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/mm.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/mman.h>
  11 #include <linux/slab.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/swap.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/namei.h>
  17 #include <linux/shm.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/writeback.h>
  20 #include <linux/proc_fs.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/init.h>
  23 #include <linux/module.h>
  24 #include <linux/rmap.h>
  25 #include <linux/security.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/mutex.h>
  28 #include <linux/capability.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/memcontrol.h>
  31
  32 #include <asm/pgtable.h>
  33 #include <asm/tlbflush.h>
  34 #include <linux/swapops.h>
  35
  36 static DEFINE_SPINLOCK(swap_lock);
  37 static unsigned int nr_swapfiles;
  38 long total_swap_pages;
  39 static int swap_overflow;
  40 static int least_priority;
  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 static struct swap_list_t swap_list = {-1, -1};
  48
  49 static 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                 spin_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                 spin_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         if (is_migration_entry(entry))
 400                 return;
 401
 402         p = swap_info_get(entry);
 403         if (p) {
 404                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 405                         page = find_get_page(&swapper_space, entry.val);
 406                         if (page && unlikely(TestSetPageLocked(page))) {
 407                                 page_cache_release(page);
 408                                 page = NULL;
 409                         }
 410                 }
 411                 spin_unlock(&swap_lock);
 412         }
 413         if (page) {
 414                 int one_user;
 415
 416                 BUG_ON(PagePrivate(page));
 417                 one_user = (page_count(page) == 2);
 418                 /* Only cache user (+us), or swap space full? Free it! */
 419                 /* Also recheck PageSwapCache after page is locked (above) */
 420                 if (PageSwapCache(page) && !PageWriteback(page) &&
 421                                         (one_user || vm_swap_full())) {
 422                         delete_from_swap_cache(page);
 423                         SetPageDirty(page);
 424                 }
 425                 unlock_page(page);
 426                 page_cache_release(page);
 427         }
 428 }
 429
 430 #ifdef CONFIG_HIBERNATION
 431 /*
 432  * Find the swap type that corresponds to given device (if any).
 433  *
 434  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 435  * from 0, in which the swap header is expected to be located.
 436  *
 437  * This is needed for the suspend to disk (aka swsusp).
 438  */
 439 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 440 {
 441         struct block_device *bdev = NULL;
 442         int i;
 443
 444         if (device)
 445                 bdev = bdget(device);
 446
 447         spin_lock(&swap_lock);
 448         for (i = 0; i < nr_swapfiles; i++) {
 449                 struct swap_info_struct *sis = swap_info + i;
 450
 451                 if (!(sis->flags & SWP_WRITEOK))
 452                         continue;
 453
 454                 if (!bdev) {
 455                         if (bdev_p)
 456                                 *bdev_p = sis->bdev;
 457
 458                         spin_unlock(&swap_lock);
 459                         return i;
 460                 }
 461                 if (bdev == sis->bdev) {
 462                         struct swap_extent *se;
 463
 464                         se = list_entry(sis->extent_list.next,
 465                                         struct swap_extent, list);
 466                         if (se->start_block == offset) {
 467                                 if (bdev_p)
 468                                         *bdev_p = sis->bdev;
 469
 470                                 spin_unlock(&swap_lock);
 471                                 bdput(bdev);
 472                                 return i;
 473                         }
 474                 }
 475         }
 476         spin_unlock(&swap_lock);
 477         if (bdev)
 478                 bdput(bdev);
 479
 480         return -ENODEV;
 481 }
 482
 483 /*
 484  * Return either the total number of swap pages of given type, or the number
 485  * of free pages of that type (depending on @free)
 486  *
 487  * This is needed for software suspend
 488  */
 489 unsigned int count_swap_pages(int type, int free)
 490 {
 491         unsigned int n = 0;
 492
 493         if (type < nr_swapfiles) {
 494                 spin_lock(&swap_lock);
 495                 if (swap_info[type].flags & SWP_WRITEOK) {
 496                         n = swap_info[type].pages;
 497                         if (free)
 498                                 n -= swap_info[type].inuse_pages;
 499                 }
 500                 spin_unlock(&swap_lock);
 501         }
 502         return n;
 503 }
 504 #endif
 505
 506 /*
 507  * No need to decide whether this PTE shares the swap entry with others,
 508  * just let do_wp_page work it out if a write is requested later - to
 509  * force COW, vm_page_prot omits write permission from any private vma.
 510  */
 511 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 512                 unsigned long addr, swp_entry_t entry, struct page *page)
 513 {
 514         spinlock_t *ptl;
 515         pte_t *pte;
 516         int ret = 1;
 517
 518         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
 519                 ret = -ENOMEM;
 520
 521         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 522         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 523                 if (ret > 0)
 524                         mem_cgroup_uncharge_page(page);
 525                 ret = 0;
 526                 goto out;
 527         }
 528
 529         inc_mm_counter(vma->vm_mm, anon_rss);
 530         get_page(page);
 531         set_pte_at(vma->vm_mm, addr, pte,
 532                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 533         page_add_anon_rmap(page, vma, addr);
 534         swap_free(entry);
 535         /*
 536          * Move the page to the active list so it is not
 537          * immediately swapped out again after swapon.
 538          */
 539         activate_page(page);
 540 out:
 541         pte_unmap_unlock(pte, ptl);
 542         return ret;
 543 }
 544
 545 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 546                                 unsigned long addr, unsigned long end,
 547                                 swp_entry_t entry, struct page *page)
 548 {
 549         pte_t swp_pte = swp_entry_to_pte(entry);
 550         pte_t *pte;
 551         int ret = 0;
 552
 553         /*
 554          * We don't actually need pte lock while scanning for swp_pte: since
 555          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 556          * page table while we're scanning; though it could get zapped, and on
 557          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 558          * of unmatched parts which look like swp_pte, so unuse_pte must
 559          * recheck under pte lock.  Scanning without pte lock lets it be
 560          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 561          */
 562         pte = pte_offset_map(pmd, addr);
 563         do {
 564                 /*
 565                  * swapoff spends a _lot_ of time in this loop!
 566                  * Test inline before going to call unuse_pte.
 567                  */
 568                 if (unlikely(pte_same(*pte, swp_pte))) {
 569                         pte_unmap(pte);
 570                         ret = unuse_pte(vma, pmd, addr, entry, page);
 571                         if (ret)
 572                                 goto out;
 573                         pte = pte_offset_map(pmd, addr);
 574                 }
 575         } while (pte++, addr += PAGE_SIZE, addr != end);
 576         pte_unmap(pte - 1);
 577 out:
 578         return ret;
 579 }
 580
 581 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 582                                 unsigned long addr, unsigned long end,
 583                                 swp_entry_t entry, struct page *page)
 584 {
 585         pmd_t *pmd;
 586         unsigned long next;
 587         int ret;
 588
 589         pmd = pmd_offset(pud, addr);
 590         do {
 591                 next = pmd_addr_end(addr, end);
 592                 if (pmd_none_or_clear_bad(pmd))
 593                         continue;
 594                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 595                 if (ret)
 596                         return ret;
 597         } while (pmd++, addr = next, addr != end);
 598         return 0;
 599 }
 600
 601 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 602                                 unsigned long addr, unsigned long end,
 603                                 swp_entry_t entry, struct page *page)
 604 {
 605         pud_t *pud;
 606         unsigned long next;
 607         int ret;
 608
 609         pud = pud_offset(pgd, addr);
 610         do {
 611                 next = pud_addr_end(addr, end);
 612                 if (pud_none_or_clear_bad(pud))
 613                         continue;
 614                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 615                 if (ret)
 616                         return ret;
 617         } while (pud++, addr = next, addr != end);
 618         return 0;
 619 }
 620
 621 static int unuse_vma(struct vm_area_struct *vma,
 622                                 swp_entry_t entry, struct page *page)
 623 {
 624         pgd_t *pgd;
 625         unsigned long addr, end, next;
 626         int ret;
 627
 628         if (page->mapping) {
 629                 addr = page_address_in_vma(page, vma);
 630                 if (addr == -EFAULT)
 631                         return 0;
 632                 else
 633                         end = addr + PAGE_SIZE;
 634         } else {
 635                 addr = vma->vm_start;
 636                 end = vma->vm_end;
 637         }
 638
 639         pgd = pgd_offset(vma->vm_mm, addr);
 640         do {
 641                 next = pgd_addr_end(addr, end);
 642                 if (pgd_none_or_clear_bad(pgd))
 643                         continue;
 644                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 645                 if (ret)
 646                         return ret;
 647         } while (pgd++, addr = next, addr != end);
 648         return 0;
 649 }
 650
 651 static int unuse_mm(struct mm_struct *mm,
 652                                 swp_entry_t entry, struct page *page)
 653 {
 654         struct vm_area_struct *vma;
 655         int ret = 0;
 656
 657         if (!down_read_trylock(&mm->mmap_sem)) {
 658                 /*
 659                  * Activate page so shrink_cache is unlikely to unmap its
 660                  * ptes while lock is dropped, so swapoff can make progress.
 661                  */
 662                 activate_page(page);
 663                 unlock_page(page);
 664                 down_read(&mm->mmap_sem);
 665                 lock_page(page);
 666         }
 667         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 668                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
 669                         break;
 670         }
 671         up_read(&mm->mmap_sem);
 672         return (ret < 0)? ret: 0;
 673 }
 674
 675 /*
 676  * Scan swap_map from current position to next entry still in use.
 677  * Recycle to start on reaching the end, returning 0 when empty.
 678  */
 679 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 680                                         unsigned int prev)
 681 {
 682         unsigned int max = si->max;
 683         unsigned int i = prev;
 684         int count;
 685
 686         /*
 687          * No need for swap_lock here: we're just looking
 688          * for whether an entry is in use, not modifying it; false
 689          * hits are okay, and sys_swapoff() has already prevented new
 690          * allocations from this area (while holding swap_lock).
 691          */
 692         for (;;) {
 693                 if (++i >= max) {
 694                         if (!prev) {
 695                                 i = 0;
 696                                 break;
 697                         }
 698                         /*
 699                          * No entries in use at top of swap_map,
 700                          * loop back to start and recheck there.
 701                          */
 702                         max = prev + 1;
 703                         prev = 0;
 704                         i = 1;
 705                 }
 706                 count = si->swap_map[i];
 707                 if (count && count != SWAP_MAP_BAD)
 708                         break;
 709         }
 710         return i;
 711 }
 712
 713 /*
 714  * We completely avoid races by reading each swap page in advance,
 715  * and then search for the process using it.  All the necessary
 716  * page table adjustments can then be made atomically.
 717  */
 718 static int try_to_unuse(unsigned int type)
 719 {
 720         struct swap_info_struct * si = &swap_info[type];
 721         struct mm_struct *start_mm;
 722         unsigned short *swap_map;
 723         unsigned short swcount;
 724         struct page *page;
 725         swp_entry_t entry;
 726         unsigned int i = 0;
 727         int retval = 0;
 728         int reset_overflow = 0;
 729         int shmem;
 730
 731         /*
 732          * When searching mms for an entry, a good strategy is to
 733          * start at the first mm we freed the previous entry from
 734          * (though actually we don't notice whether we or coincidence
 735          * freed the entry).  Initialize this start_mm with a hold.
 736          *
 737          * A simpler strategy would be to start at the last mm we
 738          * freed the previous entry from; but that would take less
 739          * advantage of mmlist ordering, which clusters forked mms
 740          * together, child after parent.  If we race with dup_mmap(), we
 741          * prefer to resolve parent before child, lest we miss entries
 742          * duplicated after we scanned child: using last mm would invert
 743          * that.  Though it's only a serious concern when an overflowed
 744          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 745          */
 746         start_mm = &init_mm;
 747         atomic_inc(&init_mm.mm_users);
 748
 749         /*
 750          * Keep on scanning until all entries have gone.  Usually,
 751          * one pass through swap_map is enough, but not necessarily:
 752          * there are races when an instance of an entry might be missed.
 753          */
 754         while ((i = find_next_to_unuse(si, i)) != 0) {
 755                 if (signal_pending(current)) {
 756                         retval = -EINTR;
 757                         break;
 758                 }
 759
 760                 /*
 761                  * Get a page for the entry, using the existing swap
 762                  * cache page if there is one.  Otherwise, get a clean
 763                  * page and read the swap into it.
 764                  */
 765                 swap_map = &si->swap_map[i];
 766                 entry = swp_entry(type, i);
 767                 page = read_swap_cache_async(entry,
 768                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
 769                 if (!page) {
 770                         /*
 771                          * Either swap_duplicate() failed because entry
 772                          * has been freed independently, and will not be
 773                          * reused since sys_swapoff() already disabled
 774                          * allocation from here, or alloc_page() failed.
 775                          */
 776                         if (!*swap_map)
 777                                 continue;
 778                         retval = -ENOMEM;
 779                         break;
 780                 }
 781
 782                 /*
 783                  * Don't hold on to start_mm if it looks like exiting.
 784                  */
 785                 if (atomic_read(&start_mm->mm_users) == 1) {
 786                         mmput(start_mm);
 787                         start_mm = &init_mm;
 788                         atomic_inc(&init_mm.mm_users);
 789                 }
 790
 791                 /*
 792                  * Wait for and lock page.  When do_swap_page races with
 793                  * try_to_unuse, do_swap_page can handle the fault much
 794                  * faster than try_to_unuse can locate the entry.  This
 795                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 796                  * defer to do_swap_page in such a case - in some tests,
 797                  * do_swap_page and try_to_unuse repeatedly compete.
 798                  */
 799                 wait_on_page_locked(page);
 800                 wait_on_page_writeback(page);
 801                 lock_page(page);
 802                 wait_on_page_writeback(page);
 803
 804                 /*
 805                  * Remove all references to entry.
 806                  * Whenever we reach init_mm, there's no address space
 807                  * to search, but use it as a reminder to search shmem.
 808                  */
 809                 shmem = 0;
 810                 swcount = *swap_map;
 811                 if (swcount > 1) {
 812                         if (start_mm == &init_mm)
 813                                 shmem = shmem_unuse(entry, page);
 814                         else
 815                                 retval = unuse_mm(start_mm, entry, page);
 816                 }
 817                 if (*swap_map > 1) {
 818                         int set_start_mm = (*swap_map >= swcount);
 819                         struct list_head *p = &start_mm->mmlist;
 820                         struct mm_struct *new_start_mm = start_mm;
 821                         struct mm_struct *prev_mm = start_mm;
 822                         struct mm_struct *mm;
 823
 824                         atomic_inc(&new_start_mm->mm_users);
 825                         atomic_inc(&prev_mm->mm_users);
 826                         spin_lock(&mmlist_lock);
 827                         while (*swap_map > 1 && !retval && !shmem &&
 828                                         (p = p->next) != &start_mm->mmlist) {
 829                                 mm = list_entry(p, struct mm_struct, mmlist);
 830                                 if (!atomic_inc_not_zero(&mm->mm_users))
 831                                         continue;
 832                                 spin_unlock(&mmlist_lock);
 833                                 mmput(prev_mm);
 834                                 prev_mm = mm;
 835
 836                                 cond_resched();
 837
 838                                 swcount = *swap_map;
 839                                 if (swcount <= 1)
 840                                         ;
 841                                 else if (mm == &init_mm) {
 842                                         set_start_mm = 1;
 843                                         shmem = shmem_unuse(entry, page);
 844                                 } else
 845                                         retval = unuse_mm(mm, entry, page);
 846                                 if (set_start_mm && *swap_map < swcount) {
 847                                         mmput(new_start_mm);
 848                                         atomic_inc(&mm->mm_users);
 849                                         new_start_mm = mm;
 850                                         set_start_mm = 0;
 851                                 }
 852                                 spin_lock(&mmlist_lock);
 853                         }
 854                         spin_unlock(&mmlist_lock);
 855                         mmput(prev_mm);
 856                         mmput(start_mm);
 857                         start_mm = new_start_mm;
 858                 }
 859                 if (shmem) {
 860                         /* page has already been unlocked and released */
 861                         if (shmem > 0)
 862                                 continue;
 863                         retval = shmem;
 864                         break;
 865                 }
 866                 if (retval) {
 867                         unlock_page(page);
 868                         page_cache_release(page);
 869                         break;
 870                 }
 871
 872                 /*
 873                  * How could swap count reach 0x7fff when the maximum
 874                  * pid is 0x7fff, and there's no way to repeat a swap
 875                  * page within an mm (except in shmem, where it's the
 876                  * shared object which takes the reference count)?
 877                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 878                  *
 879                  * If that's wrong, then we should worry more about
 880                  * exit_mmap() and do_munmap() cases described above:
 881                  * we might be resetting SWAP_MAP_MAX too early here.
 882                  * We know "Undead"s can happen, they're okay, so don't
 883                  * report them; but do report if we reset SWAP_MAP_MAX.
 884                  */
 885                 if (*swap_map == SWAP_MAP_MAX) {
 886                         spin_lock(&swap_lock);
 887                         *swap_map = 1;
 888                         spin_unlock(&swap_lock);
 889                         reset_overflow = 1;
 890                 }
 891
 892                 /*
 893                  * If a reference remains (rare), we would like to leave
 894                  * the page in the swap cache; but try_to_unmap could
 895                  * then re-duplicate the entry once we drop page lock,
 896                  * so we might loop indefinitely; also, that page could
 897                  * not be swapped out to other storage meanwhile.  So:
 898                  * delete from cache even if there's another reference,
 899                  * after ensuring that the data has been saved to disk -
 900                  * since if the reference remains (rarer), it will be
 901                  * read from disk into another page.  Splitting into two
 902                  * pages would be incorrect if swap supported "shared
 903                  * private" pages, but they are handled by tmpfs files.
 904                  */
 905                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 906                         struct writeback_control wbc = {
 907                                 .sync_mode = WB_SYNC_NONE,
 908                         };
 909
 910                         swap_writepage(page, &wbc);
 911                         lock_page(page);
 912                         wait_on_page_writeback(page);
 913                 }
 914                 if (PageSwapCache(page))
 915                         delete_from_swap_cache(page);
 916
 917                 /*
 918                  * So we could skip searching mms once swap count went
 919                  * to 1, we did not mark any present ptes as dirty: must
 920                  * mark page dirty so shrink_page_list will preserve it.
 921                  */
 922                 SetPageDirty(page);
 923                 unlock_page(page);
 924                 page_cache_release(page);
 925
 926                 /*
 927                  * Make sure that we aren't completely killing
 928                  * interactive performance.
 929                  */
 930                 cond_resched();
 931         }
 932
 933         mmput(start_mm);
 934         if (reset_overflow) {
 935                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 936                 swap_overflow = 0;
 937         }
 938         return retval;
 939 }
 940
 941 /*
 942  * After a successful try_to_unuse, if no swap is now in use, we know
 943  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 944  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 945  * added to the mmlist just after page_duplicate - before would be racy.
 946  */
 947 static void drain_mmlist(void)
 948 {
 949         struct list_head *p, *next;
 950         unsigned int i;
 951
 952         for (i = 0; i < nr_swapfiles; i++)
 953                 if (swap_info[i].inuse_pages)
 954                         return;
 955         spin_lock(&mmlist_lock);
 956         list_for_each_safe(p, next, &init_mm.mmlist)
 957                 list_del_init(p);
 958         spin_unlock(&mmlist_lock);
 959 }
 960
 961 /*
 962  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 963  * corresponds to page offset `offset'.
 964  */
 965 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 966 {
 967         struct swap_extent *se = sis->curr_swap_extent;
 968         struct swap_extent *start_se = se;
 969
 970         for ( ; ; ) {
 971                 struct list_head *lh;
 972
 973                 if (se->start_page <= offset &&
 974                                 offset < (se->start_page + se->nr_pages)) {
 975                         return se->start_block + (offset - se->start_page);
 976                 }
 977                 lh = se->list.next;
 978                 if (lh == &sis->extent_list)
 979                         lh = lh->next;
 980                 se = list_entry(lh, struct swap_extent, list);
 981                 sis->curr_swap_extent = se;
 982                 BUG_ON(se == start_se);         /* It *must* be present */
 983         }
 984 }
 985
 986 #ifdef CONFIG_HIBERNATION
 987 /*
 988  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 989  * corresponding to given index in swap_info (swap type).
 990  */
 991 sector_t swapdev_block(int swap_type, pgoff_t offset)
 992 {
 993         struct swap_info_struct *sis;
 994
 995         if (swap_type >= nr_swapfiles)
 996                 return 0;
 997
 998         sis = swap_info + swap_type;
 999         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
1000 }
1001 #endif /* CONFIG_HIBERNATION */
1002
1003 /*
1004  * Free all of a swapdev's extent information
1005  */
1006 static void destroy_swap_extents(struct swap_info_struct *sis)
1007 {
1008         while (!list_empty(&sis->extent_list)) {
1009                 struct swap_extent *se;
1010
1011                 se = list_entry(sis->extent_list.next,
1012                                 struct swap_extent, list);
1013                 list_del(&se->list);
1014                 kfree(se);
1015         }
1016 }
1017
1018 /*
1019  * Add a block range (and the corresponding page range) into this swapdev's
1020  * extent list.  The extent list is kept sorted in page order.
1021  *
1022  * This function rather assumes that it is called in ascending page order.
1023  */
1024 static int
1025 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1026                 unsigned long nr_pages, sector_t start_block)
1027 {
1028         struct swap_extent *se;
1029         struct swap_extent *new_se;
1030         struct list_head *lh;
1031
1032         lh = sis->extent_list.prev;     /* The highest page extent */
1033         if (lh != &sis->extent_list) {
1034                 se = list_entry(lh, struct swap_extent, list);
1035                 BUG_ON(se->start_page + se->nr_pages != start_page);
1036                 if (se->start_block + se->nr_pages == start_block) {
1037                         /* Merge it */
1038                         se->nr_pages += nr_pages;
1039                         return 0;
1040                 }
1041         }
1042
1043         /*
1044          * No merge.  Insert a new extent, preserving ordering.
1045          */
1046         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1047         if (new_se == NULL)
1048                 return -ENOMEM;
1049         new_se->start_page = start_page;
1050         new_se->nr_pages = nr_pages;
1051         new_se->start_block = start_block;
1052
1053         list_add_tail(&new_se->list, &sis->extent_list);
1054         return 1;
1055 }
1056
1057 /*
1058  * A `swap extent' is a simple thing which maps a contiguous range of pages
1059  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1060  * is built at swapon time and is then used at swap_writepage/swap_readpage
1061  * time for locating where on disk a page belongs.
1062  *
1063  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1064  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1065  * swap files identically.
1066  *
1067  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1068  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1069  * swapfiles are handled *identically* after swapon time.
1070  *
1071  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1072  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1073  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1074  * requirements, they are simply tossed out - we will never use those blocks
1075  * for swapping.
1076  *
1077  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1078  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1079  * which will scribble on the fs.
1080  *
1081  * The amount of disk space which a single swap extent represents varies.
1082  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1083  * extents in the list.  To avoid much list walking, we cache the previous
1084  * search location in `curr_swap_extent', and start new searches from there.
1085  * This is extremely effective.  The average number of iterations in
1086  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1087  */
1088 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1089 {
1090         struct inode *inode;
1091         unsigned blocks_per_page;
1092         unsigned long page_no;
1093         unsigned blkbits;
1094         sector_t probe_block;
1095         sector_t last_block;
1096         sector_t lowest_block = -1;
1097         sector_t highest_block = 0;
1098         int nr_extents = 0;
1099         int ret;
1100
1101         inode = sis->swap_file->f_mapping->host;
1102         if (S_ISBLK(inode->i_mode)) {
1103                 ret = add_swap_extent(sis, 0, sis->max, 0);
1104                 *span = sis->pages;
1105                 goto done;
1106         }
1107
1108         blkbits = inode->i_blkbits;
1109         blocks_per_page = PAGE_SIZE >> blkbits;
1110
1111         /*
1112          * Map all the blocks into the extent list.  This code doesn't try
1113          * to be very smart.
1114          */
1115         probe_block = 0;
1116         page_no = 0;
1117         last_block = i_size_read(inode) >> blkbits;
1118         while ((probe_block + blocks_per_page) <= last_block &&
1119                         page_no < sis->max) {
1120                 unsigned block_in_page;
1121                 sector_t first_block;
1122
1123                 first_block = bmap(inode, probe_block);
1124                 if (first_block == 0)
1125                         goto bad_bmap;
1126
1127                 /*
1128                  * It must be PAGE_SIZE aligned on-disk
1129                  */
1130                 if (first_block & (blocks_per_page - 1)) {
1131                         probe_block++;
1132                         goto reprobe;
1133                 }
1134
1135                 for (block_in_page = 1; block_in_page < blocks_per_page;
1136                                         block_in_page++) {
1137                         sector_t block;
1138
1139                         block = bmap(inode, probe_block + block_in_page);
1140                         if (block == 0)
1141                                 goto bad_bmap;
1142                         if (block != first_block + block_in_page) {
1143                                 /* Discontiguity */
1144                                 probe_block++;
1145                                 goto reprobe;
1146                         }
1147                 }
1148
1149                 first_block >>= (PAGE_SHIFT - blkbits);
1150                 if (page_no) {  /* exclude the header page */
1151                         if (first_block < lowest_block)
1152                                 lowest_block = first_block;
1153                         if (first_block > highest_block)
1154                                 highest_block = first_block;
1155                 }
1156
1157                 /*
1158                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1159                  */
1160                 ret = add_swap_extent(sis, page_no, 1, first_block);
1161                 if (ret < 0)
1162                         goto out;
1163                 nr_extents += ret;
1164                 page_no++;
1165                 probe_block += blocks_per_page;
1166 reprobe:
1167                 continue;
1168         }
1169         ret = nr_extents;
1170         *span = 1 + highest_block - lowest_block;
1171         if (page_no == 0)
1172                 page_no = 1;    /* force Empty message */
1173         sis->max = page_no;
1174         sis->pages = page_no - 1;
1175         sis->highest_bit = page_no - 1;
1176 done:
1177         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1178                                         struct swap_extent, list);
1179         goto out;
1180 bad_bmap:
1181         printk(KERN_ERR "swapon: swapfile has holes\n");
1182         ret = -EINVAL;
1183 out:
1184         return ret;
1185 }
1186
1187 #if 0   /* We don't need this yet */
1188 #include <linux/backing-dev.h>
1189 int page_queue_congested(struct page *page)
1190 {
1191         struct backing_dev_info *bdi;
1192
1193         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1194
1195         if (PageSwapCache(page)) {
1196                 swp_entry_t entry = { .val = page_private(page) };
1197                 struct swap_info_struct *sis;
1198
1199                 sis = get_swap_info_struct(swp_type(entry));
1200                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1201         } else
1202                 bdi = page->mapping->backing_dev_info;
1203         return bdi_write_congested(bdi);
1204 }
1205 #endif
1206
1207 asmlinkage long sys_swapoff(const char __user * specialfile)
1208 {
1209         struct swap_info_struct * p = NULL;
1210         unsigned short *swap_map;
1211         struct file *swap_file, *victim;
1212         struct address_space *mapping;
1213         struct inode *inode;
1214         char * pathname;
1215         int i, type, prev;
1216         int err;
1217
1218         if (!capable(CAP_SYS_ADMIN))
1219                 return -EPERM;
1220
1221         pathname = getname(specialfile);
1222         err = PTR_ERR(pathname);
1223         if (IS_ERR(pathname))
1224                 goto out;
1225
1226         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1227         putname(pathname);
1228         err = PTR_ERR(victim);
1229         if (IS_ERR(victim))
1230                 goto out;
1231
1232         mapping = victim->f_mapping;
1233         prev = -1;
1234         spin_lock(&swap_lock);
1235         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1236                 p = swap_info + type;
1237                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1238                         if (p->swap_file->f_mapping == mapping)
1239                                 break;
1240                 }
1241                 prev = type;
1242         }
1243         if (type < 0) {
1244                 err = -EINVAL;
1245                 spin_unlock(&swap_lock);
1246                 goto out_dput;
1247         }
1248         if (!security_vm_enough_memory(p->pages))
1249                 vm_unacct_memory(p->pages);
1250         else {
1251                 err = -ENOMEM;
1252                 spin_unlock(&swap_lock);
1253                 goto out_dput;
1254         }
1255         if (prev < 0) {
1256                 swap_list.head = p->next;
1257         } else {
1258                 swap_info[prev].next = p->next;
1259         }
1260         if (type == swap_list.next) {
1261                 /* just pick something that's safe... */
1262                 swap_list.next = swap_list.head;
1263         }
1264         if (p->prio < 0) {
1265                 for (i = p->next; i >= 0; i = swap_info[i].next)
1266                         swap_info[i].prio = p->prio--;
1267                 least_priority++;
1268         }
1269         nr_swap_pages -= p->pages;
1270         total_swap_pages -= p->pages;
1271         p->flags &= ~SWP_WRITEOK;
1272         spin_unlock(&swap_lock);
1273
1274         current->flags |= PF_SWAPOFF;
1275         err = try_to_unuse(type);
1276         current->flags &= ~PF_SWAPOFF;
1277
1278         if (err) {
1279                 /* re-insert swap space back into swap_list */
1280                 spin_lock(&swap_lock);
1281                 if (p->prio < 0)
1282                         p->prio = --least_priority;
1283                 prev = -1;
1284                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1285                         if (p->prio >= swap_info[i].prio)
1286                                 break;
1287                         prev = i;
1288                 }
1289                 p->next = i;
1290                 if (prev < 0)
1291                         swap_list.head = swap_list.next = p - swap_info;
1292                 else
1293                         swap_info[prev].next = p - swap_info;
1294                 nr_swap_pages += p->pages;
1295                 total_swap_pages += p->pages;
1296                 p->flags |= SWP_WRITEOK;
1297                 spin_unlock(&swap_lock);
1298                 goto out_dput;
1299         }
1300
1301         /* wait for any unplug function to finish */
1302         down_write(&swap_unplug_sem);
1303         up_write(&swap_unplug_sem);
1304
1305         destroy_swap_extents(p);
1306         mutex_lock(&swapon_mutex);
1307         spin_lock(&swap_lock);
1308         drain_mmlist();
1309
1310         /* wait for anyone still in scan_swap_map */
1311         p->highest_bit = 0;             /* cuts scans short */
1312         while (p->flags >= SWP_SCANNING) {
1313                 spin_unlock(&swap_lock);
1314                 schedule_timeout_uninterruptible(1);
1315                 spin_lock(&swap_lock);
1316         }
1317
1318         swap_file = p->swap_file;
1319         p->swap_file = NULL;
1320         p->max = 0;
1321         swap_map = p->swap_map;
1322         p->swap_map = NULL;
1323         p->flags = 0;
1324         spin_unlock(&swap_lock);
1325         mutex_unlock(&swapon_mutex);
1326         vfree(swap_map);
1327         inode = mapping->host;
1328         if (S_ISBLK(inode->i_mode)) {
1329                 struct block_device *bdev = I_BDEV(inode);
1330                 set_blocksize(bdev, p->old_block_size);
1331                 bd_release(bdev);
1332         } else {
1333                 mutex_lock(&inode->i_mutex);
1334                 inode->i_flags &= ~S_SWAPFILE;
1335                 mutex_unlock(&inode->i_mutex);
1336         }
1337         filp_close(swap_file, NULL);
1338         err = 0;
1339
1340 out_dput:
1341         filp_close(victim, NULL);
1342 out:
1343         return err;
1344 }
1345
1346 #ifdef CONFIG_PROC_FS
1347 /* iterator */
1348 static void *swap_start(struct seq_file *swap, loff_t *pos)
1349 {
1350         struct swap_info_struct *ptr = swap_info;
1351         int i;
1352         loff_t l = *pos;
1353
1354         mutex_lock(&swapon_mutex);
1355
1356         if (!l)
1357                 return SEQ_START_TOKEN;
1358
1359         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1360                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1361                         continue;
1362                 if (!--l)
1363                         return ptr;
1364         }
1365
1366         return NULL;
1367 }
1368
1369 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1370 {
1371         struct swap_info_struct *ptr;
1372         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1373
1374         if (v == SEQ_START_TOKEN)
1375                 ptr = swap_info;
1376         else {
1377                 ptr = v;
1378                 ptr++;
1379         }
1380
1381         for (; ptr < endptr; ptr++) {
1382                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1383                         continue;
1384                 ++*pos;
1385                 return ptr;
1386         }
1387
1388         return NULL;
1389 }
1390
1391 static void swap_stop(struct seq_file *swap, void *v)
1392 {
1393         mutex_unlock(&swapon_mutex);
1394 }
1395
1396 static int swap_show(struct seq_file *swap, void *v)
1397 {
1398         struct swap_info_struct *ptr = v;
1399         struct file *file;
1400         int len;
1401
1402         if (ptr == SEQ_START_TOKEN) {
1403                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1404                 return 0;
1405         }
1406
1407         file = ptr->swap_file;
1408         len = seq_path(swap, &file->f_path, " \t\n\\");
1409         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1410                        len < 40 ? 40 - len : 1, " ",
1411                        S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1412                                 "partition" : "file\t",
1413                        ptr->pages << (PAGE_SHIFT - 10),
1414                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1415                        ptr->prio);
1416         return 0;
1417 }
1418
1419 static const struct seq_operations swaps_op = {
1420         .start =        swap_start,
1421         .next =         swap_next,
1422         .stop =         swap_stop,
1423         .show =         swap_show
1424 };
1425
1426 static int swaps_open(struct inode *inode, struct file *file)
1427 {
1428         return seq_open(file, &swaps_op);
1429 }
1430
1431 static const struct file_operations proc_swaps_operations = {
1432         .open           = swaps_open,
1433         .read           = seq_read,
1434         .llseek         = seq_lseek,
1435         .release        = seq_release,
1436 };
1437
1438 static int __init procswaps_init(void)
1439 {
1440         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1441         return 0;
1442 }
1443 __initcall(procswaps_init);
1444 #endif /* CONFIG_PROC_FS */
1445
1446 /*
1447  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1448  *
1449  * The swapon system call
1450  */
1451 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1452 {
1453         struct swap_info_struct * p;
1454         char *name = NULL;
1455         struct block_device *bdev = NULL;
1456         struct file *swap_file = NULL;
1457         struct address_space *mapping;
1458         unsigned int type;
1459         int i, prev;
1460         int error;
1461         union swap_header *swap_header = NULL;
1462         int swap_header_version;
1463         unsigned int nr_good_pages = 0;
1464         int nr_extents = 0;
1465         sector_t span;
1466         unsigned long maxpages = 1;
1467         int swapfilesize;
1468         unsigned short *swap_map = NULL;
1469         struct page *page = NULL;
1470         struct inode *inode = NULL;
1471         int did_down = 0;
1472
1473         if (!capable(CAP_SYS_ADMIN))
1474                 return -EPERM;
1475         spin_lock(&swap_lock);
1476         p = swap_info;
1477         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1478                 if (!(p->flags & SWP_USED))
1479                         break;
1480         error = -EPERM;
1481         if (type >= MAX_SWAPFILES) {
1482                 spin_unlock(&swap_lock);
1483                 goto out;
1484         }
1485         if (type >= nr_swapfiles)
1486                 nr_swapfiles = type+1;
1487         memset(p, 0, sizeof(*p));
1488         INIT_LIST_HEAD(&p->extent_list);
1489         p->flags = SWP_USED;
1490         p->next = -1;
1491         spin_unlock(&swap_lock);
1492         name = getname(specialfile);
1493         error = PTR_ERR(name);
1494         if (IS_ERR(name)) {
1495                 name = NULL;
1496                 goto bad_swap_2;
1497         }
1498         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1499         error = PTR_ERR(swap_file);
1500         if (IS_ERR(swap_file)) {
1501                 swap_file = NULL;
1502                 goto bad_swap_2;
1503         }
1504
1505         p->swap_file = swap_file;
1506         mapping = swap_file->f_mapping;
1507         inode = mapping->host;
1508
1509         error = -EBUSY;
1510         for (i = 0; i < nr_swapfiles; i++) {
1511                 struct swap_info_struct *q = &swap_info[i];
1512
1513                 if (i == type || !q->swap_file)
1514                         continue;
1515                 if (mapping == q->swap_file->f_mapping)
1516                         goto bad_swap;
1517         }
1518
1519         error = -EINVAL;
1520         if (S_ISBLK(inode->i_mode)) {
1521                 bdev = I_BDEV(inode);
1522                 error = bd_claim(bdev, sys_swapon);
1523                 if (error < 0) {
1524                         bdev = NULL;
1525                         error = -EINVAL;
1526                         goto bad_swap;
1527                 }
1528                 p->old_block_size = block_size(bdev);
1529                 error = set_blocksize(bdev, PAGE_SIZE);
1530                 if (error < 0)
1531                         goto bad_swap;
1532                 p->bdev = bdev;
1533         } else if (S_ISREG(inode->i_mode)) {
1534                 p->bdev = inode->i_sb->s_bdev;
1535                 mutex_lock(&inode->i_mutex);
1536                 did_down = 1;
1537                 if (IS_SWAPFILE(inode)) {
1538                         error = -EBUSY;
1539                         goto bad_swap;
1540                 }
1541         } else {
1542                 goto bad_swap;
1543         }
1544
1545         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1546
1547         /*
1548          * Read the swap header.
1549          */
1550         if (!mapping->a_ops->readpage) {
1551                 error = -EINVAL;
1552                 goto bad_swap;
1553         }
1554         page = read_mapping_page(mapping, 0, swap_file);
1555         if (IS_ERR(page)) {
1556                 error = PTR_ERR(page);
1557                 goto bad_swap;
1558         }
1559         kmap(page);
1560         swap_header = page_address(page);
1561
1562         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1563                 swap_header_version = 1;
1564         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1565                 swap_header_version = 2;
1566         else {
1567                 printk(KERN_ERR "Unable to find swap-space signature\n");
1568                 error = -EINVAL;
1569                 goto bad_swap;
1570         }
1571
1572         switch (swap_header_version) {
1573         case 1:
1574                 printk(KERN_ERR "version 0 swap is no longer supported. "
1575                         "Use mkswap -v1 %s\n", name);
1576                 error = -EINVAL;
1577                 goto bad_swap;
1578         case 2:
1579                 /* swap partition endianess hack... */
1580                 if (swab32(swap_header->info.version) == 1) {
1581                         swab32s(&swap_header->info.version);
1582                         swab32s(&swap_header->info.last_page);
1583                         swab32s(&swap_header->info.nr_badpages);
1584                         for (i = 0; i < swap_header->info.nr_badpages; i++)
1585                                 swab32s(&swap_header->info.badpages[i]);
1586                 }
1587                 /* Check the swap header's sub-version and the size of
1588                    the swap file and bad block lists */
1589                 if (swap_header->info.version != 1) {
1590                         printk(KERN_WARNING
1591                                "Unable to handle swap header version %d\n",
1592                                swap_header->info.version);
1593                         error = -EINVAL;
1594                         goto bad_swap;
1595                 }
1596
1597                 p->lowest_bit  = 1;
1598                 p->cluster_next = 1;
1599
1600                 /*
1601                  * Find out how many pages are allowed for a single swap
1602                  * device. There are two limiting factors: 1) the number of
1603                  * bits for the swap offset in the swp_entry_t type and
1604                  * 2) the number of bits in the a swap pte as defined by
1605                  * the different architectures. In order to find the
1606                  * largest possible bit mask a swap entry with swap type 0
1607                  * and swap offset ~0UL is created, encoded to a swap pte,
1608                  * decoded to a swp_entry_t again and finally the swap
1609                  * offset is extracted. This will mask all the bits from
1610                  * the initial ~0UL mask that can't be encoded in either
1611                  * the swp_entry_t or the architecture definition of a
1612                  * swap pte.
1613                  */
1614                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1615                 if (maxpages > swap_header->info.last_page)
1616                         maxpages = swap_header->info.last_page;
1617                 p->highest_bit = maxpages - 1;
1618
1619                 error = -EINVAL;
1620                 if (!maxpages)
1621                         goto bad_swap;
1622                 if (swapfilesize && maxpages > swapfilesize) {
1623                         printk(KERN_WARNING
1624                                "Swap area shorter than signature indicates\n");
1625                         goto bad_swap;
1626                 }
1627                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1628                         goto bad_swap;
1629                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1630                         goto bad_swap;
1631
1632                 /* OK, set up the swap map and apply the bad block list */
1633                 swap_map = vmalloc(maxpages * sizeof(short));
1634                 if (!swap_map) {
1635                         error = -ENOMEM;
1636                         goto bad_swap;
1637                 }
1638
1639                 error = 0;
1640                 memset(swap_map, 0, maxpages * sizeof(short));
1641                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1642                         int page_nr = swap_header->info.badpages[i];
1643                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1644                                 error = -EINVAL;
1645                         else
1646                                 swap_map[page_nr] = SWAP_MAP_BAD;
1647                 }
1648                 nr_good_pages = swap_header->info.last_page -
1649                                 swap_header->info.nr_badpages -
1650                                 1 /* header page */;
1651                 if (error)
1652                         goto bad_swap;
1653         }
1654
1655         if (nr_good_pages) {
1656                 swap_map[0] = SWAP_MAP_BAD;
1657                 p->max = maxpages;
1658                 p->pages = nr_good_pages;
1659                 nr_extents = setup_swap_extents(p, &span);
1660                 if (nr_extents < 0) {
1661                         error = nr_extents;
1662                         goto bad_swap;
1663                 }
1664                 nr_good_pages = p->pages;
1665         }
1666         if (!nr_good_pages) {
1667                 printk(KERN_WARNING "Empty swap-file\n");
1668                 error = -EINVAL;
1669                 goto bad_swap;
1670         }
1671
1672         mutex_lock(&swapon_mutex);
1673         spin_lock(&swap_lock);
1674         if (swap_flags & SWAP_FLAG_PREFER)
1675                 p->prio =
1676                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1677         else
1678                 p->prio = --least_priority;
1679         p->swap_map = swap_map;
1680         p->flags = SWP_ACTIVE;
1681         nr_swap_pages += nr_good_pages;
1682         total_swap_pages += nr_good_pages;
1683
1684         printk(KERN_INFO "Adding %uk swap on %s.  "
1685                         "Priority:%d extents:%d across:%lluk\n",
1686                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1687                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1688
1689         /* insert swap space into swap_list: */
1690         prev = -1;
1691         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1692                 if (p->prio >= swap_info[i].prio) {
1693                         break;
1694                 }
1695                 prev = i;
1696         }
1697         p->next = i;
1698         if (prev < 0) {
1699                 swap_list.head = swap_list.next = p - swap_info;
1700         } else {
1701                 swap_info[prev].next = p - swap_info;
1702         }
1703         spin_unlock(&swap_lock);
1704         mutex_unlock(&swapon_mutex);
1705         error = 0;
1706         goto out;
1707 bad_swap:
1708         if (bdev) {
1709                 set_blocksize(bdev, p->old_block_size);
1710                 bd_release(bdev);
1711         }
1712         destroy_swap_extents(p);
1713 bad_swap_2:
1714         spin_lock(&swap_lock);
1715         p->swap_file = NULL;
1716         p->flags = 0;
1717         spin_unlock(&swap_lock);
1718         vfree(swap_map);
1719         if (swap_file)
1720                 filp_close(swap_file, NULL);
1721 out:
1722         if (page && !IS_ERR(page)) {
1723                 kunmap(page);
1724                 page_cache_release(page);
1725         }
1726         if (name)
1727                 putname(name);
1728         if (did_down) {
1729                 if (!error)
1730                         inode->i_flags |= S_SWAPFILE;
1731                 mutex_unlock(&inode->i_mutex);
1732         }
1733         return error;
1734 }
1735
1736 void si_swapinfo(struct sysinfo *val)
1737 {
1738         unsigned int i;
1739         unsigned long nr_to_be_unused = 0;
1740
1741         spin_lock(&swap_lock);
1742         for (i = 0; i < nr_swapfiles; i++) {
1743                 if (!(swap_info[i].flags & SWP_USED) ||
1744                      (swap_info[i].flags & SWP_WRITEOK))
1745                         continue;
1746                 nr_to_be_unused += swap_info[i].inuse_pages;
1747         }
1748         val->freeswap = nr_swap_pages + nr_to_be_unused;
1749         val->totalswap = total_swap_pages + nr_to_be_unused;
1750         spin_unlock(&swap_lock);
1751 }
1752
1753 /*
1754  * Verify that a swap entry is valid and increment its swap map count.
1755  *
1756  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1757  * "permanent", but will be reclaimed by the next swapoff.
1758  */
1759 int swap_duplicate(swp_entry_t entry)
1760 {
1761         struct swap_info_struct * p;
1762         unsigned long offset, type;
1763         int result = 0;
1764
1765         if (is_migration_entry(entry))
1766                 return 1;
1767
1768         type = swp_type(entry);
1769         if (type >= nr_swapfiles)
1770                 goto bad_file;
1771         p = type + swap_info;
1772         offset = swp_offset(entry);
1773
1774         spin_lock(&swap_lock);
1775         if (offset < p->max && p->swap_map[offset]) {
1776                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1777                         p->swap_map[offset]++;
1778                         result = 1;
1779                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1780                         if (swap_overflow++ < 5)
1781                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1782                         p->swap_map[offset] = SWAP_MAP_MAX;
1783                         result = 1;
1784                 }
1785         }
1786         spin_unlock(&swap_lock);
1787 out:
1788         return result;
1789
1790 bad_file:
1791         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1792         goto out;
1793 }
1794
1795 struct swap_info_struct *
1796 get_swap_info_struct(unsigned type)
1797 {
1798         return &swap_info[type];
1799 }
1800
1801 /*
1802  * swap_lock prevents swap_map being freed. Don't grab an extra
1803  * reference on the swaphandle, it doesn't matter if it becomes unused.
1804  */
1805 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1806 {
1807         struct swap_info_struct *si;
1808         int our_page_cluster = page_cluster;
1809         pgoff_t target, toff;
1810         pgoff_t base, end;
1811         int nr_pages = 0;
1812
1813         if (!our_page_cluster)  /* no readahead */
1814                 return 0;
1815
1816         si = &swap_info[swp_type(entry)];
1817         target = swp_offset(entry);
1818         base = (target >> our_page_cluster) << our_page_cluster;
1819         end = base + (1 << our_page_cluster);
1820         if (!base)              /* first page is swap header */
1821                 base++;
1822
1823         spin_lock(&swap_lock);
1824         if (end > si->max)      /* don't go beyond end of map */
1825                 end = si->max;
1826
1827         /* Count contiguous allocated slots above our target */
1828         for (toff = target; ++toff < end; nr_pages++) {
1829                 /* Don't read in free or bad pages */
1830                 if (!si->swap_map[toff])
1831                         break;
1832                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1833                         break;
1834         }
1835         /* Count contiguous allocated slots below our target */
1836         for (toff = target; --toff >= base; nr_pages++) {
1837                 /* Don't read in free or bad pages */
1838                 if (!si->swap_map[toff])
1839                         break;
1840                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1841                         break;
1842         }
1843         spin_unlock(&swap_lock);
1844
1845         /*
1846          * Indicate starting offset, and return number of pages to get:
1847          * if only 1, say 0, since there's then no readahead to be done.
1848          */
1849         *offset = ++toff;
1850         return nr_pages? ++nr_pages: 0;
1851 }