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