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