mm/swapfile.c

   1 /*
   2  *  linux/mm/swapfile.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  *  Swap reorganised 29.12.95, Stephen Tweedie
   6  */
   7
   8 #include <linux/config.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/mman.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/namei.h>
  18 #include <linux/shm.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/init.h>
  24 #include <linux/module.h>
  25 #include <linux/rmap.h>
  26 #include <linux/security.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/mutex.h>
  29 #include <linux/capability.h>
  30 #include <linux/syscalls.h>
  31
  32 #include <asm/pgtable.h>
  33 #include <asm/tlbflush.h>
  34 #include <linux/swapops.h>
  35 #include <linux/vs_memory.h>
  36
  37 DEFINE_SPINLOCK(swap_lock);
  38 unsigned int nr_swapfiles;
  39 long total_swap_pages;
  40 static int swap_overflow;
  41
  42 static const char Bad_file[] = "Bad swap file entry ";
  43 static const char Unused_file[] = "Unused swap file entry ";
  44 static const char Bad_offset[] = "Bad swap offset entry ";
  45 static const char Unused_offset[] = "Unused swap offset entry ";
  46
  47 struct swap_list_t swap_list = {-1, -1};
  48
  49 static struct swap_info_struct swap_info[MAX_SWAPFILES];
  50
  51 static DEFINE_MUTEX(swapon_mutex);
  52
  53 /*
  54  * We need this because the bdev->unplug_fn can sleep and we cannot
  55  * hold swap_lock while calling the unplug_fn. And swap_lock
  56  * cannot be turned into a mutex.
  57  */
  58 static DECLARE_RWSEM(swap_unplug_sem);
  59
  60 void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
  61 {
  62         swp_entry_t entry;
  63
  64         down_read(&swap_unplug_sem);
  65         entry.val = page_private(page);
  66         if (PageSwapCache(page)) {
  67                 struct block_device *bdev = swap_info[swp_type(entry)].bdev;
  68                 struct backing_dev_info *bdi;
  69
  70                 /*
  71                  * If the page is removed from swapcache from under us (with a
  72                  * racy try_to_unuse/swapoff) we need an additional reference
  73                  * count to avoid reading garbage from page_private(page) above.
  74                  * If the WARN_ON triggers during a swapoff it maybe the race
  75                  * condition and it's harmless. However if it triggers without
  76                  * swapoff it signals a problem.
  77                  */
  78                 WARN_ON(page_count(page) <= 1);
  79
  80                 bdi = bdev->bd_inode->i_mapping->backing_dev_info;
  81                 blk_run_backing_dev(bdi, page);
  82         }
  83         up_read(&swap_unplug_sem);
  84 }
  85
  86 #define SWAPFILE_CLUSTER        256
  87 #define LATENCY_LIMIT           256
  88
  89 static inline unsigned long scan_swap_map(struct swap_info_struct *si)
  90 {
  91         unsigned long offset, last_in_cluster;
  92         int latency_ration = LATENCY_LIMIT;
  93
  94         /*
  95          * We try to cluster swap pages by allocating them sequentially
  96          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
  97          * way, however, we resort to first-free allocation, starting
  98          * a new cluster.  This prevents us from scattering swap pages
  99          * all over the entire swap partition, so that we reduce
 100          * overall disk seek times between swap pages.  -- sct
 101          * But we do now try to find an empty cluster.  -Andrea
 102          */
 103
 104         si->flags += SWP_SCANNING;
 105         if (unlikely(!si->cluster_nr)) {
 106                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
 107                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
 108                         goto lowest;
 109                 spin_unlock(&swap_lock);
 110
 111                 offset = si->lowest_bit;
 112                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 113
 114                 /* Locate the first empty (unaligned) cluster */
 115                 for (; last_in_cluster <= si->highest_bit; offset++) {
 116                         if (si->swap_map[offset])
 117                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
 118                         else if (offset == last_in_cluster) {
 119                                 spin_lock(&swap_lock);
 120                                 si->cluster_next = offset-SWAPFILE_CLUSTER+1;
 121                                 goto cluster;
 122                         }
 123                         if (unlikely(--latency_ration < 0)) {
 124                                 cond_resched();
 125                                 latency_ration = LATENCY_LIMIT;
 126                         }
 127                 }
 128                 spin_lock(&swap_lock);
 129                 goto lowest;
 130         }
 131
 132         si->cluster_nr--;
 133 cluster:
 134         offset = si->cluster_next;
 135         if (offset > si->highest_bit)
 136 lowest:         offset = si->lowest_bit;
 137 checks: if (!(si->flags & SWP_WRITEOK))
 138                 goto no_page;
 139         if (!si->highest_bit)
 140                 goto no_page;
 141         if (!si->swap_map[offset]) {
 142                 if (offset == si->lowest_bit)
 143                         si->lowest_bit++;
 144                 if (offset == si->highest_bit)
 145                         si->highest_bit--;
 146                 si->inuse_pages++;
 147                 if (si->inuse_pages == si->pages) {
 148                         si->lowest_bit = si->max;
 149                         si->highest_bit = 0;
 150                 }
 151                 si->swap_map[offset] = 1;
 152                 si->cluster_next = offset + 1;
 153                 si->flags -= SWP_SCANNING;
 154                 return offset;
 155         }
 156
 157         spin_unlock(&swap_lock);
 158         while (++offset <= si->highest_bit) {
 159                 if (!si->swap_map[offset]) {
 160                         spin_lock(&swap_lock);
 161                         goto checks;
 162                 }
 163                 if (unlikely(--latency_ration < 0)) {
 164                         cond_resched();
 165                         latency_ration = LATENCY_LIMIT;
 166                 }
 167         }
 168         spin_lock(&swap_lock);
 169         goto lowest;
 170
 171 no_page:
 172         si->flags -= SWP_SCANNING;
 173         return 0;
 174 }
 175
 176 swp_entry_t get_swap_page(void)
 177 {
 178         struct swap_info_struct *si;
 179         pgoff_t offset;
 180         int type, next;
 181         int wrapped = 0;
 182
 183         spin_lock(&swap_lock);
 184         if (nr_swap_pages <= 0)
 185                 goto noswap;
 186         nr_swap_pages--;
 187
 188         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
 189                 si = swap_info + type;
 190                 next = si->next;
 191                 if (next < 0 ||
 192                     (!wrapped && si->prio != swap_info[next].prio)) {
 193                         next = swap_list.head;
 194                         wrapped++;
 195                 }
 196
 197                 if (!si->highest_bit)
 198                         continue;
 199                 if (!(si->flags & SWP_WRITEOK))
 200                         continue;
 201
 202                 swap_list.next = next;
 203                 offset = scan_swap_map(si);
 204                 if (offset) {
 205                         spin_unlock(&swap_lock);
 206                         return swp_entry(type, offset);
 207                 }
 208                 next = swap_list.next;
 209         }
 210
 211         nr_swap_pages++;
 212 noswap:
 213         spin_unlock(&swap_lock);
 214         return (swp_entry_t) {0};
 215 }
 216
 217 swp_entry_t get_swap_page_of_type(int type)
 218 {
 219         struct swap_info_struct *si;
 220         pgoff_t offset;
 221
 222         spin_lock(&swap_lock);
 223         si = swap_info + type;
 224         if (si->flags & SWP_WRITEOK) {
 225                 nr_swap_pages--;
 226                 offset = scan_swap_map(si);
 227                 if (offset) {
 228                         spin_unlock(&swap_lock);
 229                         return swp_entry(type, offset);
 230                 }
 231                 nr_swap_pages++;
 232         }
 233         spin_unlock(&swap_lock);
 234         return (swp_entry_t) {0};
 235 }
 236
 237 static struct swap_info_struct * swap_info_get(swp_entry_t entry)
 238 {
 239         struct swap_info_struct * p;
 240         unsigned long offset, type;
 241
 242         if (!entry.val)
 243                 goto out;
 244         type = swp_type(entry);
 245         if (type >= nr_swapfiles)
 246                 goto bad_nofile;
 247         p = & swap_info[type];
 248         if (!(p->flags & SWP_USED))
 249                 goto bad_device;
 250         offset = swp_offset(entry);
 251         if (offset >= p->max)
 252                 goto bad_offset;
 253         if (!p->swap_map[offset])
 254                 goto bad_free;
 255         spin_lock(&swap_lock);
 256         return p;
 257
 258 bad_free:
 259         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
 260         goto out;
 261 bad_offset:
 262         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
 263         goto out;
 264 bad_device:
 265         printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
 266         goto out;
 267 bad_nofile:
 268         printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 269 out:
 270         return NULL;
 271 }
 272
 273 static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
 274 {
 275         int count = p->swap_map[offset];
 276
 277         if (count < SWAP_MAP_MAX) {
 278                 count--;
 279                 p->swap_map[offset] = count;
 280                 if (!count) {
 281                         if (offset < p->lowest_bit)
 282                                 p->lowest_bit = offset;
 283                         if (offset > p->highest_bit)
 284                                 p->highest_bit = offset;
 285                         if (p->prio > swap_info[swap_list.next].prio)
 286                                 swap_list.next = p - swap_info;
 287                         nr_swap_pages++;
 288                         p->inuse_pages--;
 289                 }
 290         }
 291         return count;
 292 }
 293
 294 /*
 295  * Caller has made sure that the swapdevice corresponding to entry
 296  * is still around or has not been recycled.
 297  */
 298 void swap_free(swp_entry_t entry)
 299 {
 300         struct swap_info_struct * p;
 301
 302         p = swap_info_get(entry);
 303         if (p) {
 304                 swap_entry_free(p, swp_offset(entry));
 305                 spin_unlock(&swap_lock);
 306         }
 307 }
 308
 309 /*
 310  * How many references to page are currently swapped out?
 311  */
 312 static inline int page_swapcount(struct page *page)
 313 {
 314         int count = 0;
 315         struct swap_info_struct *p;
 316         swp_entry_t entry;
 317
 318         entry.val = page_private(page);
 319         p = swap_info_get(entry);
 320         if (p) {
 321                 /* Subtract the 1 for the swap cache itself */
 322                 count = p->swap_map[swp_offset(entry)] - 1;
 323                 spin_unlock(&swap_lock);
 324         }
 325         return count;
 326 }
 327
 328 /*
 329  * We can use this swap cache entry directly
 330  * if there are no other references to it.
 331  */
 332 int can_share_swap_page(struct page *page)
 333 {
 334         int count;
 335
 336         BUG_ON(!PageLocked(page));
 337         count = page_mapcount(page);
 338         if (count <= 1 && PageSwapCache(page))
 339                 count += page_swapcount(page);
 340         return count == 1;
 341 }
 342
 343 /*
 344  * Work out if there are any other processes sharing this
 345  * swap cache page. Free it if you can. Return success.
 346  */
 347 int remove_exclusive_swap_page(struct page *page)
 348 {
 349         int retval;
 350         struct swap_info_struct * p;
 351         swp_entry_t entry;
 352
 353         BUG_ON(PagePrivate(page));
 354         BUG_ON(!PageLocked(page));
 355
 356         if (!PageSwapCache(page))
 357                 return 0;
 358         if (PageWriteback(page))
 359                 return 0;
 360         if (page_count(page) != 2) /* 2: us + cache */
 361                 return 0;
 362
 363         entry.val = page_private(page);
 364         p = swap_info_get(entry);
 365         if (!p)
 366                 return 0;
 367
 368         /* Is the only swap cache user the cache itself? */
 369         retval = 0;
 370         if (p->swap_map[swp_offset(entry)] == 1) {
 371                 /* Recheck the page count with the swapcache lock held.. */
 372                 write_lock_irq(&swapper_space.tree_lock);
 373                 if ((page_count(page) == 2) && !PageWriteback(page)) {
 374                         __delete_from_swap_cache(page);
 375                         SetPageDirty(page);
 376                         retval = 1;
 377                 }
 378                 write_unlock_irq(&swapper_space.tree_lock);
 379         }
 380         spin_unlock(&swap_lock);
 381
 382         if (retval) {
 383                 swap_free(entry);
 384                 page_cache_release(page);
 385         }
 386
 387         return retval;
 388 }
 389
 390 /*
 391  * Free the swap entry like above, but also try to
 392  * free the page cache entry if it is the last user.
 393  */
 394 void free_swap_and_cache(swp_entry_t entry)
 395 {
 396         struct swap_info_struct * p;
 397         struct page *page = NULL;
 398
 399         p = swap_info_get(entry);
 400         if (p) {
 401                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 402                         page = find_get_page(&swapper_space, entry.val);
 403                         if (page && unlikely(TestSetPageLocked(page))) {
 404                                 page_cache_release(page);
 405                                 page = NULL;
 406                         }
 407                 }
 408                 spin_unlock(&swap_lock);
 409         }
 410         if (page) {
 411                 int one_user;
 412
 413                 BUG_ON(PagePrivate(page));
 414                 one_user = (page_count(page) == 2);
 415                 /* Only cache user (+us), or swap space full? Free it! */
 416                 /* Also recheck PageSwapCache after page is locked (above) */
 417                 if (PageSwapCache(page) && !PageWriteback(page) &&
 418                                         (one_user || vm_swap_full())) {
 419                         delete_from_swap_cache(page);
 420                         SetPageDirty(page);
 421                 }
 422                 unlock_page(page);
 423                 page_cache_release(page);
 424         }
 425 }
 426
 427 #ifdef CONFIG_SOFTWARE_SUSPEND
 428 /*
 429  * Find the swap type that corresponds to given device (if any)
 430  *
 431  * This is needed for software suspend and is done in such a way that inode
 432  * aliasing is allowed.
 433  */
 434 int swap_type_of(dev_t device)
 435 {
 436         int i;
 437
 438         spin_lock(&swap_lock);
 439         for (i = 0; i < nr_swapfiles; i++) {
 440                 struct inode *inode;
 441
 442                 if (!(swap_info[i].flags & SWP_WRITEOK))
 443                         continue;
 444
 445                 if (!device) {
 446                         spin_unlock(&swap_lock);
 447                         return i;
 448                 }
 449                 inode = swap_info[i].swap_file->f_dentry->d_inode;
 450                 if (S_ISBLK(inode->i_mode) &&
 451                     device == MKDEV(imajor(inode), iminor(inode))) {
 452                         spin_unlock(&swap_lock);
 453                         return i;
 454                 }
 455         }
 456         spin_unlock(&swap_lock);
 457         return -ENODEV;
 458 }
 459
 460 /*
 461  * Return either the total number of swap pages of given type, or the number
 462  * of free pages of that type (depending on @free)
 463  *
 464  * This is needed for software suspend
 465  */
 466 unsigned int count_swap_pages(int type, int free)
 467 {
 468         unsigned int n = 0;
 469
 470         if (type < nr_swapfiles) {
 471                 spin_lock(&swap_lock);
 472                 if (swap_info[type].flags & SWP_WRITEOK) {
 473                         n = swap_info[type].pages;
 474                         if (free)
 475                                 n -= swap_info[type].inuse_pages;
 476                 }
 477                 spin_unlock(&swap_lock);
 478         }
 479         return n;
 480 }
 481 #endif
 482
 483 /*
 484  * No need to decide whether this PTE shares the swap entry with others,
 485  * just let do_wp_page work it out if a write is requested later - to
 486  * force COW, vm_page_prot omits write permission from any private vma.
 487  */
 488 static void unuse_pte(struct vm_area_struct *vma, pte_t *pte,
 489                 unsigned long addr, swp_entry_t entry, struct page *page)
 490 {
 491         inc_mm_counter(vma->vm_mm, anon_rss);
 492         get_page(page);
 493         set_pte_at(vma->vm_mm, addr, pte,
 494                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 495         page_add_anon_rmap(page, vma, addr);
 496         swap_free(entry);
 497         /*
 498          * Move the page to the active list so it is not
 499          * immediately swapped out again after swapon.
 500          */
 501         activate_page(page);
 502 }
 503
 504 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 505                                 unsigned long addr, unsigned long end,
 506                                 swp_entry_t entry, struct page *page)
 507 {
 508         pte_t swp_pte = swp_entry_to_pte(entry);
 509         pte_t *pte;
 510         spinlock_t *ptl;
 511         int found = 0;
 512
 513         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 514         do {
 515                 /*
 516                  * swapoff spends a _lot_ of time in this loop!
 517                  * Test inline before going to call unuse_pte.
 518                  */
 519                 if (unlikely(pte_same(*pte, swp_pte))) {
 520                         unuse_pte(vma, pte++, addr, entry, page);
 521                         found = 1;
 522                         break;
 523                 }
 524         } while (pte++, addr += PAGE_SIZE, addr != end);
 525         pte_unmap_unlock(pte - 1, ptl);
 526         return found;
 527 }
 528
 529 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 530                                 unsigned long addr, unsigned long end,
 531                                 swp_entry_t entry, struct page *page)
 532 {
 533         pmd_t *pmd;
 534         unsigned long next;
 535
 536         pmd = pmd_offset(pud, addr);
 537         do {
 538                 next = pmd_addr_end(addr, end);
 539                 if (pmd_none_or_clear_bad(pmd))
 540                         continue;
 541                 if (unuse_pte_range(vma, pmd, addr, next, entry, page))
 542                         return 1;
 543         } while (pmd++, addr = next, addr != end);
 544         return 0;
 545 }
 546
 547 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 548                                 unsigned long addr, unsigned long end,
 549                                 swp_entry_t entry, struct page *page)
 550 {
 551         pud_t *pud;
 552         unsigned long next;
 553
 554         pud = pud_offset(pgd, addr);
 555         do {
 556                 next = pud_addr_end(addr, end);
 557                 if (pud_none_or_clear_bad(pud))
 558                         continue;
 559                 if (unuse_pmd_range(vma, pud, addr, next, entry, page))
 560                         return 1;
 561         } while (pud++, addr = next, addr != end);
 562         return 0;
 563 }
 564
 565 static int unuse_vma(struct vm_area_struct *vma,
 566                                 swp_entry_t entry, struct page *page)
 567 {
 568         pgd_t *pgd;
 569         unsigned long addr, end, next;
 570
 571         if (page->mapping) {
 572                 addr = page_address_in_vma(page, vma);
 573                 if (addr == -EFAULT)
 574                         return 0;
 575                 else
 576                         end = addr + PAGE_SIZE;
 577         } else {
 578                 addr = vma->vm_start;
 579                 end = vma->vm_end;
 580         }
 581
 582         pgd = pgd_offset(vma->vm_mm, addr);
 583         do {
 584                 next = pgd_addr_end(addr, end);
 585                 if (pgd_none_or_clear_bad(pgd))
 586                         continue;
 587                 if (unuse_pud_range(vma, pgd, addr, next, entry, page))
 588                         return 1;
 589         } while (pgd++, addr = next, addr != end);
 590         return 0;
 591 }
 592
 593 static int unuse_mm(struct mm_struct *mm,
 594                                 swp_entry_t entry, struct page *page)
 595 {
 596         struct vm_area_struct *vma;
 597
 598         if (!down_read_trylock(&mm->mmap_sem)) {
 599                 /*
 600                  * Activate page so shrink_cache is unlikely to unmap its
 601                  * ptes while lock is dropped, so swapoff can make progress.
 602                  */
 603                 activate_page(page);
 604                 unlock_page(page);
 605                 down_read(&mm->mmap_sem);
 606                 lock_page(page);
 607         }
 608         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 609                 if (vma->anon_vma && unuse_vma(vma, entry, page))
 610                         break;
 611         }
 612         up_read(&mm->mmap_sem);
 613         /*
 614          * Currently unuse_mm cannot fail, but leave error handling
 615          * at call sites for now, since we change it from time to time.
 616          */
 617         return 0;
 618 }
 619
 620 #ifdef CONFIG_MIGRATION
 621 int remove_vma_swap(struct vm_area_struct *vma, struct page *page)
 622 {
 623         swp_entry_t entry = { .val = page_private(page) };
 624
 625         return unuse_vma(vma, entry, page);
 626 }
 627 #endif
 628
 629 /*
 630  * Scan swap_map from current position to next entry still in use.
 631  * Recycle to start on reaching the end, returning 0 when empty.
 632  */
 633 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 634                                         unsigned int prev)
 635 {
 636         unsigned int max = si->max;
 637         unsigned int i = prev;
 638         int count;
 639
 640         /*
 641          * No need for swap_lock here: we're just looking
 642          * for whether an entry is in use, not modifying it; false
 643          * hits are okay, and sys_swapoff() has already prevented new
 644          * allocations from this area (while holding swap_lock).
 645          */
 646         for (;;) {
 647                 if (++i >= max) {
 648                         if (!prev) {
 649                                 i = 0;
 650                                 break;
 651                         }
 652                         /*
 653                          * No entries in use at top of swap_map,
 654                          * loop back to start and recheck there.
 655                          */
 656                         max = prev + 1;
 657                         prev = 0;
 658                         i = 1;
 659                 }
 660                 count = si->swap_map[i];
 661                 if (count && count != SWAP_MAP_BAD)
 662                         break;
 663         }
 664         return i;
 665 }
 666
 667 /*
 668  * We completely avoid races by reading each swap page in advance,
 669  * and then search for the process using it.  All the necessary
 670  * page table adjustments can then be made atomically.
 671  */
 672 static int try_to_unuse(unsigned int type)
 673 {
 674         struct swap_info_struct * si = &swap_info[type];
 675         struct mm_struct *start_mm;
 676         unsigned short *swap_map;
 677         unsigned short swcount;
 678         struct page *page;
 679         swp_entry_t entry;
 680         unsigned int i = 0;
 681         int retval = 0;
 682         int reset_overflow = 0;
 683         int shmem;
 684
 685         /*
 686          * When searching mms for an entry, a good strategy is to
 687          * start at the first mm we freed the previous entry from
 688          * (though actually we don't notice whether we or coincidence
 689          * freed the entry).  Initialize this start_mm with a hold.
 690          *
 691          * A simpler strategy would be to start at the last mm we
 692          * freed the previous entry from; but that would take less
 693          * advantage of mmlist ordering, which clusters forked mms
 694          * together, child after parent.  If we race with dup_mmap(), we
 695          * prefer to resolve parent before child, lest we miss entries
 696          * duplicated after we scanned child: using last mm would invert
 697          * that.  Though it's only a serious concern when an overflowed
 698          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 699          */
 700         start_mm = &init_mm;
 701         atomic_inc(&init_mm.mm_users);
 702
 703         /*
 704          * Keep on scanning until all entries have gone.  Usually,
 705          * one pass through swap_map is enough, but not necessarily:
 706          * there are races when an instance of an entry might be missed.
 707          */
 708         while ((i = find_next_to_unuse(si, i)) != 0) {
 709                 if (signal_pending(current)) {
 710                         retval = -EINTR;
 711                         break;
 712                 }
 713
 714                 /*
 715                  * Get a page for the entry, using the existing swap
 716                  * cache page if there is one.  Otherwise, get a clean
 717                  * page and read the swap into it.
 718                  */
 719                 swap_map = &si->swap_map[i];
 720                 entry = swp_entry(type, i);
 721 again:
 722                 page = read_swap_cache_async(entry, NULL, 0);
 723                 if (!page) {
 724                         /*
 725                          * Either swap_duplicate() failed because entry
 726                          * has been freed independently, and will not be
 727                          * reused since sys_swapoff() already disabled
 728                          * allocation from here, or alloc_page() failed.
 729                          */
 730                         if (!*swap_map)
 731                                 continue;
 732                         retval = -ENOMEM;
 733                         break;
 734                 }
 735
 736                 /*
 737                  * Don't hold on to start_mm if it looks like exiting.
 738                  */
 739                 if (atomic_read(&start_mm->mm_users) == 1) {
 740                         mmput(start_mm);
 741                         start_mm = &init_mm;
 742                         atomic_inc(&init_mm.mm_users);
 743                 }
 744
 745                 /*
 746                  * Wait for and lock page.  When do_swap_page races with
 747                  * try_to_unuse, do_swap_page can handle the fault much
 748                  * faster than try_to_unuse can locate the entry.  This
 749                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 750                  * defer to do_swap_page in such a case - in some tests,
 751                  * do_swap_page and try_to_unuse repeatedly compete.
 752                  */
 753                 wait_on_page_locked(page);
 754                 wait_on_page_writeback(page);
 755                 lock_page(page);
 756                 if (!PageSwapCache(page)) {
 757                         /* Page migration has occured */
 758                         unlock_page(page);
 759                         page_cache_release(page);
 760                         goto again;
 761                 }
 762                 wait_on_page_writeback(page);
 763
 764                 /*
 765                  * Remove all references to entry.
 766                  * Whenever we reach init_mm, there's no address space
 767                  * to search, but use it as a reminder to search shmem.
 768                  */
 769                 shmem = 0;
 770                 swcount = *swap_map;
 771                 if (swcount > 1) {
 772                         if (start_mm == &init_mm)
 773                                 shmem = shmem_unuse(entry, page);
 774                         else
 775                                 retval = unuse_mm(start_mm, entry, page);
 776                 }
 777                 if (*swap_map > 1) {
 778                         int set_start_mm = (*swap_map >= swcount);
 779                         struct list_head *p = &start_mm->mmlist;
 780                         struct mm_struct *new_start_mm = start_mm;
 781                         struct mm_struct *prev_mm = start_mm;
 782                         struct mm_struct *mm;
 783
 784                         atomic_inc(&new_start_mm->mm_users);
 785                         atomic_inc(&prev_mm->mm_users);
 786                         spin_lock(&mmlist_lock);
 787                         while (*swap_map > 1 && !retval &&
 788                                         (p = p->next) != &start_mm->mmlist) {
 789                                 mm = list_entry(p, struct mm_struct, mmlist);
 790                                 if (atomic_inc_return(&mm->mm_users) == 1) {
 791                                         atomic_dec(&mm->mm_users);
 792                                         continue;
 793                                 }
 794                                 spin_unlock(&mmlist_lock);
 795                                 mmput(prev_mm);
 796                                 prev_mm = mm;
 797
 798                                 cond_resched();
 799
 800                                 swcount = *swap_map;
 801                                 if (swcount <= 1)
 802                                         ;
 803                                 else if (mm == &init_mm) {
 804                                         set_start_mm = 1;
 805                                         shmem = shmem_unuse(entry, page);
 806                                 } else
 807                                         retval = unuse_mm(mm, entry, page);
 808                                 if (set_start_mm && *swap_map < swcount) {
 809                                         mmput(new_start_mm);
 810                                         atomic_inc(&mm->mm_users);
 811                                         new_start_mm = mm;
 812                                         set_start_mm = 0;
 813                                 }
 814                                 spin_lock(&mmlist_lock);
 815                         }
 816                         spin_unlock(&mmlist_lock);
 817                         mmput(prev_mm);
 818                         mmput(start_mm);
 819                         start_mm = new_start_mm;
 820                 }
 821                 if (retval) {
 822                         unlock_page(page);
 823                         page_cache_release(page);
 824                         break;
 825                 }
 826
 827                 /*
 828                  * How could swap count reach 0x7fff when the maximum
 829                  * pid is 0x7fff, and there's no way to repeat a swap
 830                  * page within an mm (except in shmem, where it's the
 831                  * shared object which takes the reference count)?
 832                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 833                  *
 834                  * If that's wrong, then we should worry more about
 835                  * exit_mmap() and do_munmap() cases described above:
 836                  * we might be resetting SWAP_MAP_MAX too early here.
 837                  * We know "Undead"s can happen, they're okay, so don't
 838                  * report them; but do report if we reset SWAP_MAP_MAX.
 839                  */
 840                 if (*swap_map == SWAP_MAP_MAX) {
 841                         spin_lock(&swap_lock);
 842                         *swap_map = 1;
 843                         spin_unlock(&swap_lock);
 844                         reset_overflow = 1;
 845                 }
 846
 847                 /*
 848                  * If a reference remains (rare), we would like to leave
 849                  * the page in the swap cache; but try_to_unmap could
 850                  * then re-duplicate the entry once we drop page lock,
 851                  * so we might loop indefinitely; also, that page could
 852                  * not be swapped out to other storage meanwhile.  So:
 853                  * delete from cache even if there's another reference,
 854                  * after ensuring that the data has been saved to disk -
 855                  * since if the reference remains (rarer), it will be
 856                  * read from disk into another page.  Splitting into two
 857                  * pages would be incorrect if swap supported "shared
 858                  * private" pages, but they are handled by tmpfs files.
 859                  *
 860                  * Note shmem_unuse already deleted a swappage from
 861                  * the swap cache, unless the move to filepage failed:
 862                  * in which case it left swappage in cache, lowered its
 863                  * swap count to pass quickly through the loops above,
 864                  * and now we must reincrement count to try again later.
 865                  */
 866                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 867                         struct writeback_control wbc = {
 868                                 .sync_mode = WB_SYNC_NONE,
 869                         };
 870
 871                         swap_writepage(page, &wbc);
 872                         lock_page(page);
 873                         wait_on_page_writeback(page);
 874                 }
 875                 if (PageSwapCache(page)) {
 876                         if (shmem)
 877                                 swap_duplicate(entry);
 878                         else
 879                                 delete_from_swap_cache(page);
 880                 }
 881
 882                 /*
 883                  * So we could skip searching mms once swap count went
 884                  * to 1, we did not mark any present ptes as dirty: must
 885                  * mark page dirty so shrink_list will preserve it.
 886                  */
 887                 SetPageDirty(page);
 888                 unlock_page(page);
 889                 page_cache_release(page);
 890
 891                 /*
 892                  * Make sure that we aren't completely killing
 893                  * interactive performance.
 894                  */
 895                 cond_resched();
 896         }
 897
 898         mmput(start_mm);
 899         if (reset_overflow) {
 900                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 901                 swap_overflow = 0;
 902         }
 903         return retval;
 904 }
 905
 906 /*
 907  * After a successful try_to_unuse, if no swap is now in use, we know
 908  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 909  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 910  * added to the mmlist just after page_duplicate - before would be racy.
 911  */
 912 static void drain_mmlist(void)
 913 {
 914         struct list_head *p, *next;
 915         unsigned int i;
 916
 917         for (i = 0; i < nr_swapfiles; i++)
 918                 if (swap_info[i].inuse_pages)
 919                         return;
 920         spin_lock(&mmlist_lock);
 921         list_for_each_safe(p, next, &init_mm.mmlist)
 922                 list_del_init(p);
 923         spin_unlock(&mmlist_lock);
 924 }
 925
 926 /*
 927  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 928  * corresponds to page offset `offset'.
 929  */
 930 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 931 {
 932         struct swap_extent *se = sis->curr_swap_extent;
 933         struct swap_extent *start_se = se;
 934
 935         for ( ; ; ) {
 936                 struct list_head *lh;
 937
 938                 if (se->start_page <= offset &&
 939                                 offset < (se->start_page + se->nr_pages)) {
 940                         return se->start_block + (offset - se->start_page);
 941                 }
 942                 lh = se->list.next;
 943                 if (lh == &sis->extent_list)
 944                         lh = lh->next;
 945                 se = list_entry(lh, struct swap_extent, list);
 946                 sis->curr_swap_extent = se;
 947                 BUG_ON(se == start_se);         /* It *must* be present */
 948         }
 949 }
 950
 951 /*
 952  * Free all of a swapdev's extent information
 953  */
 954 static void destroy_swap_extents(struct swap_info_struct *sis)
 955 {
 956         while (!list_empty(&sis->extent_list)) {
 957                 struct swap_extent *se;
 958
 959                 se = list_entry(sis->extent_list.next,
 960                                 struct swap_extent, list);
 961                 list_del(&se->list);
 962                 kfree(se);
 963         }
 964 }
 965
 966 /*
 967  * Add a block range (and the corresponding page range) into this swapdev's
 968  * extent list.  The extent list is kept sorted in page order.
 969  *
 970  * This function rather assumes that it is called in ascending page order.
 971  */
 972 static int
 973 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
 974                 unsigned long nr_pages, sector_t start_block)
 975 {
 976         struct swap_extent *se;
 977         struct swap_extent *new_se;
 978         struct list_head *lh;
 979
 980         lh = sis->extent_list.prev;     /* The highest page extent */
 981         if (lh != &sis->extent_list) {
 982                 se = list_entry(lh, struct swap_extent, list);
 983                 BUG_ON(se->start_page + se->nr_pages != start_page);
 984                 if (se->start_block + se->nr_pages == start_block) {
 985                         /* Merge it */
 986                         se->nr_pages += nr_pages;
 987                         return 0;
 988                 }
 989         }
 990
 991         /*
 992          * No merge.  Insert a new extent, preserving ordering.
 993          */
 994         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
 995         if (new_se == NULL)
 996                 return -ENOMEM;
 997         new_se->start_page = start_page;
 998         new_se->nr_pages = nr_pages;
 999         new_se->start_block = start_block;
1000
1001         list_add_tail(&new_se->list, &sis->extent_list);
1002         return 1;
1003 }
1004
1005 /*
1006  * A `swap extent' is a simple thing which maps a contiguous range of pages
1007  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1008  * is built at swapon time and is then used at swap_writepage/swap_readpage
1009  * time for locating where on disk a page belongs.
1010  *
1011  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1012  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1013  * swap files identically.
1014  *
1015  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1016  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1017  * swapfiles are handled *identically* after swapon time.
1018  *
1019  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1020  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1021  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1022  * requirements, they are simply tossed out - we will never use those blocks
1023  * for swapping.
1024  *
1025  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1026  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1027  * which will scribble on the fs.
1028  *
1029  * The amount of disk space which a single swap extent represents varies.
1030  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1031  * extents in the list.  To avoid much list walking, we cache the previous
1032  * search location in `curr_swap_extent', and start new searches from there.
1033  * This is extremely effective.  The average number of iterations in
1034  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1035  */
1036 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1037 {
1038         struct inode *inode;
1039         unsigned blocks_per_page;
1040         unsigned long page_no;
1041         unsigned blkbits;
1042         sector_t probe_block;
1043         sector_t last_block;
1044         sector_t lowest_block = -1;
1045         sector_t highest_block = 0;
1046         int nr_extents = 0;
1047         int ret;
1048
1049         inode = sis->swap_file->f_mapping->host;
1050         if (S_ISBLK(inode->i_mode)) {
1051                 ret = add_swap_extent(sis, 0, sis->max, 0);
1052                 *span = sis->pages;
1053                 goto done;
1054         }
1055
1056         blkbits = inode->i_blkbits;
1057         blocks_per_page = PAGE_SIZE >> blkbits;
1058
1059         /*
1060          * Map all the blocks into the extent list.  This code doesn't try
1061          * to be very smart.
1062          */
1063         probe_block = 0;
1064         page_no = 0;
1065         last_block = i_size_read(inode) >> blkbits;
1066         while ((probe_block + blocks_per_page) <= last_block &&
1067                         page_no < sis->max) {
1068                 unsigned block_in_page;
1069                 sector_t first_block;
1070
1071                 first_block = bmap(inode, probe_block);
1072                 if (first_block == 0)
1073                         goto bad_bmap;
1074
1075                 /*
1076                  * It must be PAGE_SIZE aligned on-disk
1077                  */
1078                 if (first_block & (blocks_per_page - 1)) {
1079                         probe_block++;
1080                         goto reprobe;
1081                 }
1082
1083                 for (block_in_page = 1; block_in_page < blocks_per_page;
1084                                         block_in_page++) {
1085                         sector_t block;
1086
1087                         block = bmap(inode, probe_block + block_in_page);
1088                         if (block == 0)
1089                                 goto bad_bmap;
1090                         if (block != first_block + block_in_page) {
1091                                 /* Discontiguity */
1092                                 probe_block++;
1093                                 goto reprobe;
1094                         }
1095                 }
1096
1097                 first_block >>= (PAGE_SHIFT - blkbits);
1098                 if (page_no) {  /* exclude the header page */
1099                         if (first_block < lowest_block)
1100                                 lowest_block = first_block;
1101                         if (first_block > highest_block)
1102                                 highest_block = first_block;
1103                 }
1104
1105                 /*
1106                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1107                  */
1108                 ret = add_swap_extent(sis, page_no, 1, first_block);
1109                 if (ret < 0)
1110                         goto out;
1111                 nr_extents += ret;
1112                 page_no++;
1113                 probe_block += blocks_per_page;
1114 reprobe:
1115                 continue;
1116         }
1117         ret = nr_extents;
1118         *span = 1 + highest_block - lowest_block;
1119         if (page_no == 0)
1120                 page_no = 1;    /* force Empty message */
1121         sis->max = page_no;
1122         sis->pages = page_no - 1;
1123         sis->highest_bit = page_no - 1;
1124 done:
1125         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1126                                         struct swap_extent, list);
1127         goto out;
1128 bad_bmap:
1129         printk(KERN_ERR "swapon: swapfile has holes\n");
1130         ret = -EINVAL;
1131 out:
1132         return ret;
1133 }
1134
1135 #if 0   /* We don't need this yet */
1136 #include <linux/backing-dev.h>
1137 int page_queue_congested(struct page *page)
1138 {
1139         struct backing_dev_info *bdi;
1140
1141         BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
1142
1143         if (PageSwapCache(page)) {
1144                 swp_entry_t entry = { .val = page_private(page) };
1145                 struct swap_info_struct *sis;
1146
1147                 sis = get_swap_info_struct(swp_type(entry));
1148                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1149         } else
1150                 bdi = page->mapping->backing_dev_info;
1151         return bdi_write_congested(bdi);
1152 }
1153 #endif
1154
1155 asmlinkage long sys_swapoff(const char __user * specialfile)
1156 {
1157         struct swap_info_struct * p = NULL;
1158         unsigned short *swap_map;
1159         struct file *swap_file, *victim;
1160         struct address_space *mapping;
1161         struct inode *inode;
1162         char * pathname;
1163         int i, type, prev;
1164         int err;
1165
1166         if (!capable(CAP_SYS_ADMIN))
1167                 return -EPERM;
1168
1169         pathname = getname(specialfile);
1170         err = PTR_ERR(pathname);
1171         if (IS_ERR(pathname))
1172                 goto out;
1173
1174         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1175         putname(pathname);
1176         err = PTR_ERR(victim);
1177         if (IS_ERR(victim))
1178                 goto out;
1179
1180         mapping = victim->f_mapping;
1181         prev = -1;
1182         spin_lock(&swap_lock);
1183         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1184                 p = swap_info + type;
1185                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1186                         if (p->swap_file->f_mapping == mapping)
1187                                 break;
1188                 }
1189                 prev = type;
1190         }
1191         if (type < 0) {
1192                 err = -EINVAL;
1193                 spin_unlock(&swap_lock);
1194                 goto out_dput;
1195         }
1196         if (!security_vm_enough_memory(p->pages))
1197                 vm_unacct_memory(p->pages);
1198         else {
1199                 err = -ENOMEM;
1200                 spin_unlock(&swap_lock);
1201                 goto out_dput;
1202         }
1203         if (prev < 0) {
1204                 swap_list.head = p->next;
1205         } else {
1206                 swap_info[prev].next = p->next;
1207         }
1208         if (type == swap_list.next) {
1209                 /* just pick something that's safe... */
1210                 swap_list.next = swap_list.head;
1211         }
1212         nr_swap_pages -= p->pages;
1213         total_swap_pages -= p->pages;
1214         p->flags &= ~SWP_WRITEOK;
1215         spin_unlock(&swap_lock);
1216
1217         current->flags |= PF_SWAPOFF;
1218         err = try_to_unuse(type);
1219         current->flags &= ~PF_SWAPOFF;
1220
1221         if (err) {
1222                 /* re-insert swap space back into swap_list */
1223                 spin_lock(&swap_lock);
1224                 for (prev = -1, i = swap_list.head; i >= 0; prev = i, i = swap_info[i].next)
1225                         if (p->prio >= swap_info[i].prio)
1226                                 break;
1227                 p->next = i;
1228                 if (prev < 0)
1229                         swap_list.head = swap_list.next = p - swap_info;
1230                 else
1231                         swap_info[prev].next = p - swap_info;
1232                 nr_swap_pages += p->pages;
1233                 total_swap_pages += p->pages;
1234                 p->flags |= SWP_WRITEOK;
1235                 spin_unlock(&swap_lock);
1236                 goto out_dput;
1237         }
1238
1239         /* wait for any unplug function to finish */
1240         down_write(&swap_unplug_sem);
1241         up_write(&swap_unplug_sem);
1242
1243         destroy_swap_extents(p);
1244         mutex_lock(&swapon_mutex);
1245         spin_lock(&swap_lock);
1246         drain_mmlist();
1247
1248         /* wait for anyone still in scan_swap_map */
1249         p->highest_bit = 0;             /* cuts scans short */
1250         while (p->flags >= SWP_SCANNING) {
1251                 spin_unlock(&swap_lock);
1252                 schedule_timeout_uninterruptible(1);
1253                 spin_lock(&swap_lock);
1254         }
1255
1256         swap_file = p->swap_file;
1257         p->swap_file = NULL;
1258         p->max = 0;
1259         swap_map = p->swap_map;
1260         p->swap_map = NULL;
1261         p->flags = 0;
1262         spin_unlock(&swap_lock);
1263         mutex_unlock(&swapon_mutex);
1264         vfree(swap_map);
1265         inode = mapping->host;
1266         if (S_ISBLK(inode->i_mode)) {
1267                 struct block_device *bdev = I_BDEV(inode);
1268                 set_blocksize(bdev, p->old_block_size);
1269                 bd_release(bdev);
1270         } else {
1271                 mutex_lock(&inode->i_mutex);
1272                 inode->i_flags &= ~S_SWAPFILE;
1273                 mutex_unlock(&inode->i_mutex);
1274         }
1275         filp_close(swap_file, NULL);
1276         err = 0;
1277
1278 out_dput:
1279         filp_close(victim, NULL);
1280 out:
1281         return err;
1282 }
1283
1284 #ifdef CONFIG_PROC_FS
1285 /* iterator */
1286 static void *swap_start(struct seq_file *swap, loff_t *pos)
1287 {
1288         struct swap_info_struct *ptr = swap_info;
1289         int i;
1290         loff_t l = *pos;
1291
1292         mutex_lock(&swapon_mutex);
1293
1294         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1295                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1296                         continue;
1297                 if (!l--)
1298                         return ptr;
1299         }
1300
1301         return NULL;
1302 }
1303
1304 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1305 {
1306         struct swap_info_struct *ptr = v;
1307         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1308
1309         for (++ptr; ptr < endptr; ptr++) {
1310                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1311                         continue;
1312                 ++*pos;
1313                 return ptr;
1314         }
1315
1316         return NULL;
1317 }
1318
1319 static void swap_stop(struct seq_file *swap, void *v)
1320 {
1321         mutex_unlock(&swapon_mutex);
1322 }
1323
1324 static int swap_show(struct seq_file *swap, void *v)
1325 {
1326         struct swap_info_struct *ptr = v;
1327         struct file *file;
1328         int len;
1329
1330         if (v == swap_info)
1331                 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1332
1333         file = ptr->swap_file;
1334         len = seq_path(swap, file->f_vfsmnt, file->f_dentry, " \t\n\\");
1335         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1336                        len < 40 ? 40 - len : 1, " ",
1337                        S_ISBLK(file->f_dentry->d_inode->i_mode) ?
1338                                 "partition" : "file\t",
1339                        ptr->pages << (PAGE_SHIFT - 10),
1340                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1341                        ptr->prio);
1342         return 0;
1343 }
1344
1345 static struct seq_operations swaps_op = {
1346         .start =        swap_start,
1347         .next =         swap_next,
1348         .stop =         swap_stop,
1349         .show =         swap_show
1350 };
1351
1352 static int swaps_open(struct inode *inode, struct file *file)
1353 {
1354         return seq_open(file, &swaps_op);
1355 }
1356
1357 static struct file_operations proc_swaps_operations = {
1358         .open           = swaps_open,
1359         .read           = seq_read,
1360         .llseek         = seq_lseek,
1361         .release        = seq_release,
1362 };
1363
1364 static int __init procswaps_init(void)
1365 {
1366         struct proc_dir_entry *entry;
1367
1368         entry = create_proc_entry("swaps", 0, NULL);
1369         if (entry)
1370                 entry->proc_fops = &proc_swaps_operations;
1371         return 0;
1372 }
1373 __initcall(procswaps_init);
1374 #endif /* CONFIG_PROC_FS */
1375
1376 /*
1377  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1378  *
1379  * The swapon system call
1380  */
1381 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1382 {
1383         struct swap_info_struct * p;
1384         char *name = NULL;
1385         struct block_device *bdev = NULL;
1386         struct file *swap_file = NULL;
1387         struct address_space *mapping;
1388         unsigned int type;
1389         int i, prev;
1390         int error;
1391         static int least_priority;
1392         union swap_header *swap_header = NULL;
1393         int swap_header_version;
1394         unsigned int nr_good_pages = 0;
1395         int nr_extents = 0;
1396         sector_t span;
1397         unsigned long maxpages = 1;
1398         int swapfilesize;
1399         unsigned short *swap_map;
1400         struct page *page = NULL;
1401         struct inode *inode = NULL;
1402         int did_down = 0;
1403
1404         if (!capable(CAP_SYS_ADMIN))
1405                 return -EPERM;
1406         spin_lock(&swap_lock);
1407         p = swap_info;
1408         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1409                 if (!(p->flags & SWP_USED))
1410                         break;
1411         error = -EPERM;
1412         /*
1413          * Test if adding another swap device is possible. There are
1414          * two limiting factors: 1) the number of bits for the swap
1415          * type swp_entry_t definition and 2) the number of bits for
1416          * the swap type in the swap ptes as defined by the different
1417          * architectures. To honor both limitations a swap entry
1418          * with swap offset 0 and swap type ~0UL is created, encoded
1419          * to a swap pte, decoded to a swp_entry_t again and finally
1420          * the swap type part is extracted. This will mask all bits
1421          * from the initial ~0UL that can't be encoded in either the
1422          * swp_entry_t or the architecture definition of a swap pte.
1423          */
1424         if (type > swp_type(pte_to_swp_entry(swp_entry_to_pte(swp_entry(~0UL,0))))) {
1425                 spin_unlock(&swap_lock);
1426                 goto out;
1427         }
1428         if (type >= nr_swapfiles)
1429                 nr_swapfiles = type+1;
1430         INIT_LIST_HEAD(&p->extent_list);
1431         p->flags = SWP_USED;
1432         p->swap_file = NULL;
1433         p->old_block_size = 0;
1434         p->swap_map = NULL;
1435         p->lowest_bit = 0;
1436         p->highest_bit = 0;
1437         p->cluster_nr = 0;
1438         p->inuse_pages = 0;
1439         p->next = -1;
1440         if (swap_flags & SWAP_FLAG_PREFER) {
1441                 p->prio =
1442                   (swap_flags & SWAP_FLAG_PRIO_MASK)>>SWAP_FLAG_PRIO_SHIFT;
1443         } else {
1444                 p->prio = --least_priority;
1445         }
1446         spin_unlock(&swap_lock);
1447         name = getname(specialfile);
1448         error = PTR_ERR(name);
1449         if (IS_ERR(name)) {
1450                 name = NULL;
1451                 goto bad_swap_2;
1452         }
1453         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1454         error = PTR_ERR(swap_file);
1455         if (IS_ERR(swap_file)) {
1456                 swap_file = NULL;
1457                 goto bad_swap_2;
1458         }
1459
1460         p->swap_file = swap_file;
1461         mapping = swap_file->f_mapping;
1462         inode = mapping->host;
1463
1464         error = -EBUSY;
1465         for (i = 0; i < nr_swapfiles; i++) {
1466                 struct swap_info_struct *q = &swap_info[i];
1467
1468                 if (i == type || !q->swap_file)
1469                         continue;
1470                 if (mapping == q->swap_file->f_mapping)
1471                         goto bad_swap;
1472         }
1473
1474         error = -EINVAL;
1475         if (S_ISBLK(inode->i_mode)) {
1476                 bdev = I_BDEV(inode);
1477                 error = bd_claim(bdev, sys_swapon);
1478                 if (error < 0) {
1479                         bdev = NULL;
1480                         error = -EINVAL;
1481                         goto bad_swap;
1482                 }
1483                 p->old_block_size = block_size(bdev);
1484                 error = set_blocksize(bdev, PAGE_SIZE);
1485                 if (error < 0)
1486                         goto bad_swap;
1487                 p->bdev = bdev;
1488         } else if (S_ISREG(inode->i_mode)) {
1489                 p->bdev = inode->i_sb->s_bdev;
1490                 mutex_lock(&inode->i_mutex);
1491                 did_down = 1;
1492                 if (IS_SWAPFILE(inode)) {
1493                         error = -EBUSY;
1494                         goto bad_swap;
1495                 }
1496         } else {
1497                 goto bad_swap;
1498         }
1499
1500         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1501
1502         /*
1503          * Read the swap header.
1504          */
1505         if (!mapping->a_ops->readpage) {
1506                 error = -EINVAL;
1507                 goto bad_swap;
1508         }
1509         page = read_cache_page(mapping, 0,
1510                         (filler_t *)mapping->a_ops->readpage, swap_file);
1511         if (IS_ERR(page)) {
1512                 error = PTR_ERR(page);
1513                 goto bad_swap;
1514         }
1515         wait_on_page_locked(page);
1516         if (!PageUptodate(page))
1517                 goto bad_swap;
1518         kmap(page);
1519         swap_header = page_address(page);
1520
1521         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1522                 swap_header_version = 1;
1523         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1524                 swap_header_version = 2;
1525         else {
1526                 printk(KERN_ERR "Unable to find swap-space signature\n");
1527                 error = -EINVAL;
1528                 goto bad_swap;
1529         }
1530
1531         switch (swap_header_version) {
1532         case 1:
1533                 printk(KERN_ERR "version 0 swap is no longer supported. "
1534                         "Use mkswap -v1 %s\n", name);
1535                 error = -EINVAL;
1536                 goto bad_swap;
1537         case 2:
1538                 /* Check the swap header's sub-version and the size of
1539                    the swap file and bad block lists */
1540                 if (swap_header->info.version != 1) {
1541                         printk(KERN_WARNING
1542                                "Unable to handle swap header version %d\n",
1543                                swap_header->info.version);
1544                         error = -EINVAL;
1545                         goto bad_swap;
1546                 }
1547
1548                 p->lowest_bit  = 1;
1549                 p->cluster_next = 1;
1550
1551                 /*
1552                  * Find out how many pages are allowed for a single swap
1553                  * device. There are two limiting factors: 1) the number of
1554                  * bits for the swap offset in the swp_entry_t type and
1555                  * 2) the number of bits in the a swap pte as defined by
1556                  * the different architectures. In order to find the
1557                  * largest possible bit mask a swap entry with swap type 0
1558                  * and swap offset ~0UL is created, encoded to a swap pte,
1559                  * decoded to a swp_entry_t again and finally the swap
1560                  * offset is extracted. This will mask all the bits from
1561                  * the initial ~0UL mask that can't be encoded in either
1562                  * the swp_entry_t or the architecture definition of a
1563                  * swap pte.
1564                  */
1565                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1566                 if (maxpages > swap_header->info.last_page)
1567                         maxpages = swap_header->info.last_page;
1568                 p->highest_bit = maxpages - 1;
1569
1570                 error = -EINVAL;
1571                 if (!maxpages)
1572                         goto bad_swap;
1573                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1574                         goto bad_swap;
1575                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1576                         goto bad_swap;
1577
1578                 /* OK, set up the swap map and apply the bad block list */
1579                 if (!(p->swap_map = vmalloc(maxpages * sizeof(short)))) {
1580                         error = -ENOMEM;
1581                         goto bad_swap;
1582                 }
1583
1584                 error = 0;
1585                 memset(p->swap_map, 0, maxpages * sizeof(short));
1586                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1587                         int page_nr = swap_header->info.badpages[i];
1588                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1589                                 error = -EINVAL;
1590                         else
1591                                 p->swap_map[page_nr] = SWAP_MAP_BAD;
1592                 }
1593                 nr_good_pages = swap_header->info.last_page -
1594                                 swap_header->info.nr_badpages -
1595                                 1 /* header page */;
1596                 if (error)
1597                         goto bad_swap;
1598         }
1599
1600         if (swapfilesize && maxpages > swapfilesize) {
1601                 printk(KERN_WARNING
1602                        "Swap area shorter than signature indicates\n");
1603                 error = -EINVAL;
1604                 goto bad_swap;
1605         }
1606         if (nr_good_pages) {
1607                 p->swap_map[0] = SWAP_MAP_BAD;
1608                 p->max = maxpages;
1609                 p->pages = nr_good_pages;
1610                 nr_extents = setup_swap_extents(p, &span);
1611                 if (nr_extents < 0) {
1612                         error = nr_extents;
1613                         goto bad_swap;
1614                 }
1615                 nr_good_pages = p->pages;
1616         }
1617         if (!nr_good_pages) {
1618                 printk(KERN_WARNING "Empty swap-file\n");
1619                 error = -EINVAL;
1620                 goto bad_swap;
1621         }
1622
1623         mutex_lock(&swapon_mutex);
1624         spin_lock(&swap_lock);
1625         p->flags = SWP_ACTIVE;
1626         nr_swap_pages += nr_good_pages;
1627         total_swap_pages += nr_good_pages;
1628
1629         printk(KERN_INFO "Adding %uk swap on %s.  "
1630                         "Priority:%d extents:%d across:%lluk\n",
1631                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1632                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1633
1634         /* insert swap space into swap_list: */
1635         prev = -1;
1636         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1637                 if (p->prio >= swap_info[i].prio) {
1638                         break;
1639                 }
1640                 prev = i;
1641         }
1642         p->next = i;
1643         if (prev < 0) {
1644                 swap_list.head = swap_list.next = p - swap_info;
1645         } else {
1646                 swap_info[prev].next = p - swap_info;
1647         }
1648         spin_unlock(&swap_lock);
1649         mutex_unlock(&swapon_mutex);
1650         error = 0;
1651         goto out;
1652 bad_swap:
1653         if (bdev) {
1654                 set_blocksize(bdev, p->old_block_size);
1655                 bd_release(bdev);
1656         }
1657         destroy_swap_extents(p);
1658 bad_swap_2:
1659         spin_lock(&swap_lock);
1660         swap_map = p->swap_map;
1661         p->swap_file = NULL;
1662         p->swap_map = NULL;
1663         p->flags = 0;
1664         if (!(swap_flags & SWAP_FLAG_PREFER))
1665                 ++least_priority;
1666         spin_unlock(&swap_lock);
1667         vfree(swap_map);
1668         if (swap_file)
1669                 filp_close(swap_file, NULL);
1670 out:
1671         if (page && !IS_ERR(page)) {
1672                 kunmap(page);
1673                 page_cache_release(page);
1674         }
1675         if (name)
1676                 putname(name);
1677         if (did_down) {
1678                 if (!error)
1679                         inode->i_flags |= S_SWAPFILE;
1680                 mutex_unlock(&inode->i_mutex);
1681         }
1682         return error;
1683 }
1684
1685 void si_swapinfo(struct sysinfo *val)
1686 {
1687         unsigned int i;
1688         unsigned long nr_to_be_unused = 0;
1689
1690         spin_lock(&swap_lock);
1691         for (i = 0; i < nr_swapfiles; i++) {
1692                 if (!(swap_info[i].flags & SWP_USED) ||
1693                      (swap_info[i].flags & SWP_WRITEOK))
1694                         continue;
1695                 nr_to_be_unused += swap_info[i].inuse_pages;
1696         }
1697         val->freeswap = nr_swap_pages + nr_to_be_unused;
1698         val->totalswap = total_swap_pages + nr_to_be_unused;
1699         spin_unlock(&swap_lock);
1700         if (vx_flags(VXF_VIRT_MEM, 0))
1701                 vx_vsi_swapinfo(val);
1702 }
1703
1704 /*
1705  * Verify that a swap entry is valid and increment its swap map count.
1706  *
1707  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1708  * "permanent", but will be reclaimed by the next swapoff.
1709  */
1710 int swap_duplicate(swp_entry_t entry)
1711 {
1712         struct swap_info_struct * p;
1713         unsigned long offset, type;
1714         int result = 0;
1715
1716         type = swp_type(entry);
1717         if (type >= nr_swapfiles)
1718                 goto bad_file;
1719         p = type + swap_info;
1720         offset = swp_offset(entry);
1721
1722         spin_lock(&swap_lock);
1723         if (offset < p->max && p->swap_map[offset]) {
1724                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1725                         p->swap_map[offset]++;
1726                         result = 1;
1727                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1728                         if (swap_overflow++ < 5)
1729                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1730                         p->swap_map[offset] = SWAP_MAP_MAX;
1731                         result = 1;
1732                 }
1733         }
1734         spin_unlock(&swap_lock);
1735 out:
1736         return result;
1737
1738 bad_file:
1739         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1740         goto out;
1741 }
1742
1743 struct swap_info_struct *
1744 get_swap_info_struct(unsigned type)
1745 {
1746         return &swap_info[type];
1747 }
1748
1749 /*
1750  * swap_lock prevents swap_map being freed. Don't grab an extra
1751  * reference on the swaphandle, it doesn't matter if it becomes unused.
1752  */
1753 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1754 {
1755         int ret = 0, i = 1 << page_cluster;
1756         unsigned long toff;
1757         struct swap_info_struct *swapdev = swp_type(entry) + swap_info;
1758
1759         if (!page_cluster)      /* no readahead */
1760                 return 0;
1761         toff = (swp_offset(entry) >> page_cluster) << page_cluster;
1762         if (!toff)              /* first page is swap header */
1763                 toff++, i--;
1764         *offset = toff;
1765
1766         spin_lock(&swap_lock);
1767         do {
1768                 /* Don't read-ahead past the end of the swap area */
1769                 if (toff >= swapdev->max)
1770                         break;
1771                 /* Don't read in free or bad pages */
1772                 if (!swapdev->swap_map[toff])
1773                         break;
1774                 if (swapdev->swap_map[toff] == SWAP_MAP_BAD)
1775                         break;
1776                 toff++;
1777                 ret++;
1778         } while (--i);
1779         spin_unlock(&swap_lock);
1780         return ret;
1781 }