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