mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/config.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/mman.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/namei.h>
  18 #include <linux/shm.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28
  29 #include <asm/pgtable.h>
  30 #include <asm/tlbflush.h>
  31 #include <linux/swapops.h>
  32
  33 spinlock_t swaplock = SPIN_LOCK_UNLOCKED;
  34 unsigned int nr_swapfiles;
  35 int total_swap_pages;
  36 static int swap_overflow;
  37
  38 EXPORT_SYMBOL(total_swap_pages);
  39
  40 static const char Bad_file[] = "Bad swap file entry ";
  41 static const char Unused_file[] = "Unused swap file entry ";
  42 static const char Bad_offset[] = "Bad swap offset entry ";
  43 static const char Unused_offset[] = "Unused swap offset entry ";
  44
  45 struct swap_list_t swap_list = {-1, -1};
  46
  47 struct swap_info_struct swap_info[MAX_SWAPFILES];
  48
  49 static DECLARE_MUTEX(swapon_sem);
  50
  51 /*
  52  * We need this because the bdev->unplug_fn can sleep and we cannot
  53  * hold swap_list_lock while calling the unplug_fn. And swap_list_lock
  54  * cannot be turned into a semaphore.
  55  */
  56 static DECLARE_RWSEM(swap_unplug_sem);
  57
  58 #define SWAPFILE_CLUSTER 256
  59
  60 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  61 {
  62         swp_entry_t entry;
  63
  64         down_read(&swap_unplug_sem);
  65         entry.val = page->private;
  66         if (PageSwapCache(page)) {
  67                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  68                 struct backing_dev_info *bdi;
  69
  70                 /*
  71                  * If the page is removed from swapcache from under us (with a
  72                  * racy try_to_unuse/swapoff) we need an additional reference
  73                  * count to avoid reading garbage from page->private above. If
  74                  * the WARN_ON triggers during a swapoff it maybe the race
  75                  * condition and it's harmless. However if it triggers without
  76                  * swapoff it signals a problem.
  77                  */
  78                 WARN_ON(page_count(page) <= 1);
  79
  80                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  81                 bdi->unplug_io_fn(bdi, page);
  82         }
  83         up_read(&swap_unplug_sem);
  84 }
  85
  86 static inline int scan_swap_map(struct swap_info_struct *si)
  87 {
  88         unsigned long offset;
  89         /*
  90          * We try to cluster swap pages by allocating them
  91          * sequentially in swap.  Once we've allocated
  92          * SWAPFILE_CLUSTER pages this way, however, we resort to
  93          * first-free allocation, starting a new cluster.  This
  94          * prevents us from scattering swap pages all over the entire
  95          * swap partition, so that we reduce overall disk seek times
  96          * between swap pages.  -- sct */
  97         if (si->cluster_nr) {
  98                 while (si->cluster_next <= si->highest_bit) {
  99                         offset = si->cluster_next++;
 100                         if (si->swap_map[offset])
 101                                 continue;
 102                         si->cluster_nr--;
 103                         goto got_page;
 104                 }
 105         }
 106         si->cluster_nr = SWAPFILE_CLUSTER;
 107
 108         /* try to find an empty (even not aligned) cluster. */
 109         offset = si->lowest_bit;
 110  check_next_cluster:
 111         if (offset+SWAPFILE_CLUSTER-1 <= si->highest_bit)
 112         {
 113                 int nr;
 114                 for (nr = offset; nr < offset+SWAPFILE_CLUSTER; nr++)
 115                         if (si->swap_map[nr])
 116                         {
 117                                 offset = nr+1;
 118                                 goto check_next_cluster;
 119                         }
 120                 /* We found a completly empty cluster, so start
 121                  * using it.
 122                  */
 123                 goto got_page;
 124         }
 125         /* No luck, so now go finegrined as usual. -Andrea */
 126         for (offset = si->lowest_bit; offset <= si->highest_bit ; offset++) {
 127                 if (si->swap_map[offset])
 128                         continue;
 129                 si->lowest_bit = offset+1;
 130         got_page:
 131                 if (offset == si->lowest_bit)
 132                         si->lowest_bit++;
 133                 if (offset == si->highest_bit)
 134                         si->highest_bit--;
 135                 if (si->lowest_bit > si->highest_bit) {
 136                         si->lowest_bit = si->max;
 137                         si->highest_bit = 0;
 138                 }
 139                 si->swap_map[offset] = 1;
 140                 si->inuse_pages++;
 141                 nr_swap_pages--;
 142                 si->cluster_next = offset+1;
 143                 return offset;
 144         }
 145         si->lowest_bit = si->max;
 146         si->highest_bit = 0;
 147         return 0;
 148 }
 149
 150 swp_entry_t get_swap_page(void)
 151 {
 152         struct swap_info_struct * p;
 153         unsigned long offset;
 154         swp_entry_t entry;
 155         int type, wrapped = 0;
 156
 157         entry.val = 0;  /* Out of memory */
 158         swap_list_lock();
 159         type = swap_list.next;
 160         if (type < 0)
 161                 goto out;
 162         if (nr_swap_pages <= 0)
 163                 goto out;
 164
 165         while (1) {
 166                 p = &swap_info[type];
 167                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
 168                         swap_device_lock(p);
 169                         offset = scan_swap_map(p);
 170                         swap_device_unlock(p);
 171                         if (offset) {
 172                                 entry = swp_entry(type,offset);
 173                                 type = swap_info[type].next;
 174                                 if (type < 0 ||
 175                                         p->prio != swap_info[type].prio) {
 176                                                 swap_list.next = swap_list.head;
 177                                 } else {
 178                                         swap_list.next = type;
 179                                 }
 180                                 goto out;
 181                         }
 182                 }
 183                 type = p->next;
 184                 if (!wrapped) {
 185                         if (type < 0 || p->prio != swap_info[type].prio) {
 186                                 type = swap_list.head;
 187                                 wrapped = 1;
 188                         }
 189                 } else
 190                         if (type < 0)
 191                                 goto out;       /* out of swap space */
 192         }
 193 out:
 194         swap_list_unlock();
 195         return entry;
 196 }
 197
 198 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 199 {
 200         struct swap_info_struct * p;
 201         unsigned long offset, type;
 202
 203         if (!entry.val)
 204                 goto out;
 205         type = swp_type(entry);
 206         if (type >= nr_swapfiles)
 207                 goto bad_nofile;
 208         p = & swap_info[type];
 209         if (!(p->flags & SWP_USED))
 210                 goto bad_device;
 211         offset = swp_offset(entry);
 212         if (offset >= p->max)
 213                 goto bad_offset;
 214         if (!p->swap_map[offset])
 215                 goto bad_free;
 216         swap_list_lock();
 217         if (p->prio > swap_info[swap_list.next].prio)
 218                 swap_list.next = type;
 219         swap_device_lock(p);
 220         return p;
 221
 222 bad_free:
 223         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 224         goto out;
 225 bad_offset:
 226         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 227         goto out;
 228 bad_device:
 229         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 230         goto out;
 231 bad_nofile:
 232         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 233 out:
 234         return NULL;
 235 }
 236
 237 static void swap_info_put(struct swap_info_struct * p)
 238 {
 239         swap_device_unlock(p);
 240         swap_list_unlock();
 241 }
 242
 243 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 244 {
 245         int count = p->swap_map[offset];
 246
 247         if (count < SWAP_MAP_MAX) {
 248                 count--;
 249                 p->swap_map[offset] = count;
 250                 if (!count) {
 251                         if (offset < p->lowest_bit)
 252                                 p->lowest_bit = offset;
 253                         if (offset > p->highest_bit)
 254                                 p->highest_bit = offset;
 255                         nr_swap_pages++;
 256                         p->inuse_pages--;
 257                 }
 258         }
 259         return count;
 260 }
 261
 262 /*
 263  * Caller has made sure that the swapdevice corresponding to entry
 264  * is still around or has not been recycled.
 265  */
 266 void swap_free(swp_entry_t entry)
 267 {
 268         struct swap_info_struct * p;
 269
 270         p = swap_info_get(entry);
 271         if (p) {
 272                 swap_entry_free(p, swp_offset(entry));
 273                 swap_info_put(p);
 274         }
 275 }
 276
 277 /*
 278  * Check if we're the only user of a swap page,
 279  * when the page is locked.
 280  */
 281 static int exclusive_swap_page(struct page *page)
 282 {
 283         int retval = 0;
 284         struct swap_info_struct * p;
 285         swp_entry_t entry;
 286
 287         entry.val = page->private;
 288         p = swap_info_get(entry);
 289         if (p) {
 290                 /* Is the only swap cache user the cache itself? */
 291                 if (p->swap_map[swp_offset(entry)] == 1) {
 292                         /* Recheck the page count with the swapcache lock held.. */
 293                         spin_lock_irq(&swapper_space.tree_lock);
 294                         if (page_count(page) == 2)
 295                                 retval = 1;
 296                         spin_unlock_irq(&swapper_space.tree_lock);
 297                 }
 298                 swap_info_put(p);
 299         }
 300         return retval;
 301 }
 302
 303 /*
 304  * We can use this swap cache entry directly
 305  * if there are no other references to it.
 306  *
 307  * Here "exclusive_swap_page()" does the real
 308  * work, but we opportunistically check whether
 309  * we need to get all the locks first..
 310  */
 311 int can_share_swap_page(struct page *page)
 312 {
 313         int retval = 0;
 314
 315         if (!PageLocked(page))
 316                 BUG();
 317         switch (page_count(page)) {
 318         case 3:
 319                 if (!PagePrivate(page))
 320                         break;
 321                 /* Fallthrough */
 322         case 2:
 323                 if (!PageSwapCache(page))
 324                         break;
 325                 retval = exclusive_swap_page(page);
 326                 break;
 327         case 1:
 328                 if (PageReserved(page))
 329                         break;
 330                 retval = 1;
 331         }
 332         return retval;
 333 }
 334
 335 /*
 336  * Work out if there are any other processes sharing this
 337  * swap cache page. Free it if you can. Return success.
 338  */
 339 int remove_exclusive_swap_page(struct page *page)
 340 {
 341         int retval;
 342         struct swap_info_struct * p;
 343         swp_entry_t entry;
 344
 345         BUG_ON(PagePrivate(page));
 346         BUG_ON(!PageLocked(page));
 347
 348         if (!PageSwapCache(page))
 349                 return 0;
 350         if (PageWriteback(page))
 351                 return 0;
 352         if (page_count(page) != 2) /* 2: us + cache */
 353                 return 0;
 354
 355         entry.val = page->private;
 356         p = swap_info_get(entry);
 357         if (!p)
 358                 return 0;
 359
 360         /* Is the only swap cache user the cache itself? */
 361         retval = 0;
 362         if (p->swap_map[swp_offset(entry)] == 1) {
 363                 /* Recheck the page count with the swapcache lock held.. */
 364                 spin_lock_irq(&swapper_space.tree_lock);
 365                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 366                         __delete_from_swap_cache(page);
 367                         SetPageDirty(page);
 368                         retval = 1;
 369                 }
 370                 spin_unlock_irq(&swapper_space.tree_lock);
 371         }
 372         swap_info_put(p);
 373
 374         if (retval) {
 375                 swap_free(entry);
 376                 page_cache_release(page);
 377         }
 378
 379         return retval;
 380 }
 381
 382 /*
 383  * Free the swap entry like above, but also try to
 384  * free the page cache entry if it is the last user.
 385  */
 386 void free_swap_and_cache(swp_entry_t entry)
 387 {
 388         struct swap_info_struct * p;
 389         struct page *page = NULL;
 390
 391         p = swap_info_get(entry);
 392         if (p) {
 393                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 394                         spin_lock_irq(&swapper_space.tree_lock);
 395                         page = radix_tree_lookup(&swapper_space.page_tree,
 396                                 entry.val);
 397                         if (page && TestSetPageLocked(page))
 398                                 page = NULL;
 399                         spin_unlock_irq(&swapper_space.tree_lock);
 400                 }
 401                 swap_info_put(p);
 402         }
 403         if (page) {
 404                 int one_user;
 405
 406                 BUG_ON(PagePrivate(page));
 407                 page_cache_get(page);
 408                 one_user = (page_count(page) == 2);
 409                 /* Only cache user (+us), or swap space full? Free it! */
 410                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
 411                         delete_from_swap_cache(page);
 412                         SetPageDirty(page);
 413                 }
 414                 unlock_page(page);
 415                 page_cache_release(page);
 416         }
 417 }
 418
 419 /*
 420  * The swap entry has been read in advance, and we return 1 to indicate
 421  * that the page has been used or is no longer needed.
 422  *
 423  * Always set the resulting pte to be nowrite (the same as COW pages
 424  * after one process has exited).  We don't know just how many PTEs will
 425  * share this swap entry, so be cautious and let do_wp_page work out
 426  * what to do if a write is requested later.
 427  */
 428 /* vma->vm_mm->page_table_lock is held */
 429 static void
 430 unuse_pte(struct vm_area_struct *vma, unsigned long address, pte_t *dir,
 431         swp_entry_t entry, struct page *page)
 432 {
 433         vma->vm_mm->rss++;
 434         get_page(page);
 435         set_pte(dir, pte_mkold(mk_pte(page, vma->vm_page_prot)));
 436         page_add_anon_rmap(page, vma, address);
 437         swap_free(entry);
 438 }
 439
 440 /* vma->vm_mm->page_table_lock is held */
 441 static unsigned long unuse_pmd(struct vm_area_struct * vma, pmd_t *dir,
 442         unsigned long address, unsigned long size, unsigned long offset,
 443         swp_entry_t entry, struct page *page)
 444 {
 445         pte_t * pte;
 446         unsigned long end;
 447         pte_t swp_pte = swp_entry_to_pte(entry);
 448
 449         if (pmd_none(*dir))
 450                 return 0;
 451         if (pmd_bad(*dir)) {
 452                 pmd_ERROR(*dir);
 453                 pmd_clear(dir);
 454                 return 0;
 455         }
 456         pte = pte_offset_map(dir, address);
 457         offset += address & PMD_MASK;
 458         address &= ~PMD_MASK;
 459         end = address + size;
 460         if (end > PMD_SIZE)
 461                 end = PMD_SIZE;
 462         do {
 463                 /*
 464                  * swapoff spends a _lot_ of time in this loop!
 465                  * Test inline before going to call unuse_pte.
 466                  */
 467                 if (unlikely(pte_same(*pte, swp_pte))) {
 468                         unuse_pte(vma, offset + address, pte, entry, page);
 469                         pte_unmap(pte);
 470                         /* add 1 since address may be 0 */
 471                         return 1 + offset + address;
 472                 }
 473                 address += PAGE_SIZE;
 474                 pte++;
 475         } while (address && (address < end));
 476         pte_unmap(pte - 1);
 477         return 0;
 478 }
 479
 480 /* vma->vm_mm->page_table_lock is held */
 481 static unsigned long unuse_pgd(struct vm_area_struct * vma, pgd_t *dir,
 482         unsigned long address, unsigned long size,
 483         swp_entry_t entry, struct page *page)
 484 {
 485         pmd_t * pmd;
 486         unsigned long offset, end;
 487         unsigned long foundaddr;
 488
 489         if (pgd_none(*dir))
 490                 return 0;
 491         if (pgd_bad(*dir)) {
 492                 pgd_ERROR(*dir);
 493                 pgd_clear(dir);
 494                 return 0;
 495         }
 496         pmd = pmd_offset(dir, address);
 497         offset = address & PGDIR_MASK;
 498         address &= ~PGDIR_MASK;
 499         end = address + size;
 500         if (end > PGDIR_SIZE)
 501                 end = PGDIR_SIZE;
 502         if (address >= end)
 503                 BUG();
 504         do {
 505                 foundaddr = unuse_pmd(vma, pmd, address, end - address,
 506                                                 offset, entry, page);
 507                 if (foundaddr)
 508                         return foundaddr;
 509                 address = (address + PMD_SIZE) & PMD_MASK;
 510                 pmd++;
 511         } while (address && (address < end));
 512         return 0;
 513 }
 514
 515 /* vma->vm_mm->page_table_lock is held */
 516 static unsigned long unuse_vma(struct vm_area_struct * vma, pgd_t *pgdir,
 517         swp_entry_t entry, struct page *page)
 518 {
 519         unsigned long start = vma->vm_start, end = vma->vm_end;
 520         unsigned long foundaddr;
 521
 522         if (start >= end)
 523                 BUG();
 524         do {
 525                 foundaddr = unuse_pgd(vma, pgdir, start, end - start,
 526                                                 entry, page);
 527                 if (foundaddr)
 528                         return foundaddr;
 529                 start = (start + PGDIR_SIZE) & PGDIR_MASK;
 530                 pgdir++;
 531         } while (start && (start < end));
 532         return 0;
 533 }
 534
 535 static int unuse_process(struct mm_struct * mm,
 536                         swp_entry_t entry, struct page* page)
 537 {
 538         struct vm_area_struct* vma;
 539         unsigned long foundaddr = 0;
 540
 541         /*
 542          * Go through process' page directory.
 543          */
 544         down_read(&mm->mmap_sem);
 545         spin_lock(&mm->page_table_lock);
 546         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 547                 if (!is_vm_hugetlb_page(vma)) {
 548                         pgd_t * pgd = pgd_offset(mm, vma->vm_start);
 549                         foundaddr = unuse_vma(vma, pgd, entry, page);
 550                         if (foundaddr)
 551                                 break;
 552                 }
 553         }
 554         spin_unlock(&mm->page_table_lock);
 555         up_read(&mm->mmap_sem);
 556         /*
 557          * Currently unuse_process cannot fail, but leave error handling
 558          * at call sites for now, since we change it from time to time.
 559          */
 560         return 0;
 561 }
 562
 563 /*
 564  * Scan swap_map from current position to next entry still in use.
 565  * Recycle to start on reaching the end, returning 0 when empty.
 566  */
 567 static int find_next_to_unuse(struct swap_info_struct *si, int prev)
 568 {
 569         int max = si->max;
 570         int i = prev;
 571         int count;
 572
 573         /*
 574          * No need for swap_device_lock(si) here: we're just looking
 575          * for whether an entry is in use, not modifying it; false
 576          * hits are okay, and sys_swapoff() has already prevented new
 577          * allocations from this area (while holding swap_list_lock()).
 578          */
 579         for (;;) {
 580                 if (++i >= max) {
 581                         if (!prev) {
 582                                 i = 0;
 583                                 break;
 584                         }
 585                         /*
 586                          * No entries in use at top of swap_map,
 587                          * loop back to start and recheck there.
 588                          */
 589                         max = prev + 1;
 590                         prev = 0;
 591                         i = 1;
 592                 }
 593                 count = si->swap_map[i];
 594                 if (count && count != SWAP_MAP_BAD)
 595                         break;
 596         }
 597         return i;
 598 }
 599
 600 /*
 601  * We completely avoid races by reading each swap page in advance,
 602  * and then search for the process using it.  All the necessary
 603  * page table adjustments can then be made atomically.
 604  */
 605 static int try_to_unuse(unsigned int type)
 606 {
 607         struct swap_info_struct * si = &swap_info[type];
 608         struct mm_struct *start_mm;
 609         unsigned short *swap_map;
 610         unsigned short swcount;
 611         struct page *page;
 612         swp_entry_t entry;
 613         int i = 0;
 614         int retval = 0;
 615         int reset_overflow = 0;
 616         int shmem;
 617
 618         /*
 619          * When searching mms for an entry, a good strategy is to
 620          * start at the first mm we freed the previous entry from
 621          * (though actually we don't notice whether we or coincidence
 622          * freed the entry).  Initialize this start_mm with a hold.
 623          *
 624          * A simpler strategy would be to start at the last mm we
 625          * freed the previous entry from; but that would take less
 626          * advantage of mmlist ordering (now preserved by swap_out()),
 627          * which clusters forked address spaces together, most recent
 628          * child immediately after parent.  If we race with dup_mmap(),
 629          * we very much want to resolve parent before child, otherwise
 630          * we may miss some entries: using last mm would invert that.
 631          */
 632         start_mm = &init_mm;
 633         atomic_inc(&init_mm.mm_users);
 634
 635         /*
 636          * Keep on scanning until all entries have gone.  Usually,
 637          * one pass through swap_map is enough, but not necessarily:
 638          * mmput() removes mm from mmlist before exit_mmap() and its
 639          * zap_page_range().  That's not too bad, those entries are
 640          * on their way out, and handled faster there than here.
 641          * do_munmap() behaves similarly, taking the range out of mm's
 642          * vma list before zap_page_range().  But unfortunately, when
 643          * unmapping a part of a vma, it takes the whole out first,
 644          * then reinserts what's left after (might even reschedule if
 645          * open() method called) - so swap entries may be invisible
 646          * to swapoff for a while, then reappear - but that is rare.
 647          */
 648         while ((i = find_next_to_unuse(si, i))) {
 649                 if (signal_pending(current)) {
 650                         retval = -EINTR;
 651                         break;
 652                 }
 653
 654                 /*
 655                  * Get a page for the entry, using the existing swap
 656                  * cache page if there is one.  Otherwise, get a clean
 657                  * page and read the swap into it.
 658                  */
 659                 swap_map = &si->swap_map[i];
 660                 entry = swp_entry(type, i);
 661                 page = read_swap_cache_async(entry, NULL, 0);
 662                 if (!page) {
 663                         /*
 664                          * Either swap_duplicate() failed because entry
 665                          * has been freed independently, and will not be
 666                          * reused since sys_swapoff() already disabled
 667                          * allocation from here, or alloc_page() failed.
 668                          */
 669                         if (!*swap_map)
 670                                 continue;
 671                         retval = -ENOMEM;
 672                         break;
 673                 }
 674
 675                 /*
 676                  * Don't hold on to start_mm if it looks like exiting.
 677                  */
 678                 if (atomic_read(&start_mm->mm_users) == 1) {
 679                         mmput(start_mm);
 680                         start_mm = &init_mm;
 681                         atomic_inc(&init_mm.mm_users);
 682                 }
 683
 684                 /*
 685                  * Wait for and lock page.  When do_swap_page races with
 686                  * try_to_unuse, do_swap_page can handle the fault much
 687                  * faster than try_to_unuse can locate the entry.  This
 688                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 689                  * defer to do_swap_page in such a case - in some tests,
 690                  * do_swap_page and try_to_unuse repeatedly compete.
 691                  */
 692                 wait_on_page_locked(page);
 693                 wait_on_page_writeback(page);
 694                 lock_page(page);
 695                 wait_on_page_writeback(page);
 696
 697                 /*
 698                  * Remove all references to entry, without blocking.
 699                  * Whenever we reach init_mm, there's no address space
 700                  * to search, but use it as a reminder to search shmem.
 701                  */
 702                 shmem = 0;
 703                 swcount = *swap_map;
 704                 if (swcount > 1) {
 705                         if (start_mm == &init_mm)
 706                                 shmem = shmem_unuse(entry, page);
 707                         else
 708                                 retval = unuse_process(start_mm, entry, page);
 709                 }
 710                 if (*swap_map > 1) {
 711                         int set_start_mm = (*swap_map >= swcount);
 712                         struct list_head *p = &start_mm->mmlist;
 713                         struct mm_struct *new_start_mm = start_mm;
 714                         struct mm_struct *prev_mm = start_mm;
 715                         struct mm_struct *mm;
 716
 717                         atomic_inc(&new_start_mm->mm_users);
 718                         atomic_inc(&prev_mm->mm_users);
 719                         spin_lock(&mmlist_lock);
 720                         while (*swap_map > 1 && !retval &&
 721                                         (p = p->next) != &start_mm->mmlist) {
 722                                 mm = list_entry(p, struct mm_struct, mmlist);
 723                                 atomic_inc(&mm->mm_users);
 724                                 spin_unlock(&mmlist_lock);
 725                                 mmput(prev_mm);
 726                                 prev_mm = mm;
 727
 728                                 cond_resched();
 729
 730                                 swcount = *swap_map;
 731                                 if (swcount <= 1)
 732                                         ;
 733                                 else if (mm == &init_mm) {
 734                                         set_start_mm = 1;
 735                                         shmem = shmem_unuse(entry, page);
 736                                 } else
 737                                         retval = unuse_process(mm, entry, page);
 738                                 if (set_start_mm && *swap_map < swcount) {
 739                                         mmput(new_start_mm);
 740                                         atomic_inc(&mm->mm_users);
 741                                         new_start_mm = mm;
 742                                         set_start_mm = 0;
 743                                 }
 744                                 spin_lock(&mmlist_lock);
 745                         }
 746                         spin_unlock(&mmlist_lock);
 747                         mmput(prev_mm);
 748                         mmput(start_mm);
 749                         start_mm = new_start_mm;
 750                 }
 751                 if (retval) {
 752                         unlock_page(page);
 753                         page_cache_release(page);
 754                         break;
 755                 }
 756
 757                 /*
 758                  * How could swap count reach 0x7fff when the maximum
 759                  * pid is 0x7fff, and there's no way to repeat a swap
 760                  * page within an mm (except in shmem, where it's the
 761                  * shared object which takes the reference count)?
 762                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 763                  *
 764                  * If that's wrong, then we should worry more about
 765                  * exit_mmap() and do_munmap() cases described above:
 766                  * we might be resetting SWAP_MAP_MAX too early here.
 767                  * We know "Undead"s can happen, they're okay, so don't
 768                  * report them; but do report if we reset SWAP_MAP_MAX.
 769                  */
 770                 if (*swap_map == SWAP_MAP_MAX) {
 771                         swap_device_lock(si);
 772                         *swap_map = 1;
 773                         swap_device_unlock(si);
 774                         reset_overflow = 1;
 775                 }
 776
 777                 /*
 778                  * If a reference remains (rare), we would like to leave
 779                  * the page in the swap cache; but try_to_unmap could
 780                  * then re-duplicate the entry once we drop page lock,
 781                  * so we might loop indefinitely; also, that page could
 782                  * not be swapped out to other storage meanwhile.  So:
 783                  * delete from cache even if there's another reference,
 784                  * after ensuring that the data has been saved to disk -
 785                  * since if the reference remains (rarer), it will be
 786                  * read from disk into another page.  Splitting into two
 787                  * pages would be incorrect if swap supported "shared
 788                  * private" pages, but they are handled by tmpfs files.
 789                  *
 790                  * Note shmem_unuse already deleted a swappage from
 791                  * the swap cache, unless the move to filepage failed:
 792                  * in which case it left swappage in cache, lowered its
 793                  * swap count to pass quickly through the loops above,
 794                  * and now we must reincrement count to try again later.
 795                  */
 796                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 797                         struct writeback_control wbc = {
 798                                 .sync_mode = WB_SYNC_NONE,
 799                         };
 800
 801                         swap_writepage(page, &wbc);
 802                         lock_page(page);
 803                         wait_on_page_writeback(page);
 804                 }
 805                 if (PageSwapCache(page)) {
 806                         if (shmem)
 807                                 swap_duplicate(entry);
 808                         else
 809                                 delete_from_swap_cache(page);
 810                 }
 811
 812                 /*
 813                  * So we could skip searching mms once swap count went
 814                  * to 1, we did not mark any present ptes as dirty: must
 815                  * mark page dirty so shrink_list will preserve it.
 816                  */
 817                 SetPageDirty(page);
 818                 unlock_page(page);
 819                 page_cache_release(page);
 820
 821                 /*
 822                  * Make sure that we aren't completely killing
 823                  * interactive performance.
 824                  */
 825                 cond_resched();
 826         }
 827
 828         mmput(start_mm);
 829         if (reset_overflow) {
 830                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 831                 swap_overflow = 0;
 832         }
 833         return retval;
 834 }
 835
 836 /*
 837  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 838  * corresponds to page offset `offset'.
 839  */
 840 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 841 {
 842         struct swap_extent *se = sis->curr_swap_extent;
 843         struct swap_extent *start_se = se;
 844
 845         for ( ; ; ) {
 846                 struct list_head *lh;
 847
 848                 if (se->start_page <= offset &&
 849                                 offset < (se->start_page + se->nr_pages)) {
 850                         return se->start_block + (offset - se->start_page);
 851                 }
 852                 lh = se->list.prev;
 853                 if (lh == &sis->extent_list)
 854                         lh = lh->prev;
 855                 se = list_entry(lh, struct swap_extent, list);
 856                 sis->curr_swap_extent = se;
 857                 BUG_ON(se == start_se);         /* It *must* be present */
 858         }
 859 }
 860
 861 /*
 862  * Free all of a swapdev's extent information
 863  */
 864 static void destroy_swap_extents(struct swap_info_struct *sis)
 865 {
 866         while (!list_empty(&sis->extent_list)) {
 867                 struct swap_extent *se;
 868
 869                 se = list_entry(sis->extent_list.next,
 870                                 struct swap_extent, list);
 871                 list_del(&se->list);
 872                 kfree(se);
 873         }
 874         sis->nr_extents = 0;
 875 }
 876
 877 /*
 878  * Add a block range (and the corresponding page range) into this swapdev's
 879  * extent list.  The extent list is kept sorted in block order.
 880  *
 881  * This function rather assumes that it is called in ascending sector_t order.
 882  * It doesn't look for extent coalescing opportunities.
 883  */
 884 static int
 885 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 886                 unsigned long nr_pages, sector_t start_block)
 887 {
 888         struct swap_extent *se;
 889         struct swap_extent *new_se;
 890         struct list_head *lh;
 891
 892         lh = sis->extent_list.next;     /* The highest-addressed block */
 893         while (lh != &sis->extent_list) {
 894                 se = list_entry(lh, struct swap_extent, list);
 895                 if (se->start_block + se->nr_pages == start_block &&
 896                     se->start_page  + se->nr_pages == start_page) {
 897                         /* Merge it */
 898                         se->nr_pages += nr_pages;
 899                         return 0;
 900                 }
 901                 lh = lh->next;
 902         }
 903
 904         /*
 905          * No merge.  Insert a new extent, preserving ordering.
 906          */
 907         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 908         if (new_se == NULL)
 909                 return -ENOMEM;
 910         new_se->start_page = start_page;
 911         new_se->nr_pages = nr_pages;
 912         new_se->start_block = start_block;
 913
 914         lh = sis->extent_list.prev;     /* The lowest block */
 915         while (lh != &sis->extent_list) {
 916                 se = list_entry(lh, struct swap_extent, list);
 917                 if (se->start_block > start_block)
 918                         break;
 919                 lh = lh->prev;
 920         }
 921         list_add_tail(&new_se->list, lh);
 922         sis->nr_extents++;
 923         return 0;
 924 }
 925
 926 /*
 927  * A `swap extent' is a simple thing which maps a contiguous range of pages
 928  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 929  * is built at swapon time and is then used at swap_writepage/swap_readpage
 930  * time for locating where on disk a page belongs.
 931  *
 932  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 933  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 934  * swap files identically.
 935  *
 936  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
 937  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
 938  * swapfiles are handled *identically* after swapon time.
 939  *
 940  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
 941  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
 942  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
 943  * requirements, they are simply tossed out - we will never use those blocks
 944  * for swapping.
 945  *
 946  * For S_ISREG swapfiles we hold i_sem across the life of the swapon.  This
 947  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
 948  * which will scribble on the fs.
 949  *
 950  * The amount of disk space which a single swap extent represents varies.
 951  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
 952  * extents in the list.  To avoid much list walking, we cache the previous
 953  * search location in `curr_swap_extent', and start new searches from there.
 954  * This is extremely effective.  The average number of iterations in
 955  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
 956  */
 957 static int setup_swap_extents(struct swap_info_struct *sis)
 958 {
 959         struct inode *inode;
 960         unsigned blocks_per_page;
 961         unsigned long page_no;
 962         unsigned blkbits;
 963         sector_t probe_block;
 964         sector_t last_block;
 965         int ret;
 966
 967         inode = sis->swap_file->f_mapping->host;
 968         if (S_ISBLK(inode->i_mode)) {
 969                 ret = add_swap_extent(sis, 0, sis->max, 0);
 970                 goto done;
 971         }
 972
 973         blkbits = inode->i_blkbits;
 974         blocks_per_page = PAGE_SIZE >> blkbits;
 975
 976         /*
 977          * Map all the blocks into the extent list.  This code doesn't try
 978          * to be very smart.
 979          */
 980         probe_block = 0;
 981         page_no = 0;
 982         last_block = i_size_read(inode) >> blkbits;
 983         while ((probe_block + blocks_per_page) <= last_block &&
 984                         page_no < sis->max) {
 985                 unsigned block_in_page;
 986                 sector_t first_block;
 987
 988                 first_block = bmap(inode, probe_block);
 989                 if (first_block == 0)
 990                         goto bad_bmap;
 991
 992                 /*
 993                  * It must be PAGE_SIZE aligned on-disk
 994                  */
 995                 if (first_block & (blocks_per_page - 1)) {
 996                         probe_block++;
 997                         goto reprobe;
 998                 }
 999
1000                 for (block_in_page = 1; block_in_page < blocks_per_page;
1001                                         block_in_page++) {
1002                         sector_t block;
1003
1004                         block = bmap(inode, probe_block + block_in_page);
1005                         if (block == 0)
1006                                 goto bad_bmap;
1007                         if (block != first_block + block_in_page) {
1008                                 /* Discontiguity */
1009                                 probe_block++;
1010                                 goto reprobe;
1011                         }
1012                 }
1013
1014                 /*
1015                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1016                  */
1017                 ret = add_swap_extent(sis, page_no, 1,
1018                                 first_block >> (PAGE_SHIFT - blkbits));
1019                 if (ret)
1020                         goto out;
1021                 page_no++;
1022                 probe_block += blocks_per_page;
1023 reprobe:
1024                 continue;
1025         }
1026         ret = 0;
1027         if (page_no == 0)
1028                 ret = -EINVAL;
1029         sis->max = page_no;
1030         sis->highest_bit = page_no - 1;
1031 done:
1032         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1033                                         struct swap_extent, list);
1034         goto out;
1035 bad_bmap:
1036         printk(KERN_ERR "swapon: swapfile has holes\n");
1037         ret = -EINVAL;
1038 out:
1039         return ret;
1040 }
1041
1042 #if 0   /* We don't need this yet */
1043 #include <linux/backing-dev.h>
1044 int page_queue_congested(struct page *page)
1045 {
1046         struct backing_dev_info *bdi;
1047
1048         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1049
1050         if (PageSwapCache(page)) {
1051                 swp_entry_t entry = { .val = page->private };
1052                 struct swap_info_struct *sis;
1053
1054                 sis = get_swap_info_struct(swp_type(entry));
1055                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1056         } else
1057                 bdi = page->mapping->backing_dev_info;
1058         return bdi_write_congested(bdi);
1059 }
1060 #endif
1061
1062 asmlinkage long sys_swapoff(const char __user * specialfile)
1063 {
1064         struct swap_info_struct * p = NULL;
1065         unsigned short *swap_map;
1066         struct file *swap_file, *victim;
1067         struct address_space *mapping;
1068         char * pathname;
1069         int i, type, prev;
1070         int err;
1071
1072         if (!capable(CAP_SYS_ADMIN))
1073                 return -EPERM;
1074
1075         pathname = getname(specialfile);
1076         err = PTR_ERR(pathname);
1077         if (IS_ERR(pathname))
1078                 goto out;
1079
1080         victim = filp_open(pathname, O_RDWR, 0);
1081         putname(pathname);
1082         err = PTR_ERR(victim);
1083         if (IS_ERR(victim))
1084                 goto out;
1085
1086         mapping = victim->f_mapping;
1087         prev = -1;
1088         swap_list_lock();
1089         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1090                 p = swap_info + type;
1091                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1092                         if (p->swap_file->f_mapping == mapping)
1093                                 break;
1094                 }
1095                 prev = type;
1096         }
1097         if (type < 0) {
1098                 err = -EINVAL;
1099                 swap_list_unlock();
1100                 goto out_dput;
1101         }
1102         if (!security_vm_enough_memory(p->pages))
1103                 vm_unacct_memory(p->pages);
1104         else {
1105                 err = -ENOMEM;
1106                 swap_list_unlock();
1107                 goto out_dput;
1108         }
1109         if (prev < 0) {
1110                 swap_list.head = p->next;
1111         } else {
1112                 swap_info[prev].next = p->next;
1113         }
1114         if (type == swap_list.next) {
1115                 /* just pick something that's safe... */
1116                 swap_list.next = swap_list.head;
1117         }
1118         nr_swap_pages -= p->pages;
1119         total_swap_pages -= p->pages;
1120         p->flags &= ~SWP_WRITEOK;
1121         swap_list_unlock();
1122         current->flags |= PF_SWAPOFF;
1123         err = try_to_unuse(type);
1124         current->flags &= ~PF_SWAPOFF;
1125
1126         /* wait for any unplug function to finish */
1127         down_write(&swap_unplug_sem);
1128         up_write(&swap_unplug_sem);
1129
1130         if (err) {
1131                 /* re-insert swap space back into swap_list */
1132                 swap_list_lock();
1133                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1134                         if (p->prio >= swap_info[i].prio)
1135                                 break;
1136                 p->next = i;
1137                 if (prev < 0)
1138                         swap_list.head = swap_list.next = p - swap_info;
1139                 else
1140                         swap_info[prev].next = p - swap_info;
1141                 nr_swap_pages += p->pages;
1142                 total_swap_pages += p->pages;
1143                 p->flags |= SWP_WRITEOK;
1144                 swap_list_unlock();
1145                 goto out_dput;
1146         }
1147         down(&swapon_sem);
1148         swap_list_lock();
1149         swap_device_lock(p);
1150         swap_file = p->swap_file;
1151         p->swap_file = NULL;
1152         p->max = 0;
1153         swap_map = p->swap_map;
1154         p->swap_map = NULL;
1155         p->flags = 0;
1156         destroy_swap_extents(p);
1157         swap_device_unlock(p);
1158         swap_list_unlock();
1159         up(&swapon_sem);
1160         vfree(swap_map);
1161         if (S_ISBLK(mapping->host->i_mode)) {
1162                 struct block_device *bdev = I_BDEV(mapping->host);
1163                 set_blocksize(bdev, p->old_block_size);
1164                 bd_release(bdev);
1165         } else {
1166                 up(&mapping->host->i_sem);
1167         }
1168         filp_close(swap_file, NULL);
1169         err = 0;
1170
1171 out_dput:
1172         filp_close(victim, NULL);
1173 out:
1174         return err;
1175 }
1176
1177 #ifdef CONFIG_PROC_FS
1178 /* iterator */
1179 static void *swap_start(struct seq_file *swap, loff_t *pos)
1180 {
1181         struct swap_info_struct *ptr = swap_info;
1182         int i;
1183         loff_t l = *pos;
1184
1185         down(&swapon_sem);
1186
1187         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1188                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1189                         continue;
1190                 if (!l--)
1191                         return ptr;
1192         }
1193
1194         return NULL;
1195 }
1196
1197 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1198 {
1199         struct swap_info_struct *ptr = v;
1200         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1201
1202         for (++ptr; ptr < endptr; ptr++) {
1203                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1204                         continue;
1205                 ++*pos;
1206                 return ptr;
1207         }
1208
1209         return NULL;
1210 }
1211
1212 static void swap_stop(struct seq_file *swap, void *v)
1213 {
1214         up(&swapon_sem);
1215 }
1216
1217 static int swap_show(struct seq_file *swap, void *v)
1218 {
1219         struct swap_info_struct *ptr = v;
1220         struct file *file;
1221         int len;
1222
1223         if (v == swap_info)
1224                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1225
1226         file = ptr->swap_file;
1227         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1228         seq_printf(swap, "%*s%s\t%d\t%ld\t%d\n",
1229                        len < 40 ? 40 - len : 1, " ",
1230                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1231                                 "partition" : "file\t",
1232                        ptr->pages << (PAGE_SHIFT - 10),
1233                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1234                        ptr->prio);
1235         return 0;
1236 }
1237
1238 static struct seq_operations swaps_op = {
1239         .start =        swap_start,
1240         .next =         swap_next,
1241         .stop =         swap_stop,
1242         .show =         swap_show
1243 };
1244
1245 static int swaps_open(struct inode *inode, struct file *file)
1246 {
1247         return seq_open(file, &swaps_op);
1248 }
1249
1250 static struct file_operations proc_swaps_operations = {
1251         .open           = swaps_open,
1252         .read           = seq_read,
1253         .llseek         = seq_lseek,
1254         .release        = seq_release,
1255 };
1256
1257 static int __init procswaps_init(void)
1258 {
1259         struct proc_dir_entry *entry;
1260
1261         entry = create_proc_entry("swaps", 0, NULL);
1262         if (entry)
1263                 entry->proc_fops = &proc_swaps_operations;
1264         return 0;
1265 }
1266 __initcall(procswaps_init);
1267 #endif /* CONFIG_PROC_FS */
1268
1269 /*
1270  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1271  *
1272  * The swapon system call
1273  */
1274 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1275 {
1276         struct swap_info_struct * p;
1277         char *name = NULL;
1278         struct block_device *bdev = NULL;
1279         struct file *swap_file = NULL;
1280         struct address_space *mapping;
1281         unsigned int type;
1282         int i, prev;
1283         int error;
1284         static int least_priority;
1285         union swap_header *swap_header = 0;
1286         int swap_header_version;
1287         int nr_good_pages = 0;
1288         unsigned long maxpages = 1;
1289         int swapfilesize;
1290         unsigned short *swap_map;
1291         struct page *page = NULL;
1292         struct inode *inode = NULL;
1293         int did_down = 0;
1294
1295         if (!capable(CAP_SYS_ADMIN))
1296                 return -EPERM;
1297         swap_list_lock();
1298         p = swap_info;
1299         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1300                 if (!(p->flags & SWP_USED))
1301                         break;
1302         error = -EPERM;
1303         /*
1304          * Test if adding another swap device is possible. There are
1305          * two limiting factors: 1) the number of bits for the swap
1306          * type swp_entry_t definition and 2) the number of bits for
1307          * the swap type in the swap ptes as defined by the different
1308          * architectures. To honor both limitations a swap entry
1309          * with swap offset 0 and swap type ~0UL is created, encoded
1310          * to a swap pte, decoded to a swp_entry_t again and finally
1311          * the swap type part is extracted. This will mask all bits
1312          * from the initial ~0UL that can't be encoded in either the
1313          * swp_entry_t or the architecture definition of a swap pte.
1314          */
1315         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1316                 swap_list_unlock();
1317                 goto out;
1318         }
1319         if (type >= nr_swapfiles)
1320                 nr_swapfiles = type+1;
1321         INIT_LIST_HEAD(&p->extent_list);
1322         p->flags = SWP_USED;
1323         p->nr_extents = 0;
1324         p->swap_file = NULL;
1325         p->old_block_size = 0;
1326         p->swap_map = NULL;
1327         p->lowest_bit = 0;
1328         p->highest_bit = 0;
1329         p->cluster_nr = 0;
1330         p->inuse_pages = 0;
1331         p->sdev_lock = SPIN_LOCK_UNLOCKED;
1332         p->next = -1;
1333         if (swap_flags & SWAP_FLAG_PREFER) {
1334                 p->prio =
1335                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1336         } else {
1337                 p->prio = --least_priority;
1338         }
1339         swap_list_unlock();
1340         name = getname(specialfile);
1341         error = PTR_ERR(name);
1342         if (IS_ERR(name)) {
1343                 name = NULL;
1344                 goto bad_swap_2;
1345         }
1346         swap_file = filp_open(name, O_RDWR, 0);
1347         error = PTR_ERR(swap_file);
1348         if (IS_ERR(swap_file)) {
1349                 swap_file = NULL;
1350                 goto bad_swap_2;
1351         }
1352
1353         p->swap_file = swap_file;
1354         mapping = swap_file->f_mapping;
1355         inode = mapping->host;
1356
1357         error = -EBUSY;
1358         for (i = 0; i < nr_swapfiles; i++) {
1359                 struct swap_info_struct *q = &swap_info[i];
1360
1361                 if (i == type || !q->swap_file)
1362                         continue;
1363                 if (mapping == q->swap_file->f_mapping)
1364                         goto bad_swap;
1365         }
1366
1367         error = -EINVAL;
1368         if (S_ISBLK(inode->i_mode)) {
1369                 bdev = I_BDEV(inode);
1370                 error = bd_claim(bdev, sys_swapon);
1371                 if (error < 0) {
1372                         bdev = NULL;
1373                         goto bad_swap;
1374                 }
1375                 p->old_block_size = block_size(bdev);
1376                 error = set_blocksize(bdev, PAGE_SIZE);
1377                 if (error < 0)
1378                         goto bad_swap;
1379                 p->bdev = bdev;
1380         } else if (S_ISREG(inode->i_mode)) {
1381                 p->bdev = inode->i_sb->s_bdev;
1382                 down(&inode->i_sem);
1383                 did_down = 1;
1384         } else {
1385                 goto bad_swap;
1386         }
1387
1388         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1389
1390         /*
1391          * Read the swap header.
1392          */
1393         if (!mapping->a_ops->readpage) {
1394                 error = -EINVAL;
1395                 goto bad_swap;
1396         }
1397         page = read_cache_page(mapping, 0,
1398                         (filler_t *)mapping->a_ops->readpage, swap_file);
1399         if (IS_ERR(page)) {
1400                 error = PTR_ERR(page);
1401                 goto bad_swap;
1402         }
1403         wait_on_page_locked(page);
1404         if (!PageUptodate(page))
1405                 goto bad_swap;
1406         kmap(page);
1407         swap_header = page_address(page);
1408
1409         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1410                 swap_header_version = 1;
1411         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1412                 swap_header_version = 2;
1413         else {
1414                 printk("Unable to find swap-space signature\n");
1415                 error = -EINVAL;
1416                 goto bad_swap;
1417         }
1418
1419         switch (swap_header_version) {
1420         case 1:
1421                 printk(KERN_ERR "version 0 swap is no longer supported. "
1422                         "Use mkswap -v1 %s\n", name);
1423                 error = -EINVAL;
1424                 goto bad_swap;
1425         case 2:
1426                 /* Check the swap header's sub-version and the size of
1427                    the swap file and bad block lists */
1428                 if (swap_header->info.version != 1) {
1429                         printk(KERN_WARNING
1430                                "Unable to handle swap header version %d\n",
1431                                swap_header->info.version);
1432                         error = -EINVAL;
1433                         goto bad_swap;
1434                 }
1435
1436                 p->lowest_bit  = 1;
1437                 /*
1438                  * Find out how many pages are allowed for a single swap
1439                  * device. There are two limiting factors: 1) the number of
1440                  * bits for the swap offset in the swp_entry_t type and
1441                  * 2) the number of bits in the a swap pte as defined by
1442                  * the different architectures. In order to find the
1443                  * largest possible bit mask a swap entry with swap type 0
1444                  * and swap offset ~0UL is created, encoded to a swap pte,
1445                  * decoded to a swp_entry_t again and finally the swap
1446                  * offset is extracted. This will mask all the bits from
1447                  * the initial ~0UL mask that can't be encoded in either
1448                  * the swp_entry_t or the architecture definition of a
1449                  * swap pte.
1450                  */
1451                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1452                 if (maxpages > swap_header->info.last_page)
1453                         maxpages = swap_header->info.last_page;
1454                 p->highest_bit = maxpages - 1;
1455
1456                 error = -EINVAL;
1457                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1458                         goto bad_swap;
1459
1460                 /* OK, set up the swap map and apply the bad block list */
1461                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1462                         error = -ENOMEM;
1463                         goto bad_swap;
1464                 }
1465
1466                 error = 0;
1467                 memset(p->swap_map, 0, maxpages * sizeof(short));
1468                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1469                         int page = swap_header->info.badpages[i];
1470                         if (page <= 0 || page >= swap_header->info.last_page)
1471                                 error = -EINVAL;
1472                         else
1473                                 p->swap_map[page] = SWAP_MAP_BAD;
1474                 }
1475                 nr_good_pages = swap_header->info.last_page -
1476                                 swap_header->info.nr_badpages -
1477                                 1 /* header page */;
1478                 if (error)
1479                         goto bad_swap;
1480         }
1481
1482         if (swapfilesize && maxpages > swapfilesize) {
1483                 printk(KERN_WARNING
1484                        "Swap area shorter than signature indicates\n");
1485                 error = -EINVAL;
1486                 goto bad_swap;
1487         }
1488         if (!nr_good_pages) {
1489                 printk(KERN_WARNING "Empty swap-file\n");
1490                 error = -EINVAL;
1491                 goto bad_swap;
1492         }
1493         p->swap_map[0] = SWAP_MAP_BAD;
1494         p->max = maxpages;
1495         p->pages = nr_good_pages;
1496
1497         error = setup_swap_extents(p);
1498         if (error)
1499                 goto bad_swap;
1500
1501         down(&swapon_sem);
1502         swap_list_lock();
1503         swap_device_lock(p);
1504         p->flags = SWP_ACTIVE;
1505         nr_swap_pages += nr_good_pages;
1506         total_swap_pages += nr_good_pages;
1507         printk(KERN_INFO "Adding %dk swap on %s.  Priority:%d extents:%d\n",
1508                 nr_good_pages<<(PAGE_SHIFT-10), name,
1509                 p->prio, p->nr_extents);
1510
1511         /* insert swap space into swap_list: */
1512         prev = -1;
1513         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1514                 if (p->prio >= swap_info[i].prio) {
1515                         break;
1516                 }
1517                 prev = i;
1518         }
1519         p->next = i;
1520         if (prev < 0) {
1521                 swap_list.head = swap_list.next = p - swap_info;
1522         } else {
1523                 swap_info[prev].next = p - swap_info;
1524         }
1525         swap_device_unlock(p);
1526         swap_list_unlock();
1527         up(&swapon_sem);
1528         error = 0;
1529         goto out;
1530 bad_swap:
1531         if (bdev) {
1532                 set_blocksize(bdev, p->old_block_size);
1533                 bd_release(bdev);
1534         }
1535 bad_swap_2:
1536         swap_list_lock();
1537         swap_map = p->swap_map;
1538         p->swap_file = NULL;
1539         p->swap_map = NULL;
1540         p->flags = 0;
1541         if (!(swap_flags & SWAP_FLAG_PREFER))
1542                 ++least_priority;
1543         swap_list_unlock();
1544         destroy_swap_extents(p);
1545         if (swap_map)
1546                 vfree(swap_map);
1547         if (swap_file)
1548                 filp_close(swap_file, NULL);
1549 out:
1550         if (page && !IS_ERR(page)) {
1551                 kunmap(page);
1552                 page_cache_release(page);
1553         }
1554         if (name)
1555                 putname(name);
1556         if (error && did_down)
1557                 up(&inode->i_sem);
1558         return error;
1559 }
1560
1561 void si_swapinfo(struct sysinfo *val)
1562 {
1563         unsigned int i;
1564         unsigned long nr_to_be_unused = 0;
1565
1566         swap_list_lock();
1567         for (i = 0; i < nr_swapfiles; i++) {
1568                 if (!(swap_info[i].flags & SWP_USED) ||
1569                      (swap_info[i].flags & SWP_WRITEOK))
1570                         continue;
1571                 nr_to_be_unused += swap_info[i].inuse_pages;
1572         }
1573         val->freeswap = nr_swap_pages + nr_to_be_unused;
1574         val->totalswap = total_swap_pages + nr_to_be_unused;
1575         swap_list_unlock();
1576 }
1577
1578 /*
1579  * Verify that a swap entry is valid and increment its swap map count.
1580  *
1581  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1582  * "permanent", but will be reclaimed by the next swapoff.
1583  */
1584 int swap_duplicate(swp_entry_t entry)
1585 {
1586         struct swap_info_struct * p;
1587         unsigned long offset, type;
1588         int result = 0;
1589
1590         type = swp_type(entry);
1591         if (type >= nr_swapfiles)
1592                 goto bad_file;
1593         p = type + swap_info;
1594         offset = swp_offset(entry);
1595
1596         swap_device_lock(p);
1597         if (offset < p->max && p->swap_map[offset]) {
1598                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1599                         p->swap_map[offset]++;
1600                         result = 1;
1601                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1602                         if (swap_overflow++ < 5)
1603                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1604                         p->swap_map[offset] = SWAP_MAP_MAX;
1605                         result = 1;
1606                 }
1607         }
1608         swap_device_unlock(p);
1609 out:
1610         return result;
1611
1612 bad_file:
1613         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1614         goto out;
1615 }
1616
1617 struct swap_info_struct *
1618 get_swap_info_struct(unsigned type)
1619 {
1620         return &swap_info[type];
1621 }
1622
1623 /*
1624  * swap_device_lock prevents swap_map being freed. Don't grab an extra
1625  * reference on the swaphandle, it doesn't matter if it becomes unused.
1626  */
1627 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1628 {
1629         int ret = 0, i = 1 << page_cluster;
1630         unsigned long toff;
1631         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1632
1633         if (!page_cluster)      /* no readahead */
1634                 return 0;
1635         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1636         if (!toff)              /* first page is swap header */
1637                 toff++, i--;
1638         *offset = toff;
1639
1640         swap_device_lock(swapdev);
1641         do {
1642                 /* Don't read-ahead past the end of the swap area */
1643                 if (toff >= swapdev->max)
1644                         break;
1645                 /* Don't read in free or bad pages */
1646                 if (!swapdev->swap_map[toff])
1647                         break;
1648                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1649                         break;
1650                 toff++;
1651                 ret++;
1652         } while (--i);
1653         swap_device_unlock(swapdev);
1654         return ret;
1655 }