mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/config.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/mman.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/namei.h>
  18 #include <linux/shm.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28
  29 #include <asm/pgtable.h>
  30 #include <asm/tlbflush.h>
  31 #include <linux/swapops.h>
  32 #include <linux/vs_base.h>
  33 #include <linux/vs_memory.h>
  34
  35 spinlock_t swaplock = SPIN_LOCK_UNLOCKED;
  36 unsigned int nr_swapfiles;
  37 long total_swap_pages;
  38 static int swap_overflow;
  39
  40 EXPORT_SYMBOL(total_swap_pages);
  41
  42 static const char Bad_file[] = "Bad swap file entry ";
  43 static const char Unused_file[] = "Unused swap file entry ";
  44 static const char Bad_offset[] = "Bad swap offset entry ";
  45 static const char Unused_offset[] = "Unused swap offset entry ";
  46
  47 struct swap_list_t swap_list = {-1, -1};
  48
  49 struct swap_info_struct swap_info[MAX_SWAPFILES];
  50
  51 static DECLARE_MUTEX(swapon_sem);
  52
  53 /*
  54  * We need this because the bdev->unplug_fn can sleep and we cannot
  55  * hold swap_list_lock while calling the unplug_fn. And swap_list_lock
  56  * cannot be turned into a semaphore.
  57  */
  58 static DECLARE_RWSEM(swap_unplug_sem);
  59
  60 #define SWAPFILE_CLUSTER 256
  61
  62 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  63 {
  64         swp_entry_t entry;
  65
  66         down_read(&swap_unplug_sem);
  67         entry.val = page->private;
  68         if (PageSwapCache(page)) {
  69                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  70                 struct backing_dev_info *bdi;
  71
  72                 /*
  73                  * If the page is removed from swapcache from under us (with a
  74                  * racy try_to_unuse/swapoff) we need an additional reference
  75                  * count to avoid reading garbage from page->private above. If
  76                  * the WARN_ON triggers during a swapoff it maybe the race
  77                  * condition and it's harmless. However if it triggers without
  78                  * swapoff it signals a problem.
  79                  */
  80                 WARN_ON(page_count(page) <= 1);
  81
  82                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  83                 bdi->unplug_io_fn(bdi, page);
  84         }
  85         up_read(&swap_unplug_sem);
  86 }
  87
  88 static inline int scan_swap_map(struct swap_info_struct *si)
  89 {
  90         unsigned long offset;
  91         /*
  92          * We try to cluster swap pages by allocating them
  93          * sequentially in swap.  Once we've allocated
  94          * SWAPFILE_CLUSTER pages this way, however, we resort to
  95          * first-free allocation, starting a new cluster.  This
  96          * prevents us from scattering swap pages all over the entire
  97          * swap partition, so that we reduce overall disk seek times
  98          * between swap pages.  -- sct */
  99         if (si->cluster_nr) {
 100                 while (si->cluster_next <= si->highest_bit) {
 101                         offset = si->cluster_next++;
 102                         if (si->swap_map[offset])
 103                                 continue;
 104                         si->cluster_nr--;
 105                         goto got_page;
 106                 }
 107         }
 108         si->cluster_nr = SWAPFILE_CLUSTER;
 109
 110         /* try to find an empty (even not aligned) cluster. */
 111         offset = si->lowest_bit;
 112  check_next_cluster:
 113         if (offset+SWAPFILE_CLUSTER-1 <= si->highest_bit)
 114         {
 115                 unsigned long nr;
 116                 for (nr = offset; nr < offset+SWAPFILE_CLUSTER; nr++)
 117                         if (si->swap_map[nr])
 118                         {
 119                                 offset = nr+1;
 120                                 goto check_next_cluster;
 121                         }
 122                 /* We found a completly empty cluster, so start
 123                  * using it.
 124                  */
 125                 goto got_page;
 126         }
 127         /* No luck, so now go finegrined as usual. -Andrea */
 128         for (offset = si->lowest_bit; offset <= si->highest_bit ; offset++) {
 129                 if (si->swap_map[offset])
 130                         continue;
 131                 si->lowest_bit = offset+1;
 132         got_page:
 133                 if (offset == si->lowest_bit)
 134                         si->lowest_bit++;
 135                 if (offset == si->highest_bit)
 136                         si->highest_bit--;
 137                 if (si->lowest_bit > si->highest_bit) {
 138                         si->lowest_bit = si->max;
 139                         si->highest_bit = 0;
 140                 }
 141                 si->swap_map[offset] = 1;
 142                 si->inuse_pages++;
 143                 nr_swap_pages--;
 144                 si->cluster_next = offset+1;
 145                 return offset;
 146         }
 147         si->lowest_bit = si->max;
 148         si->highest_bit = 0;
 149         return 0;
 150 }
 151
 152 swp_entry_t get_swap_page(void)
 153 {
 154         struct swap_info_struct * p;
 155         unsigned long offset;
 156         swp_entry_t entry;
 157         int type, wrapped = 0;
 158
 159         entry.val = 0;  /* Out of memory */
 160         swap_list_lock();
 161         type = swap_list.next;
 162         if (type < 0)
 163                 goto out;
 164         if (nr_swap_pages <= 0)
 165                 goto out;
 166
 167         while (1) {
 168                 p = &swap_info[type];
 169                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
 170                         swap_device_lock(p);
 171                         offset = scan_swap_map(p);
 172                         swap_device_unlock(p);
 173                         if (offset) {
 174                                 entry = swp_entry(type,offset);
 175                                 type = swap_info[type].next;
 176                                 if (type < 0 ||
 177                                         p->prio != swap_info[type].prio) {
 178                                                 swap_list.next = swap_list.head;
 179                                 } else {
 180                                         swap_list.next = type;
 181                                 }
 182                                 goto out;
 183                         }
 184                 }
 185                 type = p->next;
 186                 if (!wrapped) {
 187                         if (type < 0 || p->prio != swap_info[type].prio) {
 188                                 type = swap_list.head;
 189                                 wrapped = 1;
 190                         }
 191                 } else
 192                         if (type < 0)
 193                                 goto out;       /* out of swap space */
 194         }
 195 out:
 196         swap_list_unlock();
 197         return entry;
 198 }
 199
 200 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 201 {
 202         struct swap_info_struct * p;
 203         unsigned long offset, type;
 204
 205         if (!entry.val)
 206                 goto out;
 207         type = swp_type(entry);
 208         if (type >= nr_swapfiles)
 209                 goto bad_nofile;
 210         p = & swap_info[type];
 211         if (!(p->flags & SWP_USED))
 212                 goto bad_device;
 213         offset = swp_offset(entry);
 214         if (offset >= p->max)
 215                 goto bad_offset;
 216         if (!p->swap_map[offset])
 217                 goto bad_free;
 218         swap_list_lock();
 219         if (p->prio > swap_info[swap_list.next].prio)
 220                 swap_list.next = type;
 221         swap_device_lock(p);
 222         return p;
 223
 224 bad_free:
 225         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 226         goto out;
 227 bad_offset:
 228         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 229         goto out;
 230 bad_device:
 231         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 232         goto out;
 233 bad_nofile:
 234         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 235 out:
 236         return NULL;
 237 }
 238
 239 static void swap_info_put(struct swap_info_struct * p)
 240 {
 241         swap_device_unlock(p);
 242         swap_list_unlock();
 243 }
 244
 245 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 246 {
 247         int count = p->swap_map[offset];
 248
 249         if (count < SWAP_MAP_MAX) {
 250                 count--;
 251                 p->swap_map[offset] = count;
 252                 if (!count) {
 253                         if (offset < p->lowest_bit)
 254                                 p->lowest_bit = offset;
 255                         if (offset > p->highest_bit)
 256                                 p->highest_bit = offset;
 257                         nr_swap_pages++;
 258                         p->inuse_pages--;
 259                 }
 260         }
 261         return count;
 262 }
 263
 264 /*
 265  * Caller has made sure that the swapdevice corresponding to entry
 266  * is still around or has not been recycled.
 267  */
 268 void swap_free(swp_entry_t entry)
 269 {
 270         struct swap_info_struct * p;
 271
 272         p = swap_info_get(entry);
 273         if (p) {
 274                 swap_entry_free(p, swp_offset(entry));
 275                 swap_info_put(p);
 276         }
 277 }
 278
 279 /*
 280  * Check if we're the only user of a swap page,
 281  * when the page is locked.
 282  */
 283 static int exclusive_swap_page(struct page *page)
 284 {
 285         int retval = 0;
 286         struct swap_info_struct * p;
 287         swp_entry_t entry;
 288
 289         entry.val = page->private;
 290         p = swap_info_get(entry);
 291         if (p) {
 292                 /* Is the only swap cache user the cache itself? */
 293                 if (p->swap_map[swp_offset(entry)] == 1) {
 294                         /* Recheck the page count with the swapcache lock held.. */
 295                         spin_lock_irq(&swapper_space.tree_lock);
 296                         if (page_count(page) == 2)
 297                                 retval = 1;
 298                         spin_unlock_irq(&swapper_space.tree_lock);
 299                 }
 300                 swap_info_put(p);
 301         }
 302         return retval;
 303 }
 304
 305 /*
 306  * We can use this swap cache entry directly
 307  * if there are no other references to it.
 308  *
 309  * Here "exclusive_swap_page()" does the real
 310  * work, but we opportunistically check whether
 311  * we need to get all the locks first..
 312  */
 313 int can_share_swap_page(struct page *page)
 314 {
 315         int retval = 0;
 316
 317         if (!PageLocked(page))
 318                 BUG();
 319         switch (page_count(page)) {
 320         case 3:
 321                 if (!PagePrivate(page))
 322                         break;
 323                 /* Fallthrough */
 324         case 2:
 325                 if (!PageSwapCache(page))
 326                         break;
 327                 retval = exclusive_swap_page(page);
 328                 break;
 329         case 1:
 330                 if (PageReserved(page))
 331                         break;
 332                 retval = 1;
 333         }
 334         return retval;
 335 }
 336
 337 /*
 338  * Work out if there are any other processes sharing this
 339  * swap cache page. Free it if you can. Return success.
 340  */
 341 int remove_exclusive_swap_page(struct page *page)
 342 {
 343         int retval;
 344         struct swap_info_struct * p;
 345         swp_entry_t entry;
 346
 347         BUG_ON(PagePrivate(page));
 348         BUG_ON(!PageLocked(page));
 349
 350         if (!PageSwapCache(page))
 351                 return 0;
 352         if (PageWriteback(page))
 353                 return 0;
 354         if (page_count(page) != 2) /* 2: us + cache */
 355                 return 0;
 356
 357         entry.val = page->private;
 358         p = swap_info_get(entry);
 359         if (!p)
 360                 return 0;
 361
 362         /* Is the only swap cache user the cache itself? */
 363         retval = 0;
 364         if (p->swap_map[swp_offset(entry)] == 1) {
 365                 /* Recheck the page count with the swapcache lock held.. */
 366                 spin_lock_irq(&swapper_space.tree_lock);
 367                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 368                         __delete_from_swap_cache(page);
 369                         SetPageDirty(page);
 370                         retval = 1;
 371                 }
 372                 spin_unlock_irq(&swapper_space.tree_lock);
 373         }
 374         swap_info_put(p);
 375
 376         if (retval) {
 377                 swap_free(entry);
 378                 page_cache_release(page);
 379         }
 380
 381         return retval;
 382 }
 383
 384 /*
 385  * Free the swap entry like above, but also try to
 386  * free the page cache entry if it is the last user.
 387  */
 388 void free_swap_and_cache(swp_entry_t entry)
 389 {
 390         struct swap_info_struct * p;
 391         struct page *page = NULL;
 392
 393         p = swap_info_get(entry);
 394         if (p) {
 395                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 396                         spin_lock_irq(&swapper_space.tree_lock);
 397                         page = radix_tree_lookup(&swapper_space.page_tree,
 398                                 entry.val);
 399                         if (page && TestSetPageLocked(page))
 400                                 page = NULL;
 401                         spin_unlock_irq(&swapper_space.tree_lock);
 402                 }
 403                 swap_info_put(p);
 404         }
 405         if (page) {
 406                 int one_user;
 407
 408                 BUG_ON(PagePrivate(page));
 409                 page_cache_get(page);
 410                 one_user = (page_count(page) == 2);
 411                 /* Only cache user (+us), or swap space full? Free it! */
 412                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
 413                         delete_from_swap_cache(page);
 414                         SetPageDirty(page);
 415                 }
 416                 unlock_page(page);
 417                 page_cache_release(page);
 418         }
 419 }
 420
 421 /*
 422  * The swap entry has been read in advance, and we return 1 to indicate
 423  * that the page has been used or is no longer needed.
 424  *
 425  * Always set the resulting pte to be nowrite (the same as COW pages
 426  * after one process has exited).  We don't know just how many PTEs will
 427  * share this swap entry, so be cautious and let do_wp_page work out
 428  * what to do if a write is requested later.
 429  */
 430 /* vma->vm_mm->page_table_lock is held */
 431 static void
 432 unuse_pte(struct vm_area_struct *vma, unsigned long address, pte_t *dir,
 433         swp_entry_t entry, struct page *page)
 434 {
 435         // vma->vm_mm->rss++;
 436         vx_rsspages_inc(vma->vm_mm);
 437         get_page(page);
 438         set_pte(dir, pte_mkold(mk_pte(page, vma->vm_page_prot)));
 439         page_add_anon_rmap(page, vma, address);
 440         swap_free(entry);
 441 }
 442
 443 /* vma->vm_mm->page_table_lock is held */
 444 static unsigned long unuse_pmd(struct vm_area_struct * vma, pmd_t *dir,
 445         unsigned long address, unsigned long size, unsigned long offset,
 446         swp_entry_t entry, struct page *page)
 447 {
 448         pte_t * pte;
 449         unsigned long end;
 450         pte_t swp_pte = swp_entry_to_pte(entry);
 451
 452         if (pmd_none(*dir))
 453                 return 0;
 454         if (pmd_bad(*dir)) {
 455                 pmd_ERROR(*dir);
 456                 pmd_clear(dir);
 457                 return 0;
 458         }
 459         pte = pte_offset_map(dir, address);
 460         offset += address & PMD_MASK;
 461         address &= ~PMD_MASK;
 462         end = address + size;
 463         if (end > PMD_SIZE)
 464                 end = PMD_SIZE;
 465         do {
 466                 /*
 467                  * swapoff spends a _lot_ of time in this loop!
 468                  * Test inline before going to call unuse_pte.
 469                  */
 470                 if (unlikely(pte_same(*pte, swp_pte))) {
 471                         unuse_pte(vma, offset + address, pte, entry, page);
 472                         pte_unmap(pte);
 473
 474                         /*
 475                          * Move the page to the active list so it is not
 476                          * immediately swapped out again after swapon.
 477                          */
 478                         activate_page(page);
 479
 480                         /* add 1 since address may be 0 */
 481                         return 1 + offset + address;
 482                 }
 483                 address += PAGE_SIZE;
 484                 pte++;
 485         } while (address && (address < end));
 486         pte_unmap(pte - 1);
 487         return 0;
 488 }
 489
 490 /* vma->vm_mm->page_table_lock is held */
 491 static unsigned long unuse_pgd(struct vm_area_struct * vma, pgd_t *dir,
 492         unsigned long address, unsigned long size,
 493         swp_entry_t entry, struct page *page)
 494 {
 495         pmd_t * pmd;
 496         unsigned long offset, end;
 497         unsigned long foundaddr;
 498
 499         if (pgd_none(*dir))
 500                 return 0;
 501         if (pgd_bad(*dir)) {
 502                 pgd_ERROR(*dir);
 503                 pgd_clear(dir);
 504                 return 0;
 505         }
 506         pmd = pmd_offset(dir, address);
 507         offset = address & PGDIR_MASK;
 508         address &= ~PGDIR_MASK;
 509         end = address + size;
 510         if (end > PGDIR_SIZE)
 511                 end = PGDIR_SIZE;
 512         if (address >= end)
 513                 BUG();
 514         do {
 515                 foundaddr = unuse_pmd(vma, pmd, address, end - address,
 516                                                 offset, entry, page);
 517                 if (foundaddr)
 518                         return foundaddr;
 519                 address = (address + PMD_SIZE) & PMD_MASK;
 520                 pmd++;
 521         } while (address && (address < end));
 522         return 0;
 523 }
 524
 525 /* vma->vm_mm->page_table_lock is held */
 526 static unsigned long unuse_vma(struct vm_area_struct * vma,
 527         swp_entry_t entry, struct page *page)
 528 {
 529         pgd_t *pgdir;
 530         unsigned long start, end;
 531         unsigned long foundaddr;
 532
 533         if (page->mapping) {
 534                 start = page_address_in_vma(page, vma);
 535                 if (start == -EFAULT)
 536                         return 0;
 537                 else
 538                         end = start + PAGE_SIZE;
 539         } else {
 540                 start = vma->vm_start;
 541                 end = vma->vm_end;
 542         }
 543         pgdir = pgd_offset(vma->vm_mm, start);
 544         do {
 545                 foundaddr = unuse_pgd(vma, pgdir, start, end - start,
 546                                                 entry, page);
 547                 if (foundaddr)
 548                         return foundaddr;
 549                 start = (start + PGDIR_SIZE) & PGDIR_MASK;
 550                 pgdir++;
 551         } while (start && (start < end));
 552         return 0;
 553 }
 554
 555 static int unuse_process(struct mm_struct * mm,
 556                         swp_entry_t entry, struct page* page)
 557 {
 558         struct vm_area_struct* vma;
 559         unsigned long foundaddr = 0;
 560
 561         /*
 562          * Go through process' page directory.
 563          */
 564         if (!down_read_trylock(&mm->mmap_sem)) {
 565                 /*
 566                  * Our reference to the page stops try_to_unmap_one from
 567                  * unmapping its ptes, so swapoff can make progress.
 568                  */
 569                 unlock_page(page);
 570                 down_read(&mm->mmap_sem);
 571                 lock_page(page);
 572         }
 573         spin_lock(&mm->page_table_lock);
 574         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 575                 if (vma->anon_vma) {
 576                         foundaddr = unuse_vma(vma, entry, page);
 577                         if (foundaddr)
 578                                 break;
 579                 }
 580         }
 581         spin_unlock(&mm->page_table_lock);
 582         up_read(&mm->mmap_sem);
 583         /*
 584          * Currently unuse_process cannot fail, but leave error handling
 585          * at call sites for now, since we change it from time to time.
 586          */
 587         return 0;
 588 }
 589
 590 /*
 591  * Scan swap_map from current position to next entry still in use.
 592  * Recycle to start on reaching the end, returning 0 when empty.
 593  */
 594 static int find_next_to_unuse(struct swap_info_struct *si, int prev)
 595 {
 596         int max = si->max;
 597         int i = prev;
 598         int count;
 599
 600         /*
 601          * No need for swap_device_lock(si) here: we're just looking
 602          * for whether an entry is in use, not modifying it; false
 603          * hits are okay, and sys_swapoff() has already prevented new
 604          * allocations from this area (while holding swap_list_lock()).
 605          */
 606         for (;;) {
 607                 if (++i >= max) {
 608                         if (!prev) {
 609                                 i = 0;
 610                                 break;
 611                         }
 612                         /*
 613                          * No entries in use at top of swap_map,
 614                          * loop back to start and recheck there.
 615                          */
 616                         max = prev + 1;
 617                         prev = 0;
 618                         i = 1;
 619                 }
 620                 count = si->swap_map[i];
 621                 if (count && count != SWAP_MAP_BAD)
 622                         break;
 623         }
 624         return i;
 625 }
 626
 627 /*
 628  * We completely avoid races by reading each swap page in advance,
 629  * and then search for the process using it.  All the necessary
 630  * page table adjustments can then be made atomically.
 631  */
 632 static int try_to_unuse(unsigned int type)
 633 {
 634         struct swap_info_struct * si = &swap_info[type];
 635         struct mm_struct *start_mm;
 636         unsigned short *swap_map;
 637         unsigned short swcount;
 638         struct page *page;
 639         swp_entry_t entry;
 640         int i = 0;
 641         int retval = 0;
 642         int reset_overflow = 0;
 643         int shmem;
 644
 645         /*
 646          * When searching mms for an entry, a good strategy is to
 647          * start at the first mm we freed the previous entry from
 648          * (though actually we don't notice whether we or coincidence
 649          * freed the entry).  Initialize this start_mm with a hold.
 650          *
 651          * A simpler strategy would be to start at the last mm we
 652          * freed the previous entry from; but that would take less
 653          * advantage of mmlist ordering (now preserved by swap_out()),
 654          * which clusters forked address spaces together, most recent
 655          * child immediately after parent.  If we race with dup_mmap(),
 656          * we very much want to resolve parent before child, otherwise
 657          * we may miss some entries: using last mm would invert that.
 658          */
 659         start_mm = &init_mm;
 660         atomic_inc(&init_mm.mm_users);
 661
 662         /*
 663          * Keep on scanning until all entries have gone.  Usually,
 664          * one pass through swap_map is enough, but not necessarily:
 665          * mmput() removes mm from mmlist before exit_mmap() and its
 666          * zap_page_range().  That's not too bad, those entries are
 667          * on their way out, and handled faster there than here.
 668          * do_munmap() behaves similarly, taking the range out of mm's
 669          * vma list before zap_page_range().  But unfortunately, when
 670          * unmapping a part of a vma, it takes the whole out first,
 671          * then reinserts what's left after (might even reschedule if
 672          * open() method called) - so swap entries may be invisible
 673          * to swapoff for a while, then reappear - but that is rare.
 674          */
 675         while ((i = find_next_to_unuse(si, i)) != 0) {
 676                 if (signal_pending(current)) {
 677                         retval = -EINTR;
 678                         break;
 679                 }
 680
 681                 /*
 682                  * Get a page for the entry, using the existing swap
 683                  * cache page if there is one.  Otherwise, get a clean
 684                  * page and read the swap into it.
 685                  */
 686                 swap_map = &si->swap_map[i];
 687                 entry = swp_entry(type, i);
 688                 page = read_swap_cache_async(entry, NULL, 0);
 689                 if (!page) {
 690                         /*
 691                          * Either swap_duplicate() failed because entry
 692                          * has been freed independently, and will not be
 693                          * reused since sys_swapoff() already disabled
 694                          * allocation from here, or alloc_page() failed.
 695                          */
 696                         if (!*swap_map)
 697                                 continue;
 698                         retval = -ENOMEM;
 699                         break;
 700                 }
 701
 702                 /*
 703                  * Don't hold on to start_mm if it looks like exiting.
 704                  */
 705                 if (atomic_read(&start_mm->mm_users) == 1) {
 706                         mmput(start_mm);
 707                         start_mm = &init_mm;
 708                         atomic_inc(&init_mm.mm_users);
 709                 }
 710
 711                 /*
 712                  * Wait for and lock page.  When do_swap_page races with
 713                  * try_to_unuse, do_swap_page can handle the fault much
 714                  * faster than try_to_unuse can locate the entry.  This
 715                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 716                  * defer to do_swap_page in such a case - in some tests,
 717                  * do_swap_page and try_to_unuse repeatedly compete.
 718                  */
 719                 wait_on_page_locked(page);
 720                 wait_on_page_writeback(page);
 721                 lock_page(page);
 722                 wait_on_page_writeback(page);
 723
 724                 /*
 725                  * Remove all references to entry, without blocking.
 726                  * Whenever we reach init_mm, there's no address space
 727                  * to search, but use it as a reminder to search shmem.
 728                  */
 729                 shmem = 0;
 730                 swcount = *swap_map;
 731                 if (swcount > 1) {
 732                         if (start_mm == &init_mm)
 733                                 shmem = shmem_unuse(entry, page);
 734                         else
 735                                 retval = unuse_process(start_mm, entry, page);
 736                 }
 737                 if (*swap_map > 1) {
 738                         int set_start_mm = (*swap_map >= swcount);
 739                         struct list_head *p = &start_mm->mmlist;
 740                         struct mm_struct *new_start_mm = start_mm;
 741                         struct mm_struct *prev_mm = start_mm;
 742                         struct mm_struct *mm;
 743
 744                         atomic_inc(&new_start_mm->mm_users);
 745                         atomic_inc(&prev_mm->mm_users);
 746                         spin_lock(&mmlist_lock);
 747                         while (*swap_map > 1 && !retval &&
 748                                         (p = p->next) != &start_mm->mmlist) {
 749                                 mm = list_entry(p, struct mm_struct, mmlist);
 750                                 atomic_inc(&mm->mm_users);
 751                                 spin_unlock(&mmlist_lock);
 752                                 mmput(prev_mm);
 753                                 prev_mm = mm;
 754
 755                                 cond_resched();
 756
 757                                 swcount = *swap_map;
 758                                 if (swcount <= 1)
 759                                         ;
 760                                 else if (mm == &init_mm) {
 761                                         set_start_mm = 1;
 762                                         shmem = shmem_unuse(entry, page);
 763                                 } else
 764                                         retval = unuse_process(mm, entry, page);
 765                                 if (set_start_mm && *swap_map < swcount) {
 766                                         mmput(new_start_mm);
 767                                         atomic_inc(&mm->mm_users);
 768                                         new_start_mm = mm;
 769                                         set_start_mm = 0;
 770                                 }
 771                                 spin_lock(&mmlist_lock);
 772                         }
 773                         spin_unlock(&mmlist_lock);
 774                         mmput(prev_mm);
 775                         mmput(start_mm);
 776                         start_mm = new_start_mm;
 777                 }
 778                 if (retval) {
 779                         unlock_page(page);
 780                         page_cache_release(page);
 781                         break;
 782                 }
 783
 784                 /*
 785                  * How could swap count reach 0x7fff when the maximum
 786                  * pid is 0x7fff, and there's no way to repeat a swap
 787                  * page within an mm (except in shmem, where it's the
 788                  * shared object which takes the reference count)?
 789                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 790                  *
 791                  * If that's wrong, then we should worry more about
 792                  * exit_mmap() and do_munmap() cases described above:
 793                  * we might be resetting SWAP_MAP_MAX too early here.
 794                  * We know "Undead"s can happen, they're okay, so don't
 795                  * report them; but do report if we reset SWAP_MAP_MAX.
 796                  */
 797                 if (*swap_map == SWAP_MAP_MAX) {
 798                         swap_device_lock(si);
 799                         *swap_map = 1;
 800                         swap_device_unlock(si);
 801                         reset_overflow = 1;
 802                 }
 803
 804                 /*
 805                  * If a reference remains (rare), we would like to leave
 806                  * the page in the swap cache; but try_to_unmap could
 807                  * then re-duplicate the entry once we drop page lock,
 808                  * so we might loop indefinitely; also, that page could
 809                  * not be swapped out to other storage meanwhile.  So:
 810                  * delete from cache even if there's another reference,
 811                  * after ensuring that the data has been saved to disk -
 812                  * since if the reference remains (rarer), it will be
 813                  * read from disk into another page.  Splitting into two
 814                  * pages would be incorrect if swap supported "shared
 815                  * private" pages, but they are handled by tmpfs files.
 816                  *
 817                  * Note shmem_unuse already deleted a swappage from
 818                  * the swap cache, unless the move to filepage failed:
 819                  * in which case it left swappage in cache, lowered its
 820                  * swap count to pass quickly through the loops above,
 821                  * and now we must reincrement count to try again later.
 822                  */
 823                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 824                         struct writeback_control wbc = {
 825                                 .sync_mode = WB_SYNC_NONE,
 826                         };
 827
 828                         swap_writepage(page, &wbc);
 829                         lock_page(page);
 830                         wait_on_page_writeback(page);
 831                 }
 832                 if (PageSwapCache(page)) {
 833                         if (shmem)
 834                                 swap_duplicate(entry);
 835                         else
 836                                 delete_from_swap_cache(page);
 837                 }
 838
 839                 /*
 840                  * So we could skip searching mms once swap count went
 841                  * to 1, we did not mark any present ptes as dirty: must
 842                  * mark page dirty so shrink_list will preserve it.
 843                  */
 844                 SetPageDirty(page);
 845                 unlock_page(page);
 846                 page_cache_release(page);
 847
 848                 /*
 849                  * Make sure that we aren't completely killing
 850                  * interactive performance.
 851                  */
 852                 cond_resched();
 853         }
 854
 855         mmput(start_mm);
 856         if (reset_overflow) {
 857                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 858                 swap_overflow = 0;
 859         }
 860         return retval;
 861 }
 862
 863 /*
 864  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 865  * corresponds to page offset `offset'.
 866  */
 867 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 868 {
 869         struct swap_extent *se = sis->curr_swap_extent;
 870         struct swap_extent *start_se = se;
 871
 872         for ( ; ; ) {
 873                 struct list_head *lh;
 874
 875                 if (se->start_page <= offset &&
 876                                 offset < (se->start_page + se->nr_pages)) {
 877                         return se->start_block + (offset - se->start_page);
 878                 }
 879                 lh = se->list.prev;
 880                 if (lh == &sis->extent_list)
 881                         lh = lh->prev;
 882                 se = list_entry(lh, struct swap_extent, list);
 883                 sis->curr_swap_extent = se;
 884                 BUG_ON(se == start_se);         /* It *must* be present */
 885         }
 886 }
 887
 888 /*
 889  * Free all of a swapdev's extent information
 890  */
 891 static void destroy_swap_extents(struct swap_info_struct *sis)
 892 {
 893         while (!list_empty(&sis->extent_list)) {
 894                 struct swap_extent *se;
 895
 896                 se = list_entry(sis->extent_list.next,
 897                                 struct swap_extent, list);
 898                 list_del(&se->list);
 899                 kfree(se);
 900         }
 901         sis->nr_extents = 0;
 902 }
 903
 904 /*
 905  * Add a block range (and the corresponding page range) into this swapdev's
 906  * extent list.  The extent list is kept sorted in block order.
 907  *
 908  * This function rather assumes that it is called in ascending sector_t order.
 909  * It doesn't look for extent coalescing opportunities.
 910  */
 911 static int
 912 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 913                 unsigned long nr_pages, sector_t start_block)
 914 {
 915         struct swap_extent *se;
 916         struct swap_extent *new_se;
 917         struct list_head *lh;
 918
 919         lh = sis->extent_list.next;     /* The highest-addressed block */
 920         while (lh != &sis->extent_list) {
 921                 se = list_entry(lh, struct swap_extent, list);
 922                 if (se->start_block + se->nr_pages == start_block &&
 923                     se->start_page  + se->nr_pages == start_page) {
 924                         /* Merge it */
 925                         se->nr_pages += nr_pages;
 926                         return 0;
 927                 }
 928                 lh = lh->next;
 929         }
 930
 931         /*
 932          * No merge.  Insert a new extent, preserving ordering.
 933          */
 934         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 935         if (new_se == NULL)
 936                 return -ENOMEM;
 937         new_se->start_page = start_page;
 938         new_se->nr_pages = nr_pages;
 939         new_se->start_block = start_block;
 940
 941         lh = sis->extent_list.prev;     /* The lowest block */
 942         while (lh != &sis->extent_list) {
 943                 se = list_entry(lh, struct swap_extent, list);
 944                 if (se->start_block > start_block)
 945                         break;
 946                 lh = lh->prev;
 947         }
 948         list_add_tail(&new_se->list, lh);
 949         sis->nr_extents++;
 950         return 0;
 951 }
 952
 953 /*
 954  * A `swap extent' is a simple thing which maps a contiguous range of pages
 955  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 956  * is built at swapon time and is then used at swap_writepage/swap_readpage
 957  * time for locating where on disk a page belongs.
 958  *
 959  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 960  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 961  * swap files identically.
 962  *
 963  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
 964  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
 965  * swapfiles are handled *identically* after swapon time.
 966  *
 967  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
 968  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
 969  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
 970  * requirements, they are simply tossed out - we will never use those blocks
 971  * for swapping.
 972  *
 973  * For S_ISREG swapfiles we hold i_sem across the life of the swapon.  This
 974  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
 975  * which will scribble on the fs.
 976  *
 977  * The amount of disk space which a single swap extent represents varies.
 978  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
 979  * extents in the list.  To avoid much list walking, we cache the previous
 980  * search location in `curr_swap_extent', and start new searches from there.
 981  * This is extremely effective.  The average number of iterations in
 982  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
 983  */
 984 static int setup_swap_extents(struct swap_info_struct *sis)
 985 {
 986         struct inode *inode;
 987         unsigned blocks_per_page;
 988         unsigned long page_no;
 989         unsigned blkbits;
 990         sector_t probe_block;
 991         sector_t last_block;
 992         int ret;
 993
 994         inode = sis->swap_file->f_mapping->host;
 995         if (S_ISBLK(inode->i_mode)) {
 996                 ret = add_swap_extent(sis, 0, sis->max, 0);
 997                 goto done;
 998         }
 999
1000         blkbits = inode->i_blkbits;
1001         blocks_per_page = PAGE_SIZE >> blkbits;
1002
1003         /*
1004          * Map all the blocks into the extent list.  This code doesn't try
1005          * to be very smart.
1006          */
1007         probe_block = 0;
1008         page_no = 0;
1009         last_block = i_size_read(inode) >> blkbits;
1010         while ((probe_block + blocks_per_page) <= last_block &&
1011                         page_no < sis->max) {
1012                 unsigned block_in_page;
1013                 sector_t first_block;
1014
1015                 first_block = bmap(inode, probe_block);
1016                 if (first_block == 0)
1017                         goto bad_bmap;
1018
1019                 /*
1020                  * It must be PAGE_SIZE aligned on-disk
1021                  */
1022                 if (first_block & (blocks_per_page - 1)) {
1023                         probe_block++;
1024                         goto reprobe;
1025                 }
1026
1027                 for (block_in_page = 1; block_in_page < blocks_per_page;
1028                                         block_in_page++) {
1029                         sector_t block;
1030
1031                         block = bmap(inode, probe_block + block_in_page);
1032                         if (block == 0)
1033                                 goto bad_bmap;
1034                         if (block != first_block + block_in_page) {
1035                                 /* Discontiguity */
1036                                 probe_block++;
1037                                 goto reprobe;
1038                         }
1039                 }
1040
1041                 /*
1042                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1043                  */
1044                 ret = add_swap_extent(sis, page_no, 1,
1045                                 first_block >> (PAGE_SHIFT - blkbits));
1046                 if (ret)
1047                         goto out;
1048                 page_no++;
1049                 probe_block += blocks_per_page;
1050 reprobe:
1051                 continue;
1052         }
1053         ret = 0;
1054         if (page_no == 0)
1055                 ret = -EINVAL;
1056         sis->max = page_no;
1057         sis->highest_bit = page_no - 1;
1058 done:
1059         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1060                                         struct swap_extent, list);
1061         goto out;
1062 bad_bmap:
1063         printk(KERN_ERR "swapon: swapfile has holes\n");
1064         ret = -EINVAL;
1065 out:
1066         return ret;
1067 }
1068
1069 #if 0   /* We don't need this yet */
1070 #include <linux/backing-dev.h>
1071 int page_queue_congested(struct page *page)
1072 {
1073         struct backing_dev_info *bdi;
1074
1075         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1076
1077         if (PageSwapCache(page)) {
1078                 swp_entry_t entry = { .val = page->private };
1079                 struct swap_info_struct *sis;
1080
1081                 sis = get_swap_info_struct(swp_type(entry));
1082                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1083         } else
1084                 bdi = page->mapping->backing_dev_info;
1085         return bdi_write_congested(bdi);
1086 }
1087 #endif
1088
1089 asmlinkage long sys_swapoff(const char __user * specialfile)
1090 {
1091         struct swap_info_struct * p = NULL;
1092         unsigned short *swap_map;
1093         struct file *swap_file, *victim;
1094         struct address_space *mapping;
1095         struct inode *inode;
1096         char * pathname;
1097         int i, type, prev;
1098         int err;
1099
1100         if (!capable(CAP_SYS_ADMIN))
1101                 return -EPERM;
1102
1103         pathname = getname(specialfile);
1104         err = PTR_ERR(pathname);
1105         if (IS_ERR(pathname))
1106                 goto out;
1107
1108         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1109         putname(pathname);
1110         err = PTR_ERR(victim);
1111         if (IS_ERR(victim))
1112                 goto out;
1113
1114         mapping = victim->f_mapping;
1115         prev = -1;
1116         swap_list_lock();
1117         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1118                 p = swap_info + type;
1119                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1120                         if (p->swap_file->f_mapping == mapping)
1121                                 break;
1122                 }
1123                 prev = type;
1124         }
1125         if (type < 0) {
1126                 err = -EINVAL;
1127                 swap_list_unlock();
1128                 goto out_dput;
1129         }
1130         if (!security_vm_enough_memory(p->pages))
1131                 vm_unacct_memory(p->pages);
1132         else {
1133                 err = -ENOMEM;
1134                 swap_list_unlock();
1135                 goto out_dput;
1136         }
1137         if (prev < 0) {
1138                 swap_list.head = p->next;
1139         } else {
1140                 swap_info[prev].next = p->next;
1141         }
1142         if (type == swap_list.next) {
1143                 /* just pick something that's safe... */
1144                 swap_list.next = swap_list.head;
1145         }
1146         nr_swap_pages -= p->pages;
1147         total_swap_pages -= p->pages;
1148         p->flags &= ~SWP_WRITEOK;
1149         swap_list_unlock();
1150         current->flags |= PF_SWAPOFF;
1151         err = try_to_unuse(type);
1152         current->flags &= ~PF_SWAPOFF;
1153
1154         /* wait for any unplug function to finish */
1155         down_write(&swap_unplug_sem);
1156         up_write(&swap_unplug_sem);
1157
1158         if (err) {
1159                 /* re-insert swap space back into swap_list */
1160                 swap_list_lock();
1161                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1162                         if (p->prio >= swap_info[i].prio)
1163                                 break;
1164                 p->next = i;
1165                 if (prev < 0)
1166                         swap_list.head = swap_list.next = p - swap_info;
1167                 else
1168                         swap_info[prev].next = p - swap_info;
1169                 nr_swap_pages += p->pages;
1170                 total_swap_pages += p->pages;
1171                 p->flags |= SWP_WRITEOK;
1172                 swap_list_unlock();
1173                 goto out_dput;
1174         }
1175         down(&swapon_sem);
1176         swap_list_lock();
1177         swap_device_lock(p);
1178         swap_file = p->swap_file;
1179         p->swap_file = NULL;
1180         p->max = 0;
1181         swap_map = p->swap_map;
1182         p->swap_map = NULL;
1183         p->flags = 0;
1184         destroy_swap_extents(p);
1185         swap_device_unlock(p);
1186         swap_list_unlock();
1187         up(&swapon_sem);
1188         vfree(swap_map);
1189         inode = mapping->host;
1190         if (S_ISBLK(inode->i_mode)) {
1191                 struct block_device *bdev = I_BDEV(inode);
1192                 set_blocksize(bdev, p->old_block_size);
1193                 bd_release(bdev);
1194         } else {
1195                 down(&inode->i_sem);
1196                 inode->i_flags &= ~S_SWAPFILE;
1197                 up(&inode->i_sem);
1198         }
1199         filp_close(swap_file, NULL);
1200         err = 0;
1201
1202 out_dput:
1203         filp_close(victim, NULL);
1204 out:
1205         return err;
1206 }
1207
1208 #ifdef CONFIG_PROC_FS
1209 /* iterator */
1210 static void *swap_start(struct seq_file *swap, loff_t *pos)
1211 {
1212         struct swap_info_struct *ptr = swap_info;
1213         int i;
1214         loff_t l = *pos;
1215
1216         down(&swapon_sem);
1217
1218         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1219                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1220                         continue;
1221                 if (!l--)
1222                         return ptr;
1223         }
1224
1225         return NULL;
1226 }
1227
1228 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1229 {
1230         struct swap_info_struct *ptr = v;
1231         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1232
1233         for (++ptr; ptr < endptr; ptr++) {
1234                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1235                         continue;
1236                 ++*pos;
1237                 return ptr;
1238         }
1239
1240         return NULL;
1241 }
1242
1243 static void swap_stop(struct seq_file *swap, void *v)
1244 {
1245         up(&swapon_sem);
1246 }
1247
1248 static int swap_show(struct seq_file *swap, void *v)
1249 {
1250         struct swap_info_struct *ptr = v;
1251         struct file *file;
1252         int len;
1253
1254         if (v == swap_info)
1255                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1256
1257         file = ptr->swap_file;
1258         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1259         seq_printf(swap, "%*s%s\t%d\t%ld\t%d\n",
1260                        len < 40 ? 40 - len : 1, " ",
1261                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1262                                 "partition" : "file\t",
1263                        ptr->pages << (PAGE_SHIFT - 10),
1264                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1265                        ptr->prio);
1266         return 0;
1267 }
1268
1269 static struct seq_operations swaps_op = {
1270         .start =        swap_start,
1271         .next =         swap_next,
1272         .stop =         swap_stop,
1273         .show =         swap_show
1274 };
1275
1276 static int swaps_open(struct inode *inode, struct file *file)
1277 {
1278         return seq_open(file, &swaps_op);
1279 }
1280
1281 static struct file_operations proc_swaps_operations = {
1282         .open           = swaps_open,
1283         .read           = seq_read,
1284         .llseek         = seq_lseek,
1285         .release        = seq_release,
1286 };
1287
1288 static int __init procswaps_init(void)
1289 {
1290         struct proc_dir_entry *entry;
1291
1292         entry = create_proc_entry("swaps", 0, NULL);
1293         if (entry)
1294                 entry->proc_fops = &proc_swaps_operations;
1295         return 0;
1296 }
1297 __initcall(procswaps_init);
1298 #endif /* CONFIG_PROC_FS */
1299
1300 /*
1301  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1302  *
1303  * The swapon system call
1304  */
1305 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1306 {
1307         struct swap_info_struct * p;
1308         char *name = NULL;
1309         struct block_device *bdev = NULL;
1310         struct file *swap_file = NULL;
1311         struct address_space *mapping;
1312         unsigned int type;
1313         int i, prev;
1314         int error;
1315         static int least_priority;
1316         union swap_header *swap_header = NULL;
1317         int swap_header_version;
1318         int nr_good_pages = 0;
1319         unsigned long maxpages = 1;
1320         int swapfilesize;
1321         unsigned short *swap_map;
1322         struct page *page = NULL;
1323         struct inode *inode = NULL;
1324         int did_down = 0;
1325
1326         if (!capable(CAP_SYS_ADMIN))
1327                 return -EPERM;
1328         swap_list_lock();
1329         p = swap_info;
1330         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1331                 if (!(p->flags & SWP_USED))
1332                         break;
1333         error = -EPERM;
1334         /*
1335          * Test if adding another swap device is possible. There are
1336          * two limiting factors: 1) the number of bits for the swap
1337          * type swp_entry_t definition and 2) the number of bits for
1338          * the swap type in the swap ptes as defined by the different
1339          * architectures. To honor both limitations a swap entry
1340          * with swap offset 0 and swap type ~0UL is created, encoded
1341          * to a swap pte, decoded to a swp_entry_t again and finally
1342          * the swap type part is extracted. This will mask all bits
1343          * from the initial ~0UL that can't be encoded in either the
1344          * swp_entry_t or the architecture definition of a swap pte.
1345          */
1346         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1347                 swap_list_unlock();
1348                 goto out;
1349         }
1350         if (type >= nr_swapfiles)
1351                 nr_swapfiles = type+1;
1352         INIT_LIST_HEAD(&p->extent_list);
1353         p->flags = SWP_USED;
1354         p->nr_extents = 0;
1355         p->swap_file = NULL;
1356         p->old_block_size = 0;
1357         p->swap_map = NULL;
1358         p->lowest_bit = 0;
1359         p->highest_bit = 0;
1360         p->cluster_nr = 0;
1361         p->inuse_pages = 0;
1362         p->sdev_lock = SPIN_LOCK_UNLOCKED;
1363         p->next = -1;
1364         if (swap_flags & SWAP_FLAG_PREFER) {
1365                 p->prio =
1366                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1367         } else {
1368                 p->prio = --least_priority;
1369         }
1370         swap_list_unlock();
1371         name = getname(specialfile);
1372         error = PTR_ERR(name);
1373         if (IS_ERR(name)) {
1374                 name = NULL;
1375                 goto bad_swap_2;
1376         }
1377         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1378         error = PTR_ERR(swap_file);
1379         if (IS_ERR(swap_file)) {
1380                 swap_file = NULL;
1381                 goto bad_swap_2;
1382         }
1383
1384         p->swap_file = swap_file;
1385         mapping = swap_file->f_mapping;
1386         inode = mapping->host;
1387
1388         error = -EBUSY;
1389         for (i = 0; i < nr_swapfiles; i++) {
1390                 struct swap_info_struct *q = &swap_info[i];
1391
1392                 if (i == type || !q->swap_file)
1393                         continue;
1394                 if (mapping == q->swap_file->f_mapping)
1395                         goto bad_swap;
1396         }
1397
1398         error = -EINVAL;
1399         if (S_ISBLK(inode->i_mode)) {
1400                 bdev = I_BDEV(inode);
1401                 error = bd_claim(bdev, sys_swapon);
1402                 if (error < 0) {
1403                         bdev = NULL;
1404                         goto bad_swap;
1405                 }
1406                 p->old_block_size = block_size(bdev);
1407                 error = set_blocksize(bdev, PAGE_SIZE);
1408                 if (error < 0)
1409                         goto bad_swap;
1410                 p->bdev = bdev;
1411         } else if (S_ISREG(inode->i_mode)) {
1412                 p->bdev = inode->i_sb->s_bdev;
1413                 down(&inode->i_sem);
1414                 did_down = 1;
1415                 if (IS_SWAPFILE(inode)) {
1416                         error = -EBUSY;
1417                         goto bad_swap;
1418                 }
1419         } else {
1420                 goto bad_swap;
1421         }
1422
1423         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1424
1425         /*
1426          * Read the swap header.
1427          */
1428         if (!mapping->a_ops->readpage) {
1429                 error = -EINVAL;
1430                 goto bad_swap;
1431         }
1432         page = read_cache_page(mapping, 0,
1433                         (filler_t *)mapping->a_ops->readpage, swap_file);
1434         if (IS_ERR(page)) {
1435                 error = PTR_ERR(page);
1436                 goto bad_swap;
1437         }
1438         wait_on_page_locked(page);
1439         if (!PageUptodate(page))
1440                 goto bad_swap;
1441         kmap(page);
1442         swap_header = page_address(page);
1443
1444         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1445                 swap_header_version = 1;
1446         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1447                 swap_header_version = 2;
1448         else {
1449                 printk("Unable to find swap-space signature\n");
1450                 error = -EINVAL;
1451                 goto bad_swap;
1452         }
1453
1454         switch (swap_header_version) {
1455         case 1:
1456                 printk(KERN_ERR "version 0 swap is no longer supported. "
1457                         "Use mkswap -v1 %s\n", name);
1458                 error = -EINVAL;
1459                 goto bad_swap;
1460         case 2:
1461                 /* Check the swap header's sub-version and the size of
1462                    the swap file and bad block lists */
1463                 if (swap_header->info.version != 1) {
1464                         printk(KERN_WARNING
1465                                "Unable to handle swap header version %d\n",
1466                                swap_header->info.version);
1467                         error = -EINVAL;
1468                         goto bad_swap;
1469                 }
1470
1471                 p->lowest_bit  = 1;
1472                 /*
1473                  * Find out how many pages are allowed for a single swap
1474                  * device. There are two limiting factors: 1) the number of
1475                  * bits for the swap offset in the swp_entry_t type and
1476                  * 2) the number of bits in the a swap pte as defined by
1477                  * the different architectures. In order to find the
1478                  * largest possible bit mask a swap entry with swap type 0
1479                  * and swap offset ~0UL is created, encoded to a swap pte,
1480                  * decoded to a swp_entry_t again and finally the swap
1481                  * offset is extracted. This will mask all the bits from
1482                  * the initial ~0UL mask that can't be encoded in either
1483                  * the swp_entry_t or the architecture definition of a
1484                  * swap pte.
1485                  */
1486                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1487                 if (maxpages > swap_header->info.last_page)
1488                         maxpages = swap_header->info.last_page;
1489                 p->highest_bit = maxpages - 1;
1490
1491                 error = -EINVAL;
1492                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1493                         goto bad_swap;
1494
1495                 /* OK, set up the swap map and apply the bad block list */
1496                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1497                         error = -ENOMEM;
1498                         goto bad_swap;
1499                 }
1500
1501                 error = 0;
1502                 memset(p->swap_map, 0, maxpages * sizeof(short));
1503                 for (i=0; i<swap_header->info.nr_badpages; i++) {
1504                         int page = swap_header->info.badpages[i];
1505                         if (page <= 0 || page >= swap_header->info.last_page)
1506                                 error = -EINVAL;
1507                         else
1508                                 p->swap_map[page] = SWAP_MAP_BAD;
1509                 }
1510                 nr_good_pages = swap_header->info.last_page -
1511                                 swap_header->info.nr_badpages -
1512                                 1 /* header page */;
1513                 if (error)
1514                         goto bad_swap;
1515         }
1516
1517         if (swapfilesize && maxpages > swapfilesize) {
1518                 printk(KERN_WARNING
1519                        "Swap area shorter than signature indicates\n");
1520                 error = -EINVAL;
1521                 goto bad_swap;
1522         }
1523         if (!nr_good_pages) {
1524                 printk(KERN_WARNING "Empty swap-file\n");
1525                 error = -EINVAL;
1526                 goto bad_swap;
1527         }
1528         p->swap_map[0] = SWAP_MAP_BAD;
1529         p->max = maxpages;
1530         p->pages = nr_good_pages;
1531
1532         error = setup_swap_extents(p);
1533         if (error)
1534                 goto bad_swap;
1535
1536         down(&swapon_sem);
1537         swap_list_lock();
1538         swap_device_lock(p);
1539         p->flags = SWP_ACTIVE;
1540         nr_swap_pages += nr_good_pages;
1541         total_swap_pages += nr_good_pages;
1542         printk(KERN_INFO "Adding %dk swap on %s.  Priority:%d extents:%d\n",
1543                 nr_good_pages<<(PAGE_SHIFT-10), name,
1544                 p->prio, p->nr_extents);
1545
1546         /* insert swap space into swap_list: */
1547         prev = -1;
1548         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1549                 if (p->prio >= swap_info[i].prio) {
1550                         break;
1551                 }
1552                 prev = i;
1553         }
1554         p->next = i;
1555         if (prev < 0) {
1556                 swap_list.head = swap_list.next = p - swap_info;
1557         } else {
1558                 swap_info[prev].next = p - swap_info;
1559         }
1560         swap_device_unlock(p);
1561         swap_list_unlock();
1562         up(&swapon_sem);
1563         error = 0;
1564         goto out;
1565 bad_swap:
1566         if (bdev) {
1567                 set_blocksize(bdev, p->old_block_size);
1568                 bd_release(bdev);
1569         }
1570 bad_swap_2:
1571         swap_list_lock();
1572         swap_map = p->swap_map;
1573         p->swap_file = NULL;
1574         p->swap_map = NULL;
1575         p->flags = 0;
1576         if (!(swap_flags & SWAP_FLAG_PREFER))
1577                 ++least_priority;
1578         swap_list_unlock();
1579         destroy_swap_extents(p);
1580         if (swap_map)
1581                 vfree(swap_map);
1582         if (swap_file)
1583                 filp_close(swap_file, NULL);
1584 out:
1585         if (page && !IS_ERR(page)) {
1586                 kunmap(page);
1587                 page_cache_release(page);
1588         }
1589         if (name)
1590                 putname(name);
1591         if (did_down) {
1592                 if (!error)
1593                         inode->i_flags |= S_SWAPFILE;
1594                 up(&inode->i_sem);
1595         }
1596         return error;
1597 }
1598
1599 void si_swapinfo(struct sysinfo *val)
1600 {
1601         unsigned int i;
1602         unsigned long nr_to_be_unused = 0;
1603
1604         swap_list_lock();
1605         for (i = 0; i < nr_swapfiles; i++) {
1606                 if (!(swap_info[i].flags & SWP_USED) ||
1607                      (swap_info[i].flags & SWP_WRITEOK))
1608                         continue;
1609                 nr_to_be_unused += swap_info[i].inuse_pages;
1610         }
1611         val->freeswap = nr_swap_pages + nr_to_be_unused;
1612         val->totalswap = total_swap_pages + nr_to_be_unused;
1613         swap_list_unlock();
1614         if (vx_flags(VXF_VIRT_MEM, 0))
1615                 vx_vsi_swapinfo(val);
1616 }
1617
1618 /*
1619  * Verify that a swap entry is valid and increment its swap map count.
1620  *
1621  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1622  * "permanent", but will be reclaimed by the next swapoff.
1623  */
1624 int swap_duplicate(swp_entry_t entry)
1625 {
1626         struct swap_info_struct * p;
1627         unsigned long offset, type;
1628         int result = 0;
1629
1630         type = swp_type(entry);
1631         if (type >= nr_swapfiles)
1632                 goto bad_file;
1633         p = type + swap_info;
1634         offset = swp_offset(entry);
1635
1636         swap_device_lock(p);
1637         if (offset < p->max && p->swap_map[offset]) {
1638                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1639                         p->swap_map[offset]++;
1640                         result = 1;
1641                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1642                         if (swap_overflow++ < 5)
1643                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1644                         p->swap_map[offset] = SWAP_MAP_MAX;
1645                         result = 1;
1646                 }
1647         }
1648         swap_device_unlock(p);
1649 out:
1650         return result;
1651
1652 bad_file:
1653         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1654         goto out;
1655 }
1656
1657 struct swap_info_struct *
1658 get_swap_info_struct(unsigned type)
1659 {
1660         return &swap_info[type];
1661 }
1662
1663 /*
1664  * swap_device_lock prevents swap_map being freed. Don't grab an extra
1665  * reference on the swaphandle, it doesn't matter if it becomes unused.
1666  */
1667 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1668 {
1669         int ret = 0, i = 1 << page_cluster;
1670         unsigned long toff;
1671         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1672
1673         if (!page_cluster)      /* no readahead */
1674                 return 0;
1675         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1676         if (!toff)              /* first page is swap header */
1677                 toff++, i--;
1678         *offset = toff;
1679
1680         swap_device_lock(swapdev);
1681         do {
1682                 /* Don't read-ahead past the end of the swap area */
1683                 if (toff >= swapdev->max)
1684                         break;
1685                 /* Don't read in free or bad pages */
1686                 if (!swapdev->swap_map[toff])
1687                         break;
1688                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1689                         break;
1690                 toff++;
1691                 ret++;
1692         } while (--i);
1693         swap_device_unlock(swapdev);
1694         return ret;
1695 }