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