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