mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/config.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/mman.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/namei.h>
  18 #include <linux/shm.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/mutex.h>
  29 #include <linux/capability.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/vs_base.h>
  32 #include <linux/vs_memory.h>
  33
  34 #include <asm/pgtable.h>
  35 #include <asm/tlbflush.h>
  36 #include <linux/swapops.h>
  37
  38 DEFINE_SPINLOCK(swap_lock);
  39 unsigned int nr_swapfiles;
  40 long total_swap_pages;
  41 static int swap_overflow;
  42
  43 static const char Bad_file[] = "Bad swap file entry ";
  44 static const char Unused_file[] = "Unused swap file entry ";
  45 static const char Bad_offset[] = "Bad swap offset entry ";
  46 static const char Unused_offset[] = "Unused swap offset entry ";
  47
  48 struct swap_list_t swap_list = {-1, -1};
  49
  50 static struct swap_info_struct swap_info[MAX_SWAPFILES];
  51
  52 static DEFINE_MUTEX(swapon_mutex);
  53
  54 /*
  55  * We need this because the bdev->unplug_fn can sleep and we cannot
  56  * hold swap_lock while calling the unplug_fn. And swap_lock
  57  * cannot be turned into a mutex.
  58  */
  59 static DECLARE_RWSEM(swap_unplug_sem);
  60
  61 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  62 {
  63         swp_entry_t entry;
  64
  65         down_read(&swap_unplug_sem);
  66         entry.val = page_private(page);
  67         if (PageSwapCache(page)) {
  68                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  69                 struct backing_dev_info *bdi;
  70
  71                 /*
  72                  * If the page is removed from swapcache from under us (with a
  73                  * racy try_to_unuse/swapoff) we need an additional reference
  74                  * count to avoid reading garbage from page_private(page) above.
  75                  * If the WARN_ON triggers during a swapoff it maybe the race
  76                  * condition and it's harmless. However if it triggers without
  77                  * swapoff it signals a problem.
  78                  */
  79                 WARN_ON(page_count(page) <= 1);
  80
  81                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  82                 blk_run_backing_dev(bdi, page);
  83         }
  84         up_read(&swap_unplug_sem);
  85 }
  86
  87 #define SWAPFILE_CLUSTER        256
  88 #define LATENCY_LIMIT           256
  89
  90 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
  91 {
  92         unsigned long offset, last_in_cluster;
  93         int latency_ration = LATENCY_LIMIT;
  94
  95         /*
  96          * We try to cluster swap pages by allocating them sequentially
  97          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
  98          * way, however, we resort to first-free allocation, starting
  99          * a new cluster.  This prevents us from scattering swap pages
 100          * all over the entire swap partition, so that we reduce
 101          * overall disk seek times between swap pages.  -- sct
 102          * But we do now try to find an empty cluster.  -Andrea
 103          */
 104
 105         si->flags += SWP_SCANNING;
 106         if (unlikely(!si->cluster_nr)) {
 107                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 108                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
 109                         goto lowest;
 110                 spin_unlock(&swap_lock);
 111
 112                 offset = si->lowest_bit;
 113                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 114
 115                 /* Locate the first empty (unaligned) cluster */
 116                 for (; last_in_cluster <= si->highest_bit; offset++) {
 117                         if (si->swap_map[offset])
 118                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 119                         else if (offset == last_in_cluster) {
 120                                 spin_lock(&swap_lock);
 121                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
 122                                 goto cluster;
 123                         }
 124                         if (unlikely(--latency_ration < 0)) {
 125                                 cond_resched();
 126                                 latency_ration = LATENCY_LIMIT;
 127                         }
 128                 }
 129                 spin_lock(&swap_lock);
 130                 goto lowest;
 131         }
 132
 133         si->cluster_nr--;
 134 cluster:
 135         offset = si->cluster_next;
 136         if (offset > si->highest_bit)
 137 lowest:         offset = si->lowest_bit;
 138 checks: if (!(si->flags & SWP_WRITEOK))
 139                 goto no_page;
 140         if (!si->highest_bit)
 141                 goto no_page;
 142         if (!si->swap_map[offset]) {
 143                 if (offset == si->lowest_bit)
 144                         si->lowest_bit++;
 145                 if (offset == si->highest_bit)
 146                         si->highest_bit--;
 147                 si->inuse_pages++;
 148                 if (si->inuse_pages == si->pages) {
 149                         si->lowest_bit = si->max;
 150                         si->highest_bit = 0;
 151                 }
 152                 si->swap_map[offset] = 1;
 153                 si->cluster_next = offset + 1;
 154                 si->flags -= SWP_SCANNING;
 155                 return offset;
 156         }
 157
 158         spin_unlock(&swap_lock);
 159         while (++offset <= si->highest_bit) {
 160                 if (!si->swap_map[offset]) {
 161                         spin_lock(&swap_lock);
 162                         goto checks;
 163                 }
 164                 if (unlikely(--latency_ration < 0)) {
 165                         cond_resched();
 166                         latency_ration = LATENCY_LIMIT;
 167                 }
 168         }
 169         spin_lock(&swap_lock);
 170         goto lowest;
 171
 172 no_page:
 173         si->flags -= SWP_SCANNING;
 174         return 0;
 175 }
 176
 177 swp_entry_t get_swap_page(void)
 178 {
 179         struct swap_info_struct *si;
 180         pgoff_t offset;
 181         int type, next;
 182         int wrapped = 0;
 183
 184         spin_lock(&swap_lock);
 185         if (nr_swap_pages <= 0)
 186                 goto noswap;
 187         nr_swap_pages--;
 188
 189         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 190                 si = swap_info + type;
 191                 next = si->next;
 192                 if (next < 0 ||
 193                     (!wrapped && si->prio != swap_info[next].prio)) {
 194                         next = swap_list.head;
 195                         wrapped++;
 196                 }
 197
 198                 if (!si->highest_bit)
 199                         continue;
 200                 if (!(si->flags & SWP_WRITEOK))
 201                         continue;
 202
 203                 swap_list.next = next;
 204                 offset = scan_swap_map(si);
 205                 if (offset) {
 206                         spin_unlock(&swap_lock);
 207                         return swp_entry(type, offset);
 208                 }
 209                 next = swap_list.next;
 210         }
 211
 212         nr_swap_pages++;
 213 noswap:
 214         spin_unlock(&swap_lock);
 215         return (swp_entry_t) {0};
 216 }
 217
 218 swp_entry_t get_swap_page_of_type(int type)
 219 {
 220         struct swap_info_struct *si;
 221         pgoff_t offset;
 222
 223         spin_lock(&swap_lock);
 224         si = swap_info + type;
 225         if (si->flags & SWP_WRITEOK) {
 226                 nr_swap_pages--;
 227                 offset = scan_swap_map(si);
 228                 if (offset) {
 229                         spin_unlock(&swap_lock);
 230                         return swp_entry(type, offset);
 231                 }
 232                 nr_swap_pages++;
 233         }
 234         spin_unlock(&swap_lock);
 235         return (swp_entry_t) {0};
 236 }
 237
 238 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 239 {
 240         struct swap_info_struct * p;
 241         unsigned long offset, type;
 242
 243         if (!entry.val)
 244                 goto out;
 245         type = swp_type(entry);
 246         if (type >= nr_swapfiles)
 247                 goto bad_nofile;
 248         p = & swap_info[type];
 249         if (!(p->flags & SWP_USED))
 250                 goto bad_device;
 251         offset = swp_offset(entry);
 252         if (offset >= p->max)
 253                 goto bad_offset;
 254         if (!p->swap_map[offset])
 255                 goto bad_free;
 256         spin_lock(&swap_lock);
 257         return p;
 258
 259 bad_free:
 260         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 261         goto out;
 262 bad_offset:
 263         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 264         goto out;
 265 bad_device:
 266         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 267         goto out;
 268 bad_nofile:
 269         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 270 out:
 271         return NULL;
 272 }
 273
 274 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 275 {
 276         int count = p->swap_map[offset];
 277
 278         if (count < SWAP_MAP_MAX) {
 279                 count--;
 280                 p->swap_map[offset] = count;
 281                 if (!count) {
 282                         if (offset < p->lowest_bit)
 283                                 p->lowest_bit = offset;
 284                         if (offset > p->highest_bit)
 285                                 p->highest_bit = offset;
 286                         if (p->prio > swap_info[swap_list.next].prio)
 287                                 swap_list.next = p - swap_info;
 288                         nr_swap_pages++;
 289                         p->inuse_pages--;
 290                 }
 291         }
 292         return count;
 293 }
 294
 295 /*
 296  * Caller has made sure that the swapdevice corresponding to entry
 297  * is still around or has not been recycled.
 298  */
 299 void swap_free(swp_entry_t entry)
 300 {
 301         struct swap_info_struct * p;
 302
 303         p = swap_info_get(entry);
 304         if (p) {
 305                 swap_entry_free(p, swp_offset(entry));
 306                 spin_unlock(&swap_lock);
 307         }
 308 }
 309
 310 /*
 311  * How many references to page are currently swapped out?
 312  */
 313 static inline int page_swapcount(struct page *page)
 314 {
 315         int count = 0;
 316         struct swap_info_struct *p;
 317         swp_entry_t entry;
 318
 319         entry.val = page_private(page);
 320         p = swap_info_get(entry);
 321         if (p) {
 322                 /* Subtract the 1 for the swap cache itself */
 323                 count = p->swap_map[swp_offset(entry)] - 1;
 324                 spin_unlock(&swap_lock);
 325         }
 326         return count;
 327 }
 328
 329 /*
 330  * We can use this swap cache entry directly
 331  * if there are no other references to it.
 332  */
 333 int can_share_swap_page(struct page *page)
 334 {
 335         int count;
 336
 337         BUG_ON(!PageLocked(page));
 338         count = page_mapcount(page);
 339         if (count <= 1 && PageSwapCache(page))
 340                 count += page_swapcount(page);
 341         return count == 1;
 342 }
 343
 344 /*
 345  * Work out if there are any other processes sharing this
 346  * swap cache page. Free it if you can. Return success.
 347  */
 348 int remove_exclusive_swap_page(struct page *page)
 349 {
 350         int retval;
 351         struct swap_info_struct * p;
 352         swp_entry_t entry;
 353
 354         BUG_ON(PagePrivate(page));
 355         BUG_ON(!PageLocked(page));
 356
 357         if (!PageSwapCache(page))
 358                 return 0;
 359         if (PageWriteback(page))
 360                 return 0;
 361         if (page_count(page) != 2) /* 2: us + cache */
 362                 return 0;
 363
 364         entry.val = page_private(page);
 365         p = swap_info_get(entry);
 366         if (!p)
 367                 return 0;
 368
 369         /* Is the only swap cache user the cache itself? */
 370         retval = 0;
 371         if (p->swap_map[swp_offset(entry)] == 1) {
 372                 /* Recheck the page count with the swapcache lock held.. */
 373                 write_lock_irq(&swapper_space.tree_lock);
 374                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 375                         __delete_from_swap_cache(page);
 376                         SetPageDirty(page);
 377                         retval = 1;
 378                 }
 379                 write_unlock_irq(&swapper_space.tree_lock);
 380         }
 381         spin_unlock(&swap_lock);
 382
 383         if (retval) {
 384                 swap_free(entry);
 385                 page_cache_release(page);
 386         }
 387
 388         return retval;
 389 }
 390
 391 /*
 392  * Free the swap entry like above, but also try to
 393  * free the page cache entry if it is the last user.
 394  */
 395 void free_swap_and_cache(swp_entry_t entry)
 396 {
 397         struct swap_info_struct * p;
 398         struct page *page = NULL;
 399
 400         p = swap_info_get(entry);
 401         if (p) {
 402                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 403                         page = find_get_page(&swapper_space, entry.val);
 404                         if (page && unlikely(TestSetPageLocked(page))) {
 405                                 page_cache_release(page);
 406                                 page = NULL;
 407                         }
 408                 }
 409                 spin_unlock(&swap_lock);
 410         }
 411         if (page) {
 412                 int one_user;
 413
 414                 BUG_ON(PagePrivate(page));
 415                 one_user = (page_count(page) == 2);
 416                 /* Only cache user (+us), or swap space full? Free it! */
 417                 /* Also recheck PageSwapCache after page is locked (above) */
 418                 if (PageSwapCache(page) && !PageWriteback(page) &&
 419                                         (one_user || vm_swap_full())) {
 420                         delete_from_swap_cache(page);
 421                         SetPageDirty(page);
 422                 }
 423                 unlock_page(page);
 424                 page_cache_release(page);
 425         }
 426 }
 427
 428 #ifdef CONFIG_SOFTWARE_SUSPEND
 429 /*
 430  * Find the swap type that corresponds to given device (if any)
 431  *
 432  * This is needed for software suspend and is done in such a way that inode
 433  * aliasing is allowed.
 434  */
 435 int swap_type_of(dev_t device)
 436 {
 437         int i;
 438
 439         spin_lock(&swap_lock);
 440         for (i = 0; i < nr_swapfiles; i++) {
 441                 struct inode *inode;
 442
 443                 if (!(swap_info[i].flags & SWP_WRITEOK))
 444                         continue;
 445
 446                 if (!device) {
 447                         spin_unlock(&swap_lock);
 448                         return i;
 449                 }
 450                 inode = swap_info[i].swap_file->f_dentry->d_inode;
 451                 if (S_ISBLK(inode->i_mode) &&
 452                     device == MKDEV(imajor(inode), iminor(inode))) {
 453                         spin_unlock(&swap_lock);
 454                         return i;
 455                 }
 456         }
 457         spin_unlock(&swap_lock);
 458         return -ENODEV;
 459 }
 460
 461 /*
 462  * Return either the total number of swap pages of given type, or the number
 463  * of free pages of that type (depending on @free)
 464  *
 465  * This is needed for software suspend
 466  */
 467 unsigned int count_swap_pages(int type, int free)
 468 {
 469         unsigned int n = 0;
 470
 471         if (type < nr_swapfiles) {
 472                 spin_lock(&swap_lock);
 473                 if (swap_info[type].flags & SWP_WRITEOK) {
 474                         n = swap_info[type].pages;
 475                         if (free)
 476                                 n -= swap_info[type].inuse_pages;
 477                 }
 478                 spin_unlock(&swap_lock);
 479         }
 480         return n;
 481 }
 482 #endif
 483
 484 /*
 485  * No need to decide whether this PTE shares the swap entry with others,
 486  * just let do_wp_page work it out if a write is requested later - to
 487  * force COW, vm_page_prot omits write permission from any private vma.
 488  */
 489 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
 490                 unsigned long addr, swp_entry_t entry, struct page *page)
 491 {
 492         inc_mm_counter(vma->vm_mm, anon_rss);
 493         get_page(page);
 494         set_pte_at(vma->vm_mm, addr, pte,
 495                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 496         page_add_anon_rmap(page, vma, addr);
 497         swap_free(entry);
 498         /*
 499          * Move the page to the active list so it is not
 500          * immediately swapped out again after swapon.
 501          */
 502         activate_page(page);
 503 }
 504
 505 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 506                                 unsigned long addr, unsigned long end,
 507                                 swp_entry_t entry, struct page *page)
 508 {
 509         pte_t swp_pte = swp_entry_to_pte(entry);
 510         pte_t *pte;
 511         spinlock_t *ptl;
 512         int found = 0;
 513
 514         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 515         do {
 516                 /*
 517                  * swapoff spends a _lot_ of time in this loop!
 518                  * Test inline before going to call unuse_pte.
 519                  */
 520                 if (unlikely(pte_same(*pte, swp_pte))) {
 521                         unuse_pte(vma, pte++, addr, entry, page);
 522                         found = 1;
 523                         break;
 524                 }
 525         } while (pte++, addr += PAGE_SIZE, addr != end);
 526         pte_unmap_unlock(pte - 1, ptl);
 527         return found;
 528 }
 529
 530 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 531                                 unsigned long addr, unsigned long end,
 532                                 swp_entry_t entry, struct page *page)
 533 {
 534         pmd_t *pmd;
 535         unsigned long next;
 536
 537         pmd = pmd_offset(pud, addr);
 538         do {
 539                 next = pmd_addr_end(addr, end);
 540                 if (pmd_none_or_clear_bad(pmd))
 541                         continue;
 542                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 543                         return 1;
 544         } while (pmd++, addr = next, addr != end);
 545         return 0;
 546 }
 547
 548 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 549                                 unsigned long addr, unsigned long end,
 550                                 swp_entry_t entry, struct page *page)
 551 {
 552         pud_t *pud;
 553         unsigned long next;
 554
 555         pud = pud_offset(pgd, addr);
 556         do {
 557                 next = pud_addr_end(addr, end);
 558                 if (pud_none_or_clear_bad(pud))
 559                         continue;
 560                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 561                         return 1;
 562         } while (pud++, addr = next, addr != end);
 563         return 0;
 564 }
 565
 566 static int unuse_vma(struct vm_area_struct *vma,
 567                                 swp_entry_t entry, struct page *page)
 568 {
 569         pgd_t *pgd;
 570         unsigned long addr, end, next;
 571
 572         if (page->mapping) {
 573                 addr = page_address_in_vma(page, vma);
 574                 if (addr == -EFAULT)
 575                         return 0;
 576                 else
 577                         end = addr + PAGE_SIZE;
 578         } else {
 579                 addr = vma->vm_start;
 580                 end = vma->vm_end;
 581         }
 582
 583         pgd = pgd_offset(vma->vm_mm, addr);
 584         do {
 585                 next = pgd_addr_end(addr, end);
 586                 if (pgd_none_or_clear_bad(pgd))
 587                         continue;
 588                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 589                         return 1;
 590         } while (pgd++, addr = next, addr != end);
 591         return 0;
 592 }
 593
 594 static int unuse_mm(struct mm_struct *mm,
 595                                 swp_entry_t entry, struct page *page)
 596 {
 597         struct vm_area_struct *vma;
 598
 599         if (!down_read_trylock(&mm->mmap_sem)) {
 600                 /*
 601                  * Activate page so shrink_cache is unlikely to unmap its
 602                  * ptes while lock is dropped, so swapoff can make progress.
 603                  */
 604                 activate_page(page);
 605                 unlock_page(page);
 606                 down_read(&mm->mmap_sem);
 607                 lock_page(page);
 608         }
 609         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 610                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 611                         break;
 612         }
 613         up_read(&mm->mmap_sem);
 614         /*
 615          * Currently unuse_mm cannot fail, but leave error handling
 616          * at call sites for now, since we change it from time to time.
 617          */
 618         return 0;
 619 }
 620
 621 #ifdef CONFIG_MIGRATION
 622 int remove_vma_swap(struct vm_area_struct *vma, struct page *page)
 623 {
 624         swp_entry_t entry = { .val = page_private(page) };
 625
 626         return unuse_vma(vma, entry, page);
 627 }
 628 #endif
 629
 630 /*
 631  * Scan swap_map from current position to next entry still in use.
 632  * Recycle to start on reaching the end, returning 0 when empty.
 633  */
 634 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 635                                         unsigned int prev)
 636 {
 637         unsigned int max = si->max;
 638         unsigned int i = prev;
 639         int count;
 640
 641         /*
 642          * No need for swap_lock here: we're just looking
 643          * for whether an entry is in use, not modifying it; false
 644          * hits are okay, and sys_swapoff() has already prevented new
 645          * allocations from this area (while holding swap_lock).
 646          */
 647         for (;;) {
 648                 if (++i >= max) {
 649                         if (!prev) {
 650                                 i = 0;
 651                                 break;
 652                         }
 653                         /*
 654                          * No entries in use at top of swap_map,
 655                          * loop back to start and recheck there.
 656                          */
 657                         max = prev + 1;
 658                         prev = 0;
 659                         i = 1;
 660                 }
 661                 count = si->swap_map[i];
 662                 if (count && count != SWAP_MAP_BAD)
 663                         break;
 664         }
 665         return i;
 666 }
 667
 668 /*
 669  * We completely avoid races by reading each swap page in advance,
 670  * and then search for the process using it.  All the necessary
 671  * page table adjustments can then be made atomically.
 672  */
 673 static int try_to_unuse(unsigned int type)
 674 {
 675         struct swap_info_struct * si = &swap_info[type];
 676         struct mm_struct *start_mm;
 677         unsigned short *swap_map;
 678         unsigned short swcount;
 679         struct page *page;
 680         swp_entry_t entry;
 681         unsigned int i = 0;
 682         int retval = 0;
 683         int reset_overflow = 0;
 684         int shmem;
 685
 686         /*
 687          * When searching mms for an entry, a good strategy is to
 688          * start at the first mm we freed the previous entry from
 689          * (though actually we don't notice whether we or coincidence
 690          * freed the entry).  Initialize this start_mm with a hold.
 691          *
 692          * A simpler strategy would be to start at the last mm we
 693          * freed the previous entry from; but that would take less
 694          * advantage of mmlist ordering, which clusters forked mms
 695          * together, child after parent.  If we race with dup_mmap(), we
 696          * prefer to resolve parent before child, lest we miss entries
 697          * duplicated after we scanned child: using last mm would invert
 698          * that.  Though it's only a serious concern when an overflowed
 699          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 700          */
 701         start_mm = &init_mm;
 702         atomic_inc(&init_mm.mm_users);
 703
 704         /*
 705          * Keep on scanning until all entries have gone.  Usually,
 706          * one pass through swap_map is enough, but not necessarily:
 707          * there are races when an instance of an entry might be missed.
 708          */
 709         while ((i = find_next_to_unuse(si, i)) != 0) {
 710                 if (signal_pending(current)) {
 711                         retval = -EINTR;
 712                         break;
 713                 }
 714
 715                 /*
 716                  * Get a page for the entry, using the existing swap
 717                  * cache page if there is one.  Otherwise, get a clean
 718                  * page and read the swap into it.
 719                  */
 720                 swap_map = &si->swap_map[i];
 721                 entry = swp_entry(type, i);
 722 again:
 723                 page = read_swap_cache_async(entry, NULL, 0);
 724                 if (!page) {
 725                         /*
 726                          * Either swap_duplicate() failed because entry
 727                          * has been freed independently, and will not be
 728                          * reused since sys_swapoff() already disabled
 729                          * allocation from here, or alloc_page() failed.
 730                          */
 731                         if (!*swap_map)
 732                                 continue;
 733                         retval = -ENOMEM;
 734                         break;
 735                 }
 736
 737                 /*
 738                  * Don't hold on to start_mm if it looks like exiting.
 739                  */
 740                 if (atomic_read(&start_mm->mm_users) == 1) {
 741                         mmput(start_mm);
 742                         start_mm = &init_mm;
 743                         atomic_inc(&init_mm.mm_users);
 744                 }
 745
 746                 /*
 747                  * Wait for and lock page.  When do_swap_page races with
 748                  * try_to_unuse, do_swap_page can handle the fault much
 749                  * faster than try_to_unuse can locate the entry.  This
 750                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 751                  * defer to do_swap_page in such a case - in some tests,
 752                  * do_swap_page and try_to_unuse repeatedly compete.
 753                  */
 754                 wait_on_page_locked(page);
 755                 wait_on_page_writeback(page);
 756                 lock_page(page);
 757                 if (!PageSwapCache(page)) {
 758                         /* Page migration has occured */
 759                         unlock_page(page);
 760                         page_cache_release(page);
 761                         goto again;
 762                 }
 763                 wait_on_page_writeback(page);
 764
 765                 /*
 766                  * Remove all references to entry.
 767                  * Whenever we reach init_mm, there's no address space
 768                  * to search, but use it as a reminder to search shmem.
 769                  */
 770                 shmem = 0;
 771                 swcount = *swap_map;
 772                 if (swcount > 1) {
 773                         if (start_mm == &init_mm)
 774                                 shmem = shmem_unuse(entry, page);
 775                         else
 776                                 retval = unuse_mm(start_mm, entry, page);
 777                 }
 778                 if (*swap_map > 1) {
 779                         int set_start_mm = (*swap_map >= swcount);
 780                         struct list_head *p = &start_mm->mmlist;
 781                         struct mm_struct *new_start_mm = start_mm;
 782                         struct mm_struct *prev_mm = start_mm;
 783                         struct mm_struct *mm;
 784
 785                         atomic_inc(&new_start_mm->mm_users);
 786                         atomic_inc(&prev_mm->mm_users);
 787                         spin_lock(&mmlist_lock);
 788                         while (*swap_map > 1 && !retval &&
 789                                         (p = p->next) != &start_mm->mmlist) {
 790                                 mm = list_entry(p, struct mm_struct, mmlist);
 791                                 if (atomic_inc_return(&mm->mm_users) == 1) {
 792                                         atomic_dec(&mm->mm_users);
 793                                         continue;
 794                                 }
 795                                 spin_unlock(&mmlist_lock);
 796                                 mmput(prev_mm);
 797                                 prev_mm = mm;
 798
 799                                 cond_resched();
 800
 801                                 swcount = *swap_map;
 802                                 if (swcount <= 1)
 803                                         ;
 804                                 else if (mm == &init_mm) {
 805                                         set_start_mm = 1;
 806                                         shmem = shmem_unuse(entry, page);
 807                                 } else
 808                                         retval = unuse_mm(mm, entry, page);
 809                                 if (set_start_mm && *swap_map < swcount) {
 810                                         mmput(new_start_mm);
 811                                         atomic_inc(&mm->mm_users);
 812                                         new_start_mm = mm;
 813                                         set_start_mm = 0;
 814                                 }
 815                                 spin_lock(&mmlist_lock);
 816                         }
 817                         spin_unlock(&mmlist_lock);
 818                         mmput(prev_mm);
 819                         mmput(start_mm);
 820                         start_mm = new_start_mm;
 821                 }
 822                 if (retval) {
 823                         unlock_page(page);
 824                         page_cache_release(page);
 825                         break;
 826                 }
 827
 828                 /*
 829                  * How could swap count reach 0x7fff when the maximum
 830                  * pid is 0x7fff, and there's no way to repeat a swap
 831                  * page within an mm (except in shmem, where it's the
 832                  * shared object which takes the reference count)?
 833                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 834                  *
 835                  * If that's wrong, then we should worry more about
 836                  * exit_mmap() and do_munmap() cases described above:
 837                  * we might be resetting SWAP_MAP_MAX too early here.
 838                  * We know "Undead"s can happen, they're okay, so don't
 839                  * report them; but do report if we reset SWAP_MAP_MAX.
 840                  */
 841                 if (*swap_map == SWAP_MAP_MAX) {
 842                         spin_lock(&swap_lock);
 843                         *swap_map = 1;
 844                         spin_unlock(&swap_lock);
 845                         reset_overflow = 1;
 846                 }
 847
 848                 /*
 849                  * If a reference remains (rare), we would like to leave
 850                  * the page in the swap cache; but try_to_unmap could
 851                  * then re-duplicate the entry once we drop page lock,
 852                  * so we might loop indefinitely; also, that page could
 853                  * not be swapped out to other storage meanwhile.  So:
 854                  * delete from cache even if there's another reference,
 855                  * after ensuring that the data has been saved to disk -
 856                  * since if the reference remains (rarer), it will be
 857                  * read from disk into another page.  Splitting into two
 858                  * pages would be incorrect if swap supported "shared
 859                  * private" pages, but they are handled by tmpfs files.
 860                  *
 861                  * Note shmem_unuse already deleted a swappage from
 862                  * the swap cache, unless the move to filepage failed:
 863                  * in which case it left swappage in cache, lowered its
 864                  * swap count to pass quickly through the loops above,
 865                  * and now we must reincrement count to try again later.
 866                  */
 867                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 868                         struct writeback_control wbc = {
 869                                 .sync_mode = WB_SYNC_NONE,
 870                         };
 871
 872                         swap_writepage(page, &wbc);
 873                         lock_page(page);
 874                         wait_on_page_writeback(page);
 875                 }
 876                 if (PageSwapCache(page)) {
 877                         if (shmem)
 878                                 swap_duplicate(entry);
 879                         else
 880                                 delete_from_swap_cache(page);
 881                 }
 882
 883                 /*
 884                  * So we could skip searching mms once swap count went
 885                  * to 1, we did not mark any present ptes as dirty: must
 886                  * mark page dirty so shrink_list will preserve it.
 887                  */
 888                 SetPageDirty(page);
 889                 unlock_page(page);
 890                 page_cache_release(page);
 891
 892                 /*
 893                  * Make sure that we aren't completely killing
 894                  * interactive performance.
 895                  */
 896                 cond_resched();
 897         }
 898
 899         mmput(start_mm);
 900         if (reset_overflow) {
 901                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 902                 swap_overflow = 0;
 903         }
 904         return retval;
 905 }
 906
 907 /*
 908  * After a successful try_to_unuse, if no swap is now in use, we know
 909  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 910  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 911  * added to the mmlist just after page_duplicate - before would be racy.
 912  */
 913 static void drain_mmlist(void)
 914 {
 915         struct list_head *p, *next;
 916         unsigned int i;
 917
 918         for (i = 0; i < nr_swapfiles; i++)
 919                 if (swap_info[i].inuse_pages)
 920                         return;
 921         spin_lock(&mmlist_lock);
 922         list_for_each_safe(p, next, &init_mm.mmlist)
 923                 list_del_init(p);
 924         spin_unlock(&mmlist_lock);
 925 }
 926
 927 /*
 928  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 929  * corresponds to page offset `offset'.
 930  */
 931 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 932 {
 933         struct swap_extent *se = sis->curr_swap_extent;
 934         struct swap_extent *start_se = se;
 935
 936         for ( ; ; ) {
 937                 struct list_head *lh;
 938
 939                 if (se->start_page <= offset &&
 940                                 offset < (se->start_page + se->nr_pages)) {
 941                         return se->start_block + (offset - se->start_page);
 942                 }
 943                 lh = se->list.next;
 944                 if (lh == &sis->extent_list)
 945                         lh = lh->next;
 946                 se = list_entry(lh, struct swap_extent, list);
 947                 sis->curr_swap_extent = se;
 948                 BUG_ON(se == start_se);         /* It *must* be present */
 949         }
 950 }
 951
 952 /*
 953  * Free all of a swapdev's extent information
 954  */
 955 static void destroy_swap_extents(struct swap_info_struct *sis)
 956 {
 957         while (!list_empty(&sis->extent_list)) {
 958                 struct swap_extent *se;
 959
 960                 se = list_entry(sis->extent_list.next,
 961                                 struct swap_extent, list);
 962                 list_del(&se->list);
 963                 kfree(se);
 964         }
 965 }
 966
 967 /*
 968  * Add a block range (and the corresponding page range) into this swapdev's
 969  * extent list.  The extent list is kept sorted in page order.
 970  *
 971  * This function rather assumes that it is called in ascending page order.
 972  */
 973 static int
 974 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 975                 unsigned long nr_pages, sector_t start_block)
 976 {
 977         struct swap_extent *se;
 978         struct swap_extent *new_se;
 979         struct list_head *lh;
 980
 981         lh = sis->extent_list.prev;     /* The highest page extent */
 982         if (lh != &sis->extent_list) {
 983                 se = list_entry(lh, struct swap_extent, list);
 984                 BUG_ON(se->start_page + se->nr_pages != start_page);
 985                 if (se->start_block + se->nr_pages == start_block) {
 986                         /* Merge it */
 987                         se->nr_pages += nr_pages;
 988                         return 0;
 989                 }
 990         }
 991
 992         /*
 993          * No merge.  Insert a new extent, preserving ordering.
 994          */
 995         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 996         if (new_se == NULL)
 997                 return -ENOMEM;
 998         new_se->start_page = start_page;
 999         new_se->nr_pages = nr_pages;
1000         new_se->start_block = start_block;
1001
1002         list_add_tail(&new_se->list, &sis->extent_list);
1003         return 1;
1004 }
1005
1006 /*
1007  * A `swap extent' is a simple thing which maps a contiguous range of pages
1008  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1009  * is built at swapon time and is then used at swap_writepage/swap_readpage
1010  * time for locating where on disk a page belongs.
1011  *
1012  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1013  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1014  * swap files identically.
1015  *
1016  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1017  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1018  * swapfiles are handled *identically* after swapon time.
1019  *
1020  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1021  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1022  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1023  * requirements, they are simply tossed out - we will never use those blocks
1024  * for swapping.
1025  *
1026  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1027  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1028  * which will scribble on the fs.
1029  *
1030  * The amount of disk space which a single swap extent represents varies.
1031  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1032  * extents in the list.  To avoid much list walking, we cache the previous
1033  * search location in `curr_swap_extent', and start new searches from there.
1034  * This is extremely effective.  The average number of iterations in
1035  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1036  */
1037 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1038 {
1039         struct inode *inode;
1040         unsigned blocks_per_page;
1041         unsigned long page_no;
1042         unsigned blkbits;
1043         sector_t probe_block;
1044         sector_t last_block;
1045         sector_t lowest_block = -1;
1046         sector_t highest_block = 0;
1047         int nr_extents = 0;
1048         int ret;
1049
1050         inode = sis->swap_file->f_mapping->host;
1051         if (S_ISBLK(inode->i_mode)) {
1052                 ret = add_swap_extent(sis, 0, sis->max, 0);
1053                 *span = sis->pages;
1054                 goto done;
1055         }
1056
1057         blkbits = inode->i_blkbits;
1058         blocks_per_page = PAGE_SIZE >> blkbits;
1059
1060         /*
1061          * Map all the blocks into the extent list.  This code doesn't try
1062          * to be very smart.
1063          */
1064         probe_block = 0;
1065         page_no = 0;
1066         last_block = i_size_read(inode) >> blkbits;
1067         while ((probe_block + blocks_per_page) <= last_block &&
1068                         page_no < sis->max) {
1069                 unsigned block_in_page;
1070                 sector_t first_block;
1071
1072                 first_block = bmap(inode, probe_block);
1073                 if (first_block == 0)
1074                         goto bad_bmap;
1075
1076                 /*
1077                  * It must be PAGE_SIZE aligned on-disk
1078                  */
1079                 if (first_block & (blocks_per_page - 1)) {
1080                         probe_block++;
1081                         goto reprobe;
1082                 }
1083
1084                 for (block_in_page = 1; block_in_page < blocks_per_page;
1085                                         block_in_page++) {
1086                         sector_t block;
1087
1088                         block = bmap(inode, probe_block + block_in_page);
1089                         if (block == 0)
1090                                 goto bad_bmap;
1091                         if (block != first_block + block_in_page) {
1092                                 /* Discontiguity */
1093                                 probe_block++;
1094                                 goto reprobe;
1095                         }
1096                 }
1097
1098                 first_block >>= (PAGE_SHIFT - blkbits);
1099                 if (page_no) {  /* exclude the header page */
1100                         if (first_block < lowest_block)
1101                                 lowest_block = first_block;
1102                         if (first_block > highest_block)
1103                                 highest_block = first_block;
1104                 }
1105
1106                 /*
1107                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1108                  */
1109                 ret = add_swap_extent(sis, page_no, 1, first_block);
1110                 if (ret < 0)
1111                         goto out;
1112                 nr_extents += ret;
1113                 page_no++;
1114                 probe_block += blocks_per_page;
1115 reprobe:
1116                 continue;
1117         }
1118         ret = nr_extents;
1119         *span = 1 + highest_block - lowest_block;
1120         if (page_no == 0)
1121                 page_no = 1;    /* force Empty message */
1122         sis->max = page_no;
1123         sis->pages = page_no - 1;
1124         sis->highest_bit = page_no - 1;
1125 done:
1126         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1127                                         struct swap_extent, list);
1128         goto out;
1129 bad_bmap:
1130         printk(KERN_ERR "swapon: swapfile has holes\n");
1131         ret = -EINVAL;
1132 out:
1133         return ret;
1134 }
1135
1136 #if 0   /* We don't need this yet */
1137 #include <linux/backing-dev.h>
1138 int page_queue_congested(struct page *page)
1139 {
1140         struct backing_dev_info *bdi;
1141
1142         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1143
1144         if (PageSwapCache(page)) {
1145                 swp_entry_t entry = { .val = page_private(page) };
1146                 struct swap_info_struct *sis;
1147
1148                 sis = get_swap_info_struct(swp_type(entry));
1149                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1150         } else
1151                 bdi = page->mapping->backing_dev_info;
1152         return bdi_write_congested(bdi);
1153 }
1154 #endif
1155
1156 asmlinkage long sys_swapoff(const char __user * specialfile)
1157 {
1158         struct swap_info_struct * p = NULL;
1159         unsigned short *swap_map;
1160         struct file *swap_file, *victim;
1161         struct address_space *mapping;
1162         struct inode *inode;
1163         char * pathname;
1164         int i, type, prev;
1165         int err;
1166
1167         if (!capable(CAP_SYS_ADMIN))
1168                 return -EPERM;
1169
1170         pathname = getname(specialfile);
1171         err = PTR_ERR(pathname);
1172         if (IS_ERR(pathname))
1173                 goto out;
1174
1175         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1176         putname(pathname);
1177         err = PTR_ERR(victim);
1178         if (IS_ERR(victim))
1179                 goto out;
1180
1181         mapping = victim->f_mapping;
1182         prev = -1;
1183         spin_lock(&swap_lock);
1184         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1185                 p = swap_info + type;
1186                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1187                         if (p->swap_file->f_mapping == mapping)
1188                                 break;
1189                 }
1190                 prev = type;
1191         }
1192         if (type < 0) {
1193                 err = -EINVAL;
1194                 spin_unlock(&swap_lock);
1195                 goto out_dput;
1196         }
1197         if (!security_vm_enough_memory(p->pages))
1198                 vm_unacct_memory(p->pages);
1199         else {
1200                 err = -ENOMEM;
1201                 spin_unlock(&swap_lock);
1202                 goto out_dput;
1203         }
1204         if (prev < 0) {
1205                 swap_list.head = p->next;
1206         } else {
1207                 swap_info[prev].next = p->next;
1208         }
1209         if (type == swap_list.next) {
1210                 /* just pick something that's safe... */
1211                 swap_list.next = swap_list.head;
1212         }
1213         nr_swap_pages -= p->pages;
1214         total_swap_pages -= p->pages;
1215         p->flags &= ~SWP_WRITEOK;
1216         spin_unlock(&swap_lock);
1217
1218         current->flags |= PF_SWAPOFF;
1219         err = try_to_unuse(type);
1220         current->flags &= ~PF_SWAPOFF;
1221
1222         if (err) {
1223                 /* re-insert swap space back into swap_list */
1224                 spin_lock(&swap_lock);
1225                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1226                         if (p->prio >= swap_info[i].prio)
1227                                 break;
1228                 p->next = i;
1229                 if (prev < 0)
1230                         swap_list.head = swap_list.next = p - swap_info;
1231                 else
1232                         swap_info[prev].next = p - swap_info;
1233                 nr_swap_pages += p->pages;
1234                 total_swap_pages += p->pages;
1235                 p->flags |= SWP_WRITEOK;
1236                 spin_unlock(&swap_lock);
1237                 goto out_dput;
1238         }
1239
1240         /* wait for any unplug function to finish */
1241         down_write(&swap_unplug_sem);
1242         up_write(&swap_unplug_sem);
1243
1244         destroy_swap_extents(p);
1245         mutex_lock(&swapon_mutex);
1246         spin_lock(&swap_lock);
1247         drain_mmlist();
1248
1249         /* wait for anyone still in scan_swap_map */
1250         p->highest_bit = 0;             /* cuts scans short */
1251         while (p->flags >= SWP_SCANNING) {
1252                 spin_unlock(&swap_lock);
1253                 schedule_timeout_uninterruptible(1);
1254                 spin_lock(&swap_lock);
1255         }
1256
1257         swap_file = p->swap_file;
1258         p->swap_file = NULL;
1259         p->max = 0;
1260         swap_map = p->swap_map;
1261         p->swap_map = NULL;
1262         p->flags = 0;
1263         spin_unlock(&swap_lock);
1264         mutex_unlock(&swapon_mutex);
1265         vfree(swap_map);
1266         inode = mapping->host;
1267         if (S_ISBLK(inode->i_mode)) {
1268                 struct block_device *bdev = I_BDEV(inode);
1269                 set_blocksize(bdev, p->old_block_size);
1270                 bd_release(bdev);
1271         } else {
1272                 mutex_lock(&inode->i_mutex);
1273                 inode->i_flags &= ~S_SWAPFILE;
1274                 mutex_unlock(&inode->i_mutex);
1275         }
1276         filp_close(swap_file, NULL);
1277         err = 0;
1278
1279 out_dput:
1280         filp_close(victim, NULL);
1281 out:
1282         return err;
1283 }
1284
1285 #ifdef CONFIG_PROC_FS
1286 /* iterator */
1287 static void *swap_start(struct seq_file *swap, loff_t *pos)
1288 {
1289         struct swap_info_struct *ptr = swap_info;
1290         int i;
1291         loff_t l = *pos;
1292
1293         mutex_lock(&swapon_mutex);
1294
1295         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1296                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1297                         continue;
1298                 if (!l--)
1299                         return ptr;
1300         }
1301
1302         return NULL;
1303 }
1304
1305 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1306 {
1307         struct swap_info_struct *ptr = v;
1308         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1309
1310         for (++ptr; ptr < endptr; ptr++) {
1311                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1312                         continue;
1313                 ++*pos;
1314                 return ptr;
1315         }
1316
1317         return NULL;
1318 }
1319
1320 static void swap_stop(struct seq_file *swap, void *v)
1321 {
1322         mutex_unlock(&swapon_mutex);
1323 }
1324
1325 static int swap_show(struct seq_file *swap, void *v)
1326 {
1327         struct swap_info_struct *ptr = v;
1328         struct file *file;
1329         int len;
1330
1331         if (v == swap_info)
1332                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1333
1334         file = ptr->swap_file;
1335         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1336         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1337                        len < 40 ? 40 - len : 1, " ",
1338                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1339                                 "partition" : "file\t",
1340                        ptr->pages << (PAGE_SHIFT - 10),
1341                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1342                        ptr->prio);
1343         return 0;
1344 }
1345
1346 static struct seq_operations swaps_op = {
1347         .start =        swap_start,
1348         .next =         swap_next,
1349         .stop =         swap_stop,
1350         .show =         swap_show
1351 };
1352
1353 static int swaps_open(struct inode *inode, struct file *file)
1354 {
1355         return seq_open(file, &swaps_op);
1356 }
1357
1358 static struct file_operations proc_swaps_operations = {
1359         .open           = swaps_open,
1360         .read           = seq_read,
1361         .llseek         = seq_lseek,
1362         .release        = seq_release,
1363 };
1364
1365 static int __init procswaps_init(void)
1366 {
1367         struct proc_dir_entry *entry;
1368
1369         entry = create_proc_entry("swaps", 0, NULL);
1370         if (entry)
1371                 entry->proc_fops = &proc_swaps_operations;
1372         return 0;
1373 }
1374 __initcall(procswaps_init);
1375 #endif /* CONFIG_PROC_FS */
1376
1377 /*
1378  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1379  *
1380  * The swapon system call
1381  */
1382 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1383 {
1384         struct swap_info_struct * p;
1385         char *name = NULL;
1386         struct block_device *bdev = NULL;
1387         struct file *swap_file = NULL;
1388         struct address_space *mapping;
1389         unsigned int type;
1390         int i, prev;
1391         int error;
1392         static int least_priority;
1393         union swap_header *swap_header = NULL;
1394         int swap_header_version;
1395         unsigned int nr_good_pages = 0;
1396         int nr_extents = 0;
1397         sector_t span;
1398         unsigned long maxpages = 1;
1399         int swapfilesize;
1400         unsigned short *swap_map;
1401         struct page *page = NULL;
1402         struct inode *inode = NULL;
1403         int did_down = 0;
1404
1405         if (!capable(CAP_SYS_ADMIN))
1406                 return -EPERM;
1407         spin_lock(&swap_lock);
1408         p = swap_info;
1409         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1410                 if (!(p->flags & SWP_USED))
1411                         break;
1412         error = -EPERM;
1413         /*
1414          * Test if adding another swap device is possible. There are
1415          * two limiting factors: 1) the number of bits for the swap
1416          * type swp_entry_t definition and 2) the number of bits for
1417          * the swap type in the swap ptes as defined by the different
1418          * architectures. To honor both limitations a swap entry
1419          * with swap offset 0 and swap type ~0UL is created, encoded
1420          * to a swap pte, decoded to a swp_entry_t again and finally
1421          * the swap type part is extracted. This will mask all bits
1422          * from the initial ~0UL that can't be encoded in either the
1423          * swp_entry_t or the architecture definition of a swap pte.
1424          */
1425         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1426                 spin_unlock(&swap_lock);
1427                 goto out;
1428         }
1429         if (type >= nr_swapfiles)
1430                 nr_swapfiles = type+1;
1431         INIT_LIST_HEAD(&p->extent_list);
1432         p->flags = SWP_USED;
1433         p->swap_file = NULL;
1434         p->old_block_size = 0;
1435         p->swap_map = NULL;
1436         p->lowest_bit = 0;
1437         p->highest_bit = 0;
1438         p->cluster_nr = 0;
1439         p->inuse_pages = 0;
1440         p->next = -1;
1441         if (swap_flags & SWAP_FLAG_PREFER) {
1442                 p->prio =
1443                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1444         } else {
1445                 p->prio = --least_priority;
1446         }
1447         spin_unlock(&swap_lock);
1448         name = getname(specialfile);
1449         error = PTR_ERR(name);
1450         if (IS_ERR(name)) {
1451                 name = NULL;
1452                 goto bad_swap_2;
1453         }
1454         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1455         error = PTR_ERR(swap_file);
1456         if (IS_ERR(swap_file)) {
1457                 swap_file = NULL;
1458                 goto bad_swap_2;
1459         }
1460
1461         p->swap_file = swap_file;
1462         mapping = swap_file->f_mapping;
1463         inode = mapping->host;
1464
1465         error = -EBUSY;
1466         for (i = 0; i < nr_swapfiles; i++) {
1467                 struct swap_info_struct *q = &swap_info[i];
1468
1469                 if (i == type || !q->swap_file)
1470                         continue;
1471                 if (mapping == q->swap_file->f_mapping)
1472                         goto bad_swap;
1473         }
1474
1475         error = -EINVAL;
1476         if (S_ISBLK(inode->i_mode)) {
1477                 bdev = I_BDEV(inode);
1478                 error = bd_claim(bdev, sys_swapon);
1479                 if (error < 0) {
1480                         bdev = NULL;
1481                         error = -EINVAL;
1482                         goto bad_swap;
1483                 }
1484                 p->old_block_size = block_size(bdev);
1485                 error = set_blocksize(bdev, PAGE_SIZE);
1486                 if (error < 0)
1487                         goto bad_swap;
1488                 p->bdev = bdev;
1489         } else if (S_ISREG(inode->i_mode)) {
1490                 p->bdev = inode->i_sb->s_bdev;
1491                 mutex_lock(&inode->i_mutex);
1492                 did_down = 1;
1493                 if (IS_SWAPFILE(inode)) {
1494                         error = -EBUSY;
1495                         goto bad_swap;
1496                 }
1497         } else {
1498                 goto bad_swap;
1499         }
1500
1501         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1502
1503         /*
1504          * Read the swap header.
1505          */
1506         if (!mapping->a_ops->readpage) {
1507                 error = -EINVAL;
1508                 goto bad_swap;
1509         }
1510         page = read_cache_page(mapping, 0,
1511                         (filler_t *)mapping->a_ops->readpage, swap_file);
1512         if (IS_ERR(page)) {
1513                 error = PTR_ERR(page);
1514                 goto bad_swap;
1515         }
1516         wait_on_page_locked(page);
1517         if (!PageUptodate(page))
1518                 goto bad_swap;
1519         kmap(page);
1520         swap_header = page_address(page);
1521
1522         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1523                 swap_header_version = 1;
1524         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1525                 swap_header_version = 2;
1526         else {
1527                 printk(KERN_ERR "Unable to find swap-space signature\n");
1528                 error = -EINVAL;
1529                 goto bad_swap;
1530         }
1531
1532         switch (swap_header_version) {
1533         case 1:
1534                 printk(KERN_ERR "version 0 swap is no longer supported. "
1535                         "Use mkswap -v1 %s\n", name);
1536                 error = -EINVAL;
1537                 goto bad_swap;
1538         case 2:
1539                 /* Check the swap header's sub-version and the size of
1540                    the swap file and bad block lists */
1541                 if (swap_header->info.version != 1) {
1542                         printk(KERN_WARNING
1543                                "Unable to handle swap header version %d\n",
1544                                swap_header->info.version);
1545                         error = -EINVAL;
1546                         goto bad_swap;
1547                 }
1548
1549                 p->lowest_bit  = 1;
1550                 p->cluster_next = 1;
1551
1552                 /*
1553                  * Find out how many pages are allowed for a single swap
1554                  * device. There are two limiting factors: 1) the number of
1555                  * bits for the swap offset in the swp_entry_t type and
1556                  * 2) the number of bits in the a swap pte as defined by
1557                  * the different architectures. In order to find the
1558                  * largest possible bit mask a swap entry with swap type 0
1559                  * and swap offset ~0UL is created, encoded to a swap pte,
1560                  * decoded to a swp_entry_t again and finally the swap
1561                  * offset is extracted. This will mask all the bits from
1562                  * the initial ~0UL mask that can't be encoded in either
1563                  * the swp_entry_t or the architecture definition of a
1564                  * swap pte.
1565                  */
1566                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1567                 if (maxpages > swap_header->info.last_page)
1568                         maxpages = swap_header->info.last_page;
1569                 p->highest_bit = maxpages - 1;
1570
1571                 error = -EINVAL;
1572                 if (!maxpages)
1573                         goto bad_swap;
1574                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1575                         goto bad_swap;
1576                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1577                         goto bad_swap;
1578
1579                 /* OK, set up the swap map and apply the bad block list */
1580                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1581                         error = -ENOMEM;
1582                         goto bad_swap;
1583                 }
1584
1585                 error = 0;
1586                 memset(p->swap_map, 0, maxpages * sizeof(short));
1587                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1588                         int page_nr = swap_header->info.badpages[i];
1589                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1590                                 error = -EINVAL;
1591                         else
1592                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1593                 }
1594                 nr_good_pages = swap_header->info.last_page -
1595                                 swap_header->info.nr_badpages -
1596                                 1 /* header page */;
1597                 if (error)
1598                         goto bad_swap;
1599         }
1600
1601         if (swapfilesize && maxpages > swapfilesize) {
1602                 printk(KERN_WARNING
1603                        "Swap area shorter than signature indicates\n");
1604                 error = -EINVAL;
1605                 goto bad_swap;
1606         }
1607         if (nr_good_pages) {
1608                 p->swap_map[0] = SWAP_MAP_BAD;
1609                 p->max = maxpages;
1610                 p->pages = nr_good_pages;
1611                 nr_extents = setup_swap_extents(p, &span);
1612                 if (nr_extents < 0) {
1613                         error = nr_extents;
1614                         goto bad_swap;
1615                 }
1616                 nr_good_pages = p->pages;
1617         }
1618         if (!nr_good_pages) {
1619                 printk(KERN_WARNING "Empty swap-file\n");
1620                 error = -EINVAL;
1621                 goto bad_swap;
1622         }
1623
1624         mutex_lock(&swapon_mutex);
1625         spin_lock(&swap_lock);
1626         p->flags = SWP_ACTIVE;
1627         nr_swap_pages += nr_good_pages;
1628         total_swap_pages += nr_good_pages;
1629
1630         printk(KERN_INFO "Adding %uk swap on %s.  "
1631                         "Priority:%d extents:%d across:%lluk\n",
1632                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1633                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1634
1635         /* insert swap space into swap_list: */
1636         prev = -1;
1637         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1638                 if (p->prio >= swap_info[i].prio) {
1639                         break;
1640                 }
1641                 prev = i;
1642         }
1643         p->next = i;
1644         if (prev < 0) {
1645                 swap_list.head = swap_list.next = p - swap_info;
1646         } else {
1647                 swap_info[prev].next = p - swap_info;
1648         }
1649         spin_unlock(&swap_lock);
1650         mutex_unlock(&swapon_mutex);
1651         error = 0;
1652         goto out;
1653 bad_swap:
1654         if (bdev) {
1655                 set_blocksize(bdev, p->old_block_size);
1656                 bd_release(bdev);
1657         }
1658         destroy_swap_extents(p);
1659 bad_swap_2:
1660         spin_lock(&swap_lock);
1661         swap_map = p->swap_map;
1662         p->swap_file = NULL;
1663         p->swap_map = NULL;
1664         p->flags = 0;
1665         if (!(swap_flags & SWAP_FLAG_PREFER))
1666                 ++least_priority;
1667         spin_unlock(&swap_lock);
1668         vfree(swap_map);
1669         if (swap_file)
1670                 filp_close(swap_file, NULL);
1671 out:
1672         if (page && !IS_ERR(page)) {
1673                 kunmap(page);
1674                 page_cache_release(page);
1675         }
1676         if (name)
1677                 putname(name);
1678         if (did_down) {
1679                 if (!error)
1680                         inode->i_flags |= S_SWAPFILE;
1681                 mutex_unlock(&inode->i_mutex);
1682         }
1683         return error;
1684 }
1685
1686 void si_swapinfo(struct sysinfo *val)
1687 {
1688         unsigned int i;
1689         unsigned long nr_to_be_unused = 0;
1690
1691         spin_lock(&swap_lock);
1692         for (i = 0; i < nr_swapfiles; i++) {
1693                 if (!(swap_info[i].flags & SWP_USED) ||
1694                      (swap_info[i].flags & SWP_WRITEOK))
1695                         continue;
1696                 nr_to_be_unused += swap_info[i].inuse_pages;
1697         }
1698         val->freeswap = nr_swap_pages + nr_to_be_unused;
1699         val->totalswap = total_swap_pages + nr_to_be_unused;
1700         spin_unlock(&swap_lock);
1701         if (vx_flags(VXF_VIRT_MEM, 0))
1702                 vx_vsi_swapinfo(val);
1703 }
1704
1705 /*
1706  * Verify that a swap entry is valid and increment its swap map count.
1707  *
1708  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1709  * "permanent", but will be reclaimed by the next swapoff.
1710  */
1711 int swap_duplicate(swp_entry_t entry)
1712 {
1713         struct swap_info_struct * p;
1714         unsigned long offset, type;
1715         int result = 0;
1716
1717         type = swp_type(entry);
1718         if (type >= nr_swapfiles)
1719                 goto bad_file;
1720         p = type + swap_info;
1721         offset = swp_offset(entry);
1722
1723         spin_lock(&swap_lock);
1724         if (offset < p->max && p->swap_map[offset]) {
1725                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1726                         p->swap_map[offset]++;
1727                         result = 1;
1728                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1729                         if (swap_overflow++ < 5)
1730                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1731                         p->swap_map[offset] = SWAP_MAP_MAX;
1732                         result = 1;
1733                 }
1734         }
1735         spin_unlock(&swap_lock);
1736 out:
1737         return result;
1738
1739 bad_file:
1740         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1741         goto out;
1742 }
1743
1744 struct swap_info_struct *
1745 get_swap_info_struct(unsigned type)
1746 {
1747         return &swap_info[type];
1748 }
1749
1750 /*
1751  * swap_lock prevents swap_map being freed. Don't grab an extra
1752  * reference on the swaphandle, it doesn't matter if it becomes unused.
1753  */
1754 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1755 {
1756         int ret = 0, i = 1 << page_cluster;
1757         unsigned long toff;
1758         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1759
1760         if (!page_cluster)      /* no readahead */
1761                 return 0;
1762         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1763         if (!toff)              /* first page is swap header */
1764                 toff++, i--;
1765         *offset = toff;
1766
1767         spin_lock(&swap_lock);
1768         do {
1769                 /* Don't read-ahead past the end of the swap area */
1770                 if (toff >= swapdev->max)
1771                         break;
1772                 /* Don't read in free or bad pages */
1773                 if (!swapdev->swap_map[toff])
1774                         break;
1775                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1776                         break;
1777                 toff++;
1778                 ret++;
1779         } while (--i);
1780         spin_unlock(&swap_lock);
1781         return ret;
1782 }