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/vs_base.h>
  31 #include <linux/vs_memory.h>
  32
  33 #include <asm/pgtable.h>
  34 #include <asm/tlbflush.h>
  35 #include <linux/swapops.h>
  36
  37 DEFINE_SPINLOCK(swap_lock);
  38 unsigned int nr_swapfiles;
  39 long total_swap_pages;
  40 static int swap_overflow;
  41
  42 static const char Bad_file[] = "Bad swap file entry ";
  43 static const char Unused_file[] = "Unused swap file entry ";
  44 static const char Bad_offset[] = "Bad swap offset entry ";
  45 static const char Unused_offset[] = "Unused swap offset entry ";
  46
  47 struct swap_list_t swap_list = {-1, -1};
  48
  49 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                 write_lock_irq(&swapper_space.tree_lock);
 373                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 374                         __delete_from_swap_cache(page);
 375                         SetPageDirty(page);
 376                         retval = 1;
 377                 }
 378                 write_unlock_irq(&swapper_space.tree_lock);
 379         }
 380         spin_unlock(&swap_lock);
 381
 382         if (retval) {
 383                 swap_free(entry);
 384                 page_cache_release(page);
 385         }
 386
 387         return retval;
 388 }
 389
 390 /*
 391  * Free the swap entry like above, but also try to
 392  * free the page cache entry if it is the last user.
 393  */
 394 void free_swap_and_cache(swp_entry_t entry)
 395 {
 396         struct swap_info_struct * p;
 397         struct page *page = NULL;
 398
 399         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_SOFTWARE_SUSPEND
 431 /*
 432  * Find the swap type that corresponds to given device (if any)
 433  *
 434  * This is needed for software suspend and is done in such a way that inode
 435  * aliasing is allowed.
 436  */
 437 int swap_type_of(dev_t device)
 438 {
 439         int i;
 440
 441         spin_lock(&swap_lock);
 442         for (i = 0; i < nr_swapfiles; i++) {
 443                 struct inode *inode;
 444
 445                 if (!(swap_info[i].flags & SWP_WRITEOK))
 446                         continue;
 447
 448                 if (!device) {
 449                         spin_unlock(&swap_lock);
 450                         return i;
 451                 }
 452                 inode = swap_info[i].swap_file->f_dentry->d_inode;
 453                 if (S_ISBLK(inode->i_mode) &&
 454                     device == MKDEV(imajor(inode), iminor(inode))) {
 455                         spin_unlock(&swap_lock);
 456                         return i;
 457                 }
 458         }
 459         spin_unlock(&swap_lock);
 460         return -ENODEV;
 461 }
 462
 463 /*
 464  * Return either the total number of swap pages of given type, or the number
 465  * of free pages of that type (depending on @free)
 466  *
 467  * This is needed for software suspend
 468  */
 469 unsigned int count_swap_pages(int type, int free)
 470 {
 471         unsigned int n = 0;
 472
 473         if (type < nr_swapfiles) {
 474                 spin_lock(&swap_lock);
 475                 if (swap_info[type].flags & SWP_WRITEOK) {
 476                         n = swap_info[type].pages;
 477                         if (free)
 478                                 n -= swap_info[type].inuse_pages;
 479                 }
 480                 spin_unlock(&swap_lock);
 481         }
 482         return n;
 483 }
 484 #endif
 485
 486 /*
 487  * No need to decide whether this PTE shares the swap entry with others,
 488  * just let do_wp_page work it out if a write is requested later - to
 489  * force COW, vm_page_prot omits write permission from any private vma.
 490  */
 491 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
 492                 unsigned long addr, swp_entry_t entry, struct page *page)
 493 {
 494         inc_mm_counter(vma->vm_mm, anon_rss);
 495         get_page(page);
 496         set_pte_at(vma->vm_mm, addr, pte,
 497                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 498         page_add_anon_rmap(page, vma, addr);
 499         swap_free(entry);
 500         /*
 501          * Move the page to the active list so it is not
 502          * immediately swapped out again after swapon.
 503          */
 504         activate_page(page);
 505 }
 506
 507 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 508                                 unsigned long addr, unsigned long end,
 509                                 swp_entry_t entry, struct page *page)
 510 {
 511         pte_t swp_pte = swp_entry_to_pte(entry);
 512         pte_t *pte;
 513         spinlock_t *ptl;
 514         int found = 0;
 515
 516         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 517         do {
 518                 /*
 519                  * swapoff spends a _lot_ of time in this loop!
 520                  * Test inline before going to call unuse_pte.
 521                  */
 522                 if (unlikely(pte_same(*pte, swp_pte))) {
 523                         unuse_pte(vma, pte++, addr, entry, page);
 524                         found = 1;
 525                         break;
 526                 }
 527         } while (pte++, addr += PAGE_SIZE, addr != end);
 528         pte_unmap_unlock(pte - 1, ptl);
 529         return found;
 530 }
 531
 532 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 533                                 unsigned long addr, unsigned long end,
 534                                 swp_entry_t entry, struct page *page)
 535 {
 536         pmd_t *pmd;
 537         unsigned long next;
 538
 539         pmd = pmd_offset(pud, addr);
 540         do {
 541                 next = pmd_addr_end(addr, end);
 542                 if (pmd_none_or_clear_bad(pmd))
 543                         continue;
 544                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 545                         return 1;
 546         } while (pmd++, addr = next, addr != end);
 547         return 0;
 548 }
 549
 550 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 551                                 unsigned long addr, unsigned long end,
 552                                 swp_entry_t entry, struct page *page)
 553 {
 554         pud_t *pud;
 555         unsigned long next;
 556
 557         pud = pud_offset(pgd, addr);
 558         do {
 559                 next = pud_addr_end(addr, end);
 560                 if (pud_none_or_clear_bad(pud))
 561                         continue;
 562                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 563                         return 1;
 564         } while (pud++, addr = next, addr != end);
 565         return 0;
 566 }
 567
 568 static int unuse_vma(struct vm_area_struct *vma,
 569                                 swp_entry_t entry, struct page *page)
 570 {
 571         pgd_t *pgd;
 572         unsigned long addr, end, next;
 573
 574         if (page->mapping) {
 575                 addr = page_address_in_vma(page, vma);
 576                 if (addr == -EFAULT)
 577                         return 0;
 578                 else
 579                         end = addr + PAGE_SIZE;
 580         } else {
 581                 addr = vma->vm_start;
 582                 end = vma->vm_end;
 583         }
 584
 585         pgd = pgd_offset(vma->vm_mm, addr);
 586         do {
 587                 next = pgd_addr_end(addr, end);
 588                 if (pgd_none_or_clear_bad(pgd))
 589                         continue;
 590                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 591                         return 1;
 592         } while (pgd++, addr = next, addr != end);
 593         return 0;
 594 }
 595
 596 static int unuse_mm(struct mm_struct *mm,
 597                                 swp_entry_t entry, struct page *page)
 598 {
 599         struct vm_area_struct *vma;
 600
 601         if (!down_read_trylock(&mm->mmap_sem)) {
 602                 /*
 603                  * Activate page so shrink_cache is unlikely to unmap its
 604                  * ptes while lock is dropped, so swapoff can make progress.
 605                  */
 606                 activate_page(page);
 607                 unlock_page(page);
 608                 down_read(&mm->mmap_sem);
 609                 lock_page(page);
 610         }
 611         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 612                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 613                         break;
 614         }
 615         up_read(&mm->mmap_sem);
 616         /*
 617          * Currently unuse_mm cannot fail, but leave error handling
 618          * at call sites for now, since we change it from time to time.
 619          */
 620         return 0;
 621 }
 622
 623 /*
 624  * Scan swap_map from current position to next entry still in use.
 625  * Recycle to start on reaching the end, returning 0 when empty.
 626  */
 627 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 628                                         unsigned int prev)
 629 {
 630         unsigned int max = si->max;
 631         unsigned int i = prev;
 632         int count;
 633
 634         /*
 635          * No need for swap_lock here: we're just looking
 636          * for whether an entry is in use, not modifying it; false
 637          * hits are okay, and sys_swapoff() has already prevented new
 638          * allocations from this area (while holding swap_lock).
 639          */
 640         for (;;) {
 641                 if (++i >= max) {
 642                         if (!prev) {
 643                                 i = 0;
 644                                 break;
 645                         }
 646                         /*
 647                          * No entries in use at top of swap_map,
 648                          * loop back to start and recheck there.
 649                          */
 650                         max = prev + 1;
 651                         prev = 0;
 652                         i = 1;
 653                 }
 654                 count = si->swap_map[i];
 655                 if (count && count != SWAP_MAP_BAD)
 656                         break;
 657         }
 658         return i;
 659 }
 660
 661 /*
 662  * We completely avoid races by reading each swap page in advance,
 663  * and then search for the process using it.  All the necessary
 664  * page table adjustments can then be made atomically.
 665  */
 666 static int try_to_unuse(unsigned int type)
 667 {
 668         struct swap_info_struct * si = &swap_info[type];
 669         struct mm_struct *start_mm;
 670         unsigned short *swap_map;
 671         unsigned short swcount;
 672         struct page *page;
 673         swp_entry_t entry;
 674         unsigned int i = 0;
 675         int retval = 0;
 676         int reset_overflow = 0;
 677         int shmem;
 678
 679         /*
 680          * When searching mms for an entry, a good strategy is to
 681          * start at the first mm we freed the previous entry from
 682          * (though actually we don't notice whether we or coincidence
 683          * freed the entry).  Initialize this start_mm with a hold.
 684          *
 685          * A simpler strategy would be to start at the last mm we
 686          * freed the previous entry from; but that would take less
 687          * advantage of mmlist ordering, which clusters forked mms
 688          * together, child after parent.  If we race with dup_mmap(), we
 689          * prefer to resolve parent before child, lest we miss entries
 690          * duplicated after we scanned child: using last mm would invert
 691          * that.  Though it's only a serious concern when an overflowed
 692          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 693          */
 694         start_mm = &init_mm;
 695         atomic_inc(&init_mm.mm_users);
 696
 697         /*
 698          * Keep on scanning until all entries have gone.  Usually,
 699          * one pass through swap_map is enough, but not necessarily:
 700          * there are races when an instance of an entry might be missed.
 701          */
 702         while ((i = find_next_to_unuse(si, i)) != 0) {
 703                 if (signal_pending(current)) {
 704                         retval = -EINTR;
 705                         break;
 706                 }
 707
 708                 /*
 709                  * Get a page for the entry, using the existing swap
 710                  * cache page if there is one.  Otherwise, get a clean
 711                  * page and read the swap into it.
 712                  */
 713                 swap_map = &si->swap_map[i];
 714                 entry = swp_entry(type, i);
 715                 page = read_swap_cache_async(entry, NULL, 0);
 716                 if (!page) {
 717                         /*
 718                          * Either swap_duplicate() failed because entry
 719                          * has been freed independently, and will not be
 720                          * reused since sys_swapoff() already disabled
 721                          * allocation from here, or alloc_page() failed.
 722                          */
 723                         if (!*swap_map)
 724                                 continue;
 725                         retval = -ENOMEM;
 726                         break;
 727                 }
 728
 729                 /*
 730                  * Don't hold on to start_mm if it looks like exiting.
 731                  */
 732                 if (atomic_read(&start_mm->mm_users) == 1) {
 733                         mmput(start_mm);
 734                         start_mm = &init_mm;
 735                         atomic_inc(&init_mm.mm_users);
 736                 }
 737
 738                 /*
 739                  * Wait for and lock page.  When do_swap_page races with
 740                  * try_to_unuse, do_swap_page can handle the fault much
 741                  * faster than try_to_unuse can locate the entry.  This
 742                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 743                  * defer to do_swap_page in such a case - in some tests,
 744                  * do_swap_page and try_to_unuse repeatedly compete.
 745                  */
 746                 wait_on_page_locked(page);
 747                 wait_on_page_writeback(page);
 748                 lock_page(page);
 749                 wait_on_page_writeback(page);
 750
 751                 /*
 752                  * Remove all references to entry.
 753                  * Whenever we reach init_mm, there's no address space
 754                  * to search, but use it as a reminder to search shmem.
 755                  */
 756                 shmem = 0;
 757                 swcount = *swap_map;
 758                 if (swcount > 1) {
 759                         if (start_mm == &init_mm)
 760                                 shmem = shmem_unuse(entry, page);
 761                         else
 762                                 retval = unuse_mm(start_mm, entry, page);
 763                 }
 764                 if (*swap_map > 1) {
 765                         int set_start_mm = (*swap_map >= swcount);
 766                         struct list_head *p = &start_mm->mmlist;
 767                         struct mm_struct *new_start_mm = start_mm;
 768                         struct mm_struct *prev_mm = start_mm;
 769                         struct mm_struct *mm;
 770
 771                         atomic_inc(&new_start_mm->mm_users);
 772                         atomic_inc(&prev_mm->mm_users);
 773                         spin_lock(&mmlist_lock);
 774                         while (*swap_map > 1 && !retval &&
 775                                         (p = p->next) != &start_mm->mmlist) {
 776                                 mm = list_entry(p, struct mm_struct, mmlist);
 777                                 if (!atomic_inc_not_zero(&mm->mm_users))
 778                                         continue;
 779                                 spin_unlock(&mmlist_lock);
 780                                 mmput(prev_mm);
 781                                 prev_mm = mm;
 782
 783                                 cond_resched();
 784
 785                                 swcount = *swap_map;
 786                                 if (swcount <= 1)
 787                                         ;
 788                                 else if (mm == &init_mm) {
 789                                         set_start_mm = 1;
 790                                         shmem = shmem_unuse(entry, page);
 791                                 } else
 792                                         retval = unuse_mm(mm, entry, page);
 793                                 if (set_start_mm && *swap_map < swcount) {
 794                                         mmput(new_start_mm);
 795                                         atomic_inc(&mm->mm_users);
 796                                         new_start_mm = mm;
 797                                         set_start_mm = 0;
 798                                 }
 799                                 spin_lock(&mmlist_lock);
 800                         }
 801                         spin_unlock(&mmlist_lock);
 802                         mmput(prev_mm);
 803                         mmput(start_mm);
 804                         start_mm = new_start_mm;
 805                 }
 806                 if (retval) {
 807                         unlock_page(page);
 808                         page_cache_release(page);
 809                         break;
 810                 }
 811
 812                 /*
 813                  * How could swap count reach 0x7fff when the maximum
 814                  * pid is 0x7fff, and there's no way to repeat a swap
 815                  * page within an mm (except in shmem, where it's the
 816                  * shared object which takes the reference count)?
 817                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 818                  *
 819                  * If that's wrong, then we should worry more about
 820                  * exit_mmap() and do_munmap() cases described above:
 821                  * we might be resetting SWAP_MAP_MAX too early here.
 822                  * We know "Undead"s can happen, they're okay, so don't
 823                  * report them; but do report if we reset SWAP_MAP_MAX.
 824                  */
 825                 if (*swap_map == SWAP_MAP_MAX) {
 826                         spin_lock(&swap_lock);
 827                         *swap_map = 1;
 828                         spin_unlock(&swap_lock);
 829                         reset_overflow = 1;
 830                 }
 831
 832                 /*
 833                  * If a reference remains (rare), we would like to leave
 834                  * the page in the swap cache; but try_to_unmap could
 835                  * then re-duplicate the entry once we drop page lock,
 836                  * so we might loop indefinitely; also, that page could
 837                  * not be swapped out to other storage meanwhile.  So:
 838                  * delete from cache even if there's another reference,
 839                  * after ensuring that the data has been saved to disk -
 840                  * since if the reference remains (rarer), it will be
 841                  * read from disk into another page.  Splitting into two
 842                  * pages would be incorrect if swap supported "shared
 843                  * private" pages, but they are handled by tmpfs files.
 844                  *
 845                  * Note shmem_unuse already deleted a swappage from
 846                  * the swap cache, unless the move to filepage failed:
 847                  * in which case it left swappage in cache, lowered its
 848                  * swap count to pass quickly through the loops above,
 849                  * and now we must reincrement count to try again later.
 850                  */
 851                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 852                         struct writeback_control wbc = {
 853                                 .sync_mode = WB_SYNC_NONE,
 854                         };
 855
 856                         swap_writepage(page, &wbc);
 857                         lock_page(page);
 858                         wait_on_page_writeback(page);
 859                 }
 860                 if (PageSwapCache(page)) {
 861                         if (shmem)
 862                                 swap_duplicate(entry);
 863                         else
 864                                 delete_from_swap_cache(page);
 865                 }
 866
 867                 /*
 868                  * So we could skip searching mms once swap count went
 869                  * to 1, we did not mark any present ptes as dirty: must
 870                  * mark page dirty so shrink_list will preserve it.
 871                  */
 872                 SetPageDirty(page);
 873                 unlock_page(page);
 874                 page_cache_release(page);
 875
 876                 /*
 877                  * Make sure that we aren't completely killing
 878                  * interactive performance.
 879                  */
 880                 cond_resched();
 881         }
 882
 883         mmput(start_mm);
 884         if (reset_overflow) {
 885                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 886                 swap_overflow = 0;
 887         }
 888         return retval;
 889 }
 890
 891 /*
 892  * After a successful try_to_unuse, if no swap is now in use, we know
 893  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 894  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 895  * added to the mmlist just after page_duplicate - before would be racy.
 896  */
 897 static void drain_mmlist(void)
 898 {
 899         struct list_head *p, *next;
 900         unsigned int i;
 901
 902         for (i = 0; i < nr_swapfiles; i++)
 903                 if (swap_info[i].inuse_pages)
 904                         return;
 905         spin_lock(&mmlist_lock);
 906         list_for_each_safe(p, next, &init_mm.mmlist)
 907                 list_del_init(p);
 908         spin_unlock(&mmlist_lock);
 909 }
 910
 911 /*
 912  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 913  * corresponds to page offset `offset'.
 914  */
 915 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 916 {
 917         struct swap_extent *se = sis->curr_swap_extent;
 918         struct swap_extent *start_se = se;
 919
 920         for ( ; ; ) {
 921                 struct list_head *lh;
 922
 923                 if (se->start_page <= offset &&
 924                                 offset < (se->start_page + se->nr_pages)) {
 925                         return se->start_block + (offset - se->start_page);
 926                 }
 927                 lh = se->list.next;
 928                 if (lh == &sis->extent_list)
 929                         lh = lh->next;
 930                 se = list_entry(lh, struct swap_extent, list);
 931                 sis->curr_swap_extent = se;
 932                 BUG_ON(se == start_se);         /* It *must* be present */
 933         }
 934 }
 935
 936 /*
 937  * Free all of a swapdev's extent information
 938  */
 939 static void destroy_swap_extents(struct swap_info_struct *sis)
 940 {
 941         while (!list_empty(&sis->extent_list)) {
 942                 struct swap_extent *se;
 943
 944                 se = list_entry(sis->extent_list.next,
 945                                 struct swap_extent, list);
 946                 list_del(&se->list);
 947                 kfree(se);
 948         }
 949 }
 950
 951 /*
 952  * Add a block range (and the corresponding page range) into this swapdev's
 953  * extent list.  The extent list is kept sorted in page order.
 954  *
 955  * This function rather assumes that it is called in ascending page order.
 956  */
 957 static int
 958 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 959                 unsigned long nr_pages, sector_t start_block)
 960 {
 961         struct swap_extent *se;
 962         struct swap_extent *new_se;
 963         struct list_head *lh;
 964
 965         lh = sis->extent_list.prev;     /* The highest page extent */
 966         if (lh != &sis->extent_list) {
 967                 se = list_entry(lh, struct swap_extent, list);
 968                 BUG_ON(se->start_page + se->nr_pages != start_page);
 969                 if (se->start_block + se->nr_pages == start_block) {
 970                         /* Merge it */
 971                         se->nr_pages += nr_pages;
 972                         return 0;
 973                 }
 974         }
 975
 976         /*
 977          * No merge.  Insert a new extent, preserving ordering.
 978          */
 979         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 980         if (new_se == NULL)
 981                 return -ENOMEM;
 982         new_se->start_page = start_page;
 983         new_se->nr_pages = nr_pages;
 984         new_se->start_block = start_block;
 985
 986         list_add_tail(&new_se->list, &sis->extent_list);
 987         return 1;
 988 }
 989
 990 /*
 991  * A `swap extent' is a simple thing which maps a contiguous range of pages
 992  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 993  * is built at swapon time and is then used at swap_writepage/swap_readpage
 994  * time for locating where on disk a page belongs.
 995  *
 996  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 997  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 998  * swap files identically.
 999  *
1000  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1001  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1002  * swapfiles are handled *identically* after swapon time.
1003  *
1004  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1005  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1006  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1007  * requirements, they are simply tossed out - we will never use those blocks
1008  * for swapping.
1009  *
1010  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1011  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1012  * which will scribble on the fs.
1013  *
1014  * The amount of disk space which a single swap extent represents varies.
1015  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1016  * extents in the list.  To avoid much list walking, we cache the previous
1017  * search location in `curr_swap_extent', and start new searches from there.
1018  * This is extremely effective.  The average number of iterations in
1019  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1020  */
1021 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1022 {
1023         struct inode *inode;
1024         unsigned blocks_per_page;
1025         unsigned long page_no;
1026         unsigned blkbits;
1027         sector_t probe_block;
1028         sector_t last_block;
1029         sector_t lowest_block = -1;
1030         sector_t highest_block = 0;
1031         int nr_extents = 0;
1032         int ret;
1033
1034         inode = sis->swap_file->f_mapping->host;
1035         if (S_ISBLK(inode->i_mode)) {
1036                 ret = add_swap_extent(sis, 0, sis->max, 0);
1037                 *span = sis->pages;
1038                 goto done;
1039         }
1040
1041         blkbits = inode->i_blkbits;
1042         blocks_per_page = PAGE_SIZE >> blkbits;
1043
1044         /*
1045          * Map all the blocks into the extent list.  This code doesn't try
1046          * to be very smart.
1047          */
1048         probe_block = 0;
1049         page_no = 0;
1050         last_block = i_size_read(inode) >> blkbits;
1051         while ((probe_block + blocks_per_page) <= last_block &&
1052                         page_no < sis->max) {
1053                 unsigned block_in_page;
1054                 sector_t first_block;
1055
1056                 first_block = bmap(inode, probe_block);
1057                 if (first_block == 0)
1058                         goto bad_bmap;
1059
1060                 /*
1061                  * It must be PAGE_SIZE aligned on-disk
1062                  */
1063                 if (first_block & (blocks_per_page - 1)) {
1064                         probe_block++;
1065                         goto reprobe;
1066                 }
1067
1068                 for (block_in_page = 1; block_in_page < blocks_per_page;
1069                                         block_in_page++) {
1070                         sector_t block;
1071
1072                         block = bmap(inode, probe_block + block_in_page);
1073                         if (block == 0)
1074                                 goto bad_bmap;
1075                         if (block != first_block + block_in_page) {
1076                                 /* Discontiguity */
1077                                 probe_block++;
1078                                 goto reprobe;
1079                         }
1080                 }
1081
1082                 first_block >>= (PAGE_SHIFT - blkbits);
1083                 if (page_no) {  /* exclude the header page */
1084                         if (first_block < lowest_block)
1085                                 lowest_block = first_block;
1086                         if (first_block > highest_block)
1087                                 highest_block = first_block;
1088                 }
1089
1090                 /*
1091                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1092                  */
1093                 ret = add_swap_extent(sis, page_no, 1, first_block);
1094                 if (ret < 0)
1095                         goto out;
1096                 nr_extents += ret;
1097                 page_no++;
1098                 probe_block += blocks_per_page;
1099 reprobe:
1100                 continue;
1101         }
1102         ret = nr_extents;
1103         *span = 1 + highest_block - lowest_block;
1104         if (page_no == 0)
1105                 page_no = 1;    /* force Empty message */
1106         sis->max = page_no;
1107         sis->pages = page_no - 1;
1108         sis->highest_bit = page_no - 1;
1109 done:
1110         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1111                                         struct swap_extent, list);
1112         goto out;
1113 bad_bmap:
1114         printk(KERN_ERR "swapon: swapfile has holes\n");
1115         ret = -EINVAL;
1116 out:
1117         return ret;
1118 }
1119
1120 #if 0   /* We don't need this yet */
1121 #include <linux/backing-dev.h>
1122 int page_queue_congested(struct page *page)
1123 {
1124         struct backing_dev_info *bdi;
1125
1126         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1127
1128         if (PageSwapCache(page)) {
1129                 swp_entry_t entry = { .val = page_private(page) };
1130                 struct swap_info_struct *sis;
1131
1132                 sis = get_swap_info_struct(swp_type(entry));
1133                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1134         } else
1135                 bdi = page->mapping->backing_dev_info;
1136         return bdi_write_congested(bdi);
1137 }
1138 #endif
1139
1140 asmlinkage long sys_swapoff(const char __user * specialfile)
1141 {
1142         struct swap_info_struct * p = NULL;
1143         unsigned short *swap_map;
1144         struct file *swap_file, *victim;
1145         struct address_space *mapping;
1146         struct inode *inode;
1147         char * pathname;
1148         int i, type, prev;
1149         int err;
1150
1151         if (!capable(CAP_SYS_ADMIN))
1152                 return -EPERM;
1153
1154         pathname = getname(specialfile);
1155         err = PTR_ERR(pathname);
1156         if (IS_ERR(pathname))
1157                 goto out;
1158
1159         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1160         putname(pathname);
1161         err = PTR_ERR(victim);
1162         if (IS_ERR(victim))
1163                 goto out;
1164
1165         mapping = victim->f_mapping;
1166         prev = -1;
1167         spin_lock(&swap_lock);
1168         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1169                 p = swap_info + type;
1170                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1171                         if (p->swap_file->f_mapping == mapping)
1172                                 break;
1173                 }
1174                 prev = type;
1175         }
1176         if (type < 0) {
1177                 err = -EINVAL;
1178                 spin_unlock(&swap_lock);
1179                 goto out_dput;
1180         }
1181         if (!security_vm_enough_memory(p->pages))
1182                 vm_unacct_memory(p->pages);
1183         else {
1184                 err = -ENOMEM;
1185                 spin_unlock(&swap_lock);
1186                 goto out_dput;
1187         }
1188         if (prev < 0) {
1189                 swap_list.head = p->next;
1190         } else {
1191                 swap_info[prev].next = p->next;
1192         }
1193         if (type == swap_list.next) {
1194                 /* just pick something that's safe... */
1195                 swap_list.next = swap_list.head;
1196         }
1197         nr_swap_pages -= p->pages;
1198         total_swap_pages -= p->pages;
1199         p->flags &= ~SWP_WRITEOK;
1200         spin_unlock(&swap_lock);
1201
1202         current->flags |= PF_SWAPOFF;
1203         err = try_to_unuse(type);
1204         current->flags &= ~PF_SWAPOFF;
1205
1206         if (err) {
1207                 /* re-insert swap space back into swap_list */
1208                 spin_lock(&swap_lock);
1209                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1210                         if (p->prio >= swap_info[i].prio)
1211                                 break;
1212                 p->next = i;
1213                 if (prev < 0)
1214                         swap_list.head = swap_list.next = p - swap_info;
1215                 else
1216                         swap_info[prev].next = p - swap_info;
1217                 nr_swap_pages += p->pages;
1218                 total_swap_pages += p->pages;
1219                 p->flags |= SWP_WRITEOK;
1220                 spin_unlock(&swap_lock);
1221                 goto out_dput;
1222         }
1223
1224         /* wait for any unplug function to finish */
1225         down_write(&swap_unplug_sem);
1226         up_write(&swap_unplug_sem);
1227
1228         destroy_swap_extents(p);
1229         mutex_lock(&swapon_mutex);
1230         spin_lock(&swap_lock);
1231         drain_mmlist();
1232
1233         /* wait for anyone still in scan_swap_map */
1234         p->highest_bit = 0;             /* cuts scans short */
1235         while (p->flags >= SWP_SCANNING) {
1236                 spin_unlock(&swap_lock);
1237                 schedule_timeout_uninterruptible(1);
1238                 spin_lock(&swap_lock);
1239         }
1240
1241         swap_file = p->swap_file;
1242         p->swap_file = NULL;
1243         p->max = 0;
1244         swap_map = p->swap_map;
1245         p->swap_map = NULL;
1246         p->flags = 0;
1247         spin_unlock(&swap_lock);
1248         mutex_unlock(&swapon_mutex);
1249         vfree(swap_map);
1250         inode = mapping->host;
1251         if (S_ISBLK(inode->i_mode)) {
1252                 struct block_device *bdev = I_BDEV(inode);
1253                 set_blocksize(bdev, p->old_block_size);
1254                 bd_release(bdev);
1255         } else {
1256                 mutex_lock(&inode->i_mutex);
1257                 inode->i_flags &= ~S_SWAPFILE;
1258                 mutex_unlock(&inode->i_mutex);
1259         }
1260         filp_close(swap_file, NULL);
1261         err = 0;
1262
1263 out_dput:
1264         filp_close(victim, NULL);
1265 out:
1266         return err;
1267 }
1268
1269 #ifdef CONFIG_PROC_FS
1270 /* iterator */
1271 static void *swap_start(struct seq_file *swap, loff_t *pos)
1272 {
1273         struct swap_info_struct *ptr = swap_info;
1274         int i;
1275         loff_t l = *pos;
1276
1277         mutex_lock(&swapon_mutex);
1278
1279         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1280                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1281                         continue;
1282                 if (!l--)
1283                         return ptr;
1284         }
1285
1286         return NULL;
1287 }
1288
1289 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1290 {
1291         struct swap_info_struct *ptr = v;
1292         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1293
1294         for (++ptr; ptr < endptr; ptr++) {
1295                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1296                         continue;
1297                 ++*pos;
1298                 return ptr;
1299         }
1300
1301         return NULL;
1302 }
1303
1304 static void swap_stop(struct seq_file *swap, void *v)
1305 {
1306         mutex_unlock(&swapon_mutex);
1307 }
1308
1309 static int swap_show(struct seq_file *swap, void *v)
1310 {
1311         struct swap_info_struct *ptr = v;
1312         struct file *file;
1313         int len;
1314
1315         if (v == swap_info)
1316                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1317
1318         file = ptr->swap_file;
1319         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1320         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1321                        len < 40 ? 40 - len : 1, " ",
1322                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1323                                 "partition" : "file\t",
1324                        ptr->pages << (PAGE_SHIFT - 10),
1325                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1326                        ptr->prio);
1327         return 0;
1328 }
1329
1330 static struct seq_operations swaps_op = {
1331         .start =        swap_start,
1332         .next =         swap_next,
1333         .stop =         swap_stop,
1334         .show =         swap_show
1335 };
1336
1337 static int swaps_open(struct inode *inode, struct file *file)
1338 {
1339         return seq_open(file, &swaps_op);
1340 }
1341
1342 static struct file_operations proc_swaps_operations = {
1343         .open           = swaps_open,
1344         .read           = seq_read,
1345         .llseek         = seq_lseek,
1346         .release        = seq_release,
1347 };
1348
1349 static int __init procswaps_init(void)
1350 {
1351         struct proc_dir_entry *entry;
1352
1353         entry = create_proc_entry("swaps", 0, NULL);
1354         if (entry)
1355                 entry->proc_fops = &proc_swaps_operations;
1356         return 0;
1357 }
1358 __initcall(procswaps_init);
1359 #endif /* CONFIG_PROC_FS */
1360
1361 /*
1362  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1363  *
1364  * The swapon system call
1365  */
1366 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1367 {
1368         struct swap_info_struct * p;
1369         char *name = NULL;
1370         struct block_device *bdev = NULL;
1371         struct file *swap_file = NULL;
1372         struct address_space *mapping;
1373         unsigned int type;
1374         int i, prev;
1375         int error;
1376         static int least_priority;
1377         union swap_header *swap_header = NULL;
1378         int swap_header_version;
1379         unsigned int nr_good_pages = 0;
1380         int nr_extents = 0;
1381         sector_t span;
1382         unsigned long maxpages = 1;
1383         int swapfilesize;
1384         unsigned short *swap_map;
1385         struct page *page = NULL;
1386         struct inode *inode = NULL;
1387         int did_down = 0;
1388
1389         if (!capable(CAP_SYS_ADMIN))
1390                 return -EPERM;
1391         spin_lock(&swap_lock);
1392         p = swap_info;
1393         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1394                 if (!(p->flags & SWP_USED))
1395                         break;
1396         error = -EPERM;
1397         if (type >= MAX_SWAPFILES) {
1398                 spin_unlock(&swap_lock);
1399                 goto out;
1400         }
1401         if (type >= nr_swapfiles)
1402                 nr_swapfiles = type+1;
1403         INIT_LIST_HEAD(&p->extent_list);
1404         p->flags = SWP_USED;
1405         p->swap_file = NULL;
1406         p->old_block_size = 0;
1407         p->swap_map = NULL;
1408         p->lowest_bit = 0;
1409         p->highest_bit = 0;
1410         p->cluster_nr = 0;
1411         p->inuse_pages = 0;
1412         p->next = -1;
1413         if (swap_flags & SWAP_FLAG_PREFER) {
1414                 p->prio =
1415                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1416         } else {
1417                 p->prio = --least_priority;
1418         }
1419         spin_unlock(&swap_lock);
1420         name = getname(specialfile);
1421         error = PTR_ERR(name);
1422         if (IS_ERR(name)) {
1423                 name = NULL;
1424                 goto bad_swap_2;
1425         }
1426         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1427         error = PTR_ERR(swap_file);
1428         if (IS_ERR(swap_file)) {
1429                 swap_file = NULL;
1430                 goto bad_swap_2;
1431         }
1432
1433         p->swap_file = swap_file;
1434         mapping = swap_file->f_mapping;
1435         inode = mapping->host;
1436
1437         error = -EBUSY;
1438         for (i = 0; i < nr_swapfiles; i++) {
1439                 struct swap_info_struct *q = &swap_info[i];
1440
1441                 if (i == type || !q->swap_file)
1442                         continue;
1443                 if (mapping == q->swap_file->f_mapping)
1444                         goto bad_swap;
1445         }
1446
1447         error = -EINVAL;
1448         if (S_ISBLK(inode->i_mode)) {
1449                 bdev = I_BDEV(inode);
1450                 error = bd_claim(bdev, sys_swapon);
1451                 if (error < 0) {
1452                         bdev = NULL;
1453                         error = -EINVAL;
1454                         goto bad_swap;
1455                 }
1456                 p->old_block_size = block_size(bdev);
1457                 error = set_blocksize(bdev, PAGE_SIZE);
1458                 if (error < 0)
1459                         goto bad_swap;
1460                 p->bdev = bdev;
1461         } else if (S_ISREG(inode->i_mode)) {
1462                 p->bdev = inode->i_sb->s_bdev;
1463                 mutex_lock(&inode->i_mutex);
1464                 did_down = 1;
1465                 if (IS_SWAPFILE(inode)) {
1466                         error = -EBUSY;
1467                         goto bad_swap;
1468                 }
1469         } else {
1470                 goto bad_swap;
1471         }
1472
1473         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1474
1475         /*
1476          * Read the swap header.
1477          */
1478         if (!mapping->a_ops->readpage) {
1479                 error = -EINVAL;
1480                 goto bad_swap;
1481         }
1482         page = read_mapping_page(mapping, 0, swap_file);
1483         if (IS_ERR(page)) {
1484                 error = PTR_ERR(page);
1485                 goto bad_swap;
1486         }
1487         wait_on_page_locked(page);
1488         if (!PageUptodate(page))
1489                 goto bad_swap;
1490         kmap(page);
1491         swap_header = page_address(page);
1492
1493         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1494                 swap_header_version = 1;
1495         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1496                 swap_header_version = 2;
1497         else {
1498                 printk(KERN_ERR "Unable to find swap-space signature\n");
1499                 error = -EINVAL;
1500                 goto bad_swap;
1501         }
1502
1503         switch (swap_header_version) {
1504         case 1:
1505                 printk(KERN_ERR "version 0 swap is no longer supported. "
1506                         "Use mkswap -v1 %s\n", name);
1507                 error = -EINVAL;
1508                 goto bad_swap;
1509         case 2:
1510                 /* Check the swap header's sub-version and the size of
1511                    the swap file and bad block lists */
1512                 if (swap_header->info.version != 1) {
1513                         printk(KERN_WARNING
1514                                "Unable to handle swap header version %d\n",
1515                                swap_header->info.version);
1516                         error = -EINVAL;
1517                         goto bad_swap;
1518                 }
1519
1520                 p->lowest_bit  = 1;
1521                 p->cluster_next = 1;
1522
1523                 /*
1524                  * Find out how many pages are allowed for a single swap
1525                  * device. There are two limiting factors: 1) the number of
1526                  * bits for the swap offset in the swp_entry_t type and
1527                  * 2) the number of bits in the a swap pte as defined by
1528                  * the different architectures. In order to find the
1529                  * largest possible bit mask a swap entry with swap type 0
1530                  * and swap offset ~0UL is created, encoded to a swap pte,
1531                  * decoded to a swp_entry_t again and finally the swap
1532                  * offset is extracted. This will mask all the bits from
1533                  * the initial ~0UL mask that can't be encoded in either
1534                  * the swp_entry_t or the architecture definition of a
1535                  * swap pte.
1536                  */
1537                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1538                 if (maxpages > swap_header->info.last_page)
1539                         maxpages = swap_header->info.last_page;
1540                 p->highest_bit = maxpages - 1;
1541
1542                 error = -EINVAL;
1543                 if (!maxpages)
1544                         goto bad_swap;
1545                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1546                         goto bad_swap;
1547                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1548                         goto bad_swap;
1549
1550                 /* OK, set up the swap map and apply the bad block list */
1551                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1552                         error = -ENOMEM;
1553                         goto bad_swap;
1554                 }
1555
1556                 error = 0;
1557                 memset(p->swap_map, 0, maxpages * sizeof(short));
1558                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1559                         int page_nr = swap_header->info.badpages[i];
1560                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1561                                 error = -EINVAL;
1562                         else
1563                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1564                 }
1565                 nr_good_pages = swap_header->info.last_page -
1566                                 swap_header->info.nr_badpages -
1567                                 1 /* header page */;
1568                 if (error)
1569                         goto bad_swap;
1570         }
1571
1572         if (swapfilesize && maxpages > swapfilesize) {
1573                 printk(KERN_WARNING
1574                        "Swap area shorter than signature indicates\n");
1575                 error = -EINVAL;
1576                 goto bad_swap;
1577         }
1578         if (nr_good_pages) {
1579                 p->swap_map[0] = SWAP_MAP_BAD;
1580                 p->max = maxpages;
1581                 p->pages = nr_good_pages;
1582                 nr_extents = setup_swap_extents(p, &span);
1583                 if (nr_extents < 0) {
1584                         error = nr_extents;
1585                         goto bad_swap;
1586                 }
1587                 nr_good_pages = p->pages;
1588         }
1589         if (!nr_good_pages) {
1590                 printk(KERN_WARNING "Empty swap-file\n");
1591                 error = -EINVAL;
1592                 goto bad_swap;
1593         }
1594
1595         mutex_lock(&swapon_mutex);
1596         spin_lock(&swap_lock);
1597         p->flags = SWP_ACTIVE;
1598         nr_swap_pages += nr_good_pages;
1599         total_swap_pages += nr_good_pages;
1600
1601         printk(KERN_INFO "Adding %uk swap on %s.  "
1602                         "Priority:%d extents:%d across:%lluk\n",
1603                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1604                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1605
1606         /* insert swap space into swap_list: */
1607         prev = -1;
1608         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1609                 if (p->prio >= swap_info[i].prio) {
1610                         break;
1611                 }
1612                 prev = i;
1613         }
1614         p->next = i;
1615         if (prev < 0) {
1616                 swap_list.head = swap_list.next = p - swap_info;
1617         } else {
1618                 swap_info[prev].next = p - swap_info;
1619         }
1620         spin_unlock(&swap_lock);
1621         mutex_unlock(&swapon_mutex);
1622         error = 0;
1623         goto out;
1624 bad_swap:
1625         if (bdev) {
1626                 set_blocksize(bdev, p->old_block_size);
1627                 bd_release(bdev);
1628         }
1629         destroy_swap_extents(p);
1630 bad_swap_2:
1631         spin_lock(&swap_lock);
1632         swap_map = p->swap_map;
1633         p->swap_file = NULL;
1634         p->swap_map = NULL;
1635         p->flags = 0;
1636         if (!(swap_flags & SWAP_FLAG_PREFER))
1637                 ++least_priority;
1638         spin_unlock(&swap_lock);
1639         vfree(swap_map);
1640         if (swap_file)
1641                 filp_close(swap_file, NULL);
1642 out:
1643         if (page && !IS_ERR(page)) {
1644                 kunmap(page);
1645                 page_cache_release(page);
1646         }
1647         if (name)
1648                 putname(name);
1649         if (did_down) {
1650                 if (!error)
1651                         inode->i_flags |= S_SWAPFILE;
1652                 mutex_unlock(&inode->i_mutex);
1653         }
1654         return error;
1655 }
1656
1657 void si_swapinfo(struct sysinfo *val)
1658 {
1659         unsigned int i;
1660         unsigned long nr_to_be_unused = 0;
1661
1662         spin_lock(&swap_lock);
1663         for (i = 0; i < nr_swapfiles; i++) {
1664                 if (!(swap_info[i].flags & SWP_USED) ||
1665                      (swap_info[i].flags & SWP_WRITEOK))
1666                         continue;
1667                 nr_to_be_unused += swap_info[i].inuse_pages;
1668         }
1669         val->freeswap = nr_swap_pages + nr_to_be_unused;
1670         val->totalswap = total_swap_pages + nr_to_be_unused;
1671         spin_unlock(&swap_lock);
1672         if (vx_flags(VXF_VIRT_MEM, 0))
1673                 vx_vsi_swapinfo(val);
1674 }
1675
1676 /*
1677  * Verify that a swap entry is valid and increment its swap map count.
1678  *
1679  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1680  * "permanent", but will be reclaimed by the next swapoff.
1681  */
1682 int swap_duplicate(swp_entry_t entry)
1683 {
1684         struct swap_info_struct * p;
1685         unsigned long offset, type;
1686         int result = 0;
1687
1688         if (is_migration_entry(entry))
1689                 return 1;
1690
1691         type = swp_type(entry);
1692         if (type >= nr_swapfiles)
1693                 goto bad_file;
1694         p = type + swap_info;
1695         offset = swp_offset(entry);
1696
1697         spin_lock(&swap_lock);
1698         if (offset < p->max && p->swap_map[offset]) {
1699                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1700                         p->swap_map[offset]++;
1701                         result = 1;
1702                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1703                         if (swap_overflow++ < 5)
1704                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1705                         p->swap_map[offset] = SWAP_MAP_MAX;
1706                         result = 1;
1707                 }
1708         }
1709         spin_unlock(&swap_lock);
1710 out:
1711         return result;
1712
1713 bad_file:
1714         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1715         goto out;
1716 }
1717
1718 struct swap_info_struct *
1719 get_swap_info_struct(unsigned type)
1720 {
1721         return &swap_info[type];
1722 }
1723
1724 /*
1725  * swap_lock prevents swap_map being freed. Don't grab an extra
1726  * reference on the swaphandle, it doesn't matter if it becomes unused.
1727  */
1728 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1729 {
1730         int ret = 0, i = 1 << page_cluster;
1731         unsigned long toff;
1732         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1733
1734         if (!page_cluster)      /* no readahead */
1735                 return 0;
1736         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1737         if (!toff)              /* first page is swap header */
1738                 toff++, i--;
1739         *offset = toff;
1740
1741         spin_lock(&swap_lock);
1742         do {
1743                 /* Don't read-ahead past the end of the swap area */
1744                 if (toff >= swapdev->max)
1745                         break;
1746                 /* Don't read in free or bad pages */
1747                 if (!swapdev->swap_map[toff])
1748                         break;
1749                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1750                         break;
1751                 toff++;
1752                 ret++;
1753         } while (--i);
1754         spin_unlock(&swap_lock);
1755         return ret;
1756 }