mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/config.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/mman.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/namei.h>
  18 #include <linux/shm.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/mutex.h>
  29 #include <linux/capability.h>
  30 #include <linux/syscalls.h>
  31
  32 #include <asm/pgtable.h>
  33 #include <asm/tlbflush.h>
  34 #include <linux/swapops.h>
  35
  36 DEFINE_SPINLOCK(swap_lock);
  37 unsigned int nr_swapfiles;
  38 long total_swap_pages;
  39 static int swap_overflow;
  40
  41 static const char Bad_file[] = "Bad swap file entry ";
  42 static const char Unused_file[] = "Unused swap file entry ";
  43 static const char Bad_offset[] = "Bad swap offset entry ";
  44 static const char Unused_offset[] = "Unused swap offset entry ";
  45
  46 struct swap_list_t swap_list = {-1, -1};
  47
  48 struct swap_info_struct swap_info[MAX_SWAPFILES];
  49
  50 static DEFINE_MUTEX(swapon_mutex);
  51
  52 /*
  53  * We need this because the bdev->unplug_fn can sleep and we cannot
  54  * hold swap_lock while calling the unplug_fn. And swap_lock
  55  * cannot be turned into a mutex.
  56  */
  57 static DECLARE_RWSEM(swap_unplug_sem);
  58
  59 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  60 {
  61         swp_entry_t entry;
  62
  63         down_read(&swap_unplug_sem);
  64         entry.val = page_private(page);
  65         if (PageSwapCache(page)) {
  66                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  67                 struct backing_dev_info *bdi;
  68
  69                 /*
  70                  * If the page is removed from swapcache from under us (with a
  71                  * racy try_to_unuse/swapoff) we need an additional reference
  72                  * count to avoid reading garbage from page_private(page) above.
  73                  * If the WARN_ON triggers during a swapoff it maybe the race
  74                  * condition and it's harmless. However if it triggers without
  75                  * swapoff it signals a problem.
  76                  */
  77                 WARN_ON(page_count(page) <= 1);
  78
  79                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  80                 blk_run_backing_dev(bdi, page);
  81         }
  82         up_read(&swap_unplug_sem);
  83 }
  84
  85 #define SWAPFILE_CLUSTER        256
  86 #define LATENCY_LIMIT           256
  87
  88 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
  89 {
  90         unsigned long offset, last_in_cluster;
  91         int latency_ration = LATENCY_LIMIT;
  92
  93         /*
  94          * We try to cluster swap pages by allocating them sequentially
  95          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
  96          * way, however, we resort to first-free allocation, starting
  97          * a new cluster.  This prevents us from scattering swap pages
  98          * all over the entire swap partition, so that we reduce
  99          * overall disk seek times between swap pages.  -- sct
 100          * But we do now try to find an empty cluster.  -Andrea
 101          */
 102
 103         si->flags += SWP_SCANNING;
 104         if (unlikely(!si->cluster_nr)) {
 105                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 106                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
 107                         goto lowest;
 108                 spin_unlock(&swap_lock);
 109
 110                 offset = si->lowest_bit;
 111                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 112
 113                 /* Locate the first empty (unaligned) cluster */
 114                 for (; last_in_cluster <= si->highest_bit; offset++) {
 115                         if (si->swap_map[offset])
 116                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 117                         else if (offset == last_in_cluster) {
 118                                 spin_lock(&swap_lock);
 119                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
 120                                 goto cluster;
 121                         }
 122                         if (unlikely(--latency_ration < 0)) {
 123                                 cond_resched();
 124                                 latency_ration = LATENCY_LIMIT;
 125                         }
 126                 }
 127                 spin_lock(&swap_lock);
 128                 goto lowest;
 129         }
 130
 131         si->cluster_nr--;
 132 cluster:
 133         offset = si->cluster_next;
 134         if (offset > si->highest_bit)
 135 lowest:         offset = si->lowest_bit;
 136 checks: if (!(si->flags & SWP_WRITEOK))
 137                 goto no_page;
 138         if (!si->highest_bit)
 139                 goto no_page;
 140         if (!si->swap_map[offset]) {
 141                 if (offset == si->lowest_bit)
 142                         si->lowest_bit++;
 143                 if (offset == si->highest_bit)
 144                         si->highest_bit--;
 145                 si->inuse_pages++;
 146                 if (si->inuse_pages == si->pages) {
 147                         si->lowest_bit = si->max;
 148                         si->highest_bit = 0;
 149                 }
 150                 si->swap_map[offset] = 1;
 151                 si->cluster_next = offset + 1;
 152                 si->flags -= SWP_SCANNING;
 153                 return offset;
 154         }
 155
 156         spin_unlock(&swap_lock);
 157         while (++offset <= si->highest_bit) {
 158                 if (!si->swap_map[offset]) {
 159                         spin_lock(&swap_lock);
 160                         goto checks;
 161                 }
 162                 if (unlikely(--latency_ration < 0)) {
 163                         cond_resched();
 164                         latency_ration = LATENCY_LIMIT;
 165                 }
 166         }
 167         spin_lock(&swap_lock);
 168         goto lowest;
 169
 170 no_page:
 171         si->flags -= SWP_SCANNING;
 172         return 0;
 173 }
 174
 175 swp_entry_t get_swap_page(void)
 176 {
 177         struct swap_info_struct *si;
 178         pgoff_t offset;
 179         int type, next;
 180         int wrapped = 0;
 181
 182         spin_lock(&swap_lock);
 183         if (nr_swap_pages <= 0)
 184                 goto noswap;
 185         nr_swap_pages--;
 186
 187         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 188                 si = swap_info + type;
 189                 next = si->next;
 190                 if (next < 0 ||
 191                     (!wrapped && si->prio != swap_info[next].prio)) {
 192                         next = swap_list.head;
 193                         wrapped++;
 194                 }
 195
 196                 if (!si->highest_bit)
 197                         continue;
 198                 if (!(si->flags & SWP_WRITEOK))
 199                         continue;
 200
 201                 swap_list.next = next;
 202                 offset = scan_swap_map(si);
 203                 if (offset) {
 204                         spin_unlock(&swap_lock);
 205                         return swp_entry(type, offset);
 206                 }
 207                 next = swap_list.next;
 208         }
 209
 210         nr_swap_pages++;
 211 noswap:
 212         spin_unlock(&swap_lock);
 213         return (swp_entry_t) {0};
 214 }
 215
 216 swp_entry_t get_swap_page_of_type(int type)
 217 {
 218         struct swap_info_struct *si;
 219         pgoff_t offset;
 220
 221         spin_lock(&swap_lock);
 222         si = swap_info + type;
 223         if (si->flags & SWP_WRITEOK) {
 224                 nr_swap_pages--;
 225                 offset = scan_swap_map(si);
 226                 if (offset) {
 227                         spin_unlock(&swap_lock);
 228                         return swp_entry(type, offset);
 229                 }
 230                 nr_swap_pages++;
 231         }
 232         spin_unlock(&swap_lock);
 233         return (swp_entry_t) {0};
 234 }
 235
 236 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 237 {
 238         struct swap_info_struct * p;
 239         unsigned long offset, type;
 240
 241         if (!entry.val)
 242                 goto out;
 243         type = swp_type(entry);
 244         if (type >= nr_swapfiles)
 245                 goto bad_nofile;
 246         p = & swap_info[type];
 247         if (!(p->flags & SWP_USED))
 248                 goto bad_device;
 249         offset = swp_offset(entry);
 250         if (offset >= p->max)
 251                 goto bad_offset;
 252         if (!p->swap_map[offset])
 253                 goto bad_free;
 254         spin_lock(&swap_lock);
 255         return p;
 256
 257 bad_free:
 258         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 259         goto out;
 260 bad_offset:
 261         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 262         goto out;
 263 bad_device:
 264         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 265         goto out;
 266 bad_nofile:
 267         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 268 out:
 269         return NULL;
 270 }
 271
 272 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 273 {
 274         int count = p->swap_map[offset];
 275
 276         if (count < SWAP_MAP_MAX) {
 277                 count--;
 278                 p->swap_map[offset] = count;
 279                 if (!count) {
 280                         if (offset < p->lowest_bit)
 281                                 p->lowest_bit = offset;
 282                         if (offset > p->highest_bit)
 283                                 p->highest_bit = offset;
 284                         if (p->prio > swap_info[swap_list.next].prio)
 285                                 swap_list.next = p - swap_info;
 286                         nr_swap_pages++;
 287                         p->inuse_pages--;
 288                 }
 289         }
 290         return count;
 291 }
 292
 293 /*
 294  * Caller has made sure that the swapdevice corresponding to entry
 295  * is still around or has not been recycled.
 296  */
 297 void swap_free(swp_entry_t entry)
 298 {
 299         struct swap_info_struct * p;
 300
 301         p = swap_info_get(entry);
 302         if (p) {
 303                 swap_entry_free(p, swp_offset(entry));
 304                 spin_unlock(&swap_lock);
 305         }
 306 }
 307
 308 /*
 309  * How many references to page are currently swapped out?
 310  */
 311 static inline int page_swapcount(struct page *page)
 312 {
 313         int count = 0;
 314         struct swap_info_struct *p;
 315         swp_entry_t entry;
 316
 317         entry.val = page_private(page);
 318         p = swap_info_get(entry);
 319         if (p) {
 320                 /* Subtract the 1 for the swap cache itself */
 321                 count = p->swap_map[swp_offset(entry)] - 1;
 322                 spin_unlock(&swap_lock);
 323         }
 324         return count;
 325 }
 326
 327 /*
 328  * We can use this swap cache entry directly
 329  * if there are no other references to it.
 330  */
 331 int can_share_swap_page(struct page *page)
 332 {
 333         int count;
 334
 335         BUG_ON(!PageLocked(page));
 336         count = page_mapcount(page);
 337         if (count <= 1 && PageSwapCache(page))
 338                 count += page_swapcount(page);
 339         return count == 1;
 340 }
 341
 342 /*
 343  * Work out if there are any other processes sharing this
 344  * swap cache page. Free it if you can. Return success.
 345  */
 346 int remove_exclusive_swap_page(struct page *page)
 347 {
 348         int retval;
 349         struct swap_info_struct * p;
 350         swp_entry_t entry;
 351
 352         BUG_ON(PagePrivate(page));
 353         BUG_ON(!PageLocked(page));
 354
 355         if (!PageSwapCache(page))
 356                 return 0;
 357         if (PageWriteback(page))
 358                 return 0;
 359         if (page_count(page) != 2) /* 2: us + cache */
 360                 return 0;
 361
 362         entry.val = page_private(page);
 363         p = swap_info_get(entry);
 364         if (!p)
 365                 return 0;
 366
 367         /* Is the only swap cache user the cache itself? */
 368         retval = 0;
 369         if (p->swap_map[swp_offset(entry)] == 1) {
 370                 /* Recheck the page count with the swapcache lock held.. */
 371                 write_lock_irq(&swapper_space.tree_lock);
 372                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 373                         __delete_from_swap_cache(page);
 374                         SetPageDirty(page);
 375                         retval = 1;
 376                 }
 377                 write_unlock_irq(&swapper_space.tree_lock);
 378         }
 379         spin_unlock(&swap_lock);
 380
 381         if (retval) {
 382                 swap_free(entry);
 383                 page_cache_release(page);
 384         }
 385
 386         return retval;
 387 }
 388
 389 /*
 390  * Free the swap entry like above, but also try to
 391  * free the page cache entry if it is the last user.
 392  */
 393 void free_swap_and_cache(swp_entry_t entry)
 394 {
 395         struct swap_info_struct * p;
 396         struct page *page = NULL;
 397
 398         p = swap_info_get(entry);
 399         if (p) {
 400                 if (swap_entry_free(p, swp_offset(entry)) == 1)
 401                         page = find_trylock_page(&swapper_space, entry.val);
 402                 spin_unlock(&swap_lock);
 403         }
 404         if (page) {
 405                 int one_user;
 406
 407                 BUG_ON(PagePrivate(page));
 408                 page_cache_get(page);
 409                 one_user = (page_count(page) == 2);
 410                 /* Only cache user (+us), or swap space full? Free it! */
 411                 if (!PageWriteback(page) && (one_user || vm_swap_full())) {
 412                         delete_from_swap_cache(page);
 413                         SetPageDirty(page);
 414                 }
 415                 unlock_page(page);
 416                 page_cache_release(page);
 417         }
 418 }
 419
 420 /*
 421  * No need to decide whether this PTE shares the swap entry with others,
 422  * just let do_wp_page work it out if a write is requested later - to
 423  * force COW, vm_page_prot omits write permission from any private vma.
 424  */
 425 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
 426                 unsigned long addr, swp_entry_t entry, struct page *page)
 427 {
 428         inc_mm_counter(vma->vm_mm, anon_rss);
 429         get_page(page);
 430         set_pte_at(vma->vm_mm, addr, pte,
 431                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 432         page_add_anon_rmap(page, vma, addr);
 433         swap_free(entry);
 434         /*
 435          * Move the page to the active list so it is not
 436          * immediately swapped out again after swapon.
 437          */
 438         activate_page(page);
 439 }
 440
 441 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 442                                 unsigned long addr, unsigned long end,
 443                                 swp_entry_t entry, struct page *page)
 444 {
 445         pte_t swp_pte = swp_entry_to_pte(entry);
 446         pte_t *pte;
 447         spinlock_t *ptl;
 448         int found = 0;
 449
 450         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 451         do {
 452                 /*
 453                  * swapoff spends a _lot_ of time in this loop!
 454                  * Test inline before going to call unuse_pte.
 455                  */
 456                 if (unlikely(pte_same(*pte, swp_pte))) {
 457                         unuse_pte(vma, pte++, addr, entry, page);
 458                         found = 1;
 459                         break;
 460                 }
 461         } while (pte++, addr += PAGE_SIZE, addr != end);
 462         pte_unmap_unlock(pte - 1, ptl);
 463         return found;
 464 }
 465
 466 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 467                                 unsigned long addr, unsigned long end,
 468                                 swp_entry_t entry, struct page *page)
 469 {
 470         pmd_t *pmd;
 471         unsigned long next;
 472
 473         pmd = pmd_offset(pud, addr);
 474         do {
 475                 next = pmd_addr_end(addr, end);
 476                 if (pmd_none_or_clear_bad(pmd))
 477                         continue;
 478                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 479                         return 1;
 480         } while (pmd++, addr = next, addr != end);
 481         return 0;
 482 }
 483
 484 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 485                                 unsigned long addr, unsigned long end,
 486                                 swp_entry_t entry, struct page *page)
 487 {
 488         pud_t *pud;
 489         unsigned long next;
 490
 491         pud = pud_offset(pgd, addr);
 492         do {
 493                 next = pud_addr_end(addr, end);
 494                 if (pud_none_or_clear_bad(pud))
 495                         continue;
 496                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 497                         return 1;
 498         } while (pud++, addr = next, addr != end);
 499         return 0;
 500 }
 501
 502 static int unuse_vma(struct vm_area_struct *vma,
 503                                 swp_entry_t entry, struct page *page)
 504 {
 505         pgd_t *pgd;
 506         unsigned long addr, end, next;
 507
 508         if (page->mapping) {
 509                 addr = page_address_in_vma(page, vma);
 510                 if (addr == -EFAULT)
 511                         return 0;
 512                 else
 513                         end = addr + PAGE_SIZE;
 514         } else {
 515                 addr = vma->vm_start;
 516                 end = vma->vm_end;
 517         }
 518
 519         pgd = pgd_offset(vma->vm_mm, addr);
 520         do {
 521                 next = pgd_addr_end(addr, end);
 522                 if (pgd_none_or_clear_bad(pgd))
 523                         continue;
 524                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 525                         return 1;
 526         } while (pgd++, addr = next, addr != end);
 527         return 0;
 528 }
 529
 530 static int unuse_mm(struct mm_struct *mm,
 531                                 swp_entry_t entry, struct page *page)
 532 {
 533         struct vm_area_struct *vma;
 534
 535         if (!down_read_trylock(&mm->mmap_sem)) {
 536                 /*
 537                  * Activate page so shrink_cache is unlikely to unmap its
 538                  * ptes while lock is dropped, so swapoff can make progress.
 539                  */
 540                 activate_page(page);
 541                 unlock_page(page);
 542                 down_read(&mm->mmap_sem);
 543                 lock_page(page);
 544         }
 545         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 546                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 547                         break;
 548         }
 549         up_read(&mm->mmap_sem);
 550         /*
 551          * Currently unuse_mm cannot fail, but leave error handling
 552          * at call sites for now, since we change it from time to time.
 553          */
 554         return 0;
 555 }
 556
 557 #ifdef CONFIG_MIGRATION
 558 int remove_vma_swap(struct vm_area_struct *vma, struct page *page)
 559 {
 560         swp_entry_t entry = { .val = page_private(page) };
 561
 562         return unuse_vma(vma, entry, page);
 563 }
 564 #endif
 565
 566 /*
 567  * Scan swap_map from current position to next entry still in use.
 568  * Recycle to start on reaching the end, returning 0 when empty.
 569  */
 570 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 571                                         unsigned int prev)
 572 {
 573         unsigned int max = si->max;
 574         unsigned int i = prev;
 575         int count;
 576
 577         /*
 578          * No need for swap_lock here: we're just looking
 579          * for whether an entry is in use, not modifying it; false
 580          * hits are okay, and sys_swapoff() has already prevented new
 581          * allocations from this area (while holding swap_lock).
 582          */
 583         for (;;) {
 584                 if (++i >= max) {
 585                         if (!prev) {
 586                                 i = 0;
 587                                 break;
 588                         }
 589                         /*
 590                          * No entries in use at top of swap_map,
 591                          * loop back to start and recheck there.
 592                          */
 593                         max = prev + 1;
 594                         prev = 0;
 595                         i = 1;
 596                 }
 597                 count = si->swap_map[i];
 598                 if (count && count != SWAP_MAP_BAD)
 599                         break;
 600         }
 601         return i;
 602 }
 603
 604 /*
 605  * We completely avoid races by reading each swap page in advance,
 606  * and then search for the process using it.  All the necessary
 607  * page table adjustments can then be made atomically.
 608  */
 609 static int try_to_unuse(unsigned int type)
 610 {
 611         struct swap_info_struct * si = &swap_info[type];
 612         struct mm_struct *start_mm;
 613         unsigned short *swap_map;
 614         unsigned short swcount;
 615         struct page *page;
 616         swp_entry_t entry;
 617         unsigned int i = 0;
 618         int retval = 0;
 619         int reset_overflow = 0;
 620         int shmem;
 621
 622         /*
 623          * When searching mms for an entry, a good strategy is to
 624          * start at the first mm we freed the previous entry from
 625          * (though actually we don't notice whether we or coincidence
 626          * freed the entry).  Initialize this start_mm with a hold.
 627          *
 628          * A simpler strategy would be to start at the last mm we
 629          * freed the previous entry from; but that would take less
 630          * advantage of mmlist ordering, which clusters forked mms
 631          * together, child after parent.  If we race with dup_mmap(), we
 632          * prefer to resolve parent before child, lest we miss entries
 633          * duplicated after we scanned child: using last mm would invert
 634          * that.  Though it's only a serious concern when an overflowed
 635          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 636          */
 637         start_mm = &init_mm;
 638         atomic_inc(&init_mm.mm_users);
 639
 640         /*
 641          * Keep on scanning until all entries have gone.  Usually,
 642          * one pass through swap_map is enough, but not necessarily:
 643          * there are races when an instance of an entry might be missed.
 644          */
 645         while ((i = find_next_to_unuse(si, i)) != 0) {
 646                 if (signal_pending(current)) {
 647                         retval = -EINTR;
 648                         break;
 649                 }
 650
 651                 /*
 652                  * Get a page for the entry, using the existing swap
 653                  * cache page if there is one.  Otherwise, get a clean
 654                  * page and read the swap into it.
 655                  */
 656                 swap_map = &si->swap_map[i];
 657                 entry = swp_entry(type, i);
 658 again:
 659                 page = read_swap_cache_async(entry, NULL, 0);
 660                 if (!page) {
 661                         /*
 662                          * Either swap_duplicate() failed because entry
 663                          * has been freed independently, and will not be
 664                          * reused since sys_swapoff() already disabled
 665                          * allocation from here, or alloc_page() failed.
 666                          */
 667                         if (!*swap_map)
 668                                 continue;
 669                         retval = -ENOMEM;
 670                         break;
 671                 }
 672
 673                 /*
 674                  * Don't hold on to start_mm if it looks like exiting.
 675                  */
 676                 if (atomic_read(&start_mm->mm_users) == 1) {
 677                         mmput(start_mm);
 678                         start_mm = &init_mm;
 679                         atomic_inc(&init_mm.mm_users);
 680                 }
 681
 682                 /*
 683                  * Wait for and lock page.  When do_swap_page races with
 684                  * try_to_unuse, do_swap_page can handle the fault much
 685                  * faster than try_to_unuse can locate the entry.  This
 686                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 687                  * defer to do_swap_page in such a case - in some tests,
 688                  * do_swap_page and try_to_unuse repeatedly compete.
 689                  */
 690                 wait_on_page_locked(page);
 691                 wait_on_page_writeback(page);
 692                 lock_page(page);
 693                 if (!PageSwapCache(page)) {
 694                         /* Page migration has occured */
 695                         unlock_page(page);
 696                         page_cache_release(page);
 697                         goto again;
 698                 }
 699                 wait_on_page_writeback(page);
 700
 701                 /*
 702                  * Remove all references to entry.
 703                  * Whenever we reach init_mm, there's no address space
 704                  * to search, but use it as a reminder to search shmem.
 705                  */
 706                 shmem = 0;
 707                 swcount = *swap_map;
 708                 if (swcount > 1) {
 709                         if (start_mm == &init_mm)
 710                                 shmem = shmem_unuse(entry, page);
 711                         else
 712                                 retval = unuse_mm(start_mm, entry, page);
 713                 }
 714                 if (*swap_map > 1) {
 715                         int set_start_mm = (*swap_map >= swcount);
 716                         struct list_head *p = &start_mm->mmlist;
 717                         struct mm_struct *new_start_mm = start_mm;
 718                         struct mm_struct *prev_mm = start_mm;
 719                         struct mm_struct *mm;
 720
 721                         atomic_inc(&new_start_mm->mm_users);
 722                         atomic_inc(&prev_mm->mm_users);
 723                         spin_lock(&mmlist_lock);
 724                         while (*swap_map > 1 && !retval &&
 725                                         (p = p->next) != &start_mm->mmlist) {
 726                                 mm = list_entry(p, struct mm_struct, mmlist);
 727                                 if (atomic_inc_return(&mm->mm_users) == 1) {
 728                                         atomic_dec(&mm->mm_users);
 729                                         continue;
 730                                 }
 731                                 spin_unlock(&mmlist_lock);
 732                                 mmput(prev_mm);
 733                                 prev_mm = mm;
 734
 735                                 cond_resched();
 736
 737                                 swcount = *swap_map;
 738                                 if (swcount <= 1)
 739                                         ;
 740                                 else if (mm == &init_mm) {
 741                                         set_start_mm = 1;
 742                                         shmem = shmem_unuse(entry, page);
 743                                 } else
 744                                         retval = unuse_mm(mm, entry, page);
 745                                 if (set_start_mm && *swap_map < swcount) {
 746                                         mmput(new_start_mm);
 747                                         atomic_inc(&mm->mm_users);
 748                                         new_start_mm = mm;
 749                                         set_start_mm = 0;
 750                                 }
 751                                 spin_lock(&mmlist_lock);
 752                         }
 753                         spin_unlock(&mmlist_lock);
 754                         mmput(prev_mm);
 755                         mmput(start_mm);
 756                         start_mm = new_start_mm;
 757                 }
 758                 if (retval) {
 759                         unlock_page(page);
 760                         page_cache_release(page);
 761                         break;
 762                 }
 763
 764                 /*
 765                  * How could swap count reach 0x7fff when the maximum
 766                  * pid is 0x7fff, and there's no way to repeat a swap
 767                  * page within an mm (except in shmem, where it's the
 768                  * shared object which takes the reference count)?
 769                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 770                  *
 771                  * If that's wrong, then we should worry more about
 772                  * exit_mmap() and do_munmap() cases described above:
 773                  * we might be resetting SWAP_MAP_MAX too early here.
 774                  * We know "Undead"s can happen, they're okay, so don't
 775                  * report them; but do report if we reset SWAP_MAP_MAX.
 776                  */
 777                 if (*swap_map == SWAP_MAP_MAX) {
 778                         spin_lock(&swap_lock);
 779                         *swap_map = 1;
 780                         spin_unlock(&swap_lock);
 781                         reset_overflow = 1;
 782                 }
 783
 784                 /*
 785                  * If a reference remains (rare), we would like to leave
 786                  * the page in the swap cache; but try_to_unmap could
 787                  * then re-duplicate the entry once we drop page lock,
 788                  * so we might loop indefinitely; also, that page could
 789                  * not be swapped out to other storage meanwhile.  So:
 790                  * delete from cache even if there's another reference,
 791                  * after ensuring that the data has been saved to disk -
 792                  * since if the reference remains (rarer), it will be
 793                  * read from disk into another page.  Splitting into two
 794                  * pages would be incorrect if swap supported "shared
 795                  * private" pages, but they are handled by tmpfs files.
 796                  *
 797                  * Note shmem_unuse already deleted a swappage from
 798                  * the swap cache, unless the move to filepage failed:
 799                  * in which case it left swappage in cache, lowered its
 800                  * swap count to pass quickly through the loops above,
 801                  * and now we must reincrement count to try again later.
 802                  */
 803                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 804                         struct writeback_control wbc = {
 805                                 .sync_mode = WB_SYNC_NONE,
 806                         };
 807
 808                         swap_writepage(page, &wbc);
 809                         lock_page(page);
 810                         wait_on_page_writeback(page);
 811                 }
 812                 if (PageSwapCache(page)) {
 813                         if (shmem)
 814                                 swap_duplicate(entry);
 815                         else
 816                                 delete_from_swap_cache(page);
 817                 }
 818
 819                 /*
 820                  * So we could skip searching mms once swap count went
 821                  * to 1, we did not mark any present ptes as dirty: must
 822                  * mark page dirty so shrink_list will preserve it.
 823                  */
 824                 SetPageDirty(page);
 825                 unlock_page(page);
 826                 page_cache_release(page);
 827
 828                 /*
 829                  * Make sure that we aren't completely killing
 830                  * interactive performance.
 831                  */
 832                 cond_resched();
 833         }
 834
 835         mmput(start_mm);
 836         if (reset_overflow) {
 837                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 838                 swap_overflow = 0;
 839         }
 840         return retval;
 841 }
 842
 843 /*
 844  * After a successful try_to_unuse, if no swap is now in use, we know
 845  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 846  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 847  * added to the mmlist just after page_duplicate - before would be racy.
 848  */
 849 static void drain_mmlist(void)
 850 {
 851         struct list_head *p, *next;
 852         unsigned int i;
 853
 854         for (i = 0; i < nr_swapfiles; i++)
 855                 if (swap_info[i].inuse_pages)
 856                         return;
 857         spin_lock(&mmlist_lock);
 858         list_for_each_safe(p, next, &init_mm.mmlist)
 859                 list_del_init(p);
 860         spin_unlock(&mmlist_lock);
 861 }
 862
 863 /*
 864  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 865  * corresponds to page offset `offset'.
 866  */
 867 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 868 {
 869         struct swap_extent *se = sis->curr_swap_extent;
 870         struct swap_extent *start_se = se;
 871
 872         for ( ; ; ) {
 873                 struct list_head *lh;
 874
 875                 if (se->start_page <= offset &&
 876                                 offset < (se->start_page + se->nr_pages)) {
 877                         return se->start_block + (offset - se->start_page);
 878                 }
 879                 lh = se->list.next;
 880                 if (lh == &sis->extent_list)
 881                         lh = lh->next;
 882                 se = list_entry(lh, struct swap_extent, list);
 883                 sis->curr_swap_extent = se;
 884                 BUG_ON(se == start_se);         /* It *must* be present */
 885         }
 886 }
 887
 888 /*
 889  * Free all of a swapdev's extent information
 890  */
 891 static void destroy_swap_extents(struct swap_info_struct *sis)
 892 {
 893         while (!list_empty(&sis->extent_list)) {
 894                 struct swap_extent *se;
 895
 896                 se = list_entry(sis->extent_list.next,
 897                                 struct swap_extent, list);
 898                 list_del(&se->list);
 899                 kfree(se);
 900         }
 901 }
 902
 903 /*
 904  * Add a block range (and the corresponding page range) into this swapdev's
 905  * extent list.  The extent list is kept sorted in page order.
 906  *
 907  * This function rather assumes that it is called in ascending page order.
 908  */
 909 static int
 910 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 911                 unsigned long nr_pages, sector_t start_block)
 912 {
 913         struct swap_extent *se;
 914         struct swap_extent *new_se;
 915         struct list_head *lh;
 916
 917         lh = sis->extent_list.prev;     /* The highest page extent */
 918         if (lh != &sis->extent_list) {
 919                 se = list_entry(lh, struct swap_extent, list);
 920                 BUG_ON(se->start_page + se->nr_pages != start_page);
 921                 if (se->start_block + se->nr_pages == start_block) {
 922                         /* Merge it */
 923                         se->nr_pages += nr_pages;
 924                         return 0;
 925                 }
 926         }
 927
 928         /*
 929          * No merge.  Insert a new extent, preserving ordering.
 930          */
 931         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 932         if (new_se == NULL)
 933                 return -ENOMEM;
 934         new_se->start_page = start_page;
 935         new_se->nr_pages = nr_pages;
 936         new_se->start_block = start_block;
 937
 938         list_add_tail(&new_se->list, &sis->extent_list);
 939         return 1;
 940 }
 941
 942 /*
 943  * A `swap extent' is a simple thing which maps a contiguous range of pages
 944  * onto a contiguous range of disk blocks.  An ordered list of swap extents
 945  * is built at swapon time and is then used at swap_writepage/swap_readpage
 946  * time for locating where on disk a page belongs.
 947  *
 948  * If the swapfile is an S_ISBLK block device, a single extent is installed.
 949  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
 950  * swap files identically.
 951  *
 952  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
 953  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
 954  * swapfiles are handled *identically* after swapon time.
 955  *
 956  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
 957  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
 958  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
 959  * requirements, they are simply tossed out - we will never use those blocks
 960  * for swapping.
 961  *
 962  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
 963  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
 964  * which will scribble on the fs.
 965  *
 966  * The amount of disk space which a single swap extent represents varies.
 967  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
 968  * extents in the list.  To avoid much list walking, we cache the previous
 969  * search location in `curr_swap_extent', and start new searches from there.
 970  * This is extremely effective.  The average number of iterations in
 971  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
 972  */
 973 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
 974 {
 975         struct inode *inode;
 976         unsigned blocks_per_page;
 977         unsigned long page_no;
 978         unsigned blkbits;
 979         sector_t probe_block;
 980         sector_t last_block;
 981         sector_t lowest_block = -1;
 982         sector_t highest_block = 0;
 983         int nr_extents = 0;
 984         int ret;
 985
 986         inode = sis->swap_file->f_mapping->host;
 987         if (S_ISBLK(inode->i_mode)) {
 988                 ret = add_swap_extent(sis, 0, sis->max, 0);
 989                 *span = sis->pages;
 990                 goto done;
 991         }
 992
 993         blkbits = inode->i_blkbits;
 994         blocks_per_page = PAGE_SIZE >> blkbits;
 995
 996         /*
 997          * Map all the blocks into the extent list.  This code doesn't try
 998          * to be very smart.
 999          */
1000         probe_block = 0;
1001         page_no = 0;
1002         last_block = i_size_read(inode) >> blkbits;
1003         while ((probe_block + blocks_per_page) <= last_block &&
1004                         page_no < sis->max) {
1005                 unsigned block_in_page;
1006                 sector_t first_block;
1007
1008                 first_block = bmap(inode, probe_block);
1009                 if (first_block == 0)
1010                         goto bad_bmap;
1011
1012                 /*
1013                  * It must be PAGE_SIZE aligned on-disk
1014                  */
1015                 if (first_block & (blocks_per_page - 1)) {
1016                         probe_block++;
1017                         goto reprobe;
1018                 }
1019
1020                 for (block_in_page = 1; block_in_page < blocks_per_page;
1021                                         block_in_page++) {
1022                         sector_t block;
1023
1024                         block = bmap(inode, probe_block + block_in_page);
1025                         if (block == 0)
1026                                 goto bad_bmap;
1027                         if (block != first_block + block_in_page) {
1028                                 /* Discontiguity */
1029                                 probe_block++;
1030                                 goto reprobe;
1031                         }
1032                 }
1033
1034                 first_block >>= (PAGE_SHIFT - blkbits);
1035                 if (page_no) {  /* exclude the header page */
1036                         if (first_block < lowest_block)
1037                                 lowest_block = first_block;
1038                         if (first_block > highest_block)
1039                                 highest_block = first_block;
1040                 }
1041
1042                 /*
1043                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1044                  */
1045                 ret = add_swap_extent(sis, page_no, 1, first_block);
1046                 if (ret < 0)
1047                         goto out;
1048                 nr_extents += ret;
1049                 page_no++;
1050                 probe_block += blocks_per_page;
1051 reprobe:
1052                 continue;
1053         }
1054         ret = nr_extents;
1055         *span = 1 + highest_block - lowest_block;
1056         if (page_no == 0)
1057                 page_no = 1;    /* force Empty message */
1058         sis->max = page_no;
1059         sis->pages = page_no - 1;
1060         sis->highest_bit = page_no - 1;
1061 done:
1062         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1063                                         struct swap_extent, list);
1064         goto out;
1065 bad_bmap:
1066         printk(KERN_ERR "swapon: swapfile has holes\n");
1067         ret = -EINVAL;
1068 out:
1069         return ret;
1070 }
1071
1072 #if 0   /* We don't need this yet */
1073 #include <linux/backing-dev.h>
1074 int page_queue_congested(struct page *page)
1075 {
1076         struct backing_dev_info *bdi;
1077
1078         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1079
1080         if (PageSwapCache(page)) {
1081                 swp_entry_t entry = { .val = page_private(page) };
1082                 struct swap_info_struct *sis;
1083
1084                 sis = get_swap_info_struct(swp_type(entry));
1085                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1086         } else
1087                 bdi = page->mapping->backing_dev_info;
1088         return bdi_write_congested(bdi);
1089 }
1090 #endif
1091
1092 asmlinkage long sys_swapoff(const char __user * specialfile)
1093 {
1094         struct swap_info_struct * p = NULL;
1095         unsigned short *swap_map;
1096         struct file *swap_file, *victim;
1097         struct address_space *mapping;
1098         struct inode *inode;
1099         char * pathname;
1100         int i, type, prev;
1101         int err;
1102
1103         if (!capable(CAP_SYS_ADMIN))
1104                 return -EPERM;
1105
1106         pathname = getname(specialfile);
1107         err = PTR_ERR(pathname);
1108         if (IS_ERR(pathname))
1109                 goto out;
1110
1111         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1112         putname(pathname);
1113         err = PTR_ERR(victim);
1114         if (IS_ERR(victim))
1115                 goto out;
1116
1117         mapping = victim->f_mapping;
1118         prev = -1;
1119         spin_lock(&swap_lock);
1120         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1121                 p = swap_info + type;
1122                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1123                         if (p->swap_file->f_mapping == mapping)
1124                                 break;
1125                 }
1126                 prev = type;
1127         }
1128         if (type < 0) {
1129                 err = -EINVAL;
1130                 spin_unlock(&swap_lock);
1131                 goto out_dput;
1132         }
1133         if (!security_vm_enough_memory(p->pages))
1134                 vm_unacct_memory(p->pages);
1135         else {
1136                 err = -ENOMEM;
1137                 spin_unlock(&swap_lock);
1138                 goto out_dput;
1139         }
1140         if (prev < 0) {
1141                 swap_list.head = p->next;
1142         } else {
1143                 swap_info[prev].next = p->next;
1144         }
1145         if (type == swap_list.next) {
1146                 /* just pick something that's safe... */
1147                 swap_list.next = swap_list.head;
1148         }
1149         nr_swap_pages -= p->pages;
1150         total_swap_pages -= p->pages;
1151         p->flags &= ~SWP_WRITEOK;
1152         spin_unlock(&swap_lock);
1153
1154         current->flags |= PF_SWAPOFF;
1155         err = try_to_unuse(type);
1156         current->flags &= ~PF_SWAPOFF;
1157
1158         if (err) {
1159                 /* re-insert swap space back into swap_list */
1160                 spin_lock(&swap_lock);
1161                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1162                         if (p->prio >= swap_info[i].prio)
1163                                 break;
1164                 p->next = i;
1165                 if (prev < 0)
1166                         swap_list.head = swap_list.next = p - swap_info;
1167                 else
1168                         swap_info[prev].next = p - swap_info;
1169                 nr_swap_pages += p->pages;
1170                 total_swap_pages += p->pages;
1171                 p->flags |= SWP_WRITEOK;
1172                 spin_unlock(&swap_lock);
1173                 goto out_dput;
1174         }
1175
1176         /* wait for any unplug function to finish */
1177         down_write(&swap_unplug_sem);
1178         up_write(&swap_unplug_sem);
1179
1180         destroy_swap_extents(p);
1181         mutex_lock(&swapon_mutex);
1182         spin_lock(&swap_lock);
1183         drain_mmlist();
1184
1185         /* wait for anyone still in scan_swap_map */
1186         p->highest_bit = 0;             /* cuts scans short */
1187         while (p->flags >= SWP_SCANNING) {
1188                 spin_unlock(&swap_lock);
1189                 schedule_timeout_uninterruptible(1);
1190                 spin_lock(&swap_lock);
1191         }
1192
1193         swap_file = p->swap_file;
1194         p->swap_file = NULL;
1195         p->max = 0;
1196         swap_map = p->swap_map;
1197         p->swap_map = NULL;
1198         p->flags = 0;
1199         spin_unlock(&swap_lock);
1200         mutex_unlock(&swapon_mutex);
1201         vfree(swap_map);
1202         inode = mapping->host;
1203         if (S_ISBLK(inode->i_mode)) {
1204                 struct block_device *bdev = I_BDEV(inode);
1205                 set_blocksize(bdev, p->old_block_size);
1206                 bd_release(bdev);
1207         } else {
1208                 mutex_lock(&inode->i_mutex);
1209                 inode->i_flags &= ~S_SWAPFILE;
1210                 mutex_unlock(&inode->i_mutex);
1211         }
1212         filp_close(swap_file, NULL);
1213         err = 0;
1214
1215 out_dput:
1216         filp_close(victim, NULL);
1217 out:
1218         return err;
1219 }
1220
1221 #ifdef CONFIG_PROC_FS
1222 /* iterator */
1223 static void *swap_start(struct seq_file *swap, loff_t *pos)
1224 {
1225         struct swap_info_struct *ptr = swap_info;
1226         int i;
1227         loff_t l = *pos;
1228
1229         mutex_lock(&swapon_mutex);
1230
1231         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1232                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1233                         continue;
1234                 if (!l--)
1235                         return ptr;
1236         }
1237
1238         return NULL;
1239 }
1240
1241 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1242 {
1243         struct swap_info_struct *ptr = v;
1244         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1245
1246         for (++ptr; ptr < endptr; ptr++) {
1247                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1248                         continue;
1249                 ++*pos;
1250                 return ptr;
1251         }
1252
1253         return NULL;
1254 }
1255
1256 static void swap_stop(struct seq_file *swap, void *v)
1257 {
1258         mutex_unlock(&swapon_mutex);
1259 }
1260
1261 static int swap_show(struct seq_file *swap, void *v)
1262 {
1263         struct swap_info_struct *ptr = v;
1264         struct file *file;
1265         int len;
1266
1267         if (v == swap_info)
1268                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1269
1270         file = ptr->swap_file;
1271         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1272         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1273                        len < 40 ? 40 - len : 1, " ",
1274                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1275                                 "partition" : "file\t",
1276                        ptr->pages << (PAGE_SHIFT - 10),
1277                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1278                        ptr->prio);
1279         return 0;
1280 }
1281
1282 static struct seq_operations swaps_op = {
1283         .start =        swap_start,
1284         .next =         swap_next,
1285         .stop =         swap_stop,
1286         .show =         swap_show
1287 };
1288
1289 static int swaps_open(struct inode *inode, struct file *file)
1290 {
1291         return seq_open(file, &swaps_op);
1292 }
1293
1294 static struct file_operations proc_swaps_operations = {
1295         .open           = swaps_open,
1296         .read           = seq_read,
1297         .llseek         = seq_lseek,
1298         .release        = seq_release,
1299 };
1300
1301 static int __init procswaps_init(void)
1302 {
1303         struct proc_dir_entry *entry;
1304
1305         entry = create_proc_entry("swaps", 0, NULL);
1306         if (entry)
1307                 entry->proc_fops = &proc_swaps_operations;
1308         return 0;
1309 }
1310 __initcall(procswaps_init);
1311 #endif /* CONFIG_PROC_FS */
1312
1313 /*
1314  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1315  *
1316  * The swapon system call
1317  */
1318 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1319 {
1320         struct swap_info_struct * p;
1321         char *name = NULL;
1322         struct block_device *bdev = NULL;
1323         struct file *swap_file = NULL;
1324         struct address_space *mapping;
1325         unsigned int type;
1326         int i, prev;
1327         int error;
1328         static int least_priority;
1329         union swap_header *swap_header = NULL;
1330         int swap_header_version;
1331         unsigned int nr_good_pages = 0;
1332         int nr_extents = 0;
1333         sector_t span;
1334         unsigned long maxpages = 1;
1335         int swapfilesize;
1336         unsigned short *swap_map;
1337         struct page *page = NULL;
1338         struct inode *inode = NULL;
1339         int did_down = 0;
1340
1341         if (!capable(CAP_SYS_ADMIN))
1342                 return -EPERM;
1343         spin_lock(&swap_lock);
1344         p = swap_info;
1345         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1346                 if (!(p->flags & SWP_USED))
1347                         break;
1348         error = -EPERM;
1349         /*
1350          * Test if adding another swap device is possible. There are
1351          * two limiting factors: 1) the number of bits for the swap
1352          * type swp_entry_t definition and 2) the number of bits for
1353          * the swap type in the swap ptes as defined by the different
1354          * architectures. To honor both limitations a swap entry
1355          * with swap offset 0 and swap type ~0UL is created, encoded
1356          * to a swap pte, decoded to a swp_entry_t again and finally
1357          * the swap type part is extracted. This will mask all bits
1358          * from the initial ~0UL that can't be encoded in either the
1359          * swp_entry_t or the architecture definition of a swap pte.
1360          */
1361         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1362                 spin_unlock(&swap_lock);
1363                 goto out;
1364         }
1365         if (type >= nr_swapfiles)
1366                 nr_swapfiles = type+1;
1367         INIT_LIST_HEAD(&p->extent_list);
1368         p->flags = SWP_USED;
1369         p->swap_file = NULL;
1370         p->old_block_size = 0;
1371         p->swap_map = NULL;
1372         p->lowest_bit = 0;
1373         p->highest_bit = 0;
1374         p->cluster_nr = 0;
1375         p->inuse_pages = 0;
1376         p->next = -1;
1377         if (swap_flags & SWAP_FLAG_PREFER) {
1378                 p->prio =
1379                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1380         } else {
1381                 p->prio = --least_priority;
1382         }
1383         spin_unlock(&swap_lock);
1384         name = getname(specialfile);
1385         error = PTR_ERR(name);
1386         if (IS_ERR(name)) {
1387                 name = NULL;
1388                 goto bad_swap_2;
1389         }
1390         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1391         error = PTR_ERR(swap_file);
1392         if (IS_ERR(swap_file)) {
1393                 swap_file = NULL;
1394                 goto bad_swap_2;
1395         }
1396
1397         p->swap_file = swap_file;
1398         mapping = swap_file->f_mapping;
1399         inode = mapping->host;
1400
1401         error = -EBUSY;
1402         for (i = 0; i < nr_swapfiles; i++) {
1403                 struct swap_info_struct *q = &swap_info[i];
1404
1405                 if (i == type || !q->swap_file)
1406                         continue;
1407                 if (mapping == q->swap_file->f_mapping)
1408                         goto bad_swap;
1409         }
1410
1411         error = -EINVAL;
1412         if (S_ISBLK(inode->i_mode)) {
1413                 bdev = I_BDEV(inode);
1414                 error = bd_claim(bdev, sys_swapon);
1415                 if (error < 0) {
1416                         bdev = NULL;
1417                         error = -EINVAL;
1418                         goto bad_swap;
1419                 }
1420                 p->old_block_size = block_size(bdev);
1421                 error = set_blocksize(bdev, PAGE_SIZE);
1422                 if (error < 0)
1423                         goto bad_swap;
1424                 p->bdev = bdev;
1425         } else if (S_ISREG(inode->i_mode)) {
1426                 p->bdev = inode->i_sb->s_bdev;
1427                 mutex_lock(&inode->i_mutex);
1428                 did_down = 1;
1429                 if (IS_SWAPFILE(inode)) {
1430                         error = -EBUSY;
1431                         goto bad_swap;
1432                 }
1433         } else {
1434                 goto bad_swap;
1435         }
1436
1437         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1438
1439         /*
1440          * Read the swap header.
1441          */
1442         if (!mapping->a_ops->readpage) {
1443                 error = -EINVAL;
1444                 goto bad_swap;
1445         }
1446         page = read_cache_page(mapping, 0,
1447                         (filler_t *)mapping->a_ops->readpage, swap_file);
1448         if (IS_ERR(page)) {
1449                 error = PTR_ERR(page);
1450                 goto bad_swap;
1451         }
1452         wait_on_page_locked(page);
1453         if (!PageUptodate(page))
1454                 goto bad_swap;
1455         kmap(page);
1456         swap_header = page_address(page);
1457
1458         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1459                 swap_header_version = 1;
1460         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1461                 swap_header_version = 2;
1462         else {
1463                 printk(KERN_ERR "Unable to find swap-space signature\n");
1464                 error = -EINVAL;
1465                 goto bad_swap;
1466         }
1467
1468         switch (swap_header_version) {
1469         case 1:
1470                 printk(KERN_ERR "version 0 swap is no longer supported. "
1471                         "Use mkswap -v1 %s\n", name);
1472                 error = -EINVAL;
1473                 goto bad_swap;
1474         case 2:
1475                 /* Check the swap header's sub-version and the size of
1476                    the swap file and bad block lists */
1477                 if (swap_header->info.version != 1) {
1478                         printk(KERN_WARNING
1479                                "Unable to handle swap header version %d\n",
1480                                swap_header->info.version);
1481                         error = -EINVAL;
1482                         goto bad_swap;
1483                 }
1484
1485                 p->lowest_bit  = 1;
1486                 p->cluster_next = 1;
1487
1488                 /*
1489                  * Find out how many pages are allowed for a single swap
1490                  * device. There are two limiting factors: 1) the number of
1491                  * bits for the swap offset in the swp_entry_t type and
1492                  * 2) the number of bits in the a swap pte as defined by
1493                  * the different architectures. In order to find the
1494                  * largest possible bit mask a swap entry with swap type 0
1495                  * and swap offset ~0UL is created, encoded to a swap pte,
1496                  * decoded to a swp_entry_t again and finally the swap
1497                  * offset is extracted. This will mask all the bits from
1498                  * the initial ~0UL mask that can't be encoded in either
1499                  * the swp_entry_t or the architecture definition of a
1500                  * swap pte.
1501                  */
1502                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1503                 if (maxpages > swap_header->info.last_page)
1504                         maxpages = swap_header->info.last_page;
1505                 p->highest_bit = maxpages - 1;
1506
1507                 error = -EINVAL;
1508                 if (!maxpages)
1509                         goto bad_swap;
1510                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1511                         goto bad_swap;
1512                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1513                         goto bad_swap;
1514
1515                 /* OK, set up the swap map and apply the bad block list */
1516                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1517                         error = -ENOMEM;
1518                         goto bad_swap;
1519                 }
1520
1521                 error = 0;
1522                 memset(p->swap_map, 0, maxpages * sizeof(short));
1523                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1524                         int page_nr = swap_header->info.badpages[i];
1525                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1526                                 error = -EINVAL;
1527                         else
1528                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1529                 }
1530                 nr_good_pages = swap_header->info.last_page -
1531                                 swap_header->info.nr_badpages -
1532                                 1 /* header page */;
1533                 if (error)
1534                         goto bad_swap;
1535         }
1536
1537         if (swapfilesize && maxpages > swapfilesize) {
1538                 printk(KERN_WARNING
1539                        "Swap area shorter than signature indicates\n");
1540                 error = -EINVAL;
1541                 goto bad_swap;
1542         }
1543         if (nr_good_pages) {
1544                 p->swap_map[0] = SWAP_MAP_BAD;
1545                 p->max = maxpages;
1546                 p->pages = nr_good_pages;
1547                 nr_extents = setup_swap_extents(p, &span);
1548                 if (nr_extents < 0) {
1549                         error = nr_extents;
1550                         goto bad_swap;
1551                 }
1552                 nr_good_pages = p->pages;
1553         }
1554         if (!nr_good_pages) {
1555                 printk(KERN_WARNING "Empty swap-file\n");
1556                 error = -EINVAL;
1557                 goto bad_swap;
1558         }
1559
1560         mutex_lock(&swapon_mutex);
1561         spin_lock(&swap_lock);
1562         p->flags = SWP_ACTIVE;
1563         nr_swap_pages += nr_good_pages;
1564         total_swap_pages += nr_good_pages;
1565
1566         printk(KERN_INFO "Adding %uk swap on %s.  "
1567                         "Priority:%d extents:%d across:%lluk\n",
1568                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1569                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1570
1571         /* insert swap space into swap_list: */
1572         prev = -1;
1573         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1574                 if (p->prio >= swap_info[i].prio) {
1575                         break;
1576                 }
1577                 prev = i;
1578         }
1579         p->next = i;
1580         if (prev < 0) {
1581                 swap_list.head = swap_list.next = p - swap_info;
1582         } else {
1583                 swap_info[prev].next = p - swap_info;
1584         }
1585         spin_unlock(&swap_lock);
1586         mutex_unlock(&swapon_mutex);
1587         error = 0;
1588         goto out;
1589 bad_swap:
1590         if (bdev) {
1591                 set_blocksize(bdev, p->old_block_size);
1592                 bd_release(bdev);
1593         }
1594         destroy_swap_extents(p);
1595 bad_swap_2:
1596         spin_lock(&swap_lock);
1597         swap_map = p->swap_map;
1598         p->swap_file = NULL;
1599         p->swap_map = NULL;
1600         p->flags = 0;
1601         if (!(swap_flags & SWAP_FLAG_PREFER))
1602                 ++least_priority;
1603         spin_unlock(&swap_lock);
1604         vfree(swap_map);
1605         if (swap_file)
1606                 filp_close(swap_file, NULL);
1607 out:
1608         if (page && !IS_ERR(page)) {
1609                 kunmap(page);
1610                 page_cache_release(page);
1611         }
1612         if (name)
1613                 putname(name);
1614         if (did_down) {
1615                 if (!error)
1616                         inode->i_flags |= S_SWAPFILE;
1617                 mutex_unlock(&inode->i_mutex);
1618         }
1619         return error;
1620 }
1621
1622 void si_swapinfo(struct sysinfo *val)
1623 {
1624         unsigned int i;
1625         unsigned long nr_to_be_unused = 0;
1626
1627         spin_lock(&swap_lock);
1628         for (i = 0; i < nr_swapfiles; i++) {
1629                 if (!(swap_info[i].flags & SWP_USED) ||
1630                      (swap_info[i].flags & SWP_WRITEOK))
1631                         continue;
1632                 nr_to_be_unused += swap_info[i].inuse_pages;
1633         }
1634         val->freeswap = nr_swap_pages + nr_to_be_unused;
1635         val->totalswap = total_swap_pages + nr_to_be_unused;
1636         spin_unlock(&swap_lock);
1637 }
1638
1639 /*
1640  * Verify that a swap entry is valid and increment its swap map count.
1641  *
1642  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1643  * "permanent", but will be reclaimed by the next swapoff.
1644  */
1645 int swap_duplicate(swp_entry_t entry)
1646 {
1647         struct swap_info_struct * p;
1648         unsigned long offset, type;
1649         int result = 0;
1650
1651         type = swp_type(entry);
1652         if (type >= nr_swapfiles)
1653                 goto bad_file;
1654         p = type + swap_info;
1655         offset = swp_offset(entry);
1656
1657         spin_lock(&swap_lock);
1658         if (offset < p->max && p->swap_map[offset]) {
1659                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1660                         p->swap_map[offset]++;
1661                         result = 1;
1662                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1663                         if (swap_overflow++ < 5)
1664                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1665                         p->swap_map[offset] = SWAP_MAP_MAX;
1666                         result = 1;
1667                 }
1668         }
1669         spin_unlock(&swap_lock);
1670 out:
1671         return result;
1672
1673 bad_file:
1674         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1675         goto out;
1676 }
1677
1678 struct swap_info_struct *
1679 get_swap_info_struct(unsigned type)
1680 {
1681         return &swap_info[type];
1682 }
1683
1684 /*
1685  * swap_lock prevents swap_map being freed. Don't grab an extra
1686  * reference on the swaphandle, it doesn't matter if it becomes unused.
1687  */
1688 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1689 {
1690         int ret = 0, i = 1 << page_cluster;
1691         unsigned long toff;
1692         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1693
1694         if (!page_cluster)      /* no readahead */
1695                 return 0;
1696         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1697         if (!toff)              /* first page is swap header */
1698                 toff++, i--;
1699         *offset = toff;
1700
1701         spin_lock(&swap_lock);
1702         do {
1703                 /* Don't read-ahead past the end of the swap area */
1704                 if (toff >= swapdev->max)
1705                         break;
1706                 /* Don't read in free or bad pages */
1707                 if (!swapdev->swap_map[toff])
1708                         break;
1709                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1710                         break;
1711                 toff++;
1712                 ret++;
1713         } while (--i);
1714         spin_unlock(&swap_lock);
1715         return ret;
1716 }