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