501                 next = pgd_addr_end(addr, end);
 502                 if (pgd_none_or_clear_bad(pgd))
 503                         continue;
 504                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 505                         return 1;
 506         } while (pgd++, addr = next, addr != end);
 507         return 0;
 508 }
 509
 510 static int unuse_mm(struct mm_struct *mm,
 511                                 swp_entry_t entry, struct page *page)
 512 {
 513         struct vm_area_struct *vma;
 514
 515         if (!down_read_trylock(&mm->mmap_sem)) {
 516                 /*
 517                  * Activate page so shrink_cache is unlikely to unmap its
 518                  * ptes while lock is dropped, so swapoff can make progress.
 519                  */
 520                 activate_page(page);
 521                 unlock_page(page);
 522                 down_read(&mm->mmap_sem);
 523                 lock_page(page);
 524         }
 525         spin_lock(&mm->page_table_lock);
 526         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 527                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 528                         break;
 529         }
 530         spin_unlock(&mm->page_table_lock);
 531         up_read(&mm->mmap_sem);
 532         /*
 533          * Currently unuse_mm cannot fail, but leave error handling
 534          * at call sites for now, since we change it from time to time.
 535          */
 536         return 0;
 537 }
 538
 539 /*
 540  * Scan swap_map from current position to next entry still in use.
 541  * Recycle to start on reaching the end, returning 0 when empty.
 542  */
 543 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 544                                         unsigned int prev)
 545 {
 546         unsigned int max = si->max;
 547         unsigned int i = prev;
 548         int count;
 549
 550         /*
 551          * No need for swap_device_lock(si) here: we're just looking
 552          * for whether an entry is in use, not modifying it; false
 553          * hits are okay, and sys_swapoff() has already prevented new
 554          * allocations from this area (while holding swap_list_lock()).
 555          */
 556         for (;;) {
 557                 if (++i >= max) {
 558                         if (!prev) {
 559                                 i = 0;
 560                                 break;
 561                         }
 562                         /*
 563                          * No entries in use at top of swap_map,
 564                          * loop back to start and recheck there.
 565                          */
 566                         max = prev + 1;
 567                         prev = 0;
 568                         i = 1;
 569                 }
 570                 count = si->swap_map[i];
 571                 if (count && count != SWAP_MAP_BAD)
 572                         break;
 573         }
 574         return i;
 575 }
 576
 577 /*
 578  * We completely avoid races by reading each swap page in advance,
 579  * and then search for the process using it.  All the necessary
 580  * page table adjustments can then be made atomically.
 581  */
 582 static int try_to_unuse(unsigned int type)
 583 {
 584         struct swap_info_struct * si = &swap_info[type];
 585         struct mm_struct *start_mm;
 586         unsigned short *swap_map;
 587         unsigned short swcount;
 588         struct page *page;
 589         swp_entry_t entry;
 590         unsigned int i = 0;
 591         int retval = 0;
 592         int reset_overflow = 0;
 593         int shmem;
 594
 595         /*
 596          * When searching mms for an entry, a good strategy is to
 597          * start at the first mm we freed the previous entry from
 598          * (though actually we don't notice whether we or coincidence
 599          * freed the entry).  Initialize this start_mm with a hold.
 600          *
 601          * A simpler strategy would be to start at the last mm we
 602          * freed the previous entry from; but that would take less
 603          * advantage of mmlist ordering, which clusters forked mms
 604          * together, child after parent.  If we race with dup_mmap(), we
 605          * prefer to resolve parent before child, lest we miss entries
 606          * duplicated after we scanned child: using last mm would invert
 607          * that.  Though it's only a serious concern when an overflowed
 608          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 609          */
 610         start_mm = &init_mm;
 611         atomic_inc(&init_mm.mm_users);
 612
 613         /*
 614          * Keep on scanning until all entries have gone.  Usually,
 615          * one pass through swap_map is enough, but not necessarily:
 616          * there are races when an instance of an entry might be missed.
 617          */
 618         while ((i = find_next_to_unuse(si, i)) != 0) {
 619                 if (signal_pending(current)) {
 620                         retval = -EINTR;
 621                         break;
 622                 }
 623
 624                 /*
 625                  * Get a page for the entry, using the existing swap
 626                  * cache page if there is one.  Otherwise, get a clean
 627                  * page and read the swap into it.
 628                  */
 629                 swap_map = &si->swap_map[i];
 630                 entry = swp_entry(type, i);
 631                 page = read_swap_cache_async(entry, NULL, 0);
 632                 if (!page) {
 633                         /*
 634                          * Either swap_duplicate() failed because entry
 635                          * has been freed independently, and will not be
 636                          * reused since sys_swapoff() already disabled
 637                          * allocation from here, or alloc_page() failed.
 638                          */
 639                         if (!*swap_map)
 640                                 continue;
 641                         retval = -ENOMEM;
 642                         break;
 643                 }
 644
 645                 /*
 646                  * Don't hold on to start_mm if it looks like exiting.
 647                  */
 648                 if (atomic_read(&start_mm->mm_users) == 1) {
 649                         mmput(start_mm);
 650                         start_mm = &init_mm;
 651                         atomic_inc(&init_mm.mm_users);
 652                 }
 653
 654                 /*
 655                  * Wait for and lock page.  When do_swap_page races with
 656                  * try_to_unuse, do_swap_page can handle the fault much
 657                  * faster than try_to_unuse can locate the entry.  This
 658                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 659                  * defer to do_swap_page in such a case - in some tests,
 660                  * do_swap_page and try_to_unuse repeatedly compete.
 661                  */
 662                 wait_on_page_locked(page);
 663                 wait_on_page_writeback(page);
 664                 lock_page(page);
 665                 wait_on_page_writeback(page);
 666
 667                 /*
 668                  * Remove all references to entry.
 669                  * Whenever we reach init_mm, there's no address space
 670                  * to search, but use it as a reminder to search shmem.
 671                  */
 672                 shmem = 0;
 673                 swcount = *swap_map;
 674                 if (swcount > 1) {
 675                         if (start_mm == &init_mm)
 676                                 shmem = shmem_unuse(entry, page);
 677                         else
 678                                 retval = unuse_mm(start_mm, entry, page);
 679                 }
 680                 if (*swap_map > 1) {
 681                         int set_start_mm = (*swap_map >= swcount);
 682                         struct list_head *p = &start_mm->mmlist;
 683                         struct mm_struct *new_start_mm = start_mm;
 684                         struct mm_struct *prev_mm = start_mm;
 685                         struct mm_struct *mm;
 686
 687                         atomic_inc(&new_start_mm->mm_users);
 688                         atomic_inc(&prev_mm->mm_users);
 689                         spin_lock(&mmlist_lock);
 690                         while (*swap_map > 1 && !retval &&
 691                                         (p = p->next) != &start_mm->mmlist) {
 692                                 mm = list_entry(p, struct mm_struct, mmlist);
 693                                 if (atomic_inc_return(&mm->mm_users) == 1) {
 694                                         atomic_dec(&mm->mm_users);
 695                                         continue;
 696                                 }
 697                                 spin_unlock(&mmlist_lock);
 698                                 mmput(prev_mm);
 699                                 prev_mm = mm;
 700
 701                                 cond_resched();
 702
 703                                 swcount = *swap_map;
 704                                 if (swcount <= 1)
 705                                         ;
 706                                 else if (mm == &init_mm) {
 707                                         set_start_mm = 1;
 708                                         shmem = shmem_unuse(entry, page);
 709                                 } else
 710                                         retval = unuse_mm(mm, entry, page);
 711                                 if (set_start_mm && *swap_map < swcount) {
 712                                         mmput(new_start_mm);
 713                                         atomic_inc(&mm->mm_users);
 714                                         new_start_mm = mm;
 715                                         set_start_mm = 0;
 716                                 }
 717                                 spin_lock(&mmlist_lock);
 718                         }
 719                         spin_unlock(&mmlist_lock);
 720                         mmput(prev_mm);
 721                         mmput(start_mm);
 722                         start_mm = new_start_mm;
 723                 }
 724                 if (retval) {
 725                         unlock_page(page);
 726                         page_cache_release(page);
 727                         break;
 728                 }
 729
 730                 /*
 731                  * How could swap count reach 0x7fff when the maximum
 732                  * pid is 0x7fff, and there's no way to repeat a swap
 733                  * page within an mm (except in shmem, where it's the
 734                  * shared object which takes the reference count)?
 735                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 736                  *
 737                  * If that's wrong, then we should worry more about
 738                  * exit_mmap() and do_munmap() cases described above:
 739                  * we might be resetting SWAP_MAP_MAX too early here.
 740                  * We know "Undead"s can happen, they're okay, so don't
 741                  * report them; but do report if we reset SWAP_MAP_MAX.
 742                  */
 743                 if (*swap_map == SWAP_MAP_MAX) {
 744                         swap_device_lock(si);
 745                         *swap_map = 1;
 746                         swap_device_unlock(si);
 747                         reset_overflow = 1;
 748                 }
 749
 750                 /*
 751                  * If a reference remains (rare), we would like to leave
 752                  * the page in the swap cache; but try_to_unmap could
 753                  * then re-duplicate the entry once we drop page lock,
 754                  * so we might loop indefinitely; also, that page could
 755                  * not be swapped out to other storage meanwhile.  So:
 756                  * delete from cache even if there's another reference,
 757                  * after ensuring that the data has been saved to disk -
 758                  * since if the reference remains (rarer), it will be
 759                  * read from disk into another page.  Splitting into two
 760                  * pages would be incorrect if swap supported "shared
 761                  * private" pages, but they are handled by tmpfs files.
 762                  *
 763                  * Note shmem_unuse already deleted a swappage from
 764                  * the swap cache, unless the move to filepage failed:
 765                  * in which case it left swappage in cache, lowered its
 766                  * swap count to pass quickly through the loops above,
 767                  * and now we must reincrement count to try again later.
 768                  */
 769                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 770                         struct writeback_control wbc = {
 771                                 .sync_mode = WB_SYNC_NONE,
 772                         };
 773
 774                         swap_writepage(page, &wbc);
 775                         lock_page(page);
 776                         wait_on_page_writeback(page);
 777                 }
 778                 if (PageSwapCache(page)) {
 779                         if (shmem)
 780                                 swap_duplicate(entry);
 781                         else
 782                                 delete_from_swap_cache(page);
 783                 }
 784
 785                 /*
 786                  * So we could skip searching mms once swap count went
 787                  * to 1, we did not mark any present ptes as dirty: must
 788                  * mark page dirty so shrink_list will preserve it.
 789                  */
 790                 SetPageDirty(page);
 791                 unlock_page(page);
 792                 page_cache_release(page);
 793
 794                 /*
 795                  * Make sure that we aren't completely killing
 796                  * interactive performance.
 797                  */
 798                 cond_resched();
 799         }
 800
 801         mmput(start_mm);
 802         if (reset_overflow) {
 803                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 804                 swap_overflow = 0;
 805         }
 806         return retval;
 807 }
 808
 809 /*
 810  * After a successful try_to_unuse, if no swap is now in use, we know we
 811  * can empty the mmlist.  swap_list_lock must be held on entry and exit.
 812  * Note that mmlist_lock nests inside swap_list_lock, and an mm must be
 813  * added to the mmlist just after page_duplicate - before would be racy.
 814  */
 815 static void drain_mmlist(void)
 816 {
 817         struct list_head *p, *next;
 818         unsigned int i;
 819
 820         for (i = 0; i < nr_swapfiles; i++)
 821                 if (swap_info[i].inuse_pages)
 822                         return;
 823         spin_lock(&mmlist_lock);
 824         list_for_each_safe(p, next, &init_mm.mmlist)
 825                 list_del_init(p);
 826         spin_unlock(&mmlist_lock);
 827 }
 828
 829 /*
 830  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 831  * corresponds to page offset `offset'.
 832  */
 833 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 834 {
 835         struct swap_extent *se = sis->curr_swap_extent;
 836         struct swap_extent *start_se = se;
 837
 838         for ( ; ; ) {
 839                 struct list_head *lh;
 840
 841                 if (se->start_page <= offset &&
 842                                 offset < (se->start_page + se->nr_pages)) {
 843                         return se->start_block + (offset - se->start_page);
 844                 }
 845                 lh = se->list.next;
 846                 if (lh == &sis->extent_list)
 847                         lh = lh->next;
 848                 se = list_entry(lh, struct swap_extent, list);
 849                 sis->curr_swap_extent = se;
 850                 BUG_ON(se == start_se);         /* It *must* be present */
 851         }
 852 }
 853
 854 /*
 855  * Free all of a swapdev's extent information
 856  */
 857 static void destroy_swap_extents(struct swap_info_struct *sis)
 858 {
 859         while (!list_empty(&sis->extent_list)) {
 860                 struct swap_extent *se;
 861
 862                 se = list_entry(sis->extent_list.next,
 863                                 struct swap_extent, list);
 864                 list_del(&se->list);
 865                 kfree(se);
 866         }
 867 }
 868
 869 /*
 870  * Add a block range (and the corresponding page range) into this swapdev's
 871  * extent list.  The extent list is kept sorted in page order.
 872  *
 873  * This function rather assumes that it is called in ascending page order.
 874  */
 875 static int
 876 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 877                 unsigned long nr_pages, sector_t start_block)
 878 {
 879         struct swap_extent *se;
 880         struct swap_extent *new_se;
 881         struct list_head *lh;
 882
 883         lh = sis->extent_list.prev;     /* The highest page extent */
 884         if (lh != &sis->extent_list) {
 885                 se = list_entry(lh, struct swap_extent, list);
 886                 BUG_ON(se->start_page + se->nr_pages != start_page);
 887                 if (se->start_block + se->nr_pages == start_block) {
 888                         /* Merge it */
 889                         se->nr_pages += nr_pages;
 890                         return 0;
 891                 }
 892         }
 893
 894         /*
 895          * No merge.  Insert a new extent, preserving ordering.
 896          */
 897         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 898         if (new_se == NULL)
 899                 return -ENOMEM;
 900         new_se->start_page = start_page;
 901         new_se->nr_pages = nr_pages;
 902         new_se->start_block = start_block;
 903
 904         list_add_tail(&new_se->list, &sis->extent_list);
 905         return 1;
 906 }
 907
 908 /*
 909  * A `swap extent' is a simple thing which maps a contiguous range of pages
 910  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 911  * is built at swapon time and is then used at swap_writepage/swap_readpage
 912  * time for locating where on disk a page belongs.
 913  *
 914  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 915  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 916  * swap files identically.
 917  *
 918  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
 919  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
 920  * swapfiles are handled *identically* after swapon time.
 921  *
 922  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
 923  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
 924  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
 925  * requirements, they are simply tossed out - we will never use those blocks
 926  * for swapping.
 927  *
 928  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
 929  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
 930  * which will scribble on the fs.
 931  *
 932  * The amount of disk space which a single swap extent represents varies.
 933  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
 934  * extents in the list.  To avoid much list walking, we cache the previous
 935  * search location in `curr_swap_extent', and start new searches from there.
 936  * This is extremely effective.  The average number of iterations in
 937  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
 938  */
 939 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
 940 {
 941         struct inode *inode;
 942         unsigned blocks_per_page;
 943         unsigned long page_no;
 944         unsigned blkbits;
 945         sector_t probe_block;
 946         sector_t last_block;
 947         sector_t lowest_block = -1;
 948         sector_t highest_block = 0;
 949         int nr_extents = 0;
 950         int ret;
 951
 952         inode = sis->swap_file->f_mapping->host;
 953         if (S_ISBLK(inode->i_mode)) {
 954                 ret = add_swap_extent(sis, 0, sis->max, 0);
 955                 *span = sis->pages;
 956                 goto done;
 957         }
 958
 959         blkbits = inode->i_blkbits;
 960         blocks_per_page = PAGE_SIZE >> blkbits;
 961
 962         /*
 963          * Map all the blocks into the extent list.  This code doesn't try
 964          * to be very smart.
 965          */
 966         probe_block = 0;
 967         page_no = 0;
 968         last_block = i_size_read(inode) >> blkbits;
 969         while ((probe_block + blocks_per_page) <= last_block &&
 970                         page_no < sis->max) {
 971                 unsigned block_in_page;
 972                 sector_t first_block;
 973
 974                 first_block = bmap(inode, probe_block);
 975                 if (first_block == 0)
 976                         goto bad_bmap;
 977
 978                 /*
 979                  * It must be PAGE_SIZE aligned on-disk
 980                  */
 981                 if (first_block & (blocks_per_page - 1)) {
 982                         probe_block++;
 983                         goto reprobe;
 984                 }
 985
 986                 for (block_in_page = 1; block_in_page < blocks_per_page;
 987                                         block_in_page++) {
 988                         sector_t block;
 989
 990                         block = bmap(inode, probe_block + block_in_page);
 991                         if (block == 0)
 992                                 goto bad_bmap;
 993                         if (block != first_block + block_in_page) {
 994                                 /* Discontiguity */
 995                                 probe_block++;
 996                                 goto reprobe;
 997                         }
 998                 }
 999
1000                 first_block >>= (PAGE_SHIFT - blkbits);
1001                 if (page_no) {  /* exclude the header page */
1002                         if (first_block < lowest_block)
1003                                 lowest_block = first_block;
1004                         if (first_block > highest_block)
1005                                 highest_block = first_block;
1006                 }
1007
1008                 /*
1009                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1010                  */
1011                 ret = add_swap_extent(sis, page_no, 1, first_block);
1012                 if (ret < 0)
1013                         goto out;
1014                 nr_extents += ret;
1015                 page_no++;
1016                 probe_block += blocks_per_page;
1017 reprobe:
1018                 continue;
1019         }
1020         ret = nr_extents;
1021         *span = 1 + highest_block - lowest_block;
1022         if (page_no == 0)
1023                 page_no = 1;    /* force Empty message */
1024         sis->max = page_no;
1025         sis->pages = page_no - 1;
1026         sis->highest_bit = page_no - 1;
1027 done:
1028         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1029                                         struct swap_extent, list);
1030         goto out;
1031 bad_bmap:
1032         printk(KERN_ERR "swapon: swapfile has holes\n");
1033         ret = -EINVAL;
1034 out:
1035         return ret;
1036 }
1037
1038 #if 0   /* We don't need this yet */
1039 #include <linux/backing-dev.h>
1040 int page_queue_congested(struct page *page)
1041 {
1042         struct backing_dev_info *bdi;
1043
1044         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1045
1046         if (PageSwapCache(page)) {
1047                 swp_entry_t entry = { .val = page->private };
1048                 struct swap_info_struct *sis;
1049
1050                 sis = get_swap_info_struct(swp_type(entry));
1051                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1052         } else
1053                 bdi = page->mapping->backing_dev_info;
1054         return bdi_write_congested(bdi);
1055 }
1056 #endif
1057
1058 asmlinkage long sys_swapoff(const char __user * specialfile)
1059 {
1060         struct swap_info_struct * p = NULL;
1061         unsigned short *swap_map;
1062         struct file *swap_file, *victim;
1063         struct address_space *mapping;
1064         struct inode *inode;
1065         char * pathname;
1066         int i, type, prev;
1067         int err;
1068
1069         if (!capable(CAP_SYS_ADMIN))
1070                 return -EPERM;
1071
1072         pathname = getname(specialfile);
1073         err = PTR_ERR(pathname);
1074         if (IS_ERR(pathname))
1075                 goto out;
1076
1077         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1078         putname(pathname);
1079         err = PTR_ERR(victim);
1080         if (IS_ERR(victim))
1081                 goto out;
1082
1083         mapping = victim->f_mapping;
1084         prev = -1;
1085         swap_list_lock();
1086         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1087                 p = swap_info + type;
1088                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1089                         if (p->swap_file->f_mapping == mapping)
1090                                 break;
1091                 }
1092                 prev = type;
1093         }
1094         if (type < 0) {
1095                 err = -EINVAL;
1096                 swap_list_unlock();
1097                 goto out_dput;
1098         }
1099         if (!security_vm_enough_memory(p->pages))
1100                 vm_unacct_memory(p->pages);
1101         else {
1102                 err = -ENOMEM;
1103                 swap_list_unlock();
1104                 goto out_dput;
1105         }
1106         if (prev < 0) {
1107                 swap_list.head = p->next;
1108         } else {
1109                 swap_info[prev].next = p->next;
1110         }
1111         if (type == swap_list.next) {
1112                 /* just pick something that's safe... */
1113                 swap_list.next = swap_list.head;
1114         }
1115         nr_swap_pages -= p->pages;
1116         total_swap_pages -= p->pages;
1117         swap_device_lock(p);
1118         p->flags &= ~SWP_WRITEOK;
1119         swap_device_unlock(p);
1120         swap_list_unlock();
1121
1122         current->flags |= PF_SWAPOFF;
1123         err = try_to_unuse(type);
1124         current->flags &= ~PF_SWAPOFF;
1125
1126         if (err) {
1127                 /* re-insert swap space back into swap_list */
1128                 swap_list_lock();
1129                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1130                         if (p->prio >= swap_info[i].prio)
1131                                 break;
1132                 p->next = i;
1133                 if (prev < 0)
1134                         swap_list.head = swap_list.next = p - swap_info;
1135                 else
1136                         swap_info[prev].next = p - swap_info;
1137                 nr_swap_pages += p->pages;
1138                 total_swap_pages += p->pages;
1139                 swap_device_lock(p);
1140                 p->flags |= SWP_WRITEOK;
1141                 swap_device_unlock(p);
1142                 swap_list_unlock();
1143                 goto out_dput;
1144         }
1145
1146         /* wait for any unplug function to finish */
1147         down_write(&swap_unplug_sem);
1148         up_write(&swap_unplug_sem);
1149
1150         /* wait for anyone still in scan_swap_map */
1151         swap_device_lock(p);
1152         p->highest_bit = 0;             /* cuts scans short */
1153         while (p->flags >= SWP_SCANNING) {
1154                 swap_device_unlock(p);
1155                 set_current_state(TASK_UNINTERRUPTIBLE);
1156                 schedule_timeout(1);
1157                 swap_device_lock(p);
1158         }
1159         swap_device_unlock(p);
1160
1161         destroy_swap_extents(p);
1162         down(&swapon_sem);
1163         swap_list_lock();
1164         drain_mmlist();
1165         swap_device_lock(p);
1166         swap_file = p->swap_file;
1167         p->swap_file = NULL;
1168         p->max = 0;
1169         swap_map = p->swap_map;
1170         p->swap_map = NULL;
1171         p->flags = 0;
1172         swap_device_unlock(p);
1173         swap_list_unlock();
1174         up(&swapon_sem);
1175         vfree(swap_map);
1176         inode = mapping->host;
1177         if (S_ISBLK(inode->i_mode)) {
1178                 struct block_device *bdev = I_BDEV(inode);
1179                 set_blocksize(bdev, p->old_block_size);
1180                 bd_release(bdev);
1181         } else {
1182                 down(&inode->i_sem);
1183                 inode->i_flags &= ~S_SWAPFILE;
1184                 up(&inode->i_sem);
1185         }
1186         filp_close(swap_file, NULL);
1187         err = 0;
1188
1189 out_dput:
1190         filp_close(victim, NULL);
1191 out:
1192         return err;
1193 }
1194
1195 #ifdef CONFIG_PROC_FS
1196 /* iterator */
1197 static void *swap_start(struct seq_file *swap, loff_t *pos)
1198 {
1199         struct swap_info_struct *ptr = swap_info;
1200         int i;
1201         loff_t l = *pos;
1202
1203         down(&swapon_sem);
1204
1205         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1206                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1207                         continue;
1208                 if (!l--)
1209                         return ptr;
1210         }
1211
1212         return NULL;
1213 }
1214
1215 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1216 {
1217         struct swap_info_struct *ptr = v;
1218         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1219
1220         for (++ptr; ptr < endptr; ptr++) {
1221                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1222                         continue;
1223                 ++*pos;
1224                 return ptr;
1225         }
1226
1227         return NULL;
1228 }
1229
1230 static void swap_stop(struct seq_file *swap, void *v)
1231 {
1232         up(&swapon_sem);
1233 }
1234
1235 static int swap_show(struct seq_file *swap, void *v)
1236 {
1237         struct swap_info_struct *ptr = v;
1238         struct file *file;
1239         int len;
1240
1241         if (v == swap_info)
1242                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1243
1244         file = ptr->swap_file;
1245         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1246         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1247                        len < 40 ? 40 - len : 1, " ",
1248                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1249                                 "partition" : "file\t",
1250                        ptr->pages << (PAGE_SHIFT - 10),
1251                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1252                        ptr->prio);
1253         return 0;
1254 }
1255
1256 static struct seq_operations swaps_op = {
1257         .start =        swap_start,
1258         .next =         swap_next,
1259         .stop =         swap_stop,
1260         .show =         swap_show
1261 };
1262
1263 static int swaps_open(struct inode *inode, struct file *file)
1264 {
1265         return seq_open(file, &swaps_op);
1266 }
1267
1268 static struct file_operations proc_swaps_operations = {
1269         .open           = swaps_open,
1270         .read           = seq_read,
1271         .llseek         = seq_lseek,
1272         .release        = seq_release,
1273 };
1274
1275 static int __init procswaps_init(void)
1276 {
1277         struct proc_dir_entry *entry;
1278
1279         entry = create_proc_entry("swaps", 0, NULL);
1280         if (entry)
1281                 entry->proc_fops = &proc_swaps_operations;
1282         return 0;
1283 }
1284 __initcall(procswaps_init);
1285 #endif /* CONFIG_PROC_FS */
1286
1287 /*
1288  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1289  *
1290  * The swapon system call
1291  */
1292 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1293 {
1294         struct swap_info_struct * p;
1295         char *name = NULL;
1296         struct block_device *bdev = NULL;
1297         struct file *swap_file = NULL;
1298         struct address_space *mapping;
1299         unsigned int type;
1300         int i, prev;
1301         int error;
1302         static int least_priority;
1303         union swap_header *swap_header = NULL;
1304         int swap_header_version;
1305         unsigned int nr_good_pages = 0;
1306         int nr_extents = 0;
1307         sector_t span;
1308         unsigned long maxpages = 1;
1309         int swapfilesize;
1310         unsigned short *swap_map;
1311         struct page *page = NULL;
1312         struct inode *inode = NULL;
1313         int did_down = 0;
1314
1315         if (!capable(CAP_SYS_ADMIN))
1316                 return -EPERM;
1317         swap_list_lock();
1318         p = swap_info;
1319         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1320                 if (!(p->flags & SWP_USED))
1321                         break;
1322         error = -EPERM;
1323         /*
1324          * Test if adding another swap device is possible. There are
1325          * two limiting factors: 1) the number of bits for the swap
1326          * type swp_entry_t definition and 2) the number of bits for
1327          * the swap type in the swap ptes as defined by the different
1328          * architectures. To honor both limitations a swap entry
1329          * with swap offset 0 and swap type ~0UL is created, encoded
1330          * to a swap pte, decoded to a swp_entry_t again and finally
1331          * the swap type part is extracted. This will mask all bits
1332          * from the initial ~0UL that can't be encoded in either the
1333          * swp_entry_t or the architecture definition of a swap pte.
1334          */
1335         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1336                 swap_list_unlock();
1337                 goto out;
1338         }
1339         if (type >= nr_swapfiles)
1340                 nr_swapfiles = type+1;
1341         INIT_LIST_HEAD(&p->extent_list);
1342         p->flags = SWP_USED;
1343         p->swap_file = NULL;
1344         p->old_block_size = 0;
1345         p->swap_map = NULL;
1346         p->lowest_bit = 0;
1347         p->highest_bit = 0;
1348         p->cluster_nr = 0;
1349         p->inuse_pages = 0;
1350         spin_lock_init(&p->sdev_lock);
1351         p->next = -1;
1352         if (swap_flags & SWAP_FLAG_PREFER) {
1353                 p->prio =
1354                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1355         } else {
1356                 p->prio = --least_priority;
1357         }
1358         swap_list_unlock();
1359         name = getname(specialfile);
1360         error = PTR_ERR(name);
1361         if (IS_ERR(name)) {
1362                 name = NULL;
1363                 goto bad_swap_2;
1364         }
1365         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1366         error = PTR_ERR(swap_file);
1367         if (IS_ERR(swap_file)) {
1368                 swap_file = NULL;
1369                 goto bad_swap_2;
1370         }
1371
1372         p->swap_file = swap_file;
1373         mapping = swap_file->f_mapping;
1374         inode = mapping->host;
1375
1376         error = -EBUSY;
1377         for (i = 0; i < nr_swapfiles; i++) {
1378                 struct swap_info_struct *q = &swap_info[i];
1379
1380                 if (i == type || !q->swap_file)
1381                         continue;
1382                 if (mapping == q->swap_file->f_mapping)
1383                         goto bad_swap;
1384         }
1385
1386         error = -EINVAL;
1387         if (S_ISBLK(inode->i_mode)) {
1388                 bdev = I_BDEV(inode);
1389                 error = bd_claim(bdev, sys_swapon);
1390                 if (error < 0) {
1391                         bdev = NULL;
1392                         goto bad_swap;
1393                 }
1394                 p->old_block_size = block_size(bdev);
1395                 error = set_blocksize(bdev, PAGE_SIZE);
1396                 if (error < 0)
1397                         goto bad_swap;
1398                 p->bdev = bdev;
1399         } else if (S_ISREG(inode->i_mode)) {
1400                 p->bdev = inode->i_sb->s_bdev;
1401                 down(&inode->i_sem);
1402                 did_down = 1;
1403                 if (IS_SWAPFILE(inode)) {
1404                         error = -EBUSY;
1405                         goto bad_swap;
1406                 }
1407         } else {
1408                 goto bad_swap;
1409         }
1410
1411         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1412
1413         /*
1414          * Read the swap header.
1415          */
1416         if (!mapping->a_ops->readpage) {
1417                 error = -EINVAL;
1418                 goto bad_swap;
1419         }
1420         page = read_cache_page(mapping, 0,
1421                         (filler_t *)mapping->a_ops->readpage, swap_file);
1422         if (IS_ERR(page)) {
1423                 error = PTR_ERR(page);
1424                 goto bad_swap;
1425         }
1426         wait_on_page_locked(page);
1427         if (!PageUptodate(page))
1428                 goto bad_swap;
1429         kmap(page);
1430         swap_header = page_address(page);
1431
1432         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1433                 swap_header_version = 1;
1434         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1435                 swap_header_version = 2;
1436         else {
1437                 printk("Unable to find swap-space signature\n");
1438                 error = -EINVAL;
1439                 goto bad_swap;
1440         }
1441
1442         switch (swap_header_version) {
1443         case 1:
1444                 printk(KERN_ERR "version 0 swap is no longer supported. "
1445                         "Use mkswap -v1 %s\n", name);
1446                 error = -EINVAL;
1447                 goto bad_swap;
1448         case 2:
1449                 /* Check the swap header's sub-version and the size of
1450                    the swap file and bad block lists */
1451                 if (swap_header->info.version != 1) {
1452                         printk(KERN_WARNING
1453                                "Unable to handle swap header version %d\n",
1454                                swap_header->info.version);
1455                         error = -EINVAL;
1456                         goto bad_swap;
1457                 }
1458
1459                 p->lowest_bit  = 1;
1460                 p->cluster_next = 1;
1461
1462                 /*
1463                  * Find out how many pages are allowed for a single swap
1464                  * device. There are two limiting factors: 1) the number of
1465                  * bits for the swap offset in the swp_entry_t type and
1466                  * 2) the number of bits in the a swap pte as defined by
1467                  * the different architectures. In order to find the
1468                  * largest possible bit mask a swap entry with swap type 0
1469                  * and swap offset ~0UL is created, encoded to a swap pte,
1470                  * decoded to a swp_entry_t again and finally the swap
1471                  * offset is extracted. This will mask all the bits from
1472                  * the initial ~0UL mask that can't be encoded in either
1473                  * the swp_entry_t or the architecture definition of a
1474                  * swap pte.
1475                  */
1476                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1477                 if (maxpages > swap_header->info.last_page)
1478                         maxpages = swap_header->info.last_page;
1479                 p->highest_bit = maxpages - 1;
1480
1481                 error = -EINVAL;
1482                 if (!maxpages)
1483                         goto bad_swap;
1484                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1485                         goto bad_swap;
1486                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1487                         goto bad_swap;
1488
1489                 /* OK, set up the swap map and apply the bad block list */
1490                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1491                         error = -ENOMEM;
1492                         goto bad_swap;
1493                 }
1494
1495                 error = 0;
1496                 memset(p->swap_map, 0, maxpages * sizeof(short));
1497                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1498                         int page = swap_header->info.badpages[i];
1499                         if (page <= 0 || page >= swap_header->info.last_page)
1500                                 error = -EINVAL;
1501                         else
1502                                 p->swap_map[page] = SWAP_MAP_BAD;
1503                 }
1504                 nr_good_pages = swap_header->info.last_page -
1505                                 swap_header->info.nr_badpages -
1506                                 1 /* header page */;
1507                 if (error)
1508                         goto bad_swap;
1509         }
1510
1511         if (swapfilesize && maxpages > swapfilesize) {
1512                 printk(KERN_WARNING
1513                        "Swap area shorter than signature indicates\n");
1514                 error = -EINVAL;
1515                 goto bad_swap;
1516         }
1517         if (nr_good_pages) {
1518                 p->swap_map[0] = SWAP_MAP_BAD;
1519                 p->max = maxpages;
1520                 p->pages = nr_good_pages;
1521                 nr_extents = setup_swap_extents(p, &span);
1522                 if (nr_extents < 0) {
1523                         error = nr_extents;
1524                         goto bad_swap;
1525                 }
1526                 nr_good_pages = p->pages;
1527         }
1528         if (!nr_good_pages) {
1529                 printk(KERN_WARNING "Empty swap-file\n");
1530                 error = -EINVAL;
1531                 goto bad_swap;
1532         }
1533
1534         down(&swapon_sem);
1535         swap_list_lock();
1536         swap_device_lock(p);
1537         p->flags = SWP_ACTIVE;
1538         nr_swap_pages += nr_good_pages;
1539         total_swap_pages += nr_good_pages;
1540
1541         printk(KERN_INFO "Adding %uk swap on %s.  "
1542                         "Priority:%d extents:%d across:%lluk\n",
1543                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1544                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1545
1546         /* insert swap space into swap_list: */
1547         prev = -1;
1548         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1549                 if (p->prio >= swap_info[i].prio) {
1550                         break;
1551                 }
1552                 prev = i;
1553         }
1554         p->next = i;
1555         if (prev < 0) {
1556                 swap_list.head = swap_list.next = p - swap_info;
1557         } else {
1558                 swap_info[prev].next = p - swap_info;
1559         }
1560         swap_device_unlock(p);
1561         swap_list_unlock();
1562         up(&swapon_sem);
1563         error = 0;
1564         goto out;
1565 bad_swap:
1566         if (bdev) {
1567                 set_blocksize(bdev, p->old_block_size);
1568                 bd_release(bdev);
1569         }
1570         destroy_swap_extents(p);
1571 bad_swap_2:
1572         swap_list_lock();
1573         swap_map = p->swap_map;
1574         p->swap_file = NULL;
1575         p->swap_map = NULL;
1576         p->flags = 0;
1577         if (!(swap_flags & SWAP_FLAG_PREFER))
1578                 ++least_priority;
1579         swap_list_unlock();
1580         vfree(swap_map);
1581         if (swap_file)
1582                 filp_close(swap_file, NULL);
1583 out:
1584         if (page && !IS_ERR(page)) {
1585                 kunmap(page);
1586                 page_cache_release(page);
1587         }
1588         if (name)
1589                 putname(name);
1590         if (did_down) {
1591                 if (!error)
1592                         inode->i_flags |= S_SWAPFILE;
1593                 up(&inode->i_sem);
1594         }
1595         return error;
1596 }
1597
1598 void si_swapinfo(struct sysinfo *val)
1599 {
1600         unsigned int i;
1601         unsigned long nr_to_be_unused = 0;
1602
1603         swap_list_lock();
1604         for (i = 0; i < nr_swapfiles; i++) {
1605                 if (!(swap_info[i].flags & SWP_USED) ||
1606                      (swap_info[i].flags & SWP_WRITEOK))
1607                         continue;
1608                 nr_to_be_unused += swap_info[i].inuse_pages;
1609         }
1610         val->freeswap = nr_swap_pages + nr_to_be_unused;
1611         val->totalswap = total_swap_pages + nr_to_be_unused;
1612         swap_list_unlock();
1613 }
1614
1615 /*
1616  * Verify that a swap entry is valid and increment its swap map count.
1617  *
1618  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1619  * "permanent", but will be reclaimed by the next swapoff.
1620  */
1621 int swap_duplicate(swp_entry_t entry)
1622 {
1623         struct swap_info_struct * p;
1624         unsigned long offset, type;
1625         int result = 0;
1626
1627         type = swp_type(entry);
1628         if (type >= nr_swapfiles)
1629                 goto bad_file;
1630         p = type + swap_info;
1631         offset = swp_offset(entry);
1632
1633         swap_device_lock(p);
1634         if (offset < p->max && p->swap_map[offset]) {
1635                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1636                         p->swap_map[offset]++;
1637                         result = 1;
1638                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1639                         if (swap_overflow++ < 5)
1640                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1641                         p->swap_map[offset] = SWAP_MAP_MAX;
1642                         result = 1;
1643                 }
1644         }
1645         swap_device_unlock(p);
1646 out:
1647         return result;
1648
1649 bad_file:
1650         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1651         goto out;
1652 }
1653
1654 struct swap_info_struct *
1655 get_swap_info_struct(unsigned type)
1656 {
1657         return &swap_info[type];
1658 }
1659
1660 /*
1661  * swap_device_lock prevents swap_map being freed. Don't grab an extra
1662  * reference on the swaphandle, it doesn't matter if it becomes unused.
1663  */
1664 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1665 {
1666         int ret = 0, i = 1 << page_cluster;
1667         unsigned long toff;
1668         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1669
1670         if (!page_cluster)      /* no readahead */
1671                 return 0;
1672         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1673         if (!toff)              /* first page is swap header */
1674                 toff++, i--;
1675         *offset = toff;
1676
1677         swap_device_lock(swapdev);
1678         do {
1679                 /* Don't read-ahead past the end of the swap area */
1680                 if (toff >= swapdev->max)
1681                         break;
1682                 /* Don't read in free or bad pages */
1683                 if (!swapdev->swap_map[toff])
1684                         break;
1685                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1686                         break;
1687                 toff++;
1688                 ret++;
1689         } while (--i);
1690         swap_device_unlock(swapdev);
1691         return ret;
1692 }