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