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