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/mm.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/mman.h>
  11 #include <linux/slab.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/swap.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/namei.h>
  17 #include <linux/shm.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/writeback.h>
  20 #include <linux/proc_fs.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/init.h>
  23 #include <linux/module.h>
  24 #include <linux/rmap.h>
  25 #include <linux/security.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/mutex.h>
  28 #include <linux/capability.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/memcontrol.h>
  31
  32 #include <asm/pgtable.h>
  33 #include <asm/tlbflush.h>
  34 #include <linux/swapops.h>
  35
  36 static DEFINE_SPINLOCK(swap_lock);
  37 static unsigned int nr_swapfiles;
  38 long total_swap_pages;
  39 static int swap_overflow;
  40 static int least_priority;
  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 static 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 write to an anon page without COW if there are no other references
 330  * to it.  And as a side-effect, free up its swap: because the old content
 331  * on disk will never be read, and seeking back there to write new content
 332  * later would only waste time away from clustering.
 333  */
 334 int reuse_swap_page(struct page *page)
 335 {
 336         int count;
 337
 338         VM_BUG_ON(!PageLocked(page));
 339         count = page_mapcount(page);
 340         if (count <= 1 && PageSwapCache(page)) {
 341                 count += page_swapcount(page);
 342                 if (count == 1 && !PageWriteback(page)) {
 343                         delete_from_swap_cache(page);
 344                         SetPageDirty(page);
 345                 }
 346         }
 347         return count == 1;
 348 }
 349
 350 /*
 351  * If swap is getting full, or if there are no more mappings of this page,
 352  * then try_to_free_swap is called to free its swap space.
 353  */
 354 int try_to_free_swap(struct page *page)
 355 {
 356         VM_BUG_ON(!PageLocked(page));
 357
 358         if (!PageSwapCache(page))
 359                 return 0;
 360         if (PageWriteback(page))
 361                 return 0;
 362         if (page_swapcount(page))
 363                 return 0;
 364
 365         delete_from_swap_cache(page);
 366         SetPageDirty(page);
 367         return 1;
 368 }
 369
 370 /*
 371  * Free the swap entry like above, but also try to
 372  * free the page cache entry if it is the last user.
 373  */
 374 void free_swap_and_cache(swp_entry_t entry)
 375 {
 376         struct swap_info_struct * p;
 377         struct page *page = NULL;
 378
 379         if (is_migration_entry(entry))
 380                 return;
 381
 382         p = swap_info_get(entry);
 383         if (p) {
 384                 if (swap_entry_free(p, swp_offset(entry)) == 1) {
 385                         page = find_get_page(&swapper_space, entry.val);
 386                         if (page && !trylock_page(page)) {
 387                                 page_cache_release(page);
 388                                 page = NULL;
 389                         }
 390                 }
 391                 spin_unlock(&swap_lock);
 392         }
 393         if (page) {
 394                 /*
 395                  * Not mapped elsewhere, or swap space full? Free it!
 396                  * Also recheck PageSwapCache now page is locked (above).
 397                  */
 398                 if (PageSwapCache(page) && !PageWriteback(page) &&
 399                                 (!page_mapped(page) || vm_swap_full())) {
 400                         delete_from_swap_cache(page);
 401                         SetPageDirty(page);
 402                 }
 403                 unlock_page(page);
 404                 page_cache_release(page);
 405         }
 406 }
 407
 408 #ifdef CONFIG_HIBERNATION
 409 /*
 410  * Find the swap type that corresponds to given device (if any).
 411  *
 412  * @offset - number of the PAGE_SIZE-sized block of the device, starting
 413  * from 0, in which the swap header is expected to be located.
 414  *
 415  * This is needed for the suspend to disk (aka swsusp).
 416  */
 417 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 418 {
 419         struct block_device *bdev = NULL;
 420         int i;
 421
 422         if (device)
 423                 bdev = bdget(device);
 424
 425         spin_lock(&swap_lock);
 426         for (i = 0; i < nr_swapfiles; i++) {
 427                 struct swap_info_struct *sis = swap_info + i;
 428
 429                 if (!(sis->flags & SWP_WRITEOK))
 430                         continue;
 431
 432                 if (!bdev) {
 433                         if (bdev_p)
 434                                 *bdev_p = sis->bdev;
 435
 436                         spin_unlock(&swap_lock);
 437                         return i;
 438                 }
 439                 if (bdev == sis->bdev) {
 440                         struct swap_extent *se;
 441
 442                         se = list_entry(sis->extent_list.next,
 443                                         struct swap_extent, list);
 444                         if (se->start_block == offset) {
 445                                 if (bdev_p)
 446                                         *bdev_p = sis->bdev;
 447
 448                                 spin_unlock(&swap_lock);
 449                                 bdput(bdev);
 450                                 return i;
 451                         }
 452                 }
 453         }
 454         spin_unlock(&swap_lock);
 455         if (bdev)
 456                 bdput(bdev);
 457
 458         return -ENODEV;
 459 }
 460
 461 /*
 462  * Return either the total number of swap pages of given type, or the number
 463  * of free pages of that type (depending on @free)
 464  *
 465  * This is needed for software suspend
 466  */
 467 unsigned int count_swap_pages(int type, int free)
 468 {
 469         unsigned int n = 0;
 470
 471         if (type < nr_swapfiles) {
 472                 spin_lock(&swap_lock);
 473                 if (swap_info[type].flags & SWP_WRITEOK) {
 474                         n = swap_info[type].pages;
 475                         if (free)
 476                                 n -= swap_info[type].inuse_pages;
 477                 }
 478                 spin_unlock(&swap_lock);
 479         }
 480         return n;
 481 }
 482 #endif
 483
 484 /*
 485  * No need to decide whether this PTE shares the swap entry with others,
 486  * just let do_wp_page work it out if a write is requested later - to
 487  * force COW, vm_page_prot omits write permission from any private vma.
 488  */
 489 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 490                 unsigned long addr, swp_entry_t entry, struct page *page)
 491 {
 492         spinlock_t *ptl;
 493         pte_t *pte;
 494         int ret = 1;
 495
 496         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
 497                 ret = -ENOMEM;
 498
 499         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 500         if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
 501                 if (ret > 0)
 502                         mem_cgroup_uncharge_page(page);
 503                 ret = 0;
 504                 goto out;
 505         }
 506
 507         inc_mm_counter(vma->vm_mm, anon_rss);
 508         get_page(page);
 509         set_pte_at(vma->vm_mm, addr, pte,
 510                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
 511         page_add_anon_rmap(page, vma, addr);
 512         swap_free(entry);
 513         /*
 514          * Move the page to the active list so it is not
 515          * immediately swapped out again after swapon.
 516          */
 517         activate_page(page);
 518 out:
 519         pte_unmap_unlock(pte, ptl);
 520         return ret;
 521 }
 522
 523 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 524                                 unsigned long addr, unsigned long end,
 525                                 swp_entry_t entry, struct page *page)
 526 {
 527         pte_t swp_pte = swp_entry_to_pte(entry);
 528         pte_t *pte;
 529         int ret = 0;
 530
 531         /*
 532          * We don't actually need pte lock while scanning for swp_pte: since
 533          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
 534          * page table while we're scanning; though it could get zapped, and on
 535          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
 536          * of unmatched parts which look like swp_pte, so unuse_pte must
 537          * recheck under pte lock.  Scanning without pte lock lets it be
 538          * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
 539          */
 540         pte = pte_offset_map(pmd, addr);
 541         do {
 542                 /*
 543                  * swapoff spends a _lot_ of time in this loop!
 544                  * Test inline before going to call unuse_pte.
 545                  */
 546                 if (unlikely(pte_same(*pte, swp_pte))) {
 547                         pte_unmap(pte);
 548                         ret = unuse_pte(vma, pmd, addr, entry, page);
 549                         if (ret)
 550                                 goto out;
 551                         pte = pte_offset_map(pmd, addr);
 552                 }
 553         } while (pte++, addr += PAGE_SIZE, addr != end);
 554         pte_unmap(pte - 1);
 555 out:
 556         return ret;
 557 }
 558
 559 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 560                                 unsigned long addr, unsigned long end,
 561                                 swp_entry_t entry, struct page *page)
 562 {
 563         pmd_t *pmd;
 564         unsigned long next;
 565         int ret;
 566
 567         pmd = pmd_offset(pud, addr);
 568         do {
 569                 next = pmd_addr_end(addr, end);
 570                 if (pmd_none_or_clear_bad(pmd))
 571                         continue;
 572                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
 573                 if (ret)
 574                         return ret;
 575         } while (pmd++, addr = next, addr != end);
 576         return 0;
 577 }
 578
 579 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 580                                 unsigned long addr, unsigned long end,
 581                                 swp_entry_t entry, struct page *page)
 582 {
 583         pud_t *pud;
 584         unsigned long next;
 585         int ret;
 586
 587         pud = pud_offset(pgd, addr);
 588         do {
 589                 next = pud_addr_end(addr, end);
 590                 if (pud_none_or_clear_bad(pud))
 591                         continue;
 592                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
 593                 if (ret)
 594                         return ret;
 595         } while (pud++, addr = next, addr != end);
 596         return 0;
 597 }
 598
 599 static int unuse_vma(struct vm_area_struct *vma,
 600                                 swp_entry_t entry, struct page *page)
 601 {
 602         pgd_t *pgd;
 603         unsigned long addr, end, next;
 604         int ret;
 605
 606         if (page->mapping) {
 607                 addr = page_address_in_vma(page, vma);
 608                 if (addr == -EFAULT)
 609                         return 0;
 610                 else
 611                         end = addr + PAGE_SIZE;
 612         } else {
 613                 addr = vma->vm_start;
 614                 end = vma->vm_end;
 615         }
 616
 617         pgd = pgd_offset(vma->vm_mm, addr);
 618         do {
 619                 next = pgd_addr_end(addr, end);
 620                 if (pgd_none_or_clear_bad(pgd))
 621                         continue;
 622                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
 623                 if (ret)
 624                         return ret;
 625         } while (pgd++, addr = next, addr != end);
 626         return 0;
 627 }
 628
 629 static int unuse_mm(struct mm_struct *mm,
 630                                 swp_entry_t entry, struct page *page)
 631 {
 632         struct vm_area_struct *vma;
 633         int ret = 0;
 634
 635         if (!down_read_trylock(&mm->mmap_sem)) {
 636                 /*
 637                  * Activate page so shrink_inactive_list is unlikely to unmap
 638                  * its ptes while lock is dropped, so swapoff can make progress.
 639                  */
 640                 activate_page(page);
 641                 unlock_page(page);
 642                 down_read(&mm->mmap_sem);
 643                 lock_page(page);
 644         }
 645         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 646                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
 647                         break;
 648         }
 649         up_read(&mm->mmap_sem);
 650         return (ret < 0)? ret: 0;
 651 }
 652
 653 /*
 654  * Scan swap_map from current position to next entry still in use.
 655  * Recycle to start on reaching the end, returning 0 when empty.
 656  */
 657 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 658                                         unsigned int prev)
 659 {
 660         unsigned int max = si->max;
 661         unsigned int i = prev;
 662         int count;
 663
 664         /*
 665          * No need for swap_lock here: we're just looking
 666          * for whether an entry is in use, not modifying it; false
 667          * hits are okay, and sys_swapoff() has already prevented new
 668          * allocations from this area (while holding swap_lock).
 669          */
 670         for (;;) {
 671                 if (++i >= max) {
 672                         if (!prev) {
 673                                 i = 0;
 674                                 break;
 675                         }
 676                         /*
 677                          * No entries in use at top of swap_map,
 678                          * loop back to start and recheck there.
 679                          */
 680                         max = prev + 1;
 681                         prev = 0;
 682                         i = 1;
 683                 }
 684                 count = si->swap_map[i];
 685                 if (count && count != SWAP_MAP_BAD)
 686                         break;
 687         }
 688         return i;
 689 }
 690
 691 /*
 692  * We completely avoid races by reading each swap page in advance,
 693  * and then search for the process using it.  All the necessary
 694  * page table adjustments can then be made atomically.
 695  */
 696 static int try_to_unuse(unsigned int type)
 697 {
 698         struct swap_info_struct * si = &swap_info[type];
 699         struct mm_struct *start_mm;
 700         unsigned short *swap_map;
 701         unsigned short swcount;
 702         struct page *page;
 703         swp_entry_t entry;
 704         unsigned int i = 0;
 705         int retval = 0;
 706         int reset_overflow = 0;
 707         int shmem;
 708
 709         /*
 710          * When searching mms for an entry, a good strategy is to
 711          * start at the first mm we freed the previous entry from
 712          * (though actually we don't notice whether we or coincidence
 713          * freed the entry).  Initialize this start_mm with a hold.
 714          *
 715          * A simpler strategy would be to start at the last mm we
 716          * freed the previous entry from; but that would take less
 717          * advantage of mmlist ordering, which clusters forked mms
 718          * together, child after parent.  If we race with dup_mmap(), we
 719          * prefer to resolve parent before child, lest we miss entries
 720          * duplicated after we scanned child: using last mm would invert
 721          * that.  Though it's only a serious concern when an overflowed
 722          * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
 723          */
 724         start_mm = &init_mm;
 725         atomic_inc(&init_mm.mm_users);
 726
 727         /*
 728          * Keep on scanning until all entries have gone.  Usually,
 729          * one pass through swap_map is enough, but not necessarily:
 730          * there are races when an instance of an entry might be missed.
 731          */
 732         while ((i = find_next_to_unuse(si, i)) != 0) {
 733                 if (signal_pending(current)) {
 734                         retval = -EINTR;
 735                         break;
 736                 }
 737
 738                 /*
 739                  * Get a page for the entry, using the existing swap
 740                  * cache page if there is one.  Otherwise, get a clean
 741                  * page and read the swap into it.
 742                  */
 743                 swap_map = &si->swap_map[i];
 744                 entry = swp_entry(type, i);
 745                 page = read_swap_cache_async(entry,
 746                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
 747                 if (!page) {
 748                         /*
 749                          * Either swap_duplicate() failed because entry
 750                          * has been freed independently, and will not be
 751                          * reused since sys_swapoff() already disabled
 752                          * allocation from here, or alloc_page() failed.
 753                          */
 754                         if (!*swap_map)
 755                                 continue;
 756                         retval = -ENOMEM;
 757                         break;
 758                 }
 759
 760                 /*
 761                  * Don't hold on to start_mm if it looks like exiting.
 762                  */
 763                 if (atomic_read(&start_mm->mm_users) == 1) {
 764                         mmput(start_mm);
 765                         start_mm = &init_mm;
 766                         atomic_inc(&init_mm.mm_users);
 767                 }
 768
 769                 /*
 770                  * Wait for and lock page.  When do_swap_page races with
 771                  * try_to_unuse, do_swap_page can handle the fault much
 772                  * faster than try_to_unuse can locate the entry.  This
 773                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
 774                  * defer to do_swap_page in such a case - in some tests,
 775                  * do_swap_page and try_to_unuse repeatedly compete.
 776                  */
 777                 wait_on_page_locked(page);
 778                 wait_on_page_writeback(page);
 779                 lock_page(page);
 780                 wait_on_page_writeback(page);
 781
 782                 /*
 783                  * Remove all references to entry.
 784                  * Whenever we reach init_mm, there's no address space
 785                  * to search, but use it as a reminder to search shmem.
 786                  */
 787                 shmem = 0;
 788                 swcount = *swap_map;
 789                 if (swcount > 1) {
 790                         if (start_mm == &init_mm)
 791                                 shmem = shmem_unuse(entry, page);
 792                         else
 793                                 retval = unuse_mm(start_mm, entry, page);
 794                 }
 795                 if (*swap_map > 1) {
 796                         int set_start_mm = (*swap_map >= swcount);
 797                         struct list_head *p = &start_mm->mmlist;
 798                         struct mm_struct *new_start_mm = start_mm;
 799                         struct mm_struct *prev_mm = start_mm;
 800                         struct mm_struct *mm;
 801
 802                         atomic_inc(&new_start_mm->mm_users);
 803                         atomic_inc(&prev_mm->mm_users);
 804                         spin_lock(&mmlist_lock);
 805                         while (*swap_map > 1 && !retval && !shmem &&
 806                                         (p = p->next) != &start_mm->mmlist) {
 807                                 mm = list_entry(p, struct mm_struct, mmlist);
 808                                 if (!atomic_inc_not_zero(&mm->mm_users))
 809                                         continue;
 810                                 spin_unlock(&mmlist_lock);
 811                                 mmput(prev_mm);
 812                                 prev_mm = mm;
 813
 814                                 cond_resched();
 815
 816                                 swcount = *swap_map;
 817                                 if (swcount <= 1)
 818                                         ;
 819                                 else if (mm == &init_mm) {
 820                                         set_start_mm = 1;
 821                                         shmem = shmem_unuse(entry, page);
 822                                 } else
 823                                         retval = unuse_mm(mm, entry, page);
 824                                 if (set_start_mm && *swap_map < swcount) {
 825                                         mmput(new_start_mm);
 826                                         atomic_inc(&mm->mm_users);
 827                                         new_start_mm = mm;
 828                                         set_start_mm = 0;
 829                                 }
 830                                 spin_lock(&mmlist_lock);
 831                         }
 832                         spin_unlock(&mmlist_lock);
 833                         mmput(prev_mm);
 834                         mmput(start_mm);
 835                         start_mm = new_start_mm;
 836                 }
 837                 if (shmem) {
 838                         /* page has already been unlocked and released */
 839                         if (shmem > 0)
 840                                 continue;
 841                         retval = shmem;
 842                         break;
 843                 }
 844                 if (retval) {
 845                         unlock_page(page);
 846                         page_cache_release(page);
 847                         break;
 848                 }
 849
 850                 /*
 851                  * How could swap count reach 0x7fff when the maximum
 852                  * pid is 0x7fff, and there's no way to repeat a swap
 853                  * page within an mm (except in shmem, where it's the
 854                  * shared object which takes the reference count)?
 855                  * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
 856                  *
 857                  * If that's wrong, then we should worry more about
 858                  * exit_mmap() and do_munmap() cases described above:
 859                  * we might be resetting SWAP_MAP_MAX too early here.
 860                  * We know "Undead"s can happen, they're okay, so don't
 861                  * report them; but do report if we reset SWAP_MAP_MAX.
 862                  */
 863                 if (*swap_map == SWAP_MAP_MAX) {
 864                         spin_lock(&swap_lock);
 865                         *swap_map = 1;
 866                         spin_unlock(&swap_lock);
 867                         reset_overflow = 1;
 868                 }
 869
 870                 /*
 871                  * If a reference remains (rare), we would like to leave
 872                  * the page in the swap cache; but try_to_unmap could
 873                  * then re-duplicate the entry once we drop page lock,
 874                  * so we might loop indefinitely; also, that page could
 875                  * not be swapped out to other storage meanwhile.  So:
 876                  * delete from cache even if there's another reference,
 877                  * after ensuring that the data has been saved to disk -
 878                  * since if the reference remains (rarer), it will be
 879                  * read from disk into another page.  Splitting into two
 880                  * pages would be incorrect if swap supported "shared
 881                  * private" pages, but they are handled by tmpfs files.
 882                  */
 883                 if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
 884                         struct writeback_control wbc = {
 885                                 .sync_mode = WB_SYNC_NONE,
 886                         };
 887
 888                         swap_writepage(page, &wbc);
 889                         lock_page(page);
 890                         wait_on_page_writeback(page);
 891                 }
 892
 893                 /*
 894                  * It is conceivable that a racing task removed this page from
 895                  * swap cache just before we acquired the page lock at the top,
 896                  * or while we dropped it in unuse_mm().  The page might even
 897                  * be back in swap cache on another swap area: that we must not
 898                  * delete, since it may not have been written out to swap yet.
 899                  */
 900                 if (PageSwapCache(page) &&
 901                     likely(page_private(page) == entry.val))
 902                         delete_from_swap_cache(page);
 903
 904                 /*
 905                  * So we could skip searching mms once swap count went
 906                  * to 1, we did not mark any present ptes as dirty: must
 907                  * mark page dirty so shrink_page_list will preserve it.
 908                  */
 909                 SetPageDirty(page);
 910                 unlock_page(page);
 911                 page_cache_release(page);
 912
 913                 /*
 914                  * Make sure that we aren't completely killing
 915                  * interactive performance.
 916                  */
 917                 cond_resched();
 918         }
 919
 920         mmput(start_mm);
 921         if (reset_overflow) {
 922                 printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
 923                 swap_overflow = 0;
 924         }
 925         return retval;
 926 }
 927
 928 /*
 929  * After a successful try_to_unuse, if no swap is now in use, we know
 930  * we can empty the mmlist.  swap_lock must be held on entry and exit.
 931  * Note that mmlist_lock nests inside swap_lock, and an mm must be
 932  * added to the mmlist just after page_duplicate - before would be racy.
 933  */
 934 static void drain_mmlist(void)
 935 {
 936         struct list_head *p, *next;
 937         unsigned int i;
 938
 939         for (i = 0; i < nr_swapfiles; i++)
 940                 if (swap_info[i].inuse_pages)
 941                         return;
 942         spin_lock(&mmlist_lock);
 943         list_for_each_safe(p, next, &init_mm.mmlist)
 944                 list_del_init(p);
 945         spin_unlock(&mmlist_lock);
 946 }
 947
 948 /*
 949  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
 950  * corresponds to page offset `offset'.
 951  */
 952 sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
 953 {
 954         struct swap_extent *se = sis->curr_swap_extent;
 955         struct swap_extent *start_se = se;
 956
 957         for ( ; ; ) {
 958                 struct list_head *lh;
 959
 960                 if (se->start_page <= offset &&
 961                                 offset < (se->start_page + se->nr_pages)) {
 962                         return se->start_block + (offset - se->start_page);
 963                 }
 964                 lh = se->list.next;
 965                 if (lh == &sis->extent_list)
 966                         lh = lh->next;
 967                 se = list_entry(lh, struct swap_extent, list);
 968                 sis->curr_swap_extent = se;
 969                 BUG_ON(se == start_se);         /* It *must* be present */
 970         }
 971 }
 972
 973 #ifdef CONFIG_HIBERNATION
 974 /*
 975  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
 976  * corresponding to given index in swap_info (swap type).
 977  */
 978 sector_t swapdev_block(int swap_type, pgoff_t offset)
 979 {
 980         struct swap_info_struct *sis;
 981
 982         if (swap_type >= nr_swapfiles)
 983                 return 0;
 984
 985         sis = swap_info + swap_type;
 986         return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
 987 }
 988 #endif /* CONFIG_HIBERNATION */
 989
 990 /*
 991  * Free all of a swapdev's extent information
 992  */
 993 static void destroy_swap_extents(struct swap_info_struct *sis)
 994 {
 995         while (!list_empty(&sis->extent_list)) {
 996                 struct swap_extent *se;
 997
 998                 se = list_entry(sis->extent_list.next,
 999                                 struct swap_extent, list);
1000                 list_del(&se->list);
1001                 kfree(se);
1002         }
1003 }
1004
1005 /*
1006  * Add a block range (and the corresponding page range) into this swapdev's
1007  * extent list.  The extent list is kept sorted in page order.
1008  *
1009  * This function rather assumes that it is called in ascending page order.
1010  */
1011 static int
1012 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1013                 unsigned long nr_pages, sector_t start_block)
1014 {
1015         struct swap_extent *se;
1016         struct swap_extent *new_se;
1017         struct list_head *lh;
1018
1019         lh = sis->extent_list.prev;     /* The highest page extent */
1020         if (lh != &sis->extent_list) {
1021                 se = list_entry(lh, struct swap_extent, list);
1022                 BUG_ON(se->start_page + se->nr_pages != start_page);
1023                 if (se->start_block + se->nr_pages == start_block) {
1024                         /* Merge it */
1025                         se->nr_pages += nr_pages;
1026                         return 0;
1027                 }
1028         }
1029
1030         /*
1031          * No merge.  Insert a new extent, preserving ordering.
1032          */
1033         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1034         if (new_se == NULL)
1035                 return -ENOMEM;
1036         new_se->start_page = start_page;
1037         new_se->nr_pages = nr_pages;
1038         new_se->start_block = start_block;
1039
1040         list_add_tail(&new_se->list, &sis->extent_list);
1041         return 1;
1042 }
1043
1044 /*
1045  * A `swap extent' is a simple thing which maps a contiguous range of pages
1046  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1047  * is built at swapon time and is then used at swap_writepage/swap_readpage
1048  * time for locating where on disk a page belongs.
1049  *
1050  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1051  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1052  * swap files identically.
1053  *
1054  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1055  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1056  * swapfiles are handled *identically* after swapon time.
1057  *
1058  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1059  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1060  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1061  * requirements, they are simply tossed out - we will never use those blocks
1062  * for swapping.
1063  *
1064  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1065  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1066  * which will scribble on the fs.
1067  *
1068  * The amount of disk space which a single swap extent represents varies.
1069  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1070  * extents in the list.  To avoid much list walking, we cache the previous
1071  * search location in `curr_swap_extent', and start new searches from there.
1072  * This is extremely effective.  The average number of iterations in
1073  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1074  */
1075 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1076 {
1077         struct inode *inode;
1078         unsigned blocks_per_page;
1079         unsigned long page_no;
1080         unsigned blkbits;
1081         sector_t probe_block;
1082         sector_t last_block;
1083         sector_t lowest_block = -1;
1084         sector_t highest_block = 0;
1085         int nr_extents = 0;
1086         int ret;
1087
1088         inode = sis->swap_file->f_mapping->host;
1089         if (S_ISBLK(inode->i_mode)) {
1090                 ret = add_swap_extent(sis, 0, sis->max, 0);
1091                 *span = sis->pages;
1092                 goto done;
1093         }
1094
1095         blkbits = inode->i_blkbits;
1096         blocks_per_page = PAGE_SIZE >> blkbits;
1097
1098         /*
1099          * Map all the blocks into the extent list.  This code doesn't try
1100          * to be very smart.
1101          */
1102         probe_block = 0;
1103         page_no = 0;
1104         last_block = i_size_read(inode) >> blkbits;
1105         while ((probe_block + blocks_per_page) <= last_block &&
1106                         page_no < sis->max) {
1107                 unsigned block_in_page;
1108                 sector_t first_block;
1109
1110                 first_block = bmap(inode, probe_block);
1111                 if (first_block == 0)
1112                         goto bad_bmap;
1113
1114                 /*
1115                  * It must be PAGE_SIZE aligned on-disk
1116                  */
1117                 if (first_block & (blocks_per_page - 1)) {
1118                         probe_block++;
1119                         goto reprobe;
1120                 }
1121
1122                 for (block_in_page = 1; block_in_page < blocks_per_page;
1123                                         block_in_page++) {
1124                         sector_t block;
1125
1126                         block = bmap(inode, probe_block + block_in_page);
1127                         if (block == 0)
1128                                 goto bad_bmap;
1129                         if (block != first_block + block_in_page) {
1130                                 /* Discontiguity */
1131                                 probe_block++;
1132                                 goto reprobe;
1133                         }
1134                 }
1135
1136                 first_block >>= (PAGE_SHIFT - blkbits);
1137                 if (page_no) {  /* exclude the header page */
1138                         if (first_block < lowest_block)
1139                                 lowest_block = first_block;
1140                         if (first_block > highest_block)
1141                                 highest_block = first_block;
1142                 }
1143
1144                 /*
1145                  * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1146                  */
1147                 ret = add_swap_extent(sis, page_no, 1, first_block);
1148                 if (ret < 0)
1149                         goto out;
1150                 nr_extents += ret;
1151                 page_no++;
1152                 probe_block += blocks_per_page;
1153 reprobe:
1154                 continue;
1155         }
1156         ret = nr_extents;
1157         *span = 1 + highest_block - lowest_block;
1158         if (page_no == 0)
1159                 page_no = 1;    /* force Empty message */
1160         sis->max = page_no;
1161         sis->pages = page_no - 1;
1162         sis->highest_bit = page_no - 1;
1163 done:
1164         sis->curr_swap_extent = list_entry(sis->extent_list.prev,
1165                                         struct swap_extent, list);
1166         goto out;
1167 bad_bmap:
1168         printk(KERN_ERR "swapon: swapfile has holes\n");
1169         ret = -EINVAL;
1170 out:
1171         return ret;
1172 }
1173
1174 #if 0   /* We don't need this yet */
1175 #include <linux/backing-dev.h>
1176 int page_queue_congested(struct page *page)
1177 {
1178         struct backing_dev_info *bdi;
1179
1180         VM_BUG_ON(!PageLocked(page));   /* It pins the swap_info_struct */
1181
1182         if (PageSwapCache(page)) {
1183                 swp_entry_t entry = { .val = page_private(page) };
1184                 struct swap_info_struct *sis;
1185
1186                 sis = get_swap_info_struct(swp_type(entry));
1187                 bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
1188         } else
1189                 bdi = page->mapping->backing_dev_info;
1190         return bdi_write_congested(bdi);
1191 }
1192 #endif
1193
1194 asmlinkage long sys_swapoff(const char __user * specialfile)
1195 {
1196         struct swap_info_struct * p = NULL;
1197         unsigned short *swap_map;
1198         struct file *swap_file, *victim;
1199         struct address_space *mapping;
1200         struct inode *inode;
1201         char * pathname;
1202         int i, type, prev;
1203         int err;
1204
1205         if (!capable(CAP_SYS_ADMIN))
1206                 return -EPERM;
1207
1208         pathname = getname(specialfile);
1209         err = PTR_ERR(pathname);
1210         if (IS_ERR(pathname))
1211                 goto out;
1212
1213         victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1214         putname(pathname);
1215         err = PTR_ERR(victim);
1216         if (IS_ERR(victim))
1217                 goto out;
1218
1219         mapping = victim->f_mapping;
1220         prev = -1;
1221         spin_lock(&swap_lock);
1222         for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
1223                 p = swap_info + type;
1224                 if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
1225                         if (p->swap_file->f_mapping == mapping)
1226                                 break;
1227                 }
1228                 prev = type;
1229         }
1230         if (type < 0) {
1231                 err = -EINVAL;
1232                 spin_unlock(&swap_lock);
1233                 goto out_dput;
1234         }
1235         if (!security_vm_enough_memory(p->pages))
1236                 vm_unacct_memory(p->pages);
1237         else {
1238                 err = -ENOMEM;
1239                 spin_unlock(&swap_lock);
1240                 goto out_dput;
1241         }
1242         if (prev < 0) {
1243                 swap_list.head = p->next;
1244         } else {
1245                 swap_info[prev].next = p->next;
1246         }
1247         if (type == swap_list.next) {
1248                 /* just pick something that's safe... */
1249                 swap_list.next = swap_list.head;
1250         }
1251         if (p->prio < 0) {
1252                 for (i = p->next; i >= 0; i = swap_info[i].next)
1253                         swap_info[i].prio = p->prio--;
1254                 least_priority++;
1255         }
1256         nr_swap_pages -= p->pages;
1257         total_swap_pages -= p->pages;
1258         p->flags &= ~SWP_WRITEOK;
1259         spin_unlock(&swap_lock);
1260
1261         current->flags |= PF_SWAPOFF;
1262         err = try_to_unuse(type);
1263         current->flags &= ~PF_SWAPOFF;
1264
1265         if (err) {
1266                 /* re-insert swap space back into swap_list */
1267                 spin_lock(&swap_lock);
1268                 if (p->prio < 0)
1269                         p->prio = --least_priority;
1270                 prev = -1;
1271                 for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1272                         if (p->prio >= swap_info[i].prio)
1273                                 break;
1274                         prev = i;
1275                 }
1276                 p->next = i;
1277                 if (prev < 0)
1278                         swap_list.head = swap_list.next = p - swap_info;
1279                 else
1280                         swap_info[prev].next = p - swap_info;
1281                 nr_swap_pages += p->pages;
1282                 total_swap_pages += p->pages;
1283                 p->flags |= SWP_WRITEOK;
1284                 spin_unlock(&swap_lock);
1285                 goto out_dput;
1286         }
1287
1288         /* wait for any unplug function to finish */
1289         down_write(&swap_unplug_sem);
1290         up_write(&swap_unplug_sem);
1291
1292         destroy_swap_extents(p);
1293         mutex_lock(&swapon_mutex);
1294         spin_lock(&swap_lock);
1295         drain_mmlist();
1296
1297         /* wait for anyone still in scan_swap_map */
1298         p->highest_bit = 0;             /* cuts scans short */
1299         while (p->flags >= SWP_SCANNING) {
1300                 spin_unlock(&swap_lock);
1301                 schedule_timeout_uninterruptible(1);
1302                 spin_lock(&swap_lock);
1303         }
1304
1305         swap_file = p->swap_file;
1306         p->swap_file = NULL;
1307         p->max = 0;
1308         swap_map = p->swap_map;
1309         p->swap_map = NULL;
1310         p->flags = 0;
1311         spin_unlock(&swap_lock);
1312         mutex_unlock(&swapon_mutex);
1313         vfree(swap_map);
1314         inode = mapping->host;
1315         if (S_ISBLK(inode->i_mode)) {
1316                 struct block_device *bdev = I_BDEV(inode);
1317                 set_blocksize(bdev, p->old_block_size);
1318                 bd_release(bdev);
1319         } else {
1320                 mutex_lock(&inode->i_mutex);
1321                 inode->i_flags &= ~S_SWAPFILE;
1322                 mutex_unlock(&inode->i_mutex);
1323         }
1324         filp_close(swap_file, NULL);
1325         err = 0;
1326
1327 out_dput:
1328         filp_close(victim, NULL);
1329 out:
1330         return err;
1331 }
1332
1333 #ifdef CONFIG_PROC_FS
1334 /* iterator */
1335 static void *swap_start(struct seq_file *swap, loff_t *pos)
1336 {
1337         struct swap_info_struct *ptr = swap_info;
1338         int i;
1339         loff_t l = *pos;
1340
1341         mutex_lock(&swapon_mutex);
1342
1343         if (!l)
1344                 return SEQ_START_TOKEN;
1345
1346         for (i = 0; i < nr_swapfiles; i++, ptr++) {
1347                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1348                         continue;
1349                 if (!--l)
1350                         return ptr;
1351         }
1352
1353         return NULL;
1354 }
1355
1356 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1357 {
1358         struct swap_info_struct *ptr;
1359         struct swap_info_struct *endptr = swap_info + nr_swapfiles;
1360
1361         if (v == SEQ_START_TOKEN)
1362                 ptr = swap_info;
1363         else {
1364                 ptr = v;
1365                 ptr++;
1366         }
1367
1368         for (; ptr < endptr; ptr++) {
1369                 if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
1370                         continue;
1371                 ++*pos;
1372                 return ptr;
1373         }
1374
1375         return NULL;
1376 }
1377
1378 static void swap_stop(struct seq_file *swap, void *v)
1379 {
1380         mutex_unlock(&swapon_mutex);
1381 }
1382
1383 static int swap_show(struct seq_file *swap, void *v)
1384 {
1385         struct swap_info_struct *ptr = v;
1386         struct file *file;
1387         int len;
1388
1389         if (ptr == SEQ_START_TOKEN) {
1390                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1391                 return 0;
1392         }
1393
1394         file = ptr->swap_file;
1395         len = seq_path(swap, &file->f_path, " \t\n\\");
1396         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1397                        len < 40 ? 40 - len : 1, " ",
1398                        S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1399                                 "partition" : "file\t",
1400                        ptr->pages << (PAGE_SHIFT - 10),
1401                        ptr->inuse_pages << (PAGE_SHIFT - 10),
1402                        ptr->prio);
1403         return 0;
1404 }
1405
1406 static const struct seq_operations swaps_op = {
1407         .start =        swap_start,
1408         .next =         swap_next,
1409         .stop =         swap_stop,
1410         .show =         swap_show
1411 };
1412
1413 static int swaps_open(struct inode *inode, struct file *file)
1414 {
1415         return seq_open(file, &swaps_op);
1416 }
1417
1418 static const struct file_operations proc_swaps_operations = {
1419         .open           = swaps_open,
1420         .read           = seq_read,
1421         .llseek         = seq_lseek,
1422         .release        = seq_release,
1423 };
1424
1425 static int __init procswaps_init(void)
1426 {
1427         proc_create("swaps", 0, NULL, &proc_swaps_operations);
1428         return 0;
1429 }
1430 __initcall(procswaps_init);
1431 #endif /* CONFIG_PROC_FS */
1432
1433 #ifdef MAX_SWAPFILES_CHECK
1434 static int __init max_swapfiles_check(void)
1435 {
1436         MAX_SWAPFILES_CHECK();
1437         return 0;
1438 }
1439 late_initcall(max_swapfiles_check);
1440 #endif
1441
1442 /*
1443  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
1444  *
1445  * The swapon system call
1446  */
1447 asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
1448 {
1449         struct swap_info_struct * p;
1450         char *name = NULL;
1451         struct block_device *bdev = NULL;
1452         struct file *swap_file = NULL;
1453         struct address_space *mapping;
1454         unsigned int type;
1455         int i, prev;
1456         int error;
1457         union swap_header *swap_header = NULL;
1458         int swap_header_version;
1459         unsigned int nr_good_pages = 0;
1460         int nr_extents = 0;
1461         sector_t span;
1462         unsigned long maxpages = 1;
1463         int swapfilesize;
1464         unsigned short *swap_map = NULL;
1465         struct page *page = NULL;
1466         struct inode *inode = NULL;
1467         int did_down = 0;
1468
1469         if (!capable(CAP_SYS_ADMIN))
1470                 return -EPERM;
1471         spin_lock(&swap_lock);
1472         p = swap_info;
1473         for (type = 0 ; type < nr_swapfiles ; type++,p++)
1474                 if (!(p->flags & SWP_USED))
1475                         break;
1476         error = -EPERM;
1477         if (type >= MAX_SWAPFILES) {
1478                 spin_unlock(&swap_lock);
1479                 goto out;
1480         }
1481         if (type >= nr_swapfiles)
1482                 nr_swapfiles = type+1;
1483         memset(p, 0, sizeof(*p));
1484         INIT_LIST_HEAD(&p->extent_list);
1485         p->flags = SWP_USED;
1486         p->next = -1;
1487         spin_unlock(&swap_lock);
1488         name = getname(specialfile);
1489         error = PTR_ERR(name);
1490         if (IS_ERR(name)) {
1491                 name = NULL;
1492                 goto bad_swap_2;
1493         }
1494         swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
1495         error = PTR_ERR(swap_file);
1496         if (IS_ERR(swap_file)) {
1497                 swap_file = NULL;
1498                 goto bad_swap_2;
1499         }
1500
1501         p->swap_file = swap_file;
1502         mapping = swap_file->f_mapping;
1503         inode = mapping->host;
1504
1505         error = -EBUSY;
1506         for (i = 0; i < nr_swapfiles; i++) {
1507                 struct swap_info_struct *q = &swap_info[i];
1508
1509                 if (i == type || !q->swap_file)
1510                         continue;
1511                 if (mapping == q->swap_file->f_mapping)
1512                         goto bad_swap;
1513         }
1514
1515         error = -EINVAL;
1516         if (S_ISBLK(inode->i_mode)) {
1517                 bdev = I_BDEV(inode);
1518                 error = bd_claim(bdev, sys_swapon);
1519                 if (error < 0) {
1520                         bdev = NULL;
1521                         error = -EINVAL;
1522                         goto bad_swap;
1523                 }
1524                 p->old_block_size = block_size(bdev);
1525                 error = set_blocksize(bdev, PAGE_SIZE);
1526                 if (error < 0)
1527                         goto bad_swap;
1528                 p->bdev = bdev;
1529         } else if (S_ISREG(inode->i_mode)) {
1530                 p->bdev = inode->i_sb->s_bdev;
1531                 mutex_lock(&inode->i_mutex);
1532                 did_down = 1;
1533                 if (IS_SWAPFILE(inode)) {
1534                         error = -EBUSY;
1535                         goto bad_swap;
1536                 }
1537         } else {
1538                 goto bad_swap;
1539         }
1540
1541         swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
1542
1543         /*
1544          * Read the swap header.
1545          */
1546         if (!mapping->a_ops->readpage) {
1547                 error = -EINVAL;
1548                 goto bad_swap;
1549         }
1550         page = read_mapping_page(mapping, 0, swap_file);
1551         if (IS_ERR(page)) {
1552                 error = PTR_ERR(page);
1553                 goto bad_swap;
1554         }
1555         kmap(page);
1556         swap_header = page_address(page);
1557
1558         if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
1559                 swap_header_version = 1;
1560         else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
1561                 swap_header_version = 2;
1562         else {
1563                 printk(KERN_ERR "Unable to find swap-space signature\n");
1564                 error = -EINVAL;
1565                 goto bad_swap;
1566         }
1567
1568         switch (swap_header_version) {
1569         case 1:
1570                 printk(KERN_ERR "version 0 swap is no longer supported. "
1571                         "Use mkswap -v1 %s\n", name);
1572                 error = -EINVAL;
1573                 goto bad_swap;
1574         case 2:
1575                 /* swap partition endianess hack... */
1576                 if (swab32(swap_header->info.version) == 1) {
1577                         swab32s(&swap_header->info.version);
1578                         swab32s(&swap_header->info.last_page);
1579                         swab32s(&swap_header->info.nr_badpages);
1580                         for (i = 0; i < swap_header->info.nr_badpages; i++)
1581                                 swab32s(&swap_header->info.badpages[i]);
1582                 }
1583                 /* Check the swap header's sub-version and the size of
1584                    the swap file and bad block lists */
1585                 if (swap_header->info.version != 1) {
1586                         printk(KERN_WARNING
1587                                "Unable to handle swap header version %d\n",
1588                                swap_header->info.version);
1589                         error = -EINVAL;
1590                         goto bad_swap;
1591                 }
1592
1593                 p->lowest_bit  = 1;
1594                 p->cluster_next = 1;
1595
1596                 /*
1597                  * Find out how many pages are allowed for a single swap
1598                  * device. There are two limiting factors: 1) the number of
1599                  * bits for the swap offset in the swp_entry_t type and
1600                  * 2) the number of bits in the a swap pte as defined by
1601                  * the different architectures. In order to find the
1602                  * largest possible bit mask a swap entry with swap type 0
1603                  * and swap offset ~0UL is created, encoded to a swap pte,
1604                  * decoded to a swp_entry_t again and finally the swap
1605                  * offset is extracted. This will mask all the bits from
1606                  * the initial ~0UL mask that can't be encoded in either
1607                  * the swp_entry_t or the architecture definition of a
1608                  * swap pte.
1609                  */
1610                 maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
1611                 if (maxpages > swap_header->info.last_page)
1612                         maxpages = swap_header->info.last_page;
1613                 p->highest_bit = maxpages - 1;
1614
1615                 error = -EINVAL;
1616                 if (!maxpages)
1617                         goto bad_swap;
1618                 if (swapfilesize && maxpages > swapfilesize) {
1619                         printk(KERN_WARNING
1620                                "Swap area shorter than signature indicates\n");
1621                         goto bad_swap;
1622                 }
1623                 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1624                         goto bad_swap;
1625                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1626                         goto bad_swap;
1627
1628                 /* OK, set up the swap map and apply the bad block list */
1629                 swap_map = vmalloc(maxpages * sizeof(short));
1630                 if (!swap_map) {
1631                         error = -ENOMEM;
1632                         goto bad_swap;
1633                 }
1634
1635                 error = 0;
1636                 memset(swap_map, 0, maxpages * sizeof(short));
1637                 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1638                         int page_nr = swap_header->info.badpages[i];
1639                         if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
1640                                 error = -EINVAL;
1641                         else
1642                                 swap_map[page_nr] = SWAP_MAP_BAD;
1643                 }
1644                 nr_good_pages = swap_header->info.last_page -
1645                                 swap_header->info.nr_badpages -
1646                                 1 /* header page */;
1647                 if (error)
1648                         goto bad_swap;
1649         }
1650
1651         if (nr_good_pages) {
1652                 swap_map[0] = SWAP_MAP_BAD;
1653                 p->max = maxpages;
1654                 p->pages = nr_good_pages;
1655                 nr_extents = setup_swap_extents(p, &span);
1656                 if (nr_extents < 0) {
1657                         error = nr_extents;
1658                         goto bad_swap;
1659                 }
1660                 nr_good_pages = p->pages;
1661         }
1662         if (!nr_good_pages) {
1663                 printk(KERN_WARNING "Empty swap-file\n");
1664                 error = -EINVAL;
1665                 goto bad_swap;
1666         }
1667
1668         mutex_lock(&swapon_mutex);
1669         spin_lock(&swap_lock);
1670         if (swap_flags & SWAP_FLAG_PREFER)
1671                 p->prio =
1672                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
1673         else
1674                 p->prio = --least_priority;
1675         p->swap_map = swap_map;
1676         p->flags = SWP_ACTIVE;
1677         nr_swap_pages += nr_good_pages;
1678         total_swap_pages += nr_good_pages;
1679
1680         printk(KERN_INFO "Adding %uk swap on %s.  "
1681                         "Priority:%d extents:%d across:%lluk\n",
1682                 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
1683                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
1684
1685         /* insert swap space into swap_list: */
1686         prev = -1;
1687         for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
1688                 if (p->prio >= swap_info[i].prio) {
1689                         break;
1690                 }
1691                 prev = i;
1692         }
1693         p->next = i;
1694         if (prev < 0) {
1695                 swap_list.head = swap_list.next = p - swap_info;
1696         } else {
1697                 swap_info[prev].next = p - swap_info;
1698         }
1699         spin_unlock(&swap_lock);
1700         mutex_unlock(&swapon_mutex);
1701         error = 0;
1702         goto out;
1703 bad_swap:
1704         if (bdev) {
1705                 set_blocksize(bdev, p->old_block_size);
1706                 bd_release(bdev);
1707         }
1708         destroy_swap_extents(p);
1709 bad_swap_2:
1710         spin_lock(&swap_lock);
1711         p->swap_file = NULL;
1712         p->flags = 0;
1713         spin_unlock(&swap_lock);
1714         vfree(swap_map);
1715         if (swap_file)
1716                 filp_close(swap_file, NULL);
1717 out:
1718         if (page && !IS_ERR(page)) {
1719                 kunmap(page);
1720                 page_cache_release(page);
1721         }
1722         if (name)
1723                 putname(name);
1724         if (did_down) {
1725                 if (!error)
1726                         inode->i_flags |= S_SWAPFILE;
1727                 mutex_unlock(&inode->i_mutex);
1728         }
1729         return error;
1730 }
1731
1732 void si_swapinfo(struct sysinfo *val)
1733 {
1734         unsigned int i;
1735         unsigned long nr_to_be_unused = 0;
1736
1737         spin_lock(&swap_lock);
1738         for (i = 0; i < nr_swapfiles; i++) {
1739                 if (!(swap_info[i].flags & SWP_USED) ||
1740                      (swap_info[i].flags & SWP_WRITEOK))
1741                         continue;
1742                 nr_to_be_unused += swap_info[i].inuse_pages;
1743         }
1744         val->freeswap = nr_swap_pages + nr_to_be_unused;
1745         val->totalswap = total_swap_pages + nr_to_be_unused;
1746         spin_unlock(&swap_lock);
1747 }
1748
1749 /*
1750  * Verify that a swap entry is valid and increment its swap map count.
1751  *
1752  * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
1753  * "permanent", but will be reclaimed by the next swapoff.
1754  */
1755 int swap_duplicate(swp_entry_t entry)
1756 {
1757         struct swap_info_struct * p;
1758         unsigned long offset, type;
1759         int result = 0;
1760
1761         if (is_migration_entry(entry))
1762                 return 1;
1763
1764         type = swp_type(entry);
1765         if (type >= nr_swapfiles)
1766                 goto bad_file;
1767         p = type + swap_info;
1768         offset = swp_offset(entry);
1769
1770         spin_lock(&swap_lock);
1771         if (offset < p->max && p->swap_map[offset]) {
1772                 if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
1773                         p->swap_map[offset]++;
1774                         result = 1;
1775                 } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
1776                         if (swap_overflow++ < 5)
1777                                 printk(KERN_WARNING "swap_dup: swap entry overflow\n");
1778                         p->swap_map[offset] = SWAP_MAP_MAX;
1779                         result = 1;
1780                 }
1781         }
1782         spin_unlock(&swap_lock);
1783 out:
1784         return result;
1785
1786 bad_file:
1787         printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
1788         goto out;
1789 }
1790
1791 struct swap_info_struct *
1792 get_swap_info_struct(unsigned type)
1793 {
1794         return &swap_info[type];
1795 }
1796
1797 /*
1798  * swap_lock prevents swap_map being freed. Don't grab an extra
1799  * reference on the swaphandle, it doesn't matter if it becomes unused.
1800  */
1801 int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
1802 {
1803         struct swap_info_struct *si;
1804         int our_page_cluster = page_cluster;
1805         pgoff_t target, toff;
1806         pgoff_t base, end;
1807         int nr_pages = 0;
1808
1809         if (!our_page_cluster)  /* no readahead */
1810                 return 0;
1811
1812         si = &swap_info[swp_type(entry)];
1813         target = swp_offset(entry);
1814         base = (target >> our_page_cluster) << our_page_cluster;
1815         end = base + (1 << our_page_cluster);
1816         if (!base)              /* first page is swap header */
1817                 base++;
1818
1819         spin_lock(&swap_lock);
1820         if (end > si->max)      /* don't go beyond end of map */
1821                 end = si->max;
1822
1823         /* Count contiguous allocated slots above our target */
1824         for (toff = target; ++toff < end; nr_pages++) {
1825                 /* Don't read in free or bad pages */
1826                 if (!si->swap_map[toff])
1827                         break;
1828                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1829                         break;
1830         }
1831         /* Count contiguous allocated slots below our target */
1832         for (toff = target; --toff >= base; nr_pages++) {
1833                 /* Don't read in free or bad pages */
1834                 if (!si->swap_map[toff])
1835                         break;
1836                 if (si->swap_map[toff] == SWAP_MAP_BAD)
1837                         break;
1838         }
1839         spin_unlock(&swap_lock);
1840
1841         /*
1842          * Indicate starting offset, and return number of pages to get:
1843          * if only 1, say 0, since there's then no readahead to be done.
1844          */
1845         *offset = ++toff;
1846         return nr_pages? ++nr_pages: 0;
1847 }