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