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