swapfile.c revision 570a335b8e22579e2a51a68136d2b1f907a20eec
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/random.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#include <linux/memcontrol.h> 32 33#include <asm/pgtable.h> 34#include <asm/tlbflush.h> 35#include <linux/swapops.h> 36#include <linux/page_cgroup.h> 37 38static bool swap_count_continued(struct swap_info_struct *, pgoff_t, 39 unsigned char); 40static void free_swap_count_continuations(struct swap_info_struct *); 41 42static DEFINE_SPINLOCK(swap_lock); 43static unsigned int nr_swapfiles; 44long nr_swap_pages; 45long total_swap_pages; 46static int least_priority; 47 48static const char Bad_file[] = "Bad swap file entry "; 49static const char Unused_file[] = "Unused swap file entry "; 50static const char Bad_offset[] = "Bad swap offset entry "; 51static const char Unused_offset[] = "Unused swap offset entry "; 52 53static struct swap_list_t swap_list = {-1, -1}; 54 55static struct swap_info_struct *swap_info[MAX_SWAPFILES]; 56 57static DEFINE_MUTEX(swapon_mutex); 58 59static inline unsigned char swap_count(unsigned char ent) 60{ 61 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */ 62} 63 64/* returns 1 if swap entry is freed */ 65static int 66__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset) 67{ 68 swp_entry_t entry = swp_entry(si->type, offset); 69 struct page *page; 70 int ret = 0; 71 72 page = find_get_page(&swapper_space, entry.val); 73 if (!page) 74 return 0; 75 /* 76 * This function is called from scan_swap_map() and it's called 77 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here. 78 * We have to use trylock for avoiding deadlock. This is a special 79 * case and you should use try_to_free_swap() with explicit lock_page() 80 * in usual operations. 81 */ 82 if (trylock_page(page)) { 83 ret = try_to_free_swap(page); 84 unlock_page(page); 85 } 86 page_cache_release(page); 87 return ret; 88} 89 90/* 91 * We need this because the bdev->unplug_fn can sleep and we cannot 92 * hold swap_lock while calling the unplug_fn. And swap_lock 93 * cannot be turned into a mutex. 94 */ 95static DECLARE_RWSEM(swap_unplug_sem); 96 97void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page) 98{ 99 swp_entry_t entry; 100 101 down_read(&swap_unplug_sem); 102 entry.val = page_private(page); 103 if (PageSwapCache(page)) { 104 struct block_device *bdev = swap_info[swp_type(entry)]->bdev; 105 struct backing_dev_info *bdi; 106 107 /* 108 * If the page is removed from swapcache from under us (with a 109 * racy try_to_unuse/swapoff) we need an additional reference 110 * count to avoid reading garbage from page_private(page) above. 111 * If the WARN_ON triggers during a swapoff it maybe the race 112 * condition and it's harmless. However if it triggers without 113 * swapoff it signals a problem. 114 */ 115 WARN_ON(page_count(page) <= 1); 116 117 bdi = bdev->bd_inode->i_mapping->backing_dev_info; 118 blk_run_backing_dev(bdi, page); 119 } 120 up_read(&swap_unplug_sem); 121} 122 123/* 124 * swapon tell device that all the old swap contents can be discarded, 125 * to allow the swap device to optimize its wear-levelling. 126 */ 127static int discard_swap(struct swap_info_struct *si) 128{ 129 struct swap_extent *se; 130 sector_t start_block; 131 sector_t nr_blocks; 132 int err = 0; 133 134 /* Do not discard the swap header page! */ 135 se = &si->first_swap_extent; 136 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9); 137 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9); 138 if (nr_blocks) { 139 err = blkdev_issue_discard(si->bdev, start_block, 140 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER); 141 if (err) 142 return err; 143 cond_resched(); 144 } 145 146 list_for_each_entry(se, &si->first_swap_extent.list, list) { 147 start_block = se->start_block << (PAGE_SHIFT - 9); 148 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9); 149 150 err = blkdev_issue_discard(si->bdev, start_block, 151 nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER); 152 if (err) 153 break; 154 155 cond_resched(); 156 } 157 return err; /* That will often be -EOPNOTSUPP */ 158} 159 160/* 161 * swap allocation tell device that a cluster of swap can now be discarded, 162 * to allow the swap device to optimize its wear-levelling. 163 */ 164static void discard_swap_cluster(struct swap_info_struct *si, 165 pgoff_t start_page, pgoff_t nr_pages) 166{ 167 struct swap_extent *se = si->curr_swap_extent; 168 int found_extent = 0; 169 170 while (nr_pages) { 171 struct list_head *lh; 172 173 if (se->start_page <= start_page && 174 start_page < se->start_page + se->nr_pages) { 175 pgoff_t offset = start_page - se->start_page; 176 sector_t start_block = se->start_block + offset; 177 sector_t nr_blocks = se->nr_pages - offset; 178 179 if (nr_blocks > nr_pages) 180 nr_blocks = nr_pages; 181 start_page += nr_blocks; 182 nr_pages -= nr_blocks; 183 184 if (!found_extent++) 185 si->curr_swap_extent = se; 186 187 start_block <<= PAGE_SHIFT - 9; 188 nr_blocks <<= PAGE_SHIFT - 9; 189 if (blkdev_issue_discard(si->bdev, start_block, 190 nr_blocks, GFP_NOIO, DISCARD_FL_BARRIER)) 191 break; 192 } 193 194 lh = se->list.next; 195 se = list_entry(lh, struct swap_extent, list); 196 } 197} 198 199static int wait_for_discard(void *word) 200{ 201 schedule(); 202 return 0; 203} 204 205#define SWAPFILE_CLUSTER 256 206#define LATENCY_LIMIT 256 207 208static inline unsigned long scan_swap_map(struct swap_info_struct *si, 209 unsigned char usage) 210{ 211 unsigned long offset; 212 unsigned long scan_base; 213 unsigned long last_in_cluster = 0; 214 int latency_ration = LATENCY_LIMIT; 215 int found_free_cluster = 0; 216 217 /* 218 * We try to cluster swap pages by allocating them sequentially 219 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this 220 * way, however, we resort to first-free allocation, starting 221 * a new cluster. This prevents us from scattering swap pages 222 * all over the entire swap partition, so that we reduce 223 * overall disk seek times between swap pages. -- sct 224 * But we do now try to find an empty cluster. -Andrea 225 * And we let swap pages go all over an SSD partition. Hugh 226 */ 227 228 si->flags += SWP_SCANNING; 229 scan_base = offset = si->cluster_next; 230 231 if (unlikely(!si->cluster_nr--)) { 232 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) { 233 si->cluster_nr = SWAPFILE_CLUSTER - 1; 234 goto checks; 235 } 236 if (si->flags & SWP_DISCARDABLE) { 237 /* 238 * Start range check on racing allocations, in case 239 * they overlap the cluster we eventually decide on 240 * (we scan without swap_lock to allow preemption). 241 * It's hardly conceivable that cluster_nr could be 242 * wrapped during our scan, but don't depend on it. 243 */ 244 if (si->lowest_alloc) 245 goto checks; 246 si->lowest_alloc = si->max; 247 si->highest_alloc = 0; 248 } 249 spin_unlock(&swap_lock); 250 251 /* 252 * If seek is expensive, start searching for new cluster from 253 * start of partition, to minimize the span of allocated swap. 254 * But if seek is cheap, search from our current position, so 255 * that swap is allocated from all over the partition: if the 256 * Flash Translation Layer only remaps within limited zones, 257 * we don't want to wear out the first zone too quickly. 258 */ 259 if (!(si->flags & SWP_SOLIDSTATE)) 260 scan_base = offset = si->lowest_bit; 261 last_in_cluster = offset + SWAPFILE_CLUSTER - 1; 262 263 /* Locate the first empty (unaligned) cluster */ 264 for (; last_in_cluster <= si->highest_bit; offset++) { 265 if (si->swap_map[offset]) 266 last_in_cluster = offset + SWAPFILE_CLUSTER; 267 else if (offset == last_in_cluster) { 268 spin_lock(&swap_lock); 269 offset -= SWAPFILE_CLUSTER - 1; 270 si->cluster_next = offset; 271 si->cluster_nr = SWAPFILE_CLUSTER - 1; 272 found_free_cluster = 1; 273 goto checks; 274 } 275 if (unlikely(--latency_ration < 0)) { 276 cond_resched(); 277 latency_ration = LATENCY_LIMIT; 278 } 279 } 280 281 offset = si->lowest_bit; 282 last_in_cluster = offset + SWAPFILE_CLUSTER - 1; 283 284 /* Locate the first empty (unaligned) cluster */ 285 for (; last_in_cluster < scan_base; offset++) { 286 if (si->swap_map[offset]) 287 last_in_cluster = offset + SWAPFILE_CLUSTER; 288 else if (offset == last_in_cluster) { 289 spin_lock(&swap_lock); 290 offset -= SWAPFILE_CLUSTER - 1; 291 si->cluster_next = offset; 292 si->cluster_nr = SWAPFILE_CLUSTER - 1; 293 found_free_cluster = 1; 294 goto checks; 295 } 296 if (unlikely(--latency_ration < 0)) { 297 cond_resched(); 298 latency_ration = LATENCY_LIMIT; 299 } 300 } 301 302 offset = scan_base; 303 spin_lock(&swap_lock); 304 si->cluster_nr = SWAPFILE_CLUSTER - 1; 305 si->lowest_alloc = 0; 306 } 307 308checks: 309 if (!(si->flags & SWP_WRITEOK)) 310 goto no_page; 311 if (!si->highest_bit) 312 goto no_page; 313 if (offset > si->highest_bit) 314 scan_base = offset = si->lowest_bit; 315 316 /* reuse swap entry of cache-only swap if not busy. */ 317 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) { 318 int swap_was_freed; 319 spin_unlock(&swap_lock); 320 swap_was_freed = __try_to_reclaim_swap(si, offset); 321 spin_lock(&swap_lock); 322 /* entry was freed successfully, try to use this again */ 323 if (swap_was_freed) 324 goto checks; 325 goto scan; /* check next one */ 326 } 327 328 if (si->swap_map[offset]) 329 goto scan; 330 331 if (offset == si->lowest_bit) 332 si->lowest_bit++; 333 if (offset == si->highest_bit) 334 si->highest_bit--; 335 si->inuse_pages++; 336 if (si->inuse_pages == si->pages) { 337 si->lowest_bit = si->max; 338 si->highest_bit = 0; 339 } 340 si->swap_map[offset] = usage; 341 si->cluster_next = offset + 1; 342 si->flags -= SWP_SCANNING; 343 344 if (si->lowest_alloc) { 345 /* 346 * Only set when SWP_DISCARDABLE, and there's a scan 347 * for a free cluster in progress or just completed. 348 */ 349 if (found_free_cluster) { 350 /* 351 * To optimize wear-levelling, discard the 352 * old data of the cluster, taking care not to 353 * discard any of its pages that have already 354 * been allocated by racing tasks (offset has 355 * already stepped over any at the beginning). 356 */ 357 if (offset < si->highest_alloc && 358 si->lowest_alloc <= last_in_cluster) 359 last_in_cluster = si->lowest_alloc - 1; 360 si->flags |= SWP_DISCARDING; 361 spin_unlock(&swap_lock); 362 363 if (offset < last_in_cluster) 364 discard_swap_cluster(si, offset, 365 last_in_cluster - offset + 1); 366 367 spin_lock(&swap_lock); 368 si->lowest_alloc = 0; 369 si->flags &= ~SWP_DISCARDING; 370 371 smp_mb(); /* wake_up_bit advises this */ 372 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING)); 373 374 } else if (si->flags & SWP_DISCARDING) { 375 /* 376 * Delay using pages allocated by racing tasks 377 * until the whole discard has been issued. We 378 * could defer that delay until swap_writepage, 379 * but it's easier to keep this self-contained. 380 */ 381 spin_unlock(&swap_lock); 382 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING), 383 wait_for_discard, TASK_UNINTERRUPTIBLE); 384 spin_lock(&swap_lock); 385 } else { 386 /* 387 * Note pages allocated by racing tasks while 388 * scan for a free cluster is in progress, so 389 * that its final discard can exclude them. 390 */ 391 if (offset < si->lowest_alloc) 392 si->lowest_alloc = offset; 393 if (offset > si->highest_alloc) 394 si->highest_alloc = offset; 395 } 396 } 397 return offset; 398 399scan: 400 spin_unlock(&swap_lock); 401 while (++offset <= si->highest_bit) { 402 if (!si->swap_map[offset]) { 403 spin_lock(&swap_lock); 404 goto checks; 405 } 406 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) { 407 spin_lock(&swap_lock); 408 goto checks; 409 } 410 if (unlikely(--latency_ration < 0)) { 411 cond_resched(); 412 latency_ration = LATENCY_LIMIT; 413 } 414 } 415 offset = si->lowest_bit; 416 while (++offset < scan_base) { 417 if (!si->swap_map[offset]) { 418 spin_lock(&swap_lock); 419 goto checks; 420 } 421 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) { 422 spin_lock(&swap_lock); 423 goto checks; 424 } 425 if (unlikely(--latency_ration < 0)) { 426 cond_resched(); 427 latency_ration = LATENCY_LIMIT; 428 } 429 } 430 spin_lock(&swap_lock); 431 432no_page: 433 si->flags -= SWP_SCANNING; 434 return 0; 435} 436 437swp_entry_t get_swap_page(void) 438{ 439 struct swap_info_struct *si; 440 pgoff_t offset; 441 int type, next; 442 int wrapped = 0; 443 444 spin_lock(&swap_lock); 445 if (nr_swap_pages <= 0) 446 goto noswap; 447 nr_swap_pages--; 448 449 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) { 450 si = swap_info[type]; 451 next = si->next; 452 if (next < 0 || 453 (!wrapped && si->prio != swap_info[next]->prio)) { 454 next = swap_list.head; 455 wrapped++; 456 } 457 458 if (!si->highest_bit) 459 continue; 460 if (!(si->flags & SWP_WRITEOK)) 461 continue; 462 463 swap_list.next = next; 464 /* This is called for allocating swap entry for cache */ 465 offset = scan_swap_map(si, SWAP_HAS_CACHE); 466 if (offset) { 467 spin_unlock(&swap_lock); 468 return swp_entry(type, offset); 469 } 470 next = swap_list.next; 471 } 472 473 nr_swap_pages++; 474noswap: 475 spin_unlock(&swap_lock); 476 return (swp_entry_t) {0}; 477} 478 479/* The only caller of this function is now susupend routine */ 480swp_entry_t get_swap_page_of_type(int type) 481{ 482 struct swap_info_struct *si; 483 pgoff_t offset; 484 485 spin_lock(&swap_lock); 486 si = swap_info[type]; 487 if (si && (si->flags & SWP_WRITEOK)) { 488 nr_swap_pages--; 489 /* This is called for allocating swap entry, not cache */ 490 offset = scan_swap_map(si, 1); 491 if (offset) { 492 spin_unlock(&swap_lock); 493 return swp_entry(type, offset); 494 } 495 nr_swap_pages++; 496 } 497 spin_unlock(&swap_lock); 498 return (swp_entry_t) {0}; 499} 500 501static struct swap_info_struct *swap_info_get(swp_entry_t entry) 502{ 503 struct swap_info_struct *p; 504 unsigned long offset, type; 505 506 if (!entry.val) 507 goto out; 508 type = swp_type(entry); 509 if (type >= nr_swapfiles) 510 goto bad_nofile; 511 p = swap_info[type]; 512 if (!(p->flags & SWP_USED)) 513 goto bad_device; 514 offset = swp_offset(entry); 515 if (offset >= p->max) 516 goto bad_offset; 517 if (!p->swap_map[offset]) 518 goto bad_free; 519 spin_lock(&swap_lock); 520 return p; 521 522bad_free: 523 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val); 524 goto out; 525bad_offset: 526 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val); 527 goto out; 528bad_device: 529 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val); 530 goto out; 531bad_nofile: 532 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val); 533out: 534 return NULL; 535} 536 537static unsigned char swap_entry_free(struct swap_info_struct *p, 538 swp_entry_t entry, unsigned char usage) 539{ 540 unsigned long offset = swp_offset(entry); 541 unsigned char count; 542 unsigned char has_cache; 543 544 count = p->swap_map[offset]; 545 has_cache = count & SWAP_HAS_CACHE; 546 count &= ~SWAP_HAS_CACHE; 547 548 if (usage == SWAP_HAS_CACHE) { 549 VM_BUG_ON(!has_cache); 550 has_cache = 0; 551 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) { 552 if (count == COUNT_CONTINUED) { 553 if (swap_count_continued(p, offset, count)) 554 count = SWAP_MAP_MAX | COUNT_CONTINUED; 555 else 556 count = SWAP_MAP_MAX; 557 } else 558 count--; 559 } 560 561 if (!count) 562 mem_cgroup_uncharge_swap(entry); 563 564 usage = count | has_cache; 565 p->swap_map[offset] = usage; 566 567 /* free if no reference */ 568 if (!usage) { 569 if (offset < p->lowest_bit) 570 p->lowest_bit = offset; 571 if (offset > p->highest_bit) 572 p->highest_bit = offset; 573 if (swap_list.next >= 0 && 574 p->prio > swap_info[swap_list.next]->prio) 575 swap_list.next = p->type; 576 nr_swap_pages++; 577 p->inuse_pages--; 578 } 579 580 return usage; 581} 582 583/* 584 * Caller has made sure that the swapdevice corresponding to entry 585 * is still around or has not been recycled. 586 */ 587void swap_free(swp_entry_t entry) 588{ 589 struct swap_info_struct *p; 590 591 p = swap_info_get(entry); 592 if (p) { 593 swap_entry_free(p, entry, 1); 594 spin_unlock(&swap_lock); 595 } 596} 597 598/* 599 * Called after dropping swapcache to decrease refcnt to swap entries. 600 */ 601void swapcache_free(swp_entry_t entry, struct page *page) 602{ 603 struct swap_info_struct *p; 604 unsigned char count; 605 606 p = swap_info_get(entry); 607 if (p) { 608 count = swap_entry_free(p, entry, SWAP_HAS_CACHE); 609 if (page) 610 mem_cgroup_uncharge_swapcache(page, entry, count != 0); 611 spin_unlock(&swap_lock); 612 } 613} 614 615/* 616 * How many references to page are currently swapped out? 617 * This does not give an exact answer when swap count is continued, 618 * but does include the high COUNT_CONTINUED flag to allow for that. 619 */ 620static inline int page_swapcount(struct page *page) 621{ 622 int count = 0; 623 struct swap_info_struct *p; 624 swp_entry_t entry; 625 626 entry.val = page_private(page); 627 p = swap_info_get(entry); 628 if (p) { 629 count = swap_count(p->swap_map[swp_offset(entry)]); 630 spin_unlock(&swap_lock); 631 } 632 return count; 633} 634 635/* 636 * We can write to an anon page without COW if there are no other references 637 * to it. And as a side-effect, free up its swap: because the old content 638 * on disk will never be read, and seeking back there to write new content 639 * later would only waste time away from clustering. 640 */ 641int reuse_swap_page(struct page *page) 642{ 643 int count; 644 645 VM_BUG_ON(!PageLocked(page)); 646 count = page_mapcount(page); 647 if (count <= 1 && PageSwapCache(page)) { 648 count += page_swapcount(page); 649 if (count == 1 && !PageWriteback(page)) { 650 delete_from_swap_cache(page); 651 SetPageDirty(page); 652 } 653 } 654 return count == 1; 655} 656 657/* 658 * If swap is getting full, or if there are no more mappings of this page, 659 * then try_to_free_swap is called to free its swap space. 660 */ 661int try_to_free_swap(struct page *page) 662{ 663 VM_BUG_ON(!PageLocked(page)); 664 665 if (!PageSwapCache(page)) 666 return 0; 667 if (PageWriteback(page)) 668 return 0; 669 if (page_swapcount(page)) 670 return 0; 671 672 delete_from_swap_cache(page); 673 SetPageDirty(page); 674 return 1; 675} 676 677/* 678 * Free the swap entry like above, but also try to 679 * free the page cache entry if it is the last user. 680 */ 681int free_swap_and_cache(swp_entry_t entry) 682{ 683 struct swap_info_struct *p; 684 struct page *page = NULL; 685 686 if (non_swap_entry(entry)) 687 return 1; 688 689 p = swap_info_get(entry); 690 if (p) { 691 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) { 692 page = find_get_page(&swapper_space, entry.val); 693 if (page && !trylock_page(page)) { 694 page_cache_release(page); 695 page = NULL; 696 } 697 } 698 spin_unlock(&swap_lock); 699 } 700 if (page) { 701 /* 702 * Not mapped elsewhere, or swap space full? Free it! 703 * Also recheck PageSwapCache now page is locked (above). 704 */ 705 if (PageSwapCache(page) && !PageWriteback(page) && 706 (!page_mapped(page) || vm_swap_full())) { 707 delete_from_swap_cache(page); 708 SetPageDirty(page); 709 } 710 unlock_page(page); 711 page_cache_release(page); 712 } 713 return p != NULL; 714} 715 716#ifdef CONFIG_HIBERNATION 717/* 718 * Find the swap type that corresponds to given device (if any). 719 * 720 * @offset - number of the PAGE_SIZE-sized block of the device, starting 721 * from 0, in which the swap header is expected to be located. 722 * 723 * This is needed for the suspend to disk (aka swsusp). 724 */ 725int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p) 726{ 727 struct block_device *bdev = NULL; 728 int type; 729 730 if (device) 731 bdev = bdget(device); 732 733 spin_lock(&swap_lock); 734 for (type = 0; type < nr_swapfiles; type++) { 735 struct swap_info_struct *sis = swap_info[type]; 736 737 if (!(sis->flags & SWP_WRITEOK)) 738 continue; 739 740 if (!bdev) { 741 if (bdev_p) 742 *bdev_p = bdgrab(sis->bdev); 743 744 spin_unlock(&swap_lock); 745 return type; 746 } 747 if (bdev == sis->bdev) { 748 struct swap_extent *se = &sis->first_swap_extent; 749 750 if (se->start_block == offset) { 751 if (bdev_p) 752 *bdev_p = bdgrab(sis->bdev); 753 754 spin_unlock(&swap_lock); 755 bdput(bdev); 756 return type; 757 } 758 } 759 } 760 spin_unlock(&swap_lock); 761 if (bdev) 762 bdput(bdev); 763 764 return -ENODEV; 765} 766 767/* 768 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev 769 * corresponding to given index in swap_info (swap type). 770 */ 771sector_t swapdev_block(int type, pgoff_t offset) 772{ 773 struct block_device *bdev; 774 775 if ((unsigned int)type >= nr_swapfiles) 776 return 0; 777 if (!(swap_info[type]->flags & SWP_WRITEOK)) 778 return 0; 779 return map_swap_page(swp_entry(type, offset), &bdev); 780} 781 782/* 783 * Return either the total number of swap pages of given type, or the number 784 * of free pages of that type (depending on @free) 785 * 786 * This is needed for software suspend 787 */ 788unsigned int count_swap_pages(int type, int free) 789{ 790 unsigned int n = 0; 791 792 spin_lock(&swap_lock); 793 if ((unsigned int)type < nr_swapfiles) { 794 struct swap_info_struct *sis = swap_info[type]; 795 796 if (sis->flags & SWP_WRITEOK) { 797 n = sis->pages; 798 if (free) 799 n -= sis->inuse_pages; 800 } 801 } 802 spin_unlock(&swap_lock); 803 return n; 804} 805#endif /* CONFIG_HIBERNATION */ 806 807/* 808 * No need to decide whether this PTE shares the swap entry with others, 809 * just let do_wp_page work it out if a write is requested later - to 810 * force COW, vm_page_prot omits write permission from any private vma. 811 */ 812static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd, 813 unsigned long addr, swp_entry_t entry, struct page *page) 814{ 815 struct mem_cgroup *ptr = NULL; 816 spinlock_t *ptl; 817 pte_t *pte; 818 int ret = 1; 819 820 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) { 821 ret = -ENOMEM; 822 goto out_nolock; 823 } 824 825 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 826 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) { 827 if (ret > 0) 828 mem_cgroup_cancel_charge_swapin(ptr); 829 ret = 0; 830 goto out; 831 } 832 833 inc_mm_counter(vma->vm_mm, anon_rss); 834 get_page(page); 835 set_pte_at(vma->vm_mm, addr, pte, 836 pte_mkold(mk_pte(page, vma->vm_page_prot))); 837 page_add_anon_rmap(page, vma, addr); 838 mem_cgroup_commit_charge_swapin(page, ptr); 839 swap_free(entry); 840 /* 841 * Move the page to the active list so it is not 842 * immediately swapped out again after swapon. 843 */ 844 activate_page(page); 845out: 846 pte_unmap_unlock(pte, ptl); 847out_nolock: 848 return ret; 849} 850 851static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd, 852 unsigned long addr, unsigned long end, 853 swp_entry_t entry, struct page *page) 854{ 855 pte_t swp_pte = swp_entry_to_pte(entry); 856 pte_t *pte; 857 int ret = 0; 858 859 /* 860 * We don't actually need pte lock while scanning for swp_pte: since 861 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the 862 * page table while we're scanning; though it could get zapped, and on 863 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse 864 * of unmatched parts which look like swp_pte, so unuse_pte must 865 * recheck under pte lock. Scanning without pte lock lets it be 866 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE. 867 */ 868 pte = pte_offset_map(pmd, addr); 869 do { 870 /* 871 * swapoff spends a _lot_ of time in this loop! 872 * Test inline before going to call unuse_pte. 873 */ 874 if (unlikely(pte_same(*pte, swp_pte))) { 875 pte_unmap(pte); 876 ret = unuse_pte(vma, pmd, addr, entry, page); 877 if (ret) 878 goto out; 879 pte = pte_offset_map(pmd, addr); 880 } 881 } while (pte++, addr += PAGE_SIZE, addr != end); 882 pte_unmap(pte - 1); 883out: 884 return ret; 885} 886 887static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud, 888 unsigned long addr, unsigned long end, 889 swp_entry_t entry, struct page *page) 890{ 891 pmd_t *pmd; 892 unsigned long next; 893 int ret; 894 895 pmd = pmd_offset(pud, addr); 896 do { 897 next = pmd_addr_end(addr, end); 898 if (pmd_none_or_clear_bad(pmd)) 899 continue; 900 ret = unuse_pte_range(vma, pmd, addr, next, entry, page); 901 if (ret) 902 return ret; 903 } while (pmd++, addr = next, addr != end); 904 return 0; 905} 906 907static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd, 908 unsigned long addr, unsigned long end, 909 swp_entry_t entry, struct page *page) 910{ 911 pud_t *pud; 912 unsigned long next; 913 int ret; 914 915 pud = pud_offset(pgd, addr); 916 do { 917 next = pud_addr_end(addr, end); 918 if (pud_none_or_clear_bad(pud)) 919 continue; 920 ret = unuse_pmd_range(vma, pud, addr, next, entry, page); 921 if (ret) 922 return ret; 923 } while (pud++, addr = next, addr != end); 924 return 0; 925} 926 927static int unuse_vma(struct vm_area_struct *vma, 928 swp_entry_t entry, struct page *page) 929{ 930 pgd_t *pgd; 931 unsigned long addr, end, next; 932 int ret; 933 934 if (page->mapping) { 935 addr = page_address_in_vma(page, vma); 936 if (addr == -EFAULT) 937 return 0; 938 else 939 end = addr + PAGE_SIZE; 940 } else { 941 addr = vma->vm_start; 942 end = vma->vm_end; 943 } 944 945 pgd = pgd_offset(vma->vm_mm, addr); 946 do { 947 next = pgd_addr_end(addr, end); 948 if (pgd_none_or_clear_bad(pgd)) 949 continue; 950 ret = unuse_pud_range(vma, pgd, addr, next, entry, page); 951 if (ret) 952 return ret; 953 } while (pgd++, addr = next, addr != end); 954 return 0; 955} 956 957static int unuse_mm(struct mm_struct *mm, 958 swp_entry_t entry, struct page *page) 959{ 960 struct vm_area_struct *vma; 961 int ret = 0; 962 963 if (!down_read_trylock(&mm->mmap_sem)) { 964 /* 965 * Activate page so shrink_inactive_list is unlikely to unmap 966 * its ptes while lock is dropped, so swapoff can make progress. 967 */ 968 activate_page(page); 969 unlock_page(page); 970 down_read(&mm->mmap_sem); 971 lock_page(page); 972 } 973 for (vma = mm->mmap; vma; vma = vma->vm_next) { 974 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page))) 975 break; 976 } 977 up_read(&mm->mmap_sem); 978 return (ret < 0)? ret: 0; 979} 980 981/* 982 * Scan swap_map from current position to next entry still in use. 983 * Recycle to start on reaching the end, returning 0 when empty. 984 */ 985static unsigned int find_next_to_unuse(struct swap_info_struct *si, 986 unsigned int prev) 987{ 988 unsigned int max = si->max; 989 unsigned int i = prev; 990 unsigned char count; 991 992 /* 993 * No need for swap_lock here: we're just looking 994 * for whether an entry is in use, not modifying it; false 995 * hits are okay, and sys_swapoff() has already prevented new 996 * allocations from this area (while holding swap_lock). 997 */ 998 for (;;) { 999 if (++i >= max) { 1000 if (!prev) { 1001 i = 0; 1002 break; 1003 } 1004 /* 1005 * No entries in use at top of swap_map, 1006 * loop back to start and recheck there. 1007 */ 1008 max = prev + 1; 1009 prev = 0; 1010 i = 1; 1011 } 1012 count = si->swap_map[i]; 1013 if (count && swap_count(count) != SWAP_MAP_BAD) 1014 break; 1015 } 1016 return i; 1017} 1018 1019/* 1020 * We completely avoid races by reading each swap page in advance, 1021 * and then search for the process using it. All the necessary 1022 * page table adjustments can then be made atomically. 1023 */ 1024static int try_to_unuse(unsigned int type) 1025{ 1026 struct swap_info_struct *si = swap_info[type]; 1027 struct mm_struct *start_mm; 1028 unsigned char *swap_map; 1029 unsigned char swcount; 1030 struct page *page; 1031 swp_entry_t entry; 1032 unsigned int i = 0; 1033 int retval = 0; 1034 int shmem; 1035 1036 /* 1037 * When searching mms for an entry, a good strategy is to 1038 * start at the first mm we freed the previous entry from 1039 * (though actually we don't notice whether we or coincidence 1040 * freed the entry). Initialize this start_mm with a hold. 1041 * 1042 * A simpler strategy would be to start at the last mm we 1043 * freed the previous entry from; but that would take less 1044 * advantage of mmlist ordering, which clusters forked mms 1045 * together, child after parent. If we race with dup_mmap(), we 1046 * prefer to resolve parent before child, lest we miss entries 1047 * duplicated after we scanned child: using last mm would invert 1048 * that. 1049 */ 1050 start_mm = &init_mm; 1051 atomic_inc(&init_mm.mm_users); 1052 1053 /* 1054 * Keep on scanning until all entries have gone. Usually, 1055 * one pass through swap_map is enough, but not necessarily: 1056 * there are races when an instance of an entry might be missed. 1057 */ 1058 while ((i = find_next_to_unuse(si, i)) != 0) { 1059 if (signal_pending(current)) { 1060 retval = -EINTR; 1061 break; 1062 } 1063 1064 /* 1065 * Get a page for the entry, using the existing swap 1066 * cache page if there is one. Otherwise, get a clean 1067 * page and read the swap into it. 1068 */ 1069 swap_map = &si->swap_map[i]; 1070 entry = swp_entry(type, i); 1071 page = read_swap_cache_async(entry, 1072 GFP_HIGHUSER_MOVABLE, NULL, 0); 1073 if (!page) { 1074 /* 1075 * Either swap_duplicate() failed because entry 1076 * has been freed independently, and will not be 1077 * reused since sys_swapoff() already disabled 1078 * allocation from here, or alloc_page() failed. 1079 */ 1080 if (!*swap_map) 1081 continue; 1082 retval = -ENOMEM; 1083 break; 1084 } 1085 1086 /* 1087 * Don't hold on to start_mm if it looks like exiting. 1088 */ 1089 if (atomic_read(&start_mm->mm_users) == 1) { 1090 mmput(start_mm); 1091 start_mm = &init_mm; 1092 atomic_inc(&init_mm.mm_users); 1093 } 1094 1095 /* 1096 * Wait for and lock page. When do_swap_page races with 1097 * try_to_unuse, do_swap_page can handle the fault much 1098 * faster than try_to_unuse can locate the entry. This 1099 * apparently redundant "wait_on_page_locked" lets try_to_unuse 1100 * defer to do_swap_page in such a case - in some tests, 1101 * do_swap_page and try_to_unuse repeatedly compete. 1102 */ 1103 wait_on_page_locked(page); 1104 wait_on_page_writeback(page); 1105 lock_page(page); 1106 wait_on_page_writeback(page); 1107 1108 /* 1109 * Remove all references to entry. 1110 * Whenever we reach init_mm, there's no address space 1111 * to search, but use it as a reminder to search shmem. 1112 */ 1113 shmem = 0; 1114 swcount = *swap_map; 1115 if (swap_count(swcount)) { 1116 if (start_mm == &init_mm) 1117 shmem = shmem_unuse(entry, page); 1118 else 1119 retval = unuse_mm(start_mm, entry, page); 1120 } 1121 if (swap_count(*swap_map)) { 1122 int set_start_mm = (*swap_map >= swcount); 1123 struct list_head *p = &start_mm->mmlist; 1124 struct mm_struct *new_start_mm = start_mm; 1125 struct mm_struct *prev_mm = start_mm; 1126 struct mm_struct *mm; 1127 1128 atomic_inc(&new_start_mm->mm_users); 1129 atomic_inc(&prev_mm->mm_users); 1130 spin_lock(&mmlist_lock); 1131 while (swap_count(*swap_map) && !retval && !shmem && 1132 (p = p->next) != &start_mm->mmlist) { 1133 mm = list_entry(p, struct mm_struct, mmlist); 1134 if (!atomic_inc_not_zero(&mm->mm_users)) 1135 continue; 1136 spin_unlock(&mmlist_lock); 1137 mmput(prev_mm); 1138 prev_mm = mm; 1139 1140 cond_resched(); 1141 1142 swcount = *swap_map; 1143 if (!swap_count(swcount)) /* any usage ? */ 1144 ; 1145 else if (mm == &init_mm) { 1146 set_start_mm = 1; 1147 shmem = shmem_unuse(entry, page); 1148 } else 1149 retval = unuse_mm(mm, entry, page); 1150 1151 if (set_start_mm && *swap_map < swcount) { 1152 mmput(new_start_mm); 1153 atomic_inc(&mm->mm_users); 1154 new_start_mm = mm; 1155 set_start_mm = 0; 1156 } 1157 spin_lock(&mmlist_lock); 1158 } 1159 spin_unlock(&mmlist_lock); 1160 mmput(prev_mm); 1161 mmput(start_mm); 1162 start_mm = new_start_mm; 1163 } 1164 if (shmem) { 1165 /* page has already been unlocked and released */ 1166 if (shmem > 0) 1167 continue; 1168 retval = shmem; 1169 break; 1170 } 1171 if (retval) { 1172 unlock_page(page); 1173 page_cache_release(page); 1174 break; 1175 } 1176 1177 /* 1178 * If a reference remains (rare), we would like to leave 1179 * the page in the swap cache; but try_to_unmap could 1180 * then re-duplicate the entry once we drop page lock, 1181 * so we might loop indefinitely; also, that page could 1182 * not be swapped out to other storage meanwhile. So: 1183 * delete from cache even if there's another reference, 1184 * after ensuring that the data has been saved to disk - 1185 * since if the reference remains (rarer), it will be 1186 * read from disk into another page. Splitting into two 1187 * pages would be incorrect if swap supported "shared 1188 * private" pages, but they are handled by tmpfs files. 1189 */ 1190 if (swap_count(*swap_map) && 1191 PageDirty(page) && PageSwapCache(page)) { 1192 struct writeback_control wbc = { 1193 .sync_mode = WB_SYNC_NONE, 1194 }; 1195 1196 swap_writepage(page, &wbc); 1197 lock_page(page); 1198 wait_on_page_writeback(page); 1199 } 1200 1201 /* 1202 * It is conceivable that a racing task removed this page from 1203 * swap cache just before we acquired the page lock at the top, 1204 * or while we dropped it in unuse_mm(). The page might even 1205 * be back in swap cache on another swap area: that we must not 1206 * delete, since it may not have been written out to swap yet. 1207 */ 1208 if (PageSwapCache(page) && 1209 likely(page_private(page) == entry.val)) 1210 delete_from_swap_cache(page); 1211 1212 /* 1213 * So we could skip searching mms once swap count went 1214 * to 1, we did not mark any present ptes as dirty: must 1215 * mark page dirty so shrink_page_list will preserve it. 1216 */ 1217 SetPageDirty(page); 1218 unlock_page(page); 1219 page_cache_release(page); 1220 1221 /* 1222 * Make sure that we aren't completely killing 1223 * interactive performance. 1224 */ 1225 cond_resched(); 1226 } 1227 1228 mmput(start_mm); 1229 return retval; 1230} 1231 1232/* 1233 * After a successful try_to_unuse, if no swap is now in use, we know 1234 * we can empty the mmlist. swap_lock must be held on entry and exit. 1235 * Note that mmlist_lock nests inside swap_lock, and an mm must be 1236 * added to the mmlist just after page_duplicate - before would be racy. 1237 */ 1238static void drain_mmlist(void) 1239{ 1240 struct list_head *p, *next; 1241 unsigned int type; 1242 1243 for (type = 0; type < nr_swapfiles; type++) 1244 if (swap_info[type]->inuse_pages) 1245 return; 1246 spin_lock(&mmlist_lock); 1247 list_for_each_safe(p, next, &init_mm.mmlist) 1248 list_del_init(p); 1249 spin_unlock(&mmlist_lock); 1250} 1251 1252/* 1253 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which 1254 * corresponds to page offset `offset'. Note that the type of this function 1255 * is sector_t, but it returns page offset into the bdev, not sector offset. 1256 */ 1257sector_t map_swap_page(swp_entry_t entry, struct block_device **bdev) 1258{ 1259 struct swap_info_struct *sis; 1260 struct swap_extent *start_se; 1261 struct swap_extent *se; 1262 pgoff_t offset; 1263 1264 sis = swap_info[swp_type(entry)]; 1265 *bdev = sis->bdev; 1266 1267 offset = swp_offset(entry); 1268 start_se = sis->curr_swap_extent; 1269 se = start_se; 1270 1271 for ( ; ; ) { 1272 struct list_head *lh; 1273 1274 if (se->start_page <= offset && 1275 offset < (se->start_page + se->nr_pages)) { 1276 return se->start_block + (offset - se->start_page); 1277 } 1278 lh = se->list.next; 1279 se = list_entry(lh, struct swap_extent, list); 1280 sis->curr_swap_extent = se; 1281 BUG_ON(se == start_se); /* It *must* be present */ 1282 } 1283} 1284 1285/* 1286 * Free all of a swapdev's extent information 1287 */ 1288static void destroy_swap_extents(struct swap_info_struct *sis) 1289{ 1290 while (!list_empty(&sis->first_swap_extent.list)) { 1291 struct swap_extent *se; 1292 1293 se = list_entry(sis->first_swap_extent.list.next, 1294 struct swap_extent, list); 1295 list_del(&se->list); 1296 kfree(se); 1297 } 1298} 1299 1300/* 1301 * Add a block range (and the corresponding page range) into this swapdev's 1302 * extent list. The extent list is kept sorted in page order. 1303 * 1304 * This function rather assumes that it is called in ascending page order. 1305 */ 1306static int 1307add_swap_extent(struct swap_info_struct *sis, unsigned long start_page, 1308 unsigned long nr_pages, sector_t start_block) 1309{ 1310 struct swap_extent *se; 1311 struct swap_extent *new_se; 1312 struct list_head *lh; 1313 1314 if (start_page == 0) { 1315 se = &sis->first_swap_extent; 1316 sis->curr_swap_extent = se; 1317 se->start_page = 0; 1318 se->nr_pages = nr_pages; 1319 se->start_block = start_block; 1320 return 1; 1321 } else { 1322 lh = sis->first_swap_extent.list.prev; /* Highest extent */ 1323 se = list_entry(lh, struct swap_extent, list); 1324 BUG_ON(se->start_page + se->nr_pages != start_page); 1325 if (se->start_block + se->nr_pages == start_block) { 1326 /* Merge it */ 1327 se->nr_pages += nr_pages; 1328 return 0; 1329 } 1330 } 1331 1332 /* 1333 * No merge. Insert a new extent, preserving ordering. 1334 */ 1335 new_se = kmalloc(sizeof(*se), GFP_KERNEL); 1336 if (new_se == NULL) 1337 return -ENOMEM; 1338 new_se->start_page = start_page; 1339 new_se->nr_pages = nr_pages; 1340 new_se->start_block = start_block; 1341 1342 list_add_tail(&new_se->list, &sis->first_swap_extent.list); 1343 return 1; 1344} 1345 1346/* 1347 * A `swap extent' is a simple thing which maps a contiguous range of pages 1348 * onto a contiguous range of disk blocks. An ordered list of swap extents 1349 * is built at swapon time and is then used at swap_writepage/swap_readpage 1350 * time for locating where on disk a page belongs. 1351 * 1352 * If the swapfile is an S_ISBLK block device, a single extent is installed. 1353 * This is done so that the main operating code can treat S_ISBLK and S_ISREG 1354 * swap files identically. 1355 * 1356 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap 1357 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK 1358 * swapfiles are handled *identically* after swapon time. 1359 * 1360 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks 1361 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If 1362 * some stray blocks are found which do not fall within the PAGE_SIZE alignment 1363 * requirements, they are simply tossed out - we will never use those blocks 1364 * for swapping. 1365 * 1366 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This 1367 * prevents root from shooting her foot off by ftruncating an in-use swapfile, 1368 * which will scribble on the fs. 1369 * 1370 * The amount of disk space which a single swap extent represents varies. 1371 * Typically it is in the 1-4 megabyte range. So we can have hundreds of 1372 * extents in the list. To avoid much list walking, we cache the previous 1373 * search location in `curr_swap_extent', and start new searches from there. 1374 * This is extremely effective. The average number of iterations in 1375 * map_swap_page() has been measured at about 0.3 per page. - akpm. 1376 */ 1377static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span) 1378{ 1379 struct inode *inode; 1380 unsigned blocks_per_page; 1381 unsigned long page_no; 1382 unsigned blkbits; 1383 sector_t probe_block; 1384 sector_t last_block; 1385 sector_t lowest_block = -1; 1386 sector_t highest_block = 0; 1387 int nr_extents = 0; 1388 int ret; 1389 1390 inode = sis->swap_file->f_mapping->host; 1391 if (S_ISBLK(inode->i_mode)) { 1392 ret = add_swap_extent(sis, 0, sis->max, 0); 1393 *span = sis->pages; 1394 goto out; 1395 } 1396 1397 blkbits = inode->i_blkbits; 1398 blocks_per_page = PAGE_SIZE >> blkbits; 1399 1400 /* 1401 * Map all the blocks into the extent list. This code doesn't try 1402 * to be very smart. 1403 */ 1404 probe_block = 0; 1405 page_no = 0; 1406 last_block = i_size_read(inode) >> blkbits; 1407 while ((probe_block + blocks_per_page) <= last_block && 1408 page_no < sis->max) { 1409 unsigned block_in_page; 1410 sector_t first_block; 1411 1412 first_block = bmap(inode, probe_block); 1413 if (first_block == 0) 1414 goto bad_bmap; 1415 1416 /* 1417 * It must be PAGE_SIZE aligned on-disk 1418 */ 1419 if (first_block & (blocks_per_page - 1)) { 1420 probe_block++; 1421 goto reprobe; 1422 } 1423 1424 for (block_in_page = 1; block_in_page < blocks_per_page; 1425 block_in_page++) { 1426 sector_t block; 1427 1428 block = bmap(inode, probe_block + block_in_page); 1429 if (block == 0) 1430 goto bad_bmap; 1431 if (block != first_block + block_in_page) { 1432 /* Discontiguity */ 1433 probe_block++; 1434 goto reprobe; 1435 } 1436 } 1437 1438 first_block >>= (PAGE_SHIFT - blkbits); 1439 if (page_no) { /* exclude the header page */ 1440 if (first_block < lowest_block) 1441 lowest_block = first_block; 1442 if (first_block > highest_block) 1443 highest_block = first_block; 1444 } 1445 1446 /* 1447 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks 1448 */ 1449 ret = add_swap_extent(sis, page_no, 1, first_block); 1450 if (ret < 0) 1451 goto out; 1452 nr_extents += ret; 1453 page_no++; 1454 probe_block += blocks_per_page; 1455reprobe: 1456 continue; 1457 } 1458 ret = nr_extents; 1459 *span = 1 + highest_block - lowest_block; 1460 if (page_no == 0) 1461 page_no = 1; /* force Empty message */ 1462 sis->max = page_no; 1463 sis->pages = page_no - 1; 1464 sis->highest_bit = page_no - 1; 1465out: 1466 return ret; 1467bad_bmap: 1468 printk(KERN_ERR "swapon: swapfile has holes\n"); 1469 ret = -EINVAL; 1470 goto out; 1471} 1472 1473SYSCALL_DEFINE1(swapoff, const char __user *, specialfile) 1474{ 1475 struct swap_info_struct *p = NULL; 1476 unsigned char *swap_map; 1477 struct file *swap_file, *victim; 1478 struct address_space *mapping; 1479 struct inode *inode; 1480 char *pathname; 1481 int i, type, prev; 1482 int err; 1483 1484 if (!capable(CAP_SYS_ADMIN)) 1485 return -EPERM; 1486 1487 pathname = getname(specialfile); 1488 err = PTR_ERR(pathname); 1489 if (IS_ERR(pathname)) 1490 goto out; 1491 1492 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0); 1493 putname(pathname); 1494 err = PTR_ERR(victim); 1495 if (IS_ERR(victim)) 1496 goto out; 1497 1498 mapping = victim->f_mapping; 1499 prev = -1; 1500 spin_lock(&swap_lock); 1501 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) { 1502 p = swap_info[type]; 1503 if (p->flags & SWP_WRITEOK) { 1504 if (p->swap_file->f_mapping == mapping) 1505 break; 1506 } 1507 prev = type; 1508 } 1509 if (type < 0) { 1510 err = -EINVAL; 1511 spin_unlock(&swap_lock); 1512 goto out_dput; 1513 } 1514 if (!security_vm_enough_memory(p->pages)) 1515 vm_unacct_memory(p->pages); 1516 else { 1517 err = -ENOMEM; 1518 spin_unlock(&swap_lock); 1519 goto out_dput; 1520 } 1521 if (prev < 0) 1522 swap_list.head = p->next; 1523 else 1524 swap_info[prev]->next = p->next; 1525 if (type == swap_list.next) { 1526 /* just pick something that's safe... */ 1527 swap_list.next = swap_list.head; 1528 } 1529 if (p->prio < 0) { 1530 for (i = p->next; i >= 0; i = swap_info[i]->next) 1531 swap_info[i]->prio = p->prio--; 1532 least_priority++; 1533 } 1534 nr_swap_pages -= p->pages; 1535 total_swap_pages -= p->pages; 1536 p->flags &= ~SWP_WRITEOK; 1537 spin_unlock(&swap_lock); 1538 1539 current->flags |= PF_OOM_ORIGIN; 1540 err = try_to_unuse(type); 1541 current->flags &= ~PF_OOM_ORIGIN; 1542 1543 if (err) { 1544 /* re-insert swap space back into swap_list */ 1545 spin_lock(&swap_lock); 1546 if (p->prio < 0) 1547 p->prio = --least_priority; 1548 prev = -1; 1549 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) { 1550 if (p->prio >= swap_info[i]->prio) 1551 break; 1552 prev = i; 1553 } 1554 p->next = i; 1555 if (prev < 0) 1556 swap_list.head = swap_list.next = type; 1557 else 1558 swap_info[prev]->next = type; 1559 nr_swap_pages += p->pages; 1560 total_swap_pages += p->pages; 1561 p->flags |= SWP_WRITEOK; 1562 spin_unlock(&swap_lock); 1563 goto out_dput; 1564 } 1565 1566 /* wait for any unplug function to finish */ 1567 down_write(&swap_unplug_sem); 1568 up_write(&swap_unplug_sem); 1569 1570 destroy_swap_extents(p); 1571 if (p->flags & SWP_CONTINUED) 1572 free_swap_count_continuations(p); 1573 1574 mutex_lock(&swapon_mutex); 1575 spin_lock(&swap_lock); 1576 drain_mmlist(); 1577 1578 /* wait for anyone still in scan_swap_map */ 1579 p->highest_bit = 0; /* cuts scans short */ 1580 while (p->flags >= SWP_SCANNING) { 1581 spin_unlock(&swap_lock); 1582 schedule_timeout_uninterruptible(1); 1583 spin_lock(&swap_lock); 1584 } 1585 1586 swap_file = p->swap_file; 1587 p->swap_file = NULL; 1588 p->max = 0; 1589 swap_map = p->swap_map; 1590 p->swap_map = NULL; 1591 p->flags = 0; 1592 spin_unlock(&swap_lock); 1593 mutex_unlock(&swapon_mutex); 1594 vfree(swap_map); 1595 /* Destroy swap account informatin */ 1596 swap_cgroup_swapoff(type); 1597 1598 inode = mapping->host; 1599 if (S_ISBLK(inode->i_mode)) { 1600 struct block_device *bdev = I_BDEV(inode); 1601 set_blocksize(bdev, p->old_block_size); 1602 bd_release(bdev); 1603 } else { 1604 mutex_lock(&inode->i_mutex); 1605 inode->i_flags &= ~S_SWAPFILE; 1606 mutex_unlock(&inode->i_mutex); 1607 } 1608 filp_close(swap_file, NULL); 1609 err = 0; 1610 1611out_dput: 1612 filp_close(victim, NULL); 1613out: 1614 return err; 1615} 1616 1617#ifdef CONFIG_PROC_FS 1618/* iterator */ 1619static void *swap_start(struct seq_file *swap, loff_t *pos) 1620{ 1621 struct swap_info_struct *si; 1622 int type; 1623 loff_t l = *pos; 1624 1625 mutex_lock(&swapon_mutex); 1626 1627 if (!l) 1628 return SEQ_START_TOKEN; 1629 1630 for (type = 0; type < nr_swapfiles; type++) { 1631 smp_rmb(); /* read nr_swapfiles before swap_info[type] */ 1632 si = swap_info[type]; 1633 if (!(si->flags & SWP_USED) || !si->swap_map) 1634 continue; 1635 if (!--l) 1636 return si; 1637 } 1638 1639 return NULL; 1640} 1641 1642static void *swap_next(struct seq_file *swap, void *v, loff_t *pos) 1643{ 1644 struct swap_info_struct *si = v; 1645 int type; 1646 1647 if (v == SEQ_START_TOKEN) 1648 type = 0; 1649 else 1650 type = si->type + 1; 1651 1652 for (; type < nr_swapfiles; type++) { 1653 smp_rmb(); /* read nr_swapfiles before swap_info[type] */ 1654 si = swap_info[type]; 1655 if (!(si->flags & SWP_USED) || !si->swap_map) 1656 continue; 1657 ++*pos; 1658 return si; 1659 } 1660 1661 return NULL; 1662} 1663 1664static void swap_stop(struct seq_file *swap, void *v) 1665{ 1666 mutex_unlock(&swapon_mutex); 1667} 1668 1669static int swap_show(struct seq_file *swap, void *v) 1670{ 1671 struct swap_info_struct *si = v; 1672 struct file *file; 1673 int len; 1674 1675 if (si == SEQ_START_TOKEN) { 1676 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n"); 1677 return 0; 1678 } 1679 1680 file = si->swap_file; 1681 len = seq_path(swap, &file->f_path, " \t\n\\"); 1682 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n", 1683 len < 40 ? 40 - len : 1, " ", 1684 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ? 1685 "partition" : "file\t", 1686 si->pages << (PAGE_SHIFT - 10), 1687 si->inuse_pages << (PAGE_SHIFT - 10), 1688 si->prio); 1689 return 0; 1690} 1691 1692static const struct seq_operations swaps_op = { 1693 .start = swap_start, 1694 .next = swap_next, 1695 .stop = swap_stop, 1696 .show = swap_show 1697}; 1698 1699static int swaps_open(struct inode *inode, struct file *file) 1700{ 1701 return seq_open(file, &swaps_op); 1702} 1703 1704static const struct file_operations proc_swaps_operations = { 1705 .open = swaps_open, 1706 .read = seq_read, 1707 .llseek = seq_lseek, 1708 .release = seq_release, 1709}; 1710 1711static int __init procswaps_init(void) 1712{ 1713 proc_create("swaps", 0, NULL, &proc_swaps_operations); 1714 return 0; 1715} 1716__initcall(procswaps_init); 1717#endif /* CONFIG_PROC_FS */ 1718 1719#ifdef MAX_SWAPFILES_CHECK 1720static int __init max_swapfiles_check(void) 1721{ 1722 MAX_SWAPFILES_CHECK(); 1723 return 0; 1724} 1725late_initcall(max_swapfiles_check); 1726#endif 1727 1728/* 1729 * Written 01/25/92 by Simmule Turner, heavily changed by Linus. 1730 * 1731 * The swapon system call 1732 */ 1733SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags) 1734{ 1735 struct swap_info_struct *p; 1736 char *name = NULL; 1737 struct block_device *bdev = NULL; 1738 struct file *swap_file = NULL; 1739 struct address_space *mapping; 1740 unsigned int type; 1741 int i, prev; 1742 int error; 1743 union swap_header *swap_header = NULL; 1744 unsigned int nr_good_pages = 0; 1745 int nr_extents = 0; 1746 sector_t span; 1747 unsigned long maxpages = 1; 1748 unsigned long swapfilepages; 1749 unsigned char *swap_map = NULL; 1750 struct page *page = NULL; 1751 struct inode *inode = NULL; 1752 int did_down = 0; 1753 1754 if (!capable(CAP_SYS_ADMIN)) 1755 return -EPERM; 1756 1757 p = kzalloc(sizeof(*p), GFP_KERNEL); 1758 if (!p) 1759 return -ENOMEM; 1760 1761 spin_lock(&swap_lock); 1762 for (type = 0; type < nr_swapfiles; type++) { 1763 if (!(swap_info[type]->flags & SWP_USED)) 1764 break; 1765 } 1766 error = -EPERM; 1767 if (type >= MAX_SWAPFILES) { 1768 spin_unlock(&swap_lock); 1769 kfree(p); 1770 goto out; 1771 } 1772 if (type >= nr_swapfiles) { 1773 p->type = type; 1774 swap_info[type] = p; 1775 /* 1776 * Write swap_info[type] before nr_swapfiles, in case a 1777 * racing procfs swap_start() or swap_next() is reading them. 1778 * (We never shrink nr_swapfiles, we never free this entry.) 1779 */ 1780 smp_wmb(); 1781 nr_swapfiles++; 1782 } else { 1783 kfree(p); 1784 p = swap_info[type]; 1785 /* 1786 * Do not memset this entry: a racing procfs swap_next() 1787 * would be relying on p->type to remain valid. 1788 */ 1789 } 1790 INIT_LIST_HEAD(&p->first_swap_extent.list); 1791 p->flags = SWP_USED; 1792 p->next = -1; 1793 spin_unlock(&swap_lock); 1794 1795 name = getname(specialfile); 1796 error = PTR_ERR(name); 1797 if (IS_ERR(name)) { 1798 name = NULL; 1799 goto bad_swap_2; 1800 } 1801 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0); 1802 error = PTR_ERR(swap_file); 1803 if (IS_ERR(swap_file)) { 1804 swap_file = NULL; 1805 goto bad_swap_2; 1806 } 1807 1808 p->swap_file = swap_file; 1809 mapping = swap_file->f_mapping; 1810 inode = mapping->host; 1811 1812 error = -EBUSY; 1813 for (i = 0; i < nr_swapfiles; i++) { 1814 struct swap_info_struct *q = swap_info[i]; 1815 1816 if (i == type || !q->swap_file) 1817 continue; 1818 if (mapping == q->swap_file->f_mapping) 1819 goto bad_swap; 1820 } 1821 1822 error = -EINVAL; 1823 if (S_ISBLK(inode->i_mode)) { 1824 bdev = I_BDEV(inode); 1825 error = bd_claim(bdev, sys_swapon); 1826 if (error < 0) { 1827 bdev = NULL; 1828 error = -EINVAL; 1829 goto bad_swap; 1830 } 1831 p->old_block_size = block_size(bdev); 1832 error = set_blocksize(bdev, PAGE_SIZE); 1833 if (error < 0) 1834 goto bad_swap; 1835 p->bdev = bdev; 1836 } else if (S_ISREG(inode->i_mode)) { 1837 p->bdev = inode->i_sb->s_bdev; 1838 mutex_lock(&inode->i_mutex); 1839 did_down = 1; 1840 if (IS_SWAPFILE(inode)) { 1841 error = -EBUSY; 1842 goto bad_swap; 1843 } 1844 } else { 1845 goto bad_swap; 1846 } 1847 1848 swapfilepages = i_size_read(inode) >> PAGE_SHIFT; 1849 1850 /* 1851 * Read the swap header. 1852 */ 1853 if (!mapping->a_ops->readpage) { 1854 error = -EINVAL; 1855 goto bad_swap; 1856 } 1857 page = read_mapping_page(mapping, 0, swap_file); 1858 if (IS_ERR(page)) { 1859 error = PTR_ERR(page); 1860 goto bad_swap; 1861 } 1862 swap_header = kmap(page); 1863 1864 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) { 1865 printk(KERN_ERR "Unable to find swap-space signature\n"); 1866 error = -EINVAL; 1867 goto bad_swap; 1868 } 1869 1870 /* swap partition endianess hack... */ 1871 if (swab32(swap_header->info.version) == 1) { 1872 swab32s(&swap_header->info.version); 1873 swab32s(&swap_header->info.last_page); 1874 swab32s(&swap_header->info.nr_badpages); 1875 for (i = 0; i < swap_header->info.nr_badpages; i++) 1876 swab32s(&swap_header->info.badpages[i]); 1877 } 1878 /* Check the swap header's sub-version */ 1879 if (swap_header->info.version != 1) { 1880 printk(KERN_WARNING 1881 "Unable to handle swap header version %d\n", 1882 swap_header->info.version); 1883 error = -EINVAL; 1884 goto bad_swap; 1885 } 1886 1887 p->lowest_bit = 1; 1888 p->cluster_next = 1; 1889 p->cluster_nr = 0; 1890 1891 /* 1892 * Find out how many pages are allowed for a single swap 1893 * device. There are two limiting factors: 1) the number of 1894 * bits for the swap offset in the swp_entry_t type and 1895 * 2) the number of bits in the a swap pte as defined by 1896 * the different architectures. In order to find the 1897 * largest possible bit mask a swap entry with swap type 0 1898 * and swap offset ~0UL is created, encoded to a swap pte, 1899 * decoded to a swp_entry_t again and finally the swap 1900 * offset is extracted. This will mask all the bits from 1901 * the initial ~0UL mask that can't be encoded in either 1902 * the swp_entry_t or the architecture definition of a 1903 * swap pte. 1904 */ 1905 maxpages = swp_offset(pte_to_swp_entry( 1906 swp_entry_to_pte(swp_entry(0, ~0UL)))) - 1; 1907 if (maxpages > swap_header->info.last_page) 1908 maxpages = swap_header->info.last_page; 1909 p->highest_bit = maxpages - 1; 1910 1911 error = -EINVAL; 1912 if (!maxpages) 1913 goto bad_swap; 1914 if (swapfilepages && maxpages > swapfilepages) { 1915 printk(KERN_WARNING 1916 "Swap area shorter than signature indicates\n"); 1917 goto bad_swap; 1918 } 1919 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode)) 1920 goto bad_swap; 1921 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES) 1922 goto bad_swap; 1923 1924 /* OK, set up the swap map and apply the bad block list */ 1925 swap_map = vmalloc(maxpages); 1926 if (!swap_map) { 1927 error = -ENOMEM; 1928 goto bad_swap; 1929 } 1930 1931 memset(swap_map, 0, maxpages); 1932 for (i = 0; i < swap_header->info.nr_badpages; i++) { 1933 int page_nr = swap_header->info.badpages[i]; 1934 if (page_nr <= 0 || page_nr >= swap_header->info.last_page) { 1935 error = -EINVAL; 1936 goto bad_swap; 1937 } 1938 swap_map[page_nr] = SWAP_MAP_BAD; 1939 } 1940 1941 error = swap_cgroup_swapon(type, maxpages); 1942 if (error) 1943 goto bad_swap; 1944 1945 nr_good_pages = swap_header->info.last_page - 1946 swap_header->info.nr_badpages - 1947 1 /* header page */; 1948 1949 if (nr_good_pages) { 1950 swap_map[0] = SWAP_MAP_BAD; 1951 p->max = maxpages; 1952 p->pages = nr_good_pages; 1953 nr_extents = setup_swap_extents(p, &span); 1954 if (nr_extents < 0) { 1955 error = nr_extents; 1956 goto bad_swap; 1957 } 1958 nr_good_pages = p->pages; 1959 } 1960 if (!nr_good_pages) { 1961 printk(KERN_WARNING "Empty swap-file\n"); 1962 error = -EINVAL; 1963 goto bad_swap; 1964 } 1965 1966 if (p->bdev) { 1967 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) { 1968 p->flags |= SWP_SOLIDSTATE; 1969 p->cluster_next = 1 + (random32() % p->highest_bit); 1970 } 1971 if (discard_swap(p) == 0) 1972 p->flags |= SWP_DISCARDABLE; 1973 } 1974 1975 mutex_lock(&swapon_mutex); 1976 spin_lock(&swap_lock); 1977 if (swap_flags & SWAP_FLAG_PREFER) 1978 p->prio = 1979 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT; 1980 else 1981 p->prio = --least_priority; 1982 p->swap_map = swap_map; 1983 p->flags |= SWP_WRITEOK; 1984 nr_swap_pages += nr_good_pages; 1985 total_swap_pages += nr_good_pages; 1986 1987 printk(KERN_INFO "Adding %uk swap on %s. " 1988 "Priority:%d extents:%d across:%lluk %s%s\n", 1989 nr_good_pages<<(PAGE_SHIFT-10), name, p->prio, 1990 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10), 1991 (p->flags & SWP_SOLIDSTATE) ? "SS" : "", 1992 (p->flags & SWP_DISCARDABLE) ? "D" : ""); 1993 1994 /* insert swap space into swap_list: */ 1995 prev = -1; 1996 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) { 1997 if (p->prio >= swap_info[i]->prio) 1998 break; 1999 prev = i; 2000 } 2001 p->next = i; 2002 if (prev < 0) 2003 swap_list.head = swap_list.next = type; 2004 else 2005 swap_info[prev]->next = type; 2006 spin_unlock(&swap_lock); 2007 mutex_unlock(&swapon_mutex); 2008 error = 0; 2009 goto out; 2010bad_swap: 2011 if (bdev) { 2012 set_blocksize(bdev, p->old_block_size); 2013 bd_release(bdev); 2014 } 2015 destroy_swap_extents(p); 2016 swap_cgroup_swapoff(type); 2017bad_swap_2: 2018 spin_lock(&swap_lock); 2019 p->swap_file = NULL; 2020 p->flags = 0; 2021 spin_unlock(&swap_lock); 2022 vfree(swap_map); 2023 if (swap_file) 2024 filp_close(swap_file, NULL); 2025out: 2026 if (page && !IS_ERR(page)) { 2027 kunmap(page); 2028 page_cache_release(page); 2029 } 2030 if (name) 2031 putname(name); 2032 if (did_down) { 2033 if (!error) 2034 inode->i_flags |= S_SWAPFILE; 2035 mutex_unlock(&inode->i_mutex); 2036 } 2037 return error; 2038} 2039 2040void si_swapinfo(struct sysinfo *val) 2041{ 2042 unsigned int type; 2043 unsigned long nr_to_be_unused = 0; 2044 2045 spin_lock(&swap_lock); 2046 for (type = 0; type < nr_swapfiles; type++) { 2047 struct swap_info_struct *si = swap_info[type]; 2048 2049 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK)) 2050 nr_to_be_unused += si->inuse_pages; 2051 } 2052 val->freeswap = nr_swap_pages + nr_to_be_unused; 2053 val->totalswap = total_swap_pages + nr_to_be_unused; 2054 spin_unlock(&swap_lock); 2055} 2056 2057/* 2058 * Verify that a swap entry is valid and increment its swap map count. 2059 * 2060 * Returns error code in following case. 2061 * - success -> 0 2062 * - swp_entry is invalid -> EINVAL 2063 * - swp_entry is migration entry -> EINVAL 2064 * - swap-cache reference is requested but there is already one. -> EEXIST 2065 * - swap-cache reference is requested but the entry is not used. -> ENOENT 2066 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM 2067 */ 2068static int __swap_duplicate(swp_entry_t entry, unsigned char usage) 2069{ 2070 struct swap_info_struct *p; 2071 unsigned long offset, type; 2072 unsigned char count; 2073 unsigned char has_cache; 2074 int err = -EINVAL; 2075 2076 if (non_swap_entry(entry)) 2077 goto out; 2078 2079 type = swp_type(entry); 2080 if (type >= nr_swapfiles) 2081 goto bad_file; 2082 p = swap_info[type]; 2083 offset = swp_offset(entry); 2084 2085 spin_lock(&swap_lock); 2086 if (unlikely(offset >= p->max)) 2087 goto unlock_out; 2088 2089 count = p->swap_map[offset]; 2090 has_cache = count & SWAP_HAS_CACHE; 2091 count &= ~SWAP_HAS_CACHE; 2092 err = 0; 2093 2094 if (usage == SWAP_HAS_CACHE) { 2095 2096 /* set SWAP_HAS_CACHE if there is no cache and entry is used */ 2097 if (!has_cache && count) 2098 has_cache = SWAP_HAS_CACHE; 2099 else if (has_cache) /* someone else added cache */ 2100 err = -EEXIST; 2101 else /* no users remaining */ 2102 err = -ENOENT; 2103 2104 } else if (count || has_cache) { 2105 2106 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX) 2107 count += usage; 2108 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX) 2109 err = -EINVAL; 2110 else if (swap_count_continued(p, offset, count)) 2111 count = COUNT_CONTINUED; 2112 else 2113 err = -ENOMEM; 2114 } else 2115 err = -ENOENT; /* unused swap entry */ 2116 2117 p->swap_map[offset] = count | has_cache; 2118 2119unlock_out: 2120 spin_unlock(&swap_lock); 2121out: 2122 return err; 2123 2124bad_file: 2125 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val); 2126 goto out; 2127} 2128 2129/* 2130 * increase reference count of swap entry by 1. 2131 */ 2132int swap_duplicate(swp_entry_t entry) 2133{ 2134 int err = 0; 2135 2136 while (!err && __swap_duplicate(entry, 1) == -ENOMEM) 2137 err = add_swap_count_continuation(entry, GFP_ATOMIC); 2138 return err; 2139} 2140 2141/* 2142 * @entry: swap entry for which we allocate swap cache. 2143 * 2144 * Called when allocating swap cache for existing swap entry, 2145 * This can return error codes. Returns 0 at success. 2146 * -EBUSY means there is a swap cache. 2147 * Note: return code is different from swap_duplicate(). 2148 */ 2149int swapcache_prepare(swp_entry_t entry) 2150{ 2151 return __swap_duplicate(entry, SWAP_HAS_CACHE); 2152} 2153 2154/* 2155 * swap_lock prevents swap_map being freed. Don't grab an extra 2156 * reference on the swaphandle, it doesn't matter if it becomes unused. 2157 */ 2158int valid_swaphandles(swp_entry_t entry, unsigned long *offset) 2159{ 2160 struct swap_info_struct *si; 2161 int our_page_cluster = page_cluster; 2162 pgoff_t target, toff; 2163 pgoff_t base, end; 2164 int nr_pages = 0; 2165 2166 if (!our_page_cluster) /* no readahead */ 2167 return 0; 2168 2169 si = swap_info[swp_type(entry)]; 2170 target = swp_offset(entry); 2171 base = (target >> our_page_cluster) << our_page_cluster; 2172 end = base + (1 << our_page_cluster); 2173 if (!base) /* first page is swap header */ 2174 base++; 2175 2176 spin_lock(&swap_lock); 2177 if (end > si->max) /* don't go beyond end of map */ 2178 end = si->max; 2179 2180 /* Count contiguous allocated slots above our target */ 2181 for (toff = target; ++toff < end; nr_pages++) { 2182 /* Don't read in free or bad pages */ 2183 if (!si->swap_map[toff]) 2184 break; 2185 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD) 2186 break; 2187 } 2188 /* Count contiguous allocated slots below our target */ 2189 for (toff = target; --toff >= base; nr_pages++) { 2190 /* Don't read in free or bad pages */ 2191 if (!si->swap_map[toff]) 2192 break; 2193 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD) 2194 break; 2195 } 2196 spin_unlock(&swap_lock); 2197 2198 /* 2199 * Indicate starting offset, and return number of pages to get: 2200 * if only 1, say 0, since there's then no readahead to be done. 2201 */ 2202 *offset = ++toff; 2203 return nr_pages? ++nr_pages: 0; 2204} 2205 2206/* 2207 * add_swap_count_continuation - called when a swap count is duplicated 2208 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's 2209 * page of the original vmalloc'ed swap_map, to hold the continuation count 2210 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called 2211 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc. 2212 * 2213 * These continuation pages are seldom referenced: the common paths all work 2214 * on the original swap_map, only referring to a continuation page when the 2215 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. 2216 * 2217 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding 2218 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL) 2219 * can be called after dropping locks. 2220 */ 2221int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask) 2222{ 2223 struct swap_info_struct *si; 2224 struct page *head; 2225 struct page *page; 2226 struct page *list_page; 2227 pgoff_t offset; 2228 unsigned char count; 2229 2230 /* 2231 * When debugging, it's easier to use __GFP_ZERO here; but it's better 2232 * for latency not to zero a page while GFP_ATOMIC and holding locks. 2233 */ 2234 page = alloc_page(gfp_mask | __GFP_HIGHMEM); 2235 2236 si = swap_info_get(entry); 2237 if (!si) { 2238 /* 2239 * An acceptable race has occurred since the failing 2240 * __swap_duplicate(): the swap entry has been freed, 2241 * perhaps even the whole swap_map cleared for swapoff. 2242 */ 2243 goto outer; 2244 } 2245 2246 offset = swp_offset(entry); 2247 count = si->swap_map[offset] & ~SWAP_HAS_CACHE; 2248 2249 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) { 2250 /* 2251 * The higher the swap count, the more likely it is that tasks 2252 * will race to add swap count continuation: we need to avoid 2253 * over-provisioning. 2254 */ 2255 goto out; 2256 } 2257 2258 if (!page) { 2259 spin_unlock(&swap_lock); 2260 return -ENOMEM; 2261 } 2262 2263 /* 2264 * We are fortunate that although vmalloc_to_page uses pte_offset_map, 2265 * no architecture is using highmem pages for kernel pagetables: so it 2266 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps. 2267 */ 2268 head = vmalloc_to_page(si->swap_map + offset); 2269 offset &= ~PAGE_MASK; 2270 2271 /* 2272 * Page allocation does not initialize the page's lru field, 2273 * but it does always reset its private field. 2274 */ 2275 if (!page_private(head)) { 2276 BUG_ON(count & COUNT_CONTINUED); 2277 INIT_LIST_HEAD(&head->lru); 2278 set_page_private(head, SWP_CONTINUED); 2279 si->flags |= SWP_CONTINUED; 2280 } 2281 2282 list_for_each_entry(list_page, &head->lru, lru) { 2283 unsigned char *map; 2284 2285 /* 2286 * If the previous map said no continuation, but we've found 2287 * a continuation page, free our allocation and use this one. 2288 */ 2289 if (!(count & COUNT_CONTINUED)) 2290 goto out; 2291 2292 map = kmap_atomic(list_page, KM_USER0) + offset; 2293 count = *map; 2294 kunmap_atomic(map, KM_USER0); 2295 2296 /* 2297 * If this continuation count now has some space in it, 2298 * free our allocation and use this one. 2299 */ 2300 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX) 2301 goto out; 2302 } 2303 2304 list_add_tail(&page->lru, &head->lru); 2305 page = NULL; /* now it's attached, don't free it */ 2306out: 2307 spin_unlock(&swap_lock); 2308outer: 2309 if (page) 2310 __free_page(page); 2311 return 0; 2312} 2313 2314/* 2315 * swap_count_continued - when the original swap_map count is incremented 2316 * from SWAP_MAP_MAX, check if there is already a continuation page to carry 2317 * into, carry if so, or else fail until a new continuation page is allocated; 2318 * when the original swap_map count is decremented from 0 with continuation, 2319 * borrow from the continuation and report whether it still holds more. 2320 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock. 2321 */ 2322static bool swap_count_continued(struct swap_info_struct *si, 2323 pgoff_t offset, unsigned char count) 2324{ 2325 struct page *head; 2326 struct page *page; 2327 unsigned char *map; 2328 2329 head = vmalloc_to_page(si->swap_map + offset); 2330 if (page_private(head) != SWP_CONTINUED) { 2331 BUG_ON(count & COUNT_CONTINUED); 2332 return false; /* need to add count continuation */ 2333 } 2334 2335 offset &= ~PAGE_MASK; 2336 page = list_entry(head->lru.next, struct page, lru); 2337 map = kmap_atomic(page, KM_USER0) + offset; 2338 2339 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */ 2340 goto init_map; /* jump over SWAP_CONT_MAX checks */ 2341 2342 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */ 2343 /* 2344 * Think of how you add 1 to 999 2345 */ 2346 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) { 2347 kunmap_atomic(map, KM_USER0); 2348 page = list_entry(page->lru.next, struct page, lru); 2349 BUG_ON(page == head); 2350 map = kmap_atomic(page, KM_USER0) + offset; 2351 } 2352 if (*map == SWAP_CONT_MAX) { 2353 kunmap_atomic(map, KM_USER0); 2354 page = list_entry(page->lru.next, struct page, lru); 2355 if (page == head) 2356 return false; /* add count continuation */ 2357 map = kmap_atomic(page, KM_USER0) + offset; 2358init_map: *map = 0; /* we didn't zero the page */ 2359 } 2360 *map += 1; 2361 kunmap_atomic(map, KM_USER0); 2362 page = list_entry(page->lru.prev, struct page, lru); 2363 while (page != head) { 2364 map = kmap_atomic(page, KM_USER0) + offset; 2365 *map = COUNT_CONTINUED; 2366 kunmap_atomic(map, KM_USER0); 2367 page = list_entry(page->lru.prev, struct page, lru); 2368 } 2369 return true; /* incremented */ 2370 2371 } else { /* decrementing */ 2372 /* 2373 * Think of how you subtract 1 from 1000 2374 */ 2375 BUG_ON(count != COUNT_CONTINUED); 2376 while (*map == COUNT_CONTINUED) { 2377 kunmap_atomic(map, KM_USER0); 2378 page = list_entry(page->lru.next, struct page, lru); 2379 BUG_ON(page == head); 2380 map = kmap_atomic(page, KM_USER0) + offset; 2381 } 2382 BUG_ON(*map == 0); 2383 *map -= 1; 2384 if (*map == 0) 2385 count = 0; 2386 kunmap_atomic(map, KM_USER0); 2387 page = list_entry(page->lru.prev, struct page, lru); 2388 while (page != head) { 2389 map = kmap_atomic(page, KM_USER0) + offset; 2390 *map = SWAP_CONT_MAX | count; 2391 count = COUNT_CONTINUED; 2392 kunmap_atomic(map, KM_USER0); 2393 page = list_entry(page->lru.prev, struct page, lru); 2394 } 2395 return count == COUNT_CONTINUED; 2396 } 2397} 2398 2399/* 2400 * free_swap_count_continuations - swapoff free all the continuation pages 2401 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it. 2402 */ 2403static void free_swap_count_continuations(struct swap_info_struct *si) 2404{ 2405 pgoff_t offset; 2406 2407 for (offset = 0; offset < si->max; offset += PAGE_SIZE) { 2408 struct page *head; 2409 head = vmalloc_to_page(si->swap_map + offset); 2410 if (page_private(head)) { 2411 struct list_head *this, *next; 2412 list_for_each_safe(this, next, &head->lru) { 2413 struct page *page; 2414 page = list_entry(this, struct page, lru); 2415 list_del(this); 2416 __free_page(page); 2417 } 2418 } 2419 } 2420} 2421