1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/mm/swapfile.c 4 * 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 * Swap reorganised 29.12.95, Stephen Tweedie 7 */ 8 9 #include <linux/blkdev.h> 10 #include <linux/mm.h> 11 #include <linux/sched/mm.h> 12 #include <linux/sched/task.h> 13 #include <linux/hugetlb.h> 14 #include <linux/mman.h> 15 #include <linux/slab.h> 16 #include <linux/kernel_stat.h> 17 #include <linux/swap.h> 18 #include <linux/vmalloc.h> 19 #include <linux/pagemap.h> 20 #include <linux/namei.h> 21 #include <linux/shmem_fs.h> 22 #include <linux/blk-cgroup.h> 23 #include <linux/random.h> 24 #include <linux/writeback.h> 25 #include <linux/proc_fs.h> 26 #include <linux/seq_file.h> 27 #include <linux/init.h> 28 #include <linux/ksm.h> 29 #include <linux/rmap.h> 30 #include <linux/security.h> 31 #include <linux/backing-dev.h> 32 #include <linux/mutex.h> 33 #include <linux/capability.h> 34 #include <linux/syscalls.h> 35 #include <linux/memcontrol.h> 36 #include <linux/poll.h> 37 #include <linux/oom.h> 38 #include <linux/swapfile.h> 39 #include <linux/export.h> 40 #include <linux/swap_slots.h> 41 #include <linux/sort.h> 42 #include <linux/completion.h> 43 #include <linux/suspend.h> 44 #include <linux/zswap.h> 45 #include <linux/plist.h> 46 47 #include <asm/tlbflush.h> 48 #include <linux/swapops.h> 49 #include <linux/swap_cgroup.h> 50 #include "internal.h" 51 #include "swap.h" 52 53 static bool swap_count_continued(struct swap_info_struct *, pgoff_t, 54 unsigned char); 55 static void free_swap_count_continuations(struct swap_info_struct *); 56 static void swap_entry_range_free(struct swap_info_struct *si, swp_entry_t entry, 57 unsigned int nr_pages); 58 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset, 59 unsigned int nr_entries); 60 static bool folio_swapcache_freeable(struct folio *folio); 61 static struct swap_cluster_info *lock_cluster_or_swap_info( 62 struct swap_info_struct *si, unsigned long offset); 63 static void unlock_cluster_or_swap_info(struct swap_info_struct *si, 64 struct swap_cluster_info *ci); 65 66 static DEFINE_SPINLOCK(swap_lock); 67 static unsigned int nr_swapfiles; 68 atomic_long_t nr_swap_pages; 69 /* 70 * Some modules use swappable objects and may try to swap them out under 71 * memory pressure (via the shrinker). Before doing so, they may wish to 72 * check to see if any swap space is available. 73 */ 74 EXPORT_SYMBOL_GPL(nr_swap_pages); 75 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */ 76 long total_swap_pages; 77 static int least_priority = -1; 78 unsigned long swapfile_maximum_size; 79 #ifdef CONFIG_MIGRATION 80 bool swap_migration_ad_supported; 81 #endif /* CONFIG_MIGRATION */ 82 83 static const char Bad_file[] = "Bad swap file entry "; 84 static const char Unused_file[] = "Unused swap file entry "; 85 static const char Bad_offset[] = "Bad swap offset entry "; 86 static const char Unused_offset[] = "Unused swap offset entry "; 87 88 /* 89 * all active swap_info_structs 90 * protected with swap_lock, and ordered by priority. 91 */ 92 static PLIST_HEAD(swap_active_head); 93 94 /* 95 * all available (active, not full) swap_info_structs 96 * protected with swap_avail_lock, ordered by priority. 97 * This is used by folio_alloc_swap() instead of swap_active_head 98 * because swap_active_head includes all swap_info_structs, 99 * but folio_alloc_swap() doesn't need to look at full ones. 100 * This uses its own lock instead of swap_lock because when a 101 * swap_info_struct changes between not-full/full, it needs to 102 * add/remove itself to/from this list, but the swap_info_struct->lock 103 * is held and the locking order requires swap_lock to be taken 104 * before any swap_info_struct->lock. 105 */ 106 static struct plist_head *swap_avail_heads; 107 static DEFINE_SPINLOCK(swap_avail_lock); 108 109 static struct swap_info_struct *swap_info[MAX_SWAPFILES]; 110 111 static DEFINE_MUTEX(swapon_mutex); 112 113 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait); 114 /* Activity counter to indicate that a swapon or swapoff has occurred */ 115 static atomic_t proc_poll_event = ATOMIC_INIT(0); 116 117 atomic_t nr_rotate_swap = ATOMIC_INIT(0); 118 119 static struct swap_info_struct *swap_type_to_swap_info(int type) 120 { 121 if (type >= MAX_SWAPFILES) 122 return NULL; 123 124 return READ_ONCE(swap_info[type]); /* rcu_dereference() */ 125 } 126 127 static inline unsigned char swap_count(unsigned char ent) 128 { 129 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */ 130 } 131 132 /* Reclaim the swap entry anyway if possible */ 133 #define TTRS_ANYWAY 0x1 134 /* 135 * Reclaim the swap entry if there are no more mappings of the 136 * corresponding page 137 */ 138 #define TTRS_UNMAPPED 0x2 139 /* Reclaim the swap entry if swap is getting full */ 140 #define TTRS_FULL 0x4 141 /* Reclaim directly, bypass the slot cache and don't touch device lock */ 142 #define TTRS_DIRECT 0x8 143 144 static bool swap_is_has_cache(struct swap_info_struct *si, 145 unsigned long offset, int nr_pages) 146 { 147 unsigned char *map = si->swap_map + offset; 148 unsigned char *map_end = map + nr_pages; 149 150 do { 151 VM_BUG_ON(!(*map & SWAP_HAS_CACHE)); 152 if (*map != SWAP_HAS_CACHE) 153 return false; 154 } while (++map < map_end); 155 156 return true; 157 } 158 159 static bool swap_is_last_map(struct swap_info_struct *si, 160 unsigned long offset, int nr_pages, bool *has_cache) 161 { 162 unsigned char *map = si->swap_map + offset; 163 unsigned char *map_end = map + nr_pages; 164 unsigned char count = *map; 165 166 if (swap_count(count) != 1) 167 return false; 168 169 while (++map < map_end) { 170 if (*map != count) 171 return false; 172 } 173 174 *has_cache = !!(count & SWAP_HAS_CACHE); 175 return true; 176 } 177 178 /* 179 * returns number of pages in the folio that backs the swap entry. If positive, 180 * the folio was reclaimed. If negative, the folio was not reclaimed. If 0, no 181 * folio was associated with the swap entry. 182 */ 183 static int __try_to_reclaim_swap(struct swap_info_struct *si, 184 unsigned long offset, unsigned long flags) 185 { 186 swp_entry_t entry = swp_entry(si->type, offset); 187 struct address_space *address_space = swap_address_space(entry); 188 struct swap_cluster_info *ci; 189 struct folio *folio; 190 int ret, nr_pages; 191 bool need_reclaim; 192 193 folio = filemap_get_folio(address_space, swap_cache_index(entry)); 194 if (IS_ERR(folio)) 195 return 0; 196 197 /* offset could point to the middle of a large folio */ 198 entry = folio->swap; 199 offset = swp_offset(entry); 200 nr_pages = folio_nr_pages(folio); 201 ret = -nr_pages; 202 203 /* 204 * When this function is called from scan_swap_map_slots() and it's 205 * called by vmscan.c at reclaiming folios. So we hold a folio lock 206 * here. We have to use trylock for avoiding deadlock. This is a special 207 * case and you should use folio_free_swap() with explicit folio_lock() 208 * in usual operations. 209 */ 210 if (!folio_trylock(folio)) 211 goto out; 212 213 need_reclaim = ((flags & TTRS_ANYWAY) || 214 ((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) || 215 ((flags & TTRS_FULL) && mem_cgroup_swap_full(folio))); 216 if (!need_reclaim || !folio_swapcache_freeable(folio)) 217 goto out_unlock; 218 219 /* 220 * It's safe to delete the folio from swap cache only if the folio's 221 * swap_map is HAS_CACHE only, which means the slots have no page table 222 * reference or pending writeback, and can't be allocated to others. 223 */ 224 ci = lock_cluster_or_swap_info(si, offset); 225 need_reclaim = swap_is_has_cache(si, offset, nr_pages); 226 unlock_cluster_or_swap_info(si, ci); 227 if (!need_reclaim) 228 goto out_unlock; 229 230 if (!(flags & TTRS_DIRECT)) { 231 /* Free through slot cache */ 232 delete_from_swap_cache(folio); 233 folio_set_dirty(folio); 234 ret = nr_pages; 235 goto out_unlock; 236 } 237 238 xa_lock_irq(&address_space->i_pages); 239 __delete_from_swap_cache(folio, entry, NULL); 240 xa_unlock_irq(&address_space->i_pages); 241 folio_ref_sub(folio, nr_pages); 242 folio_set_dirty(folio); 243 244 spin_lock(&si->lock); 245 /* Only sinple page folio can be backed by zswap */ 246 if (nr_pages == 1) 247 zswap_invalidate(entry); 248 swap_entry_range_free(si, entry, nr_pages); 249 spin_unlock(&si->lock); 250 ret = nr_pages; 251 out_unlock: 252 folio_unlock(folio); 253 out: 254 folio_put(folio); 255 return ret; 256 } 257 258 static inline struct swap_extent *first_se(struct swap_info_struct *sis) 259 { 260 struct rb_node *rb = rb_first(&sis->swap_extent_root); 261 return rb_entry(rb, struct swap_extent, rb_node); 262 } 263 264 static inline struct swap_extent *next_se(struct swap_extent *se) 265 { 266 struct rb_node *rb = rb_next(&se->rb_node); 267 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL; 268 } 269 270 /* 271 * swapon tell device that all the old swap contents can be discarded, 272 * to allow the swap device to optimize its wear-levelling. 273 */ 274 static int discard_swap(struct swap_info_struct *si) 275 { 276 struct swap_extent *se; 277 sector_t start_block; 278 sector_t nr_blocks; 279 int err = 0; 280 281 /* Do not discard the swap header page! */ 282 se = first_se(si); 283 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9); 284 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9); 285 if (nr_blocks) { 286 err = blkdev_issue_discard(si->bdev, start_block, 287 nr_blocks, GFP_KERNEL); 288 if (err) 289 return err; 290 cond_resched(); 291 } 292 293 for (se = next_se(se); se; se = next_se(se)) { 294 start_block = se->start_block << (PAGE_SHIFT - 9); 295 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9); 296 297 err = blkdev_issue_discard(si->bdev, start_block, 298 nr_blocks, GFP_KERNEL); 299 if (err) 300 break; 301 302 cond_resched(); 303 } 304 return err; /* That will often be -EOPNOTSUPP */ 305 } 306 307 static struct swap_extent * 308 offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset) 309 { 310 struct swap_extent *se; 311 struct rb_node *rb; 312 313 rb = sis->swap_extent_root.rb_node; 314 while (rb) { 315 se = rb_entry(rb, struct swap_extent, rb_node); 316 if (offset < se->start_page) 317 rb = rb->rb_left; 318 else if (offset >= se->start_page + se->nr_pages) 319 rb = rb->rb_right; 320 else 321 return se; 322 } 323 /* It *must* be present */ 324 BUG(); 325 } 326 327 sector_t swap_folio_sector(struct folio *folio) 328 { 329 struct swap_info_struct *sis = swp_swap_info(folio->swap); 330 struct swap_extent *se; 331 sector_t sector; 332 pgoff_t offset; 333 334 offset = swp_offset(folio->swap); 335 se = offset_to_swap_extent(sis, offset); 336 sector = se->start_block + (offset - se->start_page); 337 return sector << (PAGE_SHIFT - 9); 338 } 339 340 /* 341 * swap allocation tell device that a cluster of swap can now be discarded, 342 * to allow the swap device to optimize its wear-levelling. 343 */ 344 static void discard_swap_cluster(struct swap_info_struct *si, 345 pgoff_t start_page, pgoff_t nr_pages) 346 { 347 struct swap_extent *se = offset_to_swap_extent(si, start_page); 348 349 while (nr_pages) { 350 pgoff_t offset = start_page - se->start_page; 351 sector_t start_block = se->start_block + offset; 352 sector_t nr_blocks = se->nr_pages - offset; 353 354 if (nr_blocks > nr_pages) 355 nr_blocks = nr_pages; 356 start_page += nr_blocks; 357 nr_pages -= nr_blocks; 358 359 start_block <<= PAGE_SHIFT - 9; 360 nr_blocks <<= PAGE_SHIFT - 9; 361 if (blkdev_issue_discard(si->bdev, start_block, 362 nr_blocks, GFP_NOIO)) 363 break; 364 365 se = next_se(se); 366 } 367 } 368 369 #ifdef CONFIG_THP_SWAP 370 #define SWAPFILE_CLUSTER HPAGE_PMD_NR 371 372 #define swap_entry_order(order) (order) 373 #else 374 #define SWAPFILE_CLUSTER 256 375 376 /* 377 * Define swap_entry_order() as constant to let compiler to optimize 378 * out some code if !CONFIG_THP_SWAP 379 */ 380 #define swap_entry_order(order) 0 381 #endif 382 #define LATENCY_LIMIT 256 383 384 static inline bool cluster_is_free(struct swap_cluster_info *info) 385 { 386 return info->flags & CLUSTER_FLAG_FREE; 387 } 388 389 static inline unsigned int cluster_index(struct swap_info_struct *si, 390 struct swap_cluster_info *ci) 391 { 392 return ci - si->cluster_info; 393 } 394 395 static inline unsigned int cluster_offset(struct swap_info_struct *si, 396 struct swap_cluster_info *ci) 397 { 398 return cluster_index(si, ci) * SWAPFILE_CLUSTER; 399 } 400 401 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si, 402 unsigned long offset) 403 { 404 struct swap_cluster_info *ci; 405 406 ci = si->cluster_info; 407 if (ci) { 408 ci += offset / SWAPFILE_CLUSTER; 409 spin_lock(&ci->lock); 410 } 411 return ci; 412 } 413 414 static inline void unlock_cluster(struct swap_cluster_info *ci) 415 { 416 if (ci) 417 spin_unlock(&ci->lock); 418 } 419 420 /* 421 * Determine the locking method in use for this device. Return 422 * swap_cluster_info if SSD-style cluster-based locking is in place. 423 */ 424 static inline struct swap_cluster_info *lock_cluster_or_swap_info( 425 struct swap_info_struct *si, unsigned long offset) 426 { 427 struct swap_cluster_info *ci; 428 429 /* Try to use fine-grained SSD-style locking if available: */ 430 ci = lock_cluster(si, offset); 431 /* Otherwise, fall back to traditional, coarse locking: */ 432 if (!ci) 433 spin_lock(&si->lock); 434 435 return ci; 436 } 437 438 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si, 439 struct swap_cluster_info *ci) 440 { 441 if (ci) 442 unlock_cluster(ci); 443 else 444 spin_unlock(&si->lock); 445 } 446 447 /* Add a cluster to discard list and schedule it to do discard */ 448 static void swap_cluster_schedule_discard(struct swap_info_struct *si, 449 struct swap_cluster_info *ci) 450 { 451 unsigned int idx = cluster_index(si, ci); 452 /* 453 * If scan_swap_map_slots() can't find a free cluster, it will check 454 * si->swap_map directly. To make sure the discarding cluster isn't 455 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied). 456 * It will be cleared after discard 457 */ 458 memset(si->swap_map + idx * SWAPFILE_CLUSTER, 459 SWAP_MAP_BAD, SWAPFILE_CLUSTER); 460 461 VM_BUG_ON(ci->flags & CLUSTER_FLAG_FREE); 462 list_move_tail(&ci->list, &si->discard_clusters); 463 ci->flags = 0; 464 schedule_work(&si->discard_work); 465 } 466 467 static void __free_cluster(struct swap_info_struct *si, struct swap_cluster_info *ci) 468 { 469 lockdep_assert_held(&si->lock); 470 lockdep_assert_held(&ci->lock); 471 472 if (ci->flags) 473 list_move_tail(&ci->list, &si->free_clusters); 474 else 475 list_add_tail(&ci->list, &si->free_clusters); 476 ci->flags = CLUSTER_FLAG_FREE; 477 ci->order = 0; 478 } 479 480 /* 481 * Doing discard actually. After a cluster discard is finished, the cluster 482 * will be added to free cluster list. caller should hold si->lock. 483 */ 484 static void swap_do_scheduled_discard(struct swap_info_struct *si) 485 { 486 struct swap_cluster_info *ci; 487 unsigned int idx; 488 489 while (!list_empty(&si->discard_clusters)) { 490 ci = list_first_entry(&si->discard_clusters, struct swap_cluster_info, list); 491 list_del(&ci->list); 492 idx = cluster_index(si, ci); 493 spin_unlock(&si->lock); 494 495 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER, 496 SWAPFILE_CLUSTER); 497 498 spin_lock(&si->lock); 499 spin_lock(&ci->lock); 500 __free_cluster(si, ci); 501 memset(si->swap_map + idx * SWAPFILE_CLUSTER, 502 0, SWAPFILE_CLUSTER); 503 spin_unlock(&ci->lock); 504 } 505 } 506 507 static void swap_discard_work(struct work_struct *work) 508 { 509 struct swap_info_struct *si; 510 511 si = container_of(work, struct swap_info_struct, discard_work); 512 513 spin_lock(&si->lock); 514 swap_do_scheduled_discard(si); 515 spin_unlock(&si->lock); 516 } 517 518 static void swap_users_ref_free(struct percpu_ref *ref) 519 { 520 struct swap_info_struct *si; 521 522 si = container_of(ref, struct swap_info_struct, users); 523 complete(&si->comp); 524 } 525 526 static void free_cluster(struct swap_info_struct *si, struct swap_cluster_info *ci) 527 { 528 VM_BUG_ON(ci->count != 0); 529 lockdep_assert_held(&si->lock); 530 lockdep_assert_held(&ci->lock); 531 532 if (ci->flags & CLUSTER_FLAG_FRAG) 533 si->frag_cluster_nr[ci->order]--; 534 535 /* 536 * If the swap is discardable, prepare discard the cluster 537 * instead of free it immediately. The cluster will be freed 538 * after discard. 539 */ 540 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) == 541 (SWP_WRITEOK | SWP_PAGE_DISCARD)) { 542 swap_cluster_schedule_discard(si, ci); 543 return; 544 } 545 546 __free_cluster(si, ci); 547 } 548 549 /* 550 * The cluster corresponding to page_nr will be used. The cluster will not be 551 * added to free cluster list and its usage counter will be increased by 1. 552 * Only used for initialization. 553 */ 554 static void inc_cluster_info_page(struct swap_info_struct *si, 555 struct swap_cluster_info *cluster_info, unsigned long page_nr) 556 { 557 unsigned long idx = page_nr / SWAPFILE_CLUSTER; 558 struct swap_cluster_info *ci; 559 560 if (!cluster_info) 561 return; 562 563 ci = cluster_info + idx; 564 ci->count++; 565 566 VM_BUG_ON(ci->count > SWAPFILE_CLUSTER); 567 VM_BUG_ON(ci->flags); 568 } 569 570 /* 571 * The cluster ci decreases @nr_pages usage. If the usage counter becomes 0, 572 * which means no page in the cluster is in use, we can optionally discard 573 * the cluster and add it to free cluster list. 574 */ 575 static void dec_cluster_info_page(struct swap_info_struct *si, 576 struct swap_cluster_info *ci, int nr_pages) 577 { 578 if (!si->cluster_info) 579 return; 580 581 VM_BUG_ON(ci->count < nr_pages); 582 VM_BUG_ON(cluster_is_free(ci)); 583 lockdep_assert_held(&si->lock); 584 lockdep_assert_held(&ci->lock); 585 ci->count -= nr_pages; 586 587 if (!ci->count) { 588 free_cluster(si, ci); 589 return; 590 } 591 592 if (!(ci->flags & CLUSTER_FLAG_NONFULL)) { 593 VM_BUG_ON(ci->flags & CLUSTER_FLAG_FREE); 594 if (ci->flags & CLUSTER_FLAG_FRAG) 595 si->frag_cluster_nr[ci->order]--; 596 list_move_tail(&ci->list, &si->nonfull_clusters[ci->order]); 597 ci->flags = CLUSTER_FLAG_NONFULL; 598 } 599 } 600 601 static bool cluster_reclaim_range(struct swap_info_struct *si, 602 struct swap_cluster_info *ci, 603 unsigned long start, unsigned long end) 604 { 605 unsigned char *map = si->swap_map; 606 unsigned long offset; 607 608 spin_unlock(&ci->lock); 609 spin_unlock(&si->lock); 610 611 for (offset = start; offset < end; offset++) { 612 switch (READ_ONCE(map[offset])) { 613 case 0: 614 continue; 615 case SWAP_HAS_CACHE: 616 if (__try_to_reclaim_swap(si, offset, TTRS_ANYWAY | TTRS_DIRECT) > 0) 617 continue; 618 goto out; 619 default: 620 goto out; 621 } 622 } 623 out: 624 spin_lock(&si->lock); 625 spin_lock(&ci->lock); 626 627 /* 628 * Recheck the range no matter reclaim succeeded or not, the slot 629 * could have been be freed while we are not holding the lock. 630 */ 631 for (offset = start; offset < end; offset++) 632 if (READ_ONCE(map[offset])) 633 return false; 634 635 return true; 636 } 637 638 static bool cluster_scan_range(struct swap_info_struct *si, 639 struct swap_cluster_info *ci, 640 unsigned long start, unsigned int nr_pages) 641 { 642 unsigned long offset, end = start + nr_pages; 643 unsigned char *map = si->swap_map; 644 bool need_reclaim = false; 645 646 for (offset = start; offset < end; offset++) { 647 switch (READ_ONCE(map[offset])) { 648 case 0: 649 continue; 650 case SWAP_HAS_CACHE: 651 if (!vm_swap_full()) 652 return false; 653 need_reclaim = true; 654 continue; 655 default: 656 return false; 657 } 658 } 659 660 if (need_reclaim) 661 return cluster_reclaim_range(si, ci, start, end); 662 663 return true; 664 } 665 666 static void cluster_alloc_range(struct swap_info_struct *si, struct swap_cluster_info *ci, 667 unsigned int start, unsigned char usage, 668 unsigned int order) 669 { 670 unsigned int nr_pages = 1 << order; 671 672 if (cluster_is_free(ci)) { 673 if (nr_pages < SWAPFILE_CLUSTER) { 674 list_move_tail(&ci->list, &si->nonfull_clusters[order]); 675 ci->flags = CLUSTER_FLAG_NONFULL; 676 } 677 ci->order = order; 678 } 679 680 memset(si->swap_map + start, usage, nr_pages); 681 swap_range_alloc(si, start, nr_pages); 682 ci->count += nr_pages; 683 684 if (ci->count == SWAPFILE_CLUSTER) { 685 VM_BUG_ON(!(ci->flags & 686 (CLUSTER_FLAG_FREE | CLUSTER_FLAG_NONFULL | CLUSTER_FLAG_FRAG))); 687 if (ci->flags & CLUSTER_FLAG_FRAG) 688 si->frag_cluster_nr[ci->order]--; 689 list_move_tail(&ci->list, &si->full_clusters); 690 ci->flags = CLUSTER_FLAG_FULL; 691 } 692 } 693 694 static unsigned int alloc_swap_scan_cluster(struct swap_info_struct *si, unsigned long offset, 695 unsigned int *foundp, unsigned int order, 696 unsigned char usage) 697 { 698 unsigned long start = offset & ~(SWAPFILE_CLUSTER - 1); 699 unsigned long end = min(start + SWAPFILE_CLUSTER, si->max); 700 unsigned int nr_pages = 1 << order; 701 struct swap_cluster_info *ci; 702 703 if (end < nr_pages) 704 return SWAP_NEXT_INVALID; 705 end -= nr_pages; 706 707 ci = lock_cluster(si, offset); 708 if (ci->count + nr_pages > SWAPFILE_CLUSTER) { 709 offset = SWAP_NEXT_INVALID; 710 goto done; 711 } 712 713 while (offset <= end) { 714 if (cluster_scan_range(si, ci, offset, nr_pages)) { 715 cluster_alloc_range(si, ci, offset, usage, order); 716 *foundp = offset; 717 if (ci->count == SWAPFILE_CLUSTER) { 718 offset = SWAP_NEXT_INVALID; 719 goto done; 720 } 721 offset += nr_pages; 722 break; 723 } 724 offset += nr_pages; 725 } 726 if (offset > end) 727 offset = SWAP_NEXT_INVALID; 728 done: 729 unlock_cluster(ci); 730 return offset; 731 } 732 733 static void swap_reclaim_full_clusters(struct swap_info_struct *si) 734 { 735 long to_scan = 1; 736 unsigned long offset, end; 737 struct swap_cluster_info *ci; 738 unsigned char *map = si->swap_map; 739 int nr_reclaim, total_reclaimed = 0; 740 741 if (atomic_long_read(&nr_swap_pages) <= SWAPFILE_CLUSTER) 742 to_scan = si->inuse_pages / SWAPFILE_CLUSTER; 743 744 while (!list_empty(&si->full_clusters)) { 745 ci = list_first_entry(&si->full_clusters, struct swap_cluster_info, list); 746 list_move_tail(&ci->list, &si->full_clusters); 747 offset = cluster_offset(si, ci); 748 end = min(si->max, offset + SWAPFILE_CLUSTER); 749 to_scan--; 750 751 while (offset < end) { 752 if (READ_ONCE(map[offset]) == SWAP_HAS_CACHE) { 753 spin_unlock(&si->lock); 754 nr_reclaim = __try_to_reclaim_swap(si, offset, 755 TTRS_ANYWAY | TTRS_DIRECT); 756 spin_lock(&si->lock); 757 if (nr_reclaim > 0) { 758 offset += nr_reclaim; 759 total_reclaimed += nr_reclaim; 760 continue; 761 } else if (nr_reclaim < 0) { 762 offset += -nr_reclaim; 763 continue; 764 } 765 } 766 offset++; 767 } 768 if (to_scan <= 0 || total_reclaimed) 769 break; 770 } 771 } 772 773 /* 774 * Try to get swap entries with specified order from current cpu's swap entry 775 * pool (a cluster). This might involve allocating a new cluster for current CPU 776 * too. 777 */ 778 static unsigned long cluster_alloc_swap_entry(struct swap_info_struct *si, int order, 779 unsigned char usage) 780 { 781 struct percpu_cluster *cluster; 782 struct swap_cluster_info *ci; 783 unsigned int offset, found = 0; 784 785 new_cluster: 786 lockdep_assert_held(&si->lock); 787 cluster = this_cpu_ptr(si->percpu_cluster); 788 offset = cluster->next[order]; 789 if (offset) { 790 offset = alloc_swap_scan_cluster(si, offset, &found, order, usage); 791 if (found) 792 goto done; 793 } 794 795 if (!list_empty(&si->free_clusters)) { 796 ci = list_first_entry(&si->free_clusters, struct swap_cluster_info, list); 797 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), &found, order, usage); 798 VM_BUG_ON(!found); 799 goto done; 800 } 801 802 if (order < PMD_ORDER) { 803 unsigned int frags = 0; 804 805 while (!list_empty(&si->nonfull_clusters[order])) { 806 ci = list_first_entry(&si->nonfull_clusters[order], 807 struct swap_cluster_info, list); 808 list_move_tail(&ci->list, &si->frag_clusters[order]); 809 ci->flags = CLUSTER_FLAG_FRAG; 810 si->frag_cluster_nr[order]++; 811 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), 812 &found, order, usage); 813 frags++; 814 if (found) 815 break; 816 } 817 818 if (!found) { 819 /* 820 * Nonfull clusters are moved to frag tail if we reached 821 * here, count them too, don't over scan the frag list. 822 */ 823 while (frags < si->frag_cluster_nr[order]) { 824 ci = list_first_entry(&si->frag_clusters[order], 825 struct swap_cluster_info, list); 826 /* 827 * Rotate the frag list to iterate, they were all failing 828 * high order allocation or moved here due to per-CPU usage, 829 * this help keeping usable cluster ahead. 830 */ 831 list_move_tail(&ci->list, &si->frag_clusters[order]); 832 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), 833 &found, order, usage); 834 frags++; 835 if (found) 836 break; 837 } 838 } 839 } 840 841 if (found) 842 goto done; 843 844 if (!list_empty(&si->discard_clusters)) { 845 /* 846 * we don't have free cluster but have some clusters in 847 * discarding, do discard now and reclaim them, then 848 * reread cluster_next_cpu since we dropped si->lock 849 */ 850 swap_do_scheduled_discard(si); 851 goto new_cluster; 852 } 853 854 if (order) 855 goto done; 856 857 /* Order 0 stealing from higher order */ 858 for (int o = 1; o < SWAP_NR_ORDERS; o++) { 859 /* 860 * Clusters here have at least one usable slots and can't fail order 0 861 * allocation, but reclaim may drop si->lock and race with another user. 862 */ 863 while (!list_empty(&si->frag_clusters[o])) { 864 ci = list_first_entry(&si->frag_clusters[o], 865 struct swap_cluster_info, list); 866 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), 867 &found, 0, usage); 868 if (found) 869 goto done; 870 } 871 872 while (!list_empty(&si->nonfull_clusters[o])) { 873 ci = list_first_entry(&si->nonfull_clusters[o], 874 struct swap_cluster_info, list); 875 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), 876 &found, 0, usage); 877 if (found) 878 goto done; 879 } 880 } 881 882 done: 883 /* Try reclaim from full clusters if device is nearfull */ 884 if (vm_swap_full() && (!found || (si->pages - si->inuse_pages) < SWAPFILE_CLUSTER)) { 885 swap_reclaim_full_clusters(si); 886 if (!found && !order && si->pages != si->inuse_pages) 887 goto new_cluster; 888 } 889 890 cluster->next[order] = offset; 891 return found; 892 } 893 894 static void __del_from_avail_list(struct swap_info_struct *si) 895 { 896 int nid; 897 898 assert_spin_locked(&si->lock); 899 for_each_node(nid) 900 plist_del(&si->avail_lists[nid], &swap_avail_heads[nid]); 901 } 902 903 static void del_from_avail_list(struct swap_info_struct *si) 904 { 905 spin_lock(&swap_avail_lock); 906 __del_from_avail_list(si); 907 spin_unlock(&swap_avail_lock); 908 } 909 910 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset, 911 unsigned int nr_entries) 912 { 913 unsigned int end = offset + nr_entries - 1; 914 915 if (offset == si->lowest_bit) 916 si->lowest_bit += nr_entries; 917 if (end == si->highest_bit) 918 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries); 919 WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries); 920 if (si->inuse_pages == si->pages) { 921 si->lowest_bit = si->max; 922 si->highest_bit = 0; 923 del_from_avail_list(si); 924 } 925 } 926 927 static void add_to_avail_list(struct swap_info_struct *si) 928 { 929 int nid; 930 931 spin_lock(&swap_avail_lock); 932 for_each_node(nid) 933 plist_add(&si->avail_lists[nid], &swap_avail_heads[nid]); 934 spin_unlock(&swap_avail_lock); 935 } 936 937 static void swap_range_free(struct swap_info_struct *si, unsigned long offset, 938 unsigned int nr_entries) 939 { 940 unsigned long begin = offset; 941 unsigned long end = offset + nr_entries - 1; 942 void (*swap_slot_free_notify)(struct block_device *, unsigned long); 943 unsigned int i; 944 945 /* 946 * Use atomic clear_bit operations only on zeromap instead of non-atomic 947 * bitmap_clear to prevent adjacent bits corruption due to simultaneous writes. 948 */ 949 for (i = 0; i < nr_entries; i++) 950 clear_bit(offset + i, si->zeromap); 951 952 if (offset < si->lowest_bit) 953 si->lowest_bit = offset; 954 if (end > si->highest_bit) { 955 bool was_full = !si->highest_bit; 956 957 WRITE_ONCE(si->highest_bit, end); 958 if (was_full && (si->flags & SWP_WRITEOK)) 959 add_to_avail_list(si); 960 } 961 if (si->flags & SWP_BLKDEV) 962 swap_slot_free_notify = 963 si->bdev->bd_disk->fops->swap_slot_free_notify; 964 else 965 swap_slot_free_notify = NULL; 966 while (offset <= end) { 967 arch_swap_invalidate_page(si->type, offset); 968 if (swap_slot_free_notify) 969 swap_slot_free_notify(si->bdev, offset); 970 offset++; 971 } 972 clear_shadow_from_swap_cache(si->type, begin, end); 973 974 /* 975 * Make sure that try_to_unuse() observes si->inuse_pages reaching 0 976 * only after the above cleanups are done. 977 */ 978 smp_wmb(); 979 atomic_long_add(nr_entries, &nr_swap_pages); 980 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries); 981 } 982 983 static void set_cluster_next(struct swap_info_struct *si, unsigned long next) 984 { 985 unsigned long prev; 986 987 if (!(si->flags & SWP_SOLIDSTATE)) { 988 si->cluster_next = next; 989 return; 990 } 991 992 prev = this_cpu_read(*si->cluster_next_cpu); 993 /* 994 * Cross the swap address space size aligned trunk, choose 995 * another trunk randomly to avoid lock contention on swap 996 * address space if possible. 997 */ 998 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) != 999 (next >> SWAP_ADDRESS_SPACE_SHIFT)) { 1000 /* No free swap slots available */ 1001 if (si->highest_bit <= si->lowest_bit) 1002 return; 1003 next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit); 1004 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES); 1005 next = max_t(unsigned int, next, si->lowest_bit); 1006 } 1007 this_cpu_write(*si->cluster_next_cpu, next); 1008 } 1009 1010 static bool swap_offset_available_and_locked(struct swap_info_struct *si, 1011 unsigned long offset) 1012 { 1013 if (data_race(!si->swap_map[offset])) { 1014 spin_lock(&si->lock); 1015 return true; 1016 } 1017 1018 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) { 1019 spin_lock(&si->lock); 1020 return true; 1021 } 1022 1023 return false; 1024 } 1025 1026 static int cluster_alloc_swap(struct swap_info_struct *si, 1027 unsigned char usage, int nr, 1028 swp_entry_t slots[], int order) 1029 { 1030 int n_ret = 0; 1031 1032 VM_BUG_ON(!si->cluster_info); 1033 1034 while (n_ret < nr) { 1035 unsigned long offset = cluster_alloc_swap_entry(si, order, usage); 1036 1037 if (!offset) 1038 break; 1039 slots[n_ret++] = swp_entry(si->type, offset); 1040 } 1041 1042 return n_ret; 1043 } 1044 1045 static int scan_swap_map_slots(struct swap_info_struct *si, 1046 unsigned char usage, int nr, 1047 swp_entry_t slots[], int order) 1048 { 1049 unsigned long offset; 1050 unsigned long scan_base; 1051 unsigned long last_in_cluster = 0; 1052 int latency_ration = LATENCY_LIMIT; 1053 unsigned int nr_pages = 1 << order; 1054 int n_ret = 0; 1055 bool scanned_many = false; 1056 1057 /* 1058 * We try to cluster swap pages by allocating them sequentially 1059 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this 1060 * way, however, we resort to first-free allocation, starting 1061 * a new cluster. This prevents us from scattering swap pages 1062 * all over the entire swap partition, so that we reduce 1063 * overall disk seek times between swap pages. -- sct 1064 * But we do now try to find an empty cluster. -Andrea 1065 * And we let swap pages go all over an SSD partition. Hugh 1066 */ 1067 1068 if (order > 0) { 1069 /* 1070 * Should not even be attempting large allocations when huge 1071 * page swap is disabled. Warn and fail the allocation. 1072 */ 1073 if (!IS_ENABLED(CONFIG_THP_SWAP) || 1074 nr_pages > SWAPFILE_CLUSTER) { 1075 VM_WARN_ON_ONCE(1); 1076 return 0; 1077 } 1078 1079 /* 1080 * Swapfile is not block device or not using clusters so unable 1081 * to allocate large entries. 1082 */ 1083 if (!(si->flags & SWP_BLKDEV) || !si->cluster_info) 1084 return 0; 1085 } 1086 1087 if (si->cluster_info) 1088 return cluster_alloc_swap(si, usage, nr, slots, order); 1089 1090 si->flags += SWP_SCANNING; 1091 1092 /* For HDD, sequential access is more important. */ 1093 scan_base = si->cluster_next; 1094 offset = scan_base; 1095 1096 if (unlikely(!si->cluster_nr--)) { 1097 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) { 1098 si->cluster_nr = SWAPFILE_CLUSTER - 1; 1099 goto checks; 1100 } 1101 1102 spin_unlock(&si->lock); 1103 1104 /* 1105 * If seek is expensive, start searching for new cluster from 1106 * start of partition, to minimize the span of allocated swap. 1107 */ 1108 scan_base = offset = si->lowest_bit; 1109 last_in_cluster = offset + SWAPFILE_CLUSTER - 1; 1110 1111 /* Locate the first empty (unaligned) cluster */ 1112 for (; last_in_cluster <= READ_ONCE(si->highest_bit); offset++) { 1113 if (si->swap_map[offset]) 1114 last_in_cluster = offset + SWAPFILE_CLUSTER; 1115 else if (offset == last_in_cluster) { 1116 spin_lock(&si->lock); 1117 offset -= SWAPFILE_CLUSTER - 1; 1118 si->cluster_next = offset; 1119 si->cluster_nr = SWAPFILE_CLUSTER - 1; 1120 goto checks; 1121 } 1122 if (unlikely(--latency_ration < 0)) { 1123 cond_resched(); 1124 latency_ration = LATENCY_LIMIT; 1125 } 1126 } 1127 1128 offset = scan_base; 1129 spin_lock(&si->lock); 1130 si->cluster_nr = SWAPFILE_CLUSTER - 1; 1131 } 1132 1133 checks: 1134 if (!(si->flags & SWP_WRITEOK)) 1135 goto no_page; 1136 if (!si->highest_bit) 1137 goto no_page; 1138 if (offset > si->highest_bit) 1139 scan_base = offset = si->lowest_bit; 1140 1141 /* reuse swap entry of cache-only swap if not busy. */ 1142 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) { 1143 int swap_was_freed; 1144 spin_unlock(&si->lock); 1145 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY | TTRS_DIRECT); 1146 spin_lock(&si->lock); 1147 /* entry was freed successfully, try to use this again */ 1148 if (swap_was_freed > 0) 1149 goto checks; 1150 goto scan; /* check next one */ 1151 } 1152 1153 if (si->swap_map[offset]) { 1154 if (!n_ret) 1155 goto scan; 1156 else 1157 goto done; 1158 } 1159 memset(si->swap_map + offset, usage, nr_pages); 1160 1161 swap_range_alloc(si, offset, nr_pages); 1162 slots[n_ret++] = swp_entry(si->type, offset); 1163 1164 /* got enough slots or reach max slots? */ 1165 if ((n_ret == nr) || (offset >= si->highest_bit)) 1166 goto done; 1167 1168 /* search for next available slot */ 1169 1170 /* time to take a break? */ 1171 if (unlikely(--latency_ration < 0)) { 1172 if (n_ret) 1173 goto done; 1174 spin_unlock(&si->lock); 1175 cond_resched(); 1176 spin_lock(&si->lock); 1177 latency_ration = LATENCY_LIMIT; 1178 } 1179 1180 if (si->cluster_nr && !si->swap_map[++offset]) { 1181 /* non-ssd case, still more slots in cluster? */ 1182 --si->cluster_nr; 1183 goto checks; 1184 } 1185 1186 /* 1187 * Even if there's no free clusters available (fragmented), 1188 * try to scan a little more quickly with lock held unless we 1189 * have scanned too many slots already. 1190 */ 1191 if (!scanned_many) { 1192 unsigned long scan_limit; 1193 1194 if (offset < scan_base) 1195 scan_limit = scan_base; 1196 else 1197 scan_limit = si->highest_bit; 1198 for (; offset <= scan_limit && --latency_ration > 0; 1199 offset++) { 1200 if (!si->swap_map[offset]) 1201 goto checks; 1202 } 1203 } 1204 1205 done: 1206 if (order == 0) 1207 set_cluster_next(si, offset + 1); 1208 si->flags -= SWP_SCANNING; 1209 return n_ret; 1210 1211 scan: 1212 VM_WARN_ON(order > 0); 1213 spin_unlock(&si->lock); 1214 while (++offset <= READ_ONCE(si->highest_bit)) { 1215 if (unlikely(--latency_ration < 0)) { 1216 cond_resched(); 1217 latency_ration = LATENCY_LIMIT; 1218 scanned_many = true; 1219 } 1220 if (swap_offset_available_and_locked(si, offset)) 1221 goto checks; 1222 } 1223 offset = si->lowest_bit; 1224 while (offset < scan_base) { 1225 if (unlikely(--latency_ration < 0)) { 1226 cond_resched(); 1227 latency_ration = LATENCY_LIMIT; 1228 scanned_many = true; 1229 } 1230 if (swap_offset_available_and_locked(si, offset)) 1231 goto checks; 1232 offset++; 1233 } 1234 spin_lock(&si->lock); 1235 1236 no_page: 1237 si->flags -= SWP_SCANNING; 1238 return n_ret; 1239 } 1240 1241 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_order) 1242 { 1243 int order = swap_entry_order(entry_order); 1244 unsigned long size = 1 << order; 1245 struct swap_info_struct *si, *next; 1246 long avail_pgs; 1247 int n_ret = 0; 1248 int node; 1249 1250 spin_lock(&swap_avail_lock); 1251 1252 avail_pgs = atomic_long_read(&nr_swap_pages) / size; 1253 if (avail_pgs <= 0) { 1254 spin_unlock(&swap_avail_lock); 1255 goto noswap; 1256 } 1257 1258 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs); 1259 1260 atomic_long_sub(n_goal * size, &nr_swap_pages); 1261 1262 start_over: 1263 node = numa_node_id(); 1264 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) { 1265 /* requeue si to after same-priority siblings */ 1266 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]); 1267 spin_unlock(&swap_avail_lock); 1268 spin_lock(&si->lock); 1269 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) { 1270 spin_lock(&swap_avail_lock); 1271 if (plist_node_empty(&si->avail_lists[node])) { 1272 spin_unlock(&si->lock); 1273 goto nextsi; 1274 } 1275 WARN(!si->highest_bit, 1276 "swap_info %d in list but !highest_bit\n", 1277 si->type); 1278 WARN(!(si->flags & SWP_WRITEOK), 1279 "swap_info %d in list but !SWP_WRITEOK\n", 1280 si->type); 1281 __del_from_avail_list(si); 1282 spin_unlock(&si->lock); 1283 goto nextsi; 1284 } 1285 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE, 1286 n_goal, swp_entries, order); 1287 spin_unlock(&si->lock); 1288 if (n_ret || size > 1) 1289 goto check_out; 1290 cond_resched(); 1291 1292 spin_lock(&swap_avail_lock); 1293 nextsi: 1294 /* 1295 * if we got here, it's likely that si was almost full before, 1296 * and since scan_swap_map_slots() can drop the si->lock, 1297 * multiple callers probably all tried to get a page from the 1298 * same si and it filled up before we could get one; or, the si 1299 * filled up between us dropping swap_avail_lock and taking 1300 * si->lock. Since we dropped the swap_avail_lock, the 1301 * swap_avail_head list may have been modified; so if next is 1302 * still in the swap_avail_head list then try it, otherwise 1303 * start over if we have not gotten any slots. 1304 */ 1305 if (plist_node_empty(&next->avail_lists[node])) 1306 goto start_over; 1307 } 1308 1309 spin_unlock(&swap_avail_lock); 1310 1311 check_out: 1312 if (n_ret < n_goal) 1313 atomic_long_add((long)(n_goal - n_ret) * size, 1314 &nr_swap_pages); 1315 noswap: 1316 return n_ret; 1317 } 1318 1319 static struct swap_info_struct *_swap_info_get(swp_entry_t entry) 1320 { 1321 struct swap_info_struct *si; 1322 unsigned long offset; 1323 1324 if (!entry.val) 1325 goto out; 1326 si = swp_swap_info(entry); 1327 if (!si) 1328 goto bad_nofile; 1329 if (data_race(!(si->flags & SWP_USED))) 1330 goto bad_device; 1331 offset = swp_offset(entry); 1332 if (offset >= si->max) 1333 goto bad_offset; 1334 if (data_race(!si->swap_map[swp_offset(entry)])) 1335 goto bad_free; 1336 return si; 1337 1338 bad_free: 1339 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val); 1340 goto out; 1341 bad_offset: 1342 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val); 1343 goto out; 1344 bad_device: 1345 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val); 1346 goto out; 1347 bad_nofile: 1348 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val); 1349 out: 1350 return NULL; 1351 } 1352 1353 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry, 1354 struct swap_info_struct *q) 1355 { 1356 struct swap_info_struct *p; 1357 1358 p = _swap_info_get(entry); 1359 1360 if (p != q) { 1361 if (q != NULL) 1362 spin_unlock(&q->lock); 1363 if (p != NULL) 1364 spin_lock(&p->lock); 1365 } 1366 return p; 1367 } 1368 1369 static unsigned char __swap_entry_free_locked(struct swap_info_struct *si, 1370 unsigned long offset, 1371 unsigned char usage) 1372 { 1373 unsigned char count; 1374 unsigned char has_cache; 1375 1376 count = si->swap_map[offset]; 1377 1378 has_cache = count & SWAP_HAS_CACHE; 1379 count &= ~SWAP_HAS_CACHE; 1380 1381 if (usage == SWAP_HAS_CACHE) { 1382 VM_BUG_ON(!has_cache); 1383 has_cache = 0; 1384 } else if (count == SWAP_MAP_SHMEM) { 1385 /* 1386 * Or we could insist on shmem.c using a special 1387 * swap_shmem_free() and free_shmem_swap_and_cache()... 1388 */ 1389 count = 0; 1390 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) { 1391 if (count == COUNT_CONTINUED) { 1392 if (swap_count_continued(si, offset, count)) 1393 count = SWAP_MAP_MAX | COUNT_CONTINUED; 1394 else 1395 count = SWAP_MAP_MAX; 1396 } else 1397 count--; 1398 } 1399 1400 usage = count | has_cache; 1401 if (usage) 1402 WRITE_ONCE(si->swap_map[offset], usage); 1403 else 1404 WRITE_ONCE(si->swap_map[offset], SWAP_HAS_CACHE); 1405 1406 return usage; 1407 } 1408 1409 /* 1410 * When we get a swap entry, if there aren't some other ways to 1411 * prevent swapoff, such as the folio in swap cache is locked, RCU 1412 * reader side is locked, etc., the swap entry may become invalid 1413 * because of swapoff. Then, we need to enclose all swap related 1414 * functions with get_swap_device() and put_swap_device(), unless the 1415 * swap functions call get/put_swap_device() by themselves. 1416 * 1417 * RCU reader side lock (including any spinlock) is sufficient to 1418 * prevent swapoff, because synchronize_rcu() is called in swapoff() 1419 * before freeing data structures. 1420 * 1421 * Check whether swap entry is valid in the swap device. If so, 1422 * return pointer to swap_info_struct, and keep the swap entry valid 1423 * via preventing the swap device from being swapoff, until 1424 * put_swap_device() is called. Otherwise return NULL. 1425 * 1426 * Notice that swapoff or swapoff+swapon can still happen before the 1427 * percpu_ref_tryget_live() in get_swap_device() or after the 1428 * percpu_ref_put() in put_swap_device() if there isn't any other way 1429 * to prevent swapoff. The caller must be prepared for that. For 1430 * example, the following situation is possible. 1431 * 1432 * CPU1 CPU2 1433 * do_swap_page() 1434 * ... swapoff+swapon 1435 * __read_swap_cache_async() 1436 * swapcache_prepare() 1437 * __swap_duplicate() 1438 * // check swap_map 1439 * // verify PTE not changed 1440 * 1441 * In __swap_duplicate(), the swap_map need to be checked before 1442 * changing partly because the specified swap entry may be for another 1443 * swap device which has been swapoff. And in do_swap_page(), after 1444 * the page is read from the swap device, the PTE is verified not 1445 * changed with the page table locked to check whether the swap device 1446 * has been swapoff or swapoff+swapon. 1447 */ 1448 struct swap_info_struct *get_swap_device(swp_entry_t entry) 1449 { 1450 struct swap_info_struct *si; 1451 unsigned long offset; 1452 1453 if (!entry.val) 1454 goto out; 1455 si = swp_swap_info(entry); 1456 if (!si) 1457 goto bad_nofile; 1458 if (!percpu_ref_tryget_live(&si->users)) 1459 goto out; 1460 /* 1461 * Guarantee the si->users are checked before accessing other 1462 * fields of swap_info_struct. 1463 * 1464 * Paired with the spin_unlock() after setup_swap_info() in 1465 * enable_swap_info(). 1466 */ 1467 smp_rmb(); 1468 offset = swp_offset(entry); 1469 if (offset >= si->max) 1470 goto put_out; 1471 1472 return si; 1473 bad_nofile: 1474 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val); 1475 out: 1476 return NULL; 1477 put_out: 1478 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val); 1479 percpu_ref_put(&si->users); 1480 return NULL; 1481 } 1482 1483 static unsigned char __swap_entry_free(struct swap_info_struct *si, 1484 swp_entry_t entry) 1485 { 1486 struct swap_cluster_info *ci; 1487 unsigned long offset = swp_offset(entry); 1488 unsigned char usage; 1489 1490 ci = lock_cluster_or_swap_info(si, offset); 1491 usage = __swap_entry_free_locked(si, offset, 1); 1492 unlock_cluster_or_swap_info(si, ci); 1493 if (!usage) 1494 free_swap_slot(entry); 1495 1496 return usage; 1497 } 1498 1499 static bool __swap_entries_free(struct swap_info_struct *si, 1500 swp_entry_t entry, int nr) 1501 { 1502 unsigned long offset = swp_offset(entry); 1503 unsigned int type = swp_type(entry); 1504 struct swap_cluster_info *ci; 1505 bool has_cache = false; 1506 unsigned char count; 1507 int i; 1508 1509 if (nr <= 1 || swap_count(data_race(si->swap_map[offset])) != 1) 1510 goto fallback; 1511 /* cross into another cluster */ 1512 if (nr > SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER) 1513 goto fallback; 1514 1515 ci = lock_cluster_or_swap_info(si, offset); 1516 if (!swap_is_last_map(si, offset, nr, &has_cache)) { 1517 unlock_cluster_or_swap_info(si, ci); 1518 goto fallback; 1519 } 1520 for (i = 0; i < nr; i++) 1521 WRITE_ONCE(si->swap_map[offset + i], SWAP_HAS_CACHE); 1522 unlock_cluster_or_swap_info(si, ci); 1523 1524 if (!has_cache) { 1525 for (i = 0; i < nr; i++) 1526 zswap_invalidate(swp_entry(si->type, offset + i)); 1527 spin_lock(&si->lock); 1528 swap_entry_range_free(si, entry, nr); 1529 spin_unlock(&si->lock); 1530 } 1531 return has_cache; 1532 1533 fallback: 1534 for (i = 0; i < nr; i++) { 1535 if (data_race(si->swap_map[offset + i])) { 1536 count = __swap_entry_free(si, swp_entry(type, offset + i)); 1537 if (count == SWAP_HAS_CACHE) 1538 has_cache = true; 1539 } else { 1540 WARN_ON_ONCE(1); 1541 } 1542 } 1543 return has_cache; 1544 } 1545 1546 /* 1547 * Drop the last HAS_CACHE flag of swap entries, caller have to 1548 * ensure all entries belong to the same cgroup. 1549 */ 1550 static void swap_entry_range_free(struct swap_info_struct *si, swp_entry_t entry, 1551 unsigned int nr_pages) 1552 { 1553 unsigned long offset = swp_offset(entry); 1554 unsigned char *map = si->swap_map + offset; 1555 unsigned char *map_end = map + nr_pages; 1556 struct swap_cluster_info *ci; 1557 1558 ci = lock_cluster(si, offset); 1559 do { 1560 VM_BUG_ON(*map != SWAP_HAS_CACHE); 1561 *map = 0; 1562 } while (++map < map_end); 1563 dec_cluster_info_page(si, ci, nr_pages); 1564 unlock_cluster(ci); 1565 1566 mem_cgroup_uncharge_swap(entry, nr_pages); 1567 swap_range_free(si, offset, nr_pages); 1568 } 1569 1570 static void cluster_swap_free_nr(struct swap_info_struct *si, 1571 unsigned long offset, int nr_pages, 1572 unsigned char usage) 1573 { 1574 struct swap_cluster_info *ci; 1575 DECLARE_BITMAP(to_free, BITS_PER_LONG) = { 0 }; 1576 int i, nr; 1577 1578 ci = lock_cluster_or_swap_info(si, offset); 1579 while (nr_pages) { 1580 nr = min(BITS_PER_LONG, nr_pages); 1581 for (i = 0; i < nr; i++) { 1582 if (!__swap_entry_free_locked(si, offset + i, usage)) 1583 bitmap_set(to_free, i, 1); 1584 } 1585 if (!bitmap_empty(to_free, BITS_PER_LONG)) { 1586 unlock_cluster_or_swap_info(si, ci); 1587 for_each_set_bit(i, to_free, BITS_PER_LONG) 1588 free_swap_slot(swp_entry(si->type, offset + i)); 1589 if (nr == nr_pages) 1590 return; 1591 bitmap_clear(to_free, 0, BITS_PER_LONG); 1592 ci = lock_cluster_or_swap_info(si, offset); 1593 } 1594 offset += nr; 1595 nr_pages -= nr; 1596 } 1597 unlock_cluster_or_swap_info(si, ci); 1598 } 1599 1600 /* 1601 * Caller has made sure that the swap device corresponding to entry 1602 * is still around or has not been recycled. 1603 */ 1604 void swap_free_nr(swp_entry_t entry, int nr_pages) 1605 { 1606 int nr; 1607 struct swap_info_struct *sis; 1608 unsigned long offset = swp_offset(entry); 1609 1610 sis = _swap_info_get(entry); 1611 if (!sis) 1612 return; 1613 1614 while (nr_pages) { 1615 nr = min_t(int, nr_pages, SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER); 1616 cluster_swap_free_nr(sis, offset, nr, 1); 1617 offset += nr; 1618 nr_pages -= nr; 1619 } 1620 } 1621 1622 /* 1623 * Called after dropping swapcache to decrease refcnt to swap entries. 1624 */ 1625 void put_swap_folio(struct folio *folio, swp_entry_t entry) 1626 { 1627 unsigned long offset = swp_offset(entry); 1628 struct swap_cluster_info *ci; 1629 struct swap_info_struct *si; 1630 int size = 1 << swap_entry_order(folio_order(folio)); 1631 1632 si = _swap_info_get(entry); 1633 if (!si) 1634 return; 1635 1636 ci = lock_cluster_or_swap_info(si, offset); 1637 if (size > 1 && swap_is_has_cache(si, offset, size)) { 1638 unlock_cluster_or_swap_info(si, ci); 1639 spin_lock(&si->lock); 1640 swap_entry_range_free(si, entry, size); 1641 spin_unlock(&si->lock); 1642 return; 1643 } 1644 for (int i = 0; i < size; i++, entry.val++) { 1645 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) { 1646 unlock_cluster_or_swap_info(si, ci); 1647 free_swap_slot(entry); 1648 if (i == size - 1) 1649 return; 1650 lock_cluster_or_swap_info(si, offset); 1651 } 1652 } 1653 unlock_cluster_or_swap_info(si, ci); 1654 } 1655 1656 static int swp_entry_cmp(const void *ent1, const void *ent2) 1657 { 1658 const swp_entry_t *e1 = ent1, *e2 = ent2; 1659 1660 return (int)swp_type(*e1) - (int)swp_type(*e2); 1661 } 1662 1663 void swapcache_free_entries(swp_entry_t *entries, int n) 1664 { 1665 struct swap_info_struct *p, *prev; 1666 int i; 1667 1668 if (n <= 0) 1669 return; 1670 1671 prev = NULL; 1672 p = NULL; 1673 1674 /* 1675 * Sort swap entries by swap device, so each lock is only taken once. 1676 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is 1677 * so low that it isn't necessary to optimize further. 1678 */ 1679 if (nr_swapfiles > 1) 1680 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL); 1681 for (i = 0; i < n; ++i) { 1682 p = swap_info_get_cont(entries[i], prev); 1683 if (p) 1684 swap_entry_range_free(p, entries[i], 1); 1685 prev = p; 1686 } 1687 if (p) 1688 spin_unlock(&p->lock); 1689 } 1690 1691 int __swap_count(swp_entry_t entry) 1692 { 1693 struct swap_info_struct *si = swp_swap_info(entry); 1694 pgoff_t offset = swp_offset(entry); 1695 1696 return swap_count(si->swap_map[offset]); 1697 } 1698 1699 /* 1700 * How many references to @entry are currently swapped out? 1701 * This does not give an exact answer when swap count is continued, 1702 * but does include the high COUNT_CONTINUED flag to allow for that. 1703 */ 1704 int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry) 1705 { 1706 pgoff_t offset = swp_offset(entry); 1707 struct swap_cluster_info *ci; 1708 int count; 1709 1710 ci = lock_cluster_or_swap_info(si, offset); 1711 count = swap_count(si->swap_map[offset]); 1712 unlock_cluster_or_swap_info(si, ci); 1713 return count; 1714 } 1715 1716 /* 1717 * How many references to @entry are currently swapped out? 1718 * This considers COUNT_CONTINUED so it returns exact answer. 1719 */ 1720 int swp_swapcount(swp_entry_t entry) 1721 { 1722 int count, tmp_count, n; 1723 struct swap_info_struct *si; 1724 struct swap_cluster_info *ci; 1725 struct page *page; 1726 pgoff_t offset; 1727 unsigned char *map; 1728 1729 si = _swap_info_get(entry); 1730 if (!si) 1731 return 0; 1732 1733 offset = swp_offset(entry); 1734 1735 ci = lock_cluster_or_swap_info(si, offset); 1736 1737 count = swap_count(si->swap_map[offset]); 1738 if (!(count & COUNT_CONTINUED)) 1739 goto out; 1740 1741 count &= ~COUNT_CONTINUED; 1742 n = SWAP_MAP_MAX + 1; 1743 1744 page = vmalloc_to_page(si->swap_map + offset); 1745 offset &= ~PAGE_MASK; 1746 VM_BUG_ON(page_private(page) != SWP_CONTINUED); 1747 1748 do { 1749 page = list_next_entry(page, lru); 1750 map = kmap_local_page(page); 1751 tmp_count = map[offset]; 1752 kunmap_local(map); 1753 1754 count += (tmp_count & ~COUNT_CONTINUED) * n; 1755 n *= (SWAP_CONT_MAX + 1); 1756 } while (tmp_count & COUNT_CONTINUED); 1757 out: 1758 unlock_cluster_or_swap_info(si, ci); 1759 return count; 1760 } 1761 1762 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si, 1763 swp_entry_t entry, int order) 1764 { 1765 struct swap_cluster_info *ci; 1766 unsigned char *map = si->swap_map; 1767 unsigned int nr_pages = 1 << order; 1768 unsigned long roffset = swp_offset(entry); 1769 unsigned long offset = round_down(roffset, nr_pages); 1770 int i; 1771 bool ret = false; 1772 1773 ci = lock_cluster_or_swap_info(si, offset); 1774 if (!ci || nr_pages == 1) { 1775 if (swap_count(map[roffset])) 1776 ret = true; 1777 goto unlock_out; 1778 } 1779 for (i = 0; i < nr_pages; i++) { 1780 if (swap_count(map[offset + i])) { 1781 ret = true; 1782 break; 1783 } 1784 } 1785 unlock_out: 1786 unlock_cluster_or_swap_info(si, ci); 1787 return ret; 1788 } 1789 1790 static bool folio_swapped(struct folio *folio) 1791 { 1792 swp_entry_t entry = folio->swap; 1793 struct swap_info_struct *si = _swap_info_get(entry); 1794 1795 if (!si) 1796 return false; 1797 1798 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio))) 1799 return swap_swapcount(si, entry) != 0; 1800 1801 return swap_page_trans_huge_swapped(si, entry, folio_order(folio)); 1802 } 1803 1804 static bool folio_swapcache_freeable(struct folio *folio) 1805 { 1806 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 1807 1808 if (!folio_test_swapcache(folio)) 1809 return false; 1810 if (folio_test_writeback(folio)) 1811 return false; 1812 1813 /* 1814 * Once hibernation has begun to create its image of memory, 1815 * there's a danger that one of the calls to folio_free_swap() 1816 * - most probably a call from __try_to_reclaim_swap() while 1817 * hibernation is allocating its own swap pages for the image, 1818 * but conceivably even a call from memory reclaim - will free 1819 * the swap from a folio which has already been recorded in the 1820 * image as a clean swapcache folio, and then reuse its swap for 1821 * another page of the image. On waking from hibernation, the 1822 * original folio might be freed under memory pressure, then 1823 * later read back in from swap, now with the wrong data. 1824 * 1825 * Hibernation suspends storage while it is writing the image 1826 * to disk so check that here. 1827 */ 1828 if (pm_suspended_storage()) 1829 return false; 1830 1831 return true; 1832 } 1833 1834 /** 1835 * folio_free_swap() - Free the swap space used for this folio. 1836 * @folio: The folio to remove. 1837 * 1838 * If swap is getting full, or if there are no more mappings of this folio, 1839 * then call folio_free_swap to free its swap space. 1840 * 1841 * Return: true if we were able to release the swap space. 1842 */ 1843 bool folio_free_swap(struct folio *folio) 1844 { 1845 if (!folio_swapcache_freeable(folio)) 1846 return false; 1847 if (folio_swapped(folio)) 1848 return false; 1849 1850 delete_from_swap_cache(folio); 1851 folio_set_dirty(folio); 1852 return true; 1853 } 1854 1855 /** 1856 * free_swap_and_cache_nr() - Release reference on range of swap entries and 1857 * reclaim their cache if no more references remain. 1858 * @entry: First entry of range. 1859 * @nr: Number of entries in range. 1860 * 1861 * For each swap entry in the contiguous range, release a reference. If any swap 1862 * entries become free, try to reclaim their underlying folios, if present. The 1863 * offset range is defined by [entry.offset, entry.offset + nr). 1864 */ 1865 void free_swap_and_cache_nr(swp_entry_t entry, int nr) 1866 { 1867 const unsigned long start_offset = swp_offset(entry); 1868 const unsigned long end_offset = start_offset + nr; 1869 struct swap_info_struct *si; 1870 bool any_only_cache = false; 1871 unsigned long offset; 1872 1873 if (non_swap_entry(entry)) 1874 return; 1875 1876 si = get_swap_device(entry); 1877 if (!si) 1878 return; 1879 1880 if (WARN_ON(end_offset > si->max)) 1881 goto out; 1882 1883 /* 1884 * First free all entries in the range. 1885 */ 1886 any_only_cache = __swap_entries_free(si, entry, nr); 1887 1888 /* 1889 * Short-circuit the below loop if none of the entries had their 1890 * reference drop to zero. 1891 */ 1892 if (!any_only_cache) 1893 goto out; 1894 1895 /* 1896 * Now go back over the range trying to reclaim the swap cache. This is 1897 * more efficient for large folios because we will only try to reclaim 1898 * the swap once per folio in the common case. If we do 1899 * __swap_entry_free() and __try_to_reclaim_swap() in the same loop, the 1900 * latter will get a reference and lock the folio for every individual 1901 * page but will only succeed once the swap slot for every subpage is 1902 * zero. 1903 */ 1904 for (offset = start_offset; offset < end_offset; offset += nr) { 1905 nr = 1; 1906 if (READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) { 1907 /* 1908 * Folios are always naturally aligned in swap so 1909 * advance forward to the next boundary. Zero means no 1910 * folio was found for the swap entry, so advance by 1 1911 * in this case. Negative value means folio was found 1912 * but could not be reclaimed. Here we can still advance 1913 * to the next boundary. 1914 */ 1915 nr = __try_to_reclaim_swap(si, offset, 1916 TTRS_UNMAPPED | TTRS_FULL); 1917 if (nr == 0) 1918 nr = 1; 1919 else if (nr < 0) 1920 nr = -nr; 1921 nr = ALIGN(offset + 1, nr) - offset; 1922 } 1923 } 1924 1925 out: 1926 put_swap_device(si); 1927 } 1928 1929 #ifdef CONFIG_HIBERNATION 1930 1931 swp_entry_t get_swap_page_of_type(int type) 1932 { 1933 struct swap_info_struct *si = swap_type_to_swap_info(type); 1934 swp_entry_t entry = {0}; 1935 1936 if (!si) 1937 goto fail; 1938 1939 /* This is called for allocating swap entry, not cache */ 1940 spin_lock(&si->lock); 1941 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry, 0)) 1942 atomic_long_dec(&nr_swap_pages); 1943 spin_unlock(&si->lock); 1944 fail: 1945 return entry; 1946 } 1947 1948 /* 1949 * Find the swap type that corresponds to given device (if any). 1950 * 1951 * @offset - number of the PAGE_SIZE-sized block of the device, starting 1952 * from 0, in which the swap header is expected to be located. 1953 * 1954 * This is needed for the suspend to disk (aka swsusp). 1955 */ 1956 int swap_type_of(dev_t device, sector_t offset) 1957 { 1958 int type; 1959 1960 if (!device) 1961 return -1; 1962 1963 spin_lock(&swap_lock); 1964 for (type = 0; type < nr_swapfiles; type++) { 1965 struct swap_info_struct *sis = swap_info[type]; 1966 1967 if (!(sis->flags & SWP_WRITEOK)) 1968 continue; 1969 1970 if (device == sis->bdev->bd_dev) { 1971 struct swap_extent *se = first_se(sis); 1972 1973 if (se->start_block == offset) { 1974 spin_unlock(&swap_lock); 1975 return type; 1976 } 1977 } 1978 } 1979 spin_unlock(&swap_lock); 1980 return -ENODEV; 1981 } 1982 1983 int find_first_swap(dev_t *device) 1984 { 1985 int type; 1986 1987 spin_lock(&swap_lock); 1988 for (type = 0; type < nr_swapfiles; type++) { 1989 struct swap_info_struct *sis = swap_info[type]; 1990 1991 if (!(sis->flags & SWP_WRITEOK)) 1992 continue; 1993 *device = sis->bdev->bd_dev; 1994 spin_unlock(&swap_lock); 1995 return type; 1996 } 1997 spin_unlock(&swap_lock); 1998 return -ENODEV; 1999 } 2000 2001 /* 2002 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev 2003 * corresponding to given index in swap_info (swap type). 2004 */ 2005 sector_t swapdev_block(int type, pgoff_t offset) 2006 { 2007 struct swap_info_struct *si = swap_type_to_swap_info(type); 2008 struct swap_extent *se; 2009 2010 if (!si || !(si->flags & SWP_WRITEOK)) 2011 return 0; 2012 se = offset_to_swap_extent(si, offset); 2013 return se->start_block + (offset - se->start_page); 2014 } 2015 2016 /* 2017 * Return either the total number of swap pages of given type, or the number 2018 * of free pages of that type (depending on @free) 2019 * 2020 * This is needed for software suspend 2021 */ 2022 unsigned int count_swap_pages(int type, int free) 2023 { 2024 unsigned int n = 0; 2025 2026 spin_lock(&swap_lock); 2027 if ((unsigned int)type < nr_swapfiles) { 2028 struct swap_info_struct *sis = swap_info[type]; 2029 2030 spin_lock(&sis->lock); 2031 if (sis->flags & SWP_WRITEOK) { 2032 n = sis->pages; 2033 if (free) 2034 n -= sis->inuse_pages; 2035 } 2036 spin_unlock(&sis->lock); 2037 } 2038 spin_unlock(&swap_lock); 2039 return n; 2040 } 2041 #endif /* CONFIG_HIBERNATION */ 2042 2043 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte) 2044 { 2045 return pte_same(pte_swp_clear_flags(pte), swp_pte); 2046 } 2047 2048 /* 2049 * No need to decide whether this PTE shares the swap entry with others, 2050 * just let do_wp_page work it out if a write is requested later - to 2051 * force COW, vm_page_prot omits write permission from any private vma. 2052 */ 2053 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd, 2054 unsigned long addr, swp_entry_t entry, struct folio *folio) 2055 { 2056 struct page *page; 2057 struct folio *swapcache; 2058 spinlock_t *ptl; 2059 pte_t *pte, new_pte, old_pte; 2060 bool hwpoisoned = false; 2061 int ret = 1; 2062 2063 swapcache = folio; 2064 folio = ksm_might_need_to_copy(folio, vma, addr); 2065 if (unlikely(!folio)) 2066 return -ENOMEM; 2067 else if (unlikely(folio == ERR_PTR(-EHWPOISON))) { 2068 hwpoisoned = true; 2069 folio = swapcache; 2070 } 2071 2072 page = folio_file_page(folio, swp_offset(entry)); 2073 if (PageHWPoison(page)) 2074 hwpoisoned = true; 2075 2076 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 2077 if (unlikely(!pte || !pte_same_as_swp(ptep_get(pte), 2078 swp_entry_to_pte(entry)))) { 2079 ret = 0; 2080 goto out; 2081 } 2082 2083 old_pte = ptep_get(pte); 2084 2085 if (unlikely(hwpoisoned || !folio_test_uptodate(folio))) { 2086 swp_entry_t swp_entry; 2087 2088 dec_mm_counter(vma->vm_mm, MM_SWAPENTS); 2089 if (hwpoisoned) { 2090 swp_entry = make_hwpoison_entry(page); 2091 } else { 2092 swp_entry = make_poisoned_swp_entry(); 2093 } 2094 new_pte = swp_entry_to_pte(swp_entry); 2095 ret = 0; 2096 goto setpte; 2097 } 2098 2099 /* 2100 * Some architectures may have to restore extra metadata to the page 2101 * when reading from swap. This metadata may be indexed by swap entry 2102 * so this must be called before swap_free(). 2103 */ 2104 arch_swap_restore(folio_swap(entry, folio), folio); 2105 2106 dec_mm_counter(vma->vm_mm, MM_SWAPENTS); 2107 inc_mm_counter(vma->vm_mm, MM_ANONPAGES); 2108 folio_get(folio); 2109 if (folio == swapcache) { 2110 rmap_t rmap_flags = RMAP_NONE; 2111 2112 /* 2113 * See do_swap_page(): writeback would be problematic. 2114 * However, we do a folio_wait_writeback() just before this 2115 * call and have the folio locked. 2116 */ 2117 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio); 2118 if (pte_swp_exclusive(old_pte)) 2119 rmap_flags |= RMAP_EXCLUSIVE; 2120 /* 2121 * We currently only expect small !anon folios, which are either 2122 * fully exclusive or fully shared. If we ever get large folios 2123 * here, we have to be careful. 2124 */ 2125 if (!folio_test_anon(folio)) { 2126 VM_WARN_ON_ONCE(folio_test_large(folio)); 2127 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio); 2128 folio_add_new_anon_rmap(folio, vma, addr, rmap_flags); 2129 } else { 2130 folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags); 2131 } 2132 } else { /* ksm created a completely new copy */ 2133 folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE); 2134 folio_add_lru_vma(folio, vma); 2135 } 2136 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot)); 2137 if (pte_swp_soft_dirty(old_pte)) 2138 new_pte = pte_mksoft_dirty(new_pte); 2139 if (pte_swp_uffd_wp(old_pte)) 2140 new_pte = pte_mkuffd_wp(new_pte); 2141 setpte: 2142 set_pte_at(vma->vm_mm, addr, pte, new_pte); 2143 swap_free(entry); 2144 out: 2145 if (pte) 2146 pte_unmap_unlock(pte, ptl); 2147 if (folio != swapcache) { 2148 folio_unlock(folio); 2149 folio_put(folio); 2150 } 2151 return ret; 2152 } 2153 2154 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd, 2155 unsigned long addr, unsigned long end, 2156 unsigned int type) 2157 { 2158 pte_t *pte = NULL; 2159 struct swap_info_struct *si; 2160 2161 si = swap_info[type]; 2162 do { 2163 struct folio *folio; 2164 unsigned long offset; 2165 unsigned char swp_count; 2166 swp_entry_t entry; 2167 int ret; 2168 pte_t ptent; 2169 2170 if (!pte++) { 2171 pte = pte_offset_map(pmd, addr); 2172 if (!pte) 2173 break; 2174 } 2175 2176 ptent = ptep_get_lockless(pte); 2177 2178 if (!is_swap_pte(ptent)) 2179 continue; 2180 2181 entry = pte_to_swp_entry(ptent); 2182 if (swp_type(entry) != type) 2183 continue; 2184 2185 offset = swp_offset(entry); 2186 pte_unmap(pte); 2187 pte = NULL; 2188 2189 folio = swap_cache_get_folio(entry, vma, addr); 2190 if (!folio) { 2191 struct vm_fault vmf = { 2192 .vma = vma, 2193 .address = addr, 2194 .real_address = addr, 2195 .pmd = pmd, 2196 }; 2197 2198 folio = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 2199 &vmf); 2200 } 2201 if (!folio) { 2202 swp_count = READ_ONCE(si->swap_map[offset]); 2203 if (swp_count == 0 || swp_count == SWAP_MAP_BAD) 2204 continue; 2205 return -ENOMEM; 2206 } 2207 2208 folio_lock(folio); 2209 folio_wait_writeback(folio); 2210 ret = unuse_pte(vma, pmd, addr, entry, folio); 2211 if (ret < 0) { 2212 folio_unlock(folio); 2213 folio_put(folio); 2214 return ret; 2215 } 2216 2217 folio_free_swap(folio); 2218 folio_unlock(folio); 2219 folio_put(folio); 2220 } while (addr += PAGE_SIZE, addr != end); 2221 2222 if (pte) 2223 pte_unmap(pte); 2224 return 0; 2225 } 2226 2227 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud, 2228 unsigned long addr, unsigned long end, 2229 unsigned int type) 2230 { 2231 pmd_t *pmd; 2232 unsigned long next; 2233 int ret; 2234 2235 pmd = pmd_offset(pud, addr); 2236 do { 2237 cond_resched(); 2238 next = pmd_addr_end(addr, end); 2239 ret = unuse_pte_range(vma, pmd, addr, next, type); 2240 if (ret) 2241 return ret; 2242 } while (pmd++, addr = next, addr != end); 2243 return 0; 2244 } 2245 2246 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d, 2247 unsigned long addr, unsigned long end, 2248 unsigned int type) 2249 { 2250 pud_t *pud; 2251 unsigned long next; 2252 int ret; 2253 2254 pud = pud_offset(p4d, addr); 2255 do { 2256 next = pud_addr_end(addr, end); 2257 if (pud_none_or_clear_bad(pud)) 2258 continue; 2259 ret = unuse_pmd_range(vma, pud, addr, next, type); 2260 if (ret) 2261 return ret; 2262 } while (pud++, addr = next, addr != end); 2263 return 0; 2264 } 2265 2266 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd, 2267 unsigned long addr, unsigned long end, 2268 unsigned int type) 2269 { 2270 p4d_t *p4d; 2271 unsigned long next; 2272 int ret; 2273 2274 p4d = p4d_offset(pgd, addr); 2275 do { 2276 next = p4d_addr_end(addr, end); 2277 if (p4d_none_or_clear_bad(p4d)) 2278 continue; 2279 ret = unuse_pud_range(vma, p4d, addr, next, type); 2280 if (ret) 2281 return ret; 2282 } while (p4d++, addr = next, addr != end); 2283 return 0; 2284 } 2285 2286 static int unuse_vma(struct vm_area_struct *vma, unsigned int type) 2287 { 2288 pgd_t *pgd; 2289 unsigned long addr, end, next; 2290 int ret; 2291 2292 addr = vma->vm_start; 2293 end = vma->vm_end; 2294 2295 pgd = pgd_offset(vma->vm_mm, addr); 2296 do { 2297 next = pgd_addr_end(addr, end); 2298 if (pgd_none_or_clear_bad(pgd)) 2299 continue; 2300 ret = unuse_p4d_range(vma, pgd, addr, next, type); 2301 if (ret) 2302 return ret; 2303 } while (pgd++, addr = next, addr != end); 2304 return 0; 2305 } 2306 2307 static int unuse_mm(struct mm_struct *mm, unsigned int type) 2308 { 2309 struct vm_area_struct *vma; 2310 int ret = 0; 2311 VMA_ITERATOR(vmi, mm, 0); 2312 2313 mmap_read_lock(mm); 2314 for_each_vma(vmi, vma) { 2315 if (vma->anon_vma) { 2316 ret = unuse_vma(vma, type); 2317 if (ret) 2318 break; 2319 } 2320 2321 cond_resched(); 2322 } 2323 mmap_read_unlock(mm); 2324 return ret; 2325 } 2326 2327 /* 2328 * Scan swap_map from current position to next entry still in use. 2329 * Return 0 if there are no inuse entries after prev till end of 2330 * the map. 2331 */ 2332 static unsigned int find_next_to_unuse(struct swap_info_struct *si, 2333 unsigned int prev) 2334 { 2335 unsigned int i; 2336 unsigned char count; 2337 2338 /* 2339 * No need for swap_lock here: we're just looking 2340 * for whether an entry is in use, not modifying it; false 2341 * hits are okay, and sys_swapoff() has already prevented new 2342 * allocations from this area (while holding swap_lock). 2343 */ 2344 for (i = prev + 1; i < si->max; i++) { 2345 count = READ_ONCE(si->swap_map[i]); 2346 if (count && swap_count(count) != SWAP_MAP_BAD) 2347 break; 2348 if ((i % LATENCY_LIMIT) == 0) 2349 cond_resched(); 2350 } 2351 2352 if (i == si->max) 2353 i = 0; 2354 2355 return i; 2356 } 2357 2358 static int try_to_unuse(unsigned int type) 2359 { 2360 struct mm_struct *prev_mm; 2361 struct mm_struct *mm; 2362 struct list_head *p; 2363 int retval = 0; 2364 struct swap_info_struct *si = swap_info[type]; 2365 struct folio *folio; 2366 swp_entry_t entry; 2367 unsigned int i; 2368 2369 if (!READ_ONCE(si->inuse_pages)) 2370 goto success; 2371 2372 retry: 2373 retval = shmem_unuse(type); 2374 if (retval) 2375 return retval; 2376 2377 prev_mm = &init_mm; 2378 mmget(prev_mm); 2379 2380 spin_lock(&mmlist_lock); 2381 p = &init_mm.mmlist; 2382 while (READ_ONCE(si->inuse_pages) && 2383 !signal_pending(current) && 2384 (p = p->next) != &init_mm.mmlist) { 2385 2386 mm = list_entry(p, struct mm_struct, mmlist); 2387 if (!mmget_not_zero(mm)) 2388 continue; 2389 spin_unlock(&mmlist_lock); 2390 mmput(prev_mm); 2391 prev_mm = mm; 2392 retval = unuse_mm(mm, type); 2393 if (retval) { 2394 mmput(prev_mm); 2395 return retval; 2396 } 2397 2398 /* 2399 * Make sure that we aren't completely killing 2400 * interactive performance. 2401 */ 2402 cond_resched(); 2403 spin_lock(&mmlist_lock); 2404 } 2405 spin_unlock(&mmlist_lock); 2406 2407 mmput(prev_mm); 2408 2409 i = 0; 2410 while (READ_ONCE(si->inuse_pages) && 2411 !signal_pending(current) && 2412 (i = find_next_to_unuse(si, i)) != 0) { 2413 2414 entry = swp_entry(type, i); 2415 folio = filemap_get_folio(swap_address_space(entry), swap_cache_index(entry)); 2416 if (IS_ERR(folio)) 2417 continue; 2418 2419 /* 2420 * It is conceivable that a racing task removed this folio from 2421 * swap cache just before we acquired the page lock. The folio 2422 * might even be back in swap cache on another swap area. But 2423 * that is okay, folio_free_swap() only removes stale folios. 2424 */ 2425 folio_lock(folio); 2426 folio_wait_writeback(folio); 2427 folio_free_swap(folio); 2428 folio_unlock(folio); 2429 folio_put(folio); 2430 } 2431 2432 /* 2433 * Lets check again to see if there are still swap entries in the map. 2434 * If yes, we would need to do retry the unuse logic again. 2435 * Under global memory pressure, swap entries can be reinserted back 2436 * into process space after the mmlist loop above passes over them. 2437 * 2438 * Limit the number of retries? No: when mmget_not_zero() 2439 * above fails, that mm is likely to be freeing swap from 2440 * exit_mmap(), which proceeds at its own independent pace; 2441 * and even shmem_writepage() could have been preempted after 2442 * folio_alloc_swap(), temporarily hiding that swap. It's easy 2443 * and robust (though cpu-intensive) just to keep retrying. 2444 */ 2445 if (READ_ONCE(si->inuse_pages)) { 2446 if (!signal_pending(current)) 2447 goto retry; 2448 return -EINTR; 2449 } 2450 2451 success: 2452 /* 2453 * Make sure that further cleanups after try_to_unuse() returns happen 2454 * after swap_range_free() reduces si->inuse_pages to 0. 2455 */ 2456 smp_mb(); 2457 return 0; 2458 } 2459 2460 /* 2461 * After a successful try_to_unuse, if no swap is now in use, we know 2462 * we can empty the mmlist. swap_lock must be held on entry and exit. 2463 * Note that mmlist_lock nests inside swap_lock, and an mm must be 2464 * added to the mmlist just after page_duplicate - before would be racy. 2465 */ 2466 static void drain_mmlist(void) 2467 { 2468 struct list_head *p, *next; 2469 unsigned int type; 2470 2471 for (type = 0; type < nr_swapfiles; type++) 2472 if (swap_info[type]->inuse_pages) 2473 return; 2474 spin_lock(&mmlist_lock); 2475 list_for_each_safe(p, next, &init_mm.mmlist) 2476 list_del_init(p); 2477 spin_unlock(&mmlist_lock); 2478 } 2479 2480 /* 2481 * Free all of a swapdev's extent information 2482 */ 2483 static void destroy_swap_extents(struct swap_info_struct *sis) 2484 { 2485 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) { 2486 struct rb_node *rb = sis->swap_extent_root.rb_node; 2487 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node); 2488 2489 rb_erase(rb, &sis->swap_extent_root); 2490 kfree(se); 2491 } 2492 2493 if (sis->flags & SWP_ACTIVATED) { 2494 struct file *swap_file = sis->swap_file; 2495 struct address_space *mapping = swap_file->f_mapping; 2496 2497 sis->flags &= ~SWP_ACTIVATED; 2498 if (mapping->a_ops->swap_deactivate) 2499 mapping->a_ops->swap_deactivate(swap_file); 2500 } 2501 } 2502 2503 /* 2504 * Add a block range (and the corresponding page range) into this swapdev's 2505 * extent tree. 2506 * 2507 * This function rather assumes that it is called in ascending page order. 2508 */ 2509 int 2510 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page, 2511 unsigned long nr_pages, sector_t start_block) 2512 { 2513 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL; 2514 struct swap_extent *se; 2515 struct swap_extent *new_se; 2516 2517 /* 2518 * place the new node at the right most since the 2519 * function is called in ascending page order. 2520 */ 2521 while (*link) { 2522 parent = *link; 2523 link = &parent->rb_right; 2524 } 2525 2526 if (parent) { 2527 se = rb_entry(parent, struct swap_extent, rb_node); 2528 BUG_ON(se->start_page + se->nr_pages != start_page); 2529 if (se->start_block + se->nr_pages == start_block) { 2530 /* Merge it */ 2531 se->nr_pages += nr_pages; 2532 return 0; 2533 } 2534 } 2535 2536 /* No merge, insert a new extent. */ 2537 new_se = kmalloc(sizeof(*se), GFP_KERNEL); 2538 if (new_se == NULL) 2539 return -ENOMEM; 2540 new_se->start_page = start_page; 2541 new_se->nr_pages = nr_pages; 2542 new_se->start_block = start_block; 2543 2544 rb_link_node(&new_se->rb_node, parent, link); 2545 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root); 2546 return 1; 2547 } 2548 EXPORT_SYMBOL_GPL(add_swap_extent); 2549 2550 /* 2551 * A `swap extent' is a simple thing which maps a contiguous range of pages 2552 * onto a contiguous range of disk blocks. A rbtree of swap extents is 2553 * built at swapon time and is then used at swap_writepage/swap_read_folio 2554 * time for locating where on disk a page belongs. 2555 * 2556 * If the swapfile is an S_ISBLK block device, a single extent is installed. 2557 * This is done so that the main operating code can treat S_ISBLK and S_ISREG 2558 * swap files identically. 2559 * 2560 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap 2561 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK 2562 * swapfiles are handled *identically* after swapon time. 2563 * 2564 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks 2565 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray 2566 * blocks are found which do not fall within the PAGE_SIZE alignment 2567 * requirements, they are simply tossed out - we will never use those blocks 2568 * for swapping. 2569 * 2570 * For all swap devices we set S_SWAPFILE across the life of the swapon. This 2571 * prevents users from writing to the swap device, which will corrupt memory. 2572 * 2573 * The amount of disk space which a single swap extent represents varies. 2574 * Typically it is in the 1-4 megabyte range. So we can have hundreds of 2575 * extents in the rbtree. - akpm. 2576 */ 2577 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span) 2578 { 2579 struct file *swap_file = sis->swap_file; 2580 struct address_space *mapping = swap_file->f_mapping; 2581 struct inode *inode = mapping->host; 2582 int ret; 2583 2584 if (S_ISBLK(inode->i_mode)) { 2585 ret = add_swap_extent(sis, 0, sis->max, 0); 2586 *span = sis->pages; 2587 return ret; 2588 } 2589 2590 if (mapping->a_ops->swap_activate) { 2591 ret = mapping->a_ops->swap_activate(sis, swap_file, span); 2592 if (ret < 0) 2593 return ret; 2594 sis->flags |= SWP_ACTIVATED; 2595 if ((sis->flags & SWP_FS_OPS) && 2596 sio_pool_init() != 0) { 2597 destroy_swap_extents(sis); 2598 return -ENOMEM; 2599 } 2600 return ret; 2601 } 2602 2603 return generic_swapfile_activate(sis, swap_file, span); 2604 } 2605 2606 static int swap_node(struct swap_info_struct *si) 2607 { 2608 struct block_device *bdev; 2609 2610 if (si->bdev) 2611 bdev = si->bdev; 2612 else 2613 bdev = si->swap_file->f_inode->i_sb->s_bdev; 2614 2615 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE; 2616 } 2617 2618 static void setup_swap_info(struct swap_info_struct *si, int prio, 2619 unsigned char *swap_map, 2620 struct swap_cluster_info *cluster_info, 2621 unsigned long *zeromap) 2622 { 2623 int i; 2624 2625 if (prio >= 0) 2626 si->prio = prio; 2627 else 2628 si->prio = --least_priority; 2629 /* 2630 * the plist prio is negated because plist ordering is 2631 * low-to-high, while swap ordering is high-to-low 2632 */ 2633 si->list.prio = -si->prio; 2634 for_each_node(i) { 2635 if (si->prio >= 0) 2636 si->avail_lists[i].prio = -si->prio; 2637 else { 2638 if (swap_node(si) == i) 2639 si->avail_lists[i].prio = 1; 2640 else 2641 si->avail_lists[i].prio = -si->prio; 2642 } 2643 } 2644 si->swap_map = swap_map; 2645 si->cluster_info = cluster_info; 2646 si->zeromap = zeromap; 2647 } 2648 2649 static void _enable_swap_info(struct swap_info_struct *si) 2650 { 2651 si->flags |= SWP_WRITEOK; 2652 atomic_long_add(si->pages, &nr_swap_pages); 2653 total_swap_pages += si->pages; 2654 2655 assert_spin_locked(&swap_lock); 2656 /* 2657 * both lists are plists, and thus priority ordered. 2658 * swap_active_head needs to be priority ordered for swapoff(), 2659 * which on removal of any swap_info_struct with an auto-assigned 2660 * (i.e. negative) priority increments the auto-assigned priority 2661 * of any lower-priority swap_info_structs. 2662 * swap_avail_head needs to be priority ordered for folio_alloc_swap(), 2663 * which allocates swap pages from the highest available priority 2664 * swap_info_struct. 2665 */ 2666 plist_add(&si->list, &swap_active_head); 2667 2668 /* add to available list iff swap device is not full */ 2669 if (si->highest_bit) 2670 add_to_avail_list(si); 2671 } 2672 2673 static void enable_swap_info(struct swap_info_struct *si, int prio, 2674 unsigned char *swap_map, 2675 struct swap_cluster_info *cluster_info, 2676 unsigned long *zeromap) 2677 { 2678 spin_lock(&swap_lock); 2679 spin_lock(&si->lock); 2680 setup_swap_info(si, prio, swap_map, cluster_info, zeromap); 2681 spin_unlock(&si->lock); 2682 spin_unlock(&swap_lock); 2683 /* 2684 * Finished initializing swap device, now it's safe to reference it. 2685 */ 2686 percpu_ref_resurrect(&si->users); 2687 spin_lock(&swap_lock); 2688 spin_lock(&si->lock); 2689 _enable_swap_info(si); 2690 spin_unlock(&si->lock); 2691 spin_unlock(&swap_lock); 2692 } 2693 2694 static void reinsert_swap_info(struct swap_info_struct *si) 2695 { 2696 spin_lock(&swap_lock); 2697 spin_lock(&si->lock); 2698 setup_swap_info(si, si->prio, si->swap_map, si->cluster_info, si->zeromap); 2699 _enable_swap_info(si); 2700 spin_unlock(&si->lock); 2701 spin_unlock(&swap_lock); 2702 } 2703 2704 static bool __has_usable_swap(void) 2705 { 2706 return !plist_head_empty(&swap_active_head); 2707 } 2708 2709 bool has_usable_swap(void) 2710 { 2711 bool ret; 2712 2713 spin_lock(&swap_lock); 2714 ret = __has_usable_swap(); 2715 spin_unlock(&swap_lock); 2716 return ret; 2717 } 2718 2719 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile) 2720 { 2721 struct swap_info_struct *p = NULL; 2722 unsigned char *swap_map; 2723 unsigned long *zeromap; 2724 struct swap_cluster_info *cluster_info; 2725 struct file *swap_file, *victim; 2726 struct address_space *mapping; 2727 struct inode *inode; 2728 struct filename *pathname; 2729 int err, found = 0; 2730 2731 if (!capable(CAP_SYS_ADMIN)) 2732 return -EPERM; 2733 2734 BUG_ON(!current->mm); 2735 2736 pathname = getname(specialfile); 2737 if (IS_ERR(pathname)) 2738 return PTR_ERR(pathname); 2739 2740 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0); 2741 err = PTR_ERR(victim); 2742 if (IS_ERR(victim)) 2743 goto out; 2744 2745 mapping = victim->f_mapping; 2746 spin_lock(&swap_lock); 2747 plist_for_each_entry(p, &swap_active_head, list) { 2748 if (p->flags & SWP_WRITEOK) { 2749 if (p->swap_file->f_mapping == mapping) { 2750 found = 1; 2751 break; 2752 } 2753 } 2754 } 2755 if (!found) { 2756 err = -EINVAL; 2757 spin_unlock(&swap_lock); 2758 goto out_dput; 2759 } 2760 if (!security_vm_enough_memory_mm(current->mm, p->pages)) 2761 vm_unacct_memory(p->pages); 2762 else { 2763 err = -ENOMEM; 2764 spin_unlock(&swap_lock); 2765 goto out_dput; 2766 } 2767 spin_lock(&p->lock); 2768 del_from_avail_list(p); 2769 if (p->prio < 0) { 2770 struct swap_info_struct *si = p; 2771 int nid; 2772 2773 plist_for_each_entry_continue(si, &swap_active_head, list) { 2774 si->prio++; 2775 si->list.prio--; 2776 for_each_node(nid) { 2777 if (si->avail_lists[nid].prio != 1) 2778 si->avail_lists[nid].prio--; 2779 } 2780 } 2781 least_priority++; 2782 } 2783 plist_del(&p->list, &swap_active_head); 2784 atomic_long_sub(p->pages, &nr_swap_pages); 2785 total_swap_pages -= p->pages; 2786 p->flags &= ~SWP_WRITEOK; 2787 spin_unlock(&p->lock); 2788 spin_unlock(&swap_lock); 2789 2790 disable_swap_slots_cache_lock(); 2791 2792 set_current_oom_origin(); 2793 err = try_to_unuse(p->type); 2794 clear_current_oom_origin(); 2795 2796 if (err) { 2797 /* re-insert swap space back into swap_list */ 2798 reinsert_swap_info(p); 2799 reenable_swap_slots_cache_unlock(); 2800 goto out_dput; 2801 } 2802 2803 reenable_swap_slots_cache_unlock(); 2804 2805 /* 2806 * Wait for swap operations protected by get/put_swap_device() 2807 * to complete. Because of synchronize_rcu() here, all swap 2808 * operations protected by RCU reader side lock (including any 2809 * spinlock) will be waited too. This makes it easy to 2810 * prevent folio_test_swapcache() and the following swap cache 2811 * operations from racing with swapoff. 2812 */ 2813 percpu_ref_kill(&p->users); 2814 synchronize_rcu(); 2815 wait_for_completion(&p->comp); 2816 2817 flush_work(&p->discard_work); 2818 2819 destroy_swap_extents(p); 2820 if (p->flags & SWP_CONTINUED) 2821 free_swap_count_continuations(p); 2822 2823 if (!p->bdev || !bdev_nonrot(p->bdev)) 2824 atomic_dec(&nr_rotate_swap); 2825 2826 mutex_lock(&swapon_mutex); 2827 spin_lock(&swap_lock); 2828 spin_lock(&p->lock); 2829 drain_mmlist(); 2830 2831 /* wait for anyone still in scan_swap_map_slots */ 2832 p->highest_bit = 0; /* cuts scans short */ 2833 while (p->flags >= SWP_SCANNING) { 2834 spin_unlock(&p->lock); 2835 spin_unlock(&swap_lock); 2836 schedule_timeout_uninterruptible(1); 2837 spin_lock(&swap_lock); 2838 spin_lock(&p->lock); 2839 } 2840 2841 swap_file = p->swap_file; 2842 p->swap_file = NULL; 2843 p->max = 0; 2844 swap_map = p->swap_map; 2845 p->swap_map = NULL; 2846 zeromap = p->zeromap; 2847 p->zeromap = NULL; 2848 cluster_info = p->cluster_info; 2849 p->cluster_info = NULL; 2850 spin_unlock(&p->lock); 2851 spin_unlock(&swap_lock); 2852 arch_swap_invalidate_area(p->type); 2853 zswap_swapoff(p->type); 2854 mutex_unlock(&swapon_mutex); 2855 free_percpu(p->percpu_cluster); 2856 p->percpu_cluster = NULL; 2857 free_percpu(p->cluster_next_cpu); 2858 p->cluster_next_cpu = NULL; 2859 vfree(swap_map); 2860 kvfree(zeromap); 2861 kvfree(cluster_info); 2862 /* Destroy swap account information */ 2863 swap_cgroup_swapoff(p->type); 2864 exit_swap_address_space(p->type); 2865 2866 inode = mapping->host; 2867 2868 inode_lock(inode); 2869 inode->i_flags &= ~S_SWAPFILE; 2870 inode_unlock(inode); 2871 filp_close(swap_file, NULL); 2872 2873 /* 2874 * Clear the SWP_USED flag after all resources are freed so that swapon 2875 * can reuse this swap_info in alloc_swap_info() safely. It is ok to 2876 * not hold p->lock after we cleared its SWP_WRITEOK. 2877 */ 2878 spin_lock(&swap_lock); 2879 p->flags = 0; 2880 spin_unlock(&swap_lock); 2881 2882 err = 0; 2883 atomic_inc(&proc_poll_event); 2884 wake_up_interruptible(&proc_poll_wait); 2885 2886 out_dput: 2887 filp_close(victim, NULL); 2888 out: 2889 putname(pathname); 2890 return err; 2891 } 2892 2893 #ifdef CONFIG_PROC_FS 2894 static __poll_t swaps_poll(struct file *file, poll_table *wait) 2895 { 2896 struct seq_file *seq = file->private_data; 2897 2898 poll_wait(file, &proc_poll_wait, wait); 2899 2900 if (seq->poll_event != atomic_read(&proc_poll_event)) { 2901 seq->poll_event = atomic_read(&proc_poll_event); 2902 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI; 2903 } 2904 2905 return EPOLLIN | EPOLLRDNORM; 2906 } 2907 2908 /* iterator */ 2909 static void *swap_start(struct seq_file *swap, loff_t *pos) 2910 { 2911 struct swap_info_struct *si; 2912 int type; 2913 loff_t l = *pos; 2914 2915 mutex_lock(&swapon_mutex); 2916 2917 if (!l) 2918 return SEQ_START_TOKEN; 2919 2920 for (type = 0; (si = swap_type_to_swap_info(type)); type++) { 2921 if (!(si->flags & SWP_USED) || !si->swap_map) 2922 continue; 2923 if (!--l) 2924 return si; 2925 } 2926 2927 return NULL; 2928 } 2929 2930 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos) 2931 { 2932 struct swap_info_struct *si = v; 2933 int type; 2934 2935 if (v == SEQ_START_TOKEN) 2936 type = 0; 2937 else 2938 type = si->type + 1; 2939 2940 ++(*pos); 2941 for (; (si = swap_type_to_swap_info(type)); type++) { 2942 if (!(si->flags & SWP_USED) || !si->swap_map) 2943 continue; 2944 return si; 2945 } 2946 2947 return NULL; 2948 } 2949 2950 static void swap_stop(struct seq_file *swap, void *v) 2951 { 2952 mutex_unlock(&swapon_mutex); 2953 } 2954 2955 static int swap_show(struct seq_file *swap, void *v) 2956 { 2957 struct swap_info_struct *si = v; 2958 struct file *file; 2959 int len; 2960 unsigned long bytes, inuse; 2961 2962 if (si == SEQ_START_TOKEN) { 2963 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n"); 2964 return 0; 2965 } 2966 2967 bytes = K(si->pages); 2968 inuse = K(READ_ONCE(si->inuse_pages)); 2969 2970 file = si->swap_file; 2971 len = seq_file_path(swap, file, " \t\n\\"); 2972 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n", 2973 len < 40 ? 40 - len : 1, " ", 2974 S_ISBLK(file_inode(file)->i_mode) ? 2975 "partition" : "file\t", 2976 bytes, bytes < 10000000 ? "\t" : "", 2977 inuse, inuse < 10000000 ? "\t" : "", 2978 si->prio); 2979 return 0; 2980 } 2981 2982 static const struct seq_operations swaps_op = { 2983 .start = swap_start, 2984 .next = swap_next, 2985 .stop = swap_stop, 2986 .show = swap_show 2987 }; 2988 2989 static int swaps_open(struct inode *inode, struct file *file) 2990 { 2991 struct seq_file *seq; 2992 int ret; 2993 2994 ret = seq_open(file, &swaps_op); 2995 if (ret) 2996 return ret; 2997 2998 seq = file->private_data; 2999 seq->poll_event = atomic_read(&proc_poll_event); 3000 return 0; 3001 } 3002 3003 static const struct proc_ops swaps_proc_ops = { 3004 .proc_flags = PROC_ENTRY_PERMANENT, 3005 .proc_open = swaps_open, 3006 .proc_read = seq_read, 3007 .proc_lseek = seq_lseek, 3008 .proc_release = seq_release, 3009 .proc_poll = swaps_poll, 3010 }; 3011 3012 static int __init procswaps_init(void) 3013 { 3014 proc_create("swaps", 0, NULL, &swaps_proc_ops); 3015 return 0; 3016 } 3017 __initcall(procswaps_init); 3018 #endif /* CONFIG_PROC_FS */ 3019 3020 #ifdef MAX_SWAPFILES_CHECK 3021 static int __init max_swapfiles_check(void) 3022 { 3023 MAX_SWAPFILES_CHECK(); 3024 return 0; 3025 } 3026 late_initcall(max_swapfiles_check); 3027 #endif 3028 3029 static struct swap_info_struct *alloc_swap_info(void) 3030 { 3031 struct swap_info_struct *p; 3032 struct swap_info_struct *defer = NULL; 3033 unsigned int type; 3034 int i; 3035 3036 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL); 3037 if (!p) 3038 return ERR_PTR(-ENOMEM); 3039 3040 if (percpu_ref_init(&p->users, swap_users_ref_free, 3041 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) { 3042 kvfree(p); 3043 return ERR_PTR(-ENOMEM); 3044 } 3045 3046 spin_lock(&swap_lock); 3047 for (type = 0; type < nr_swapfiles; type++) { 3048 if (!(swap_info[type]->flags & SWP_USED)) 3049 break; 3050 } 3051 if (type >= MAX_SWAPFILES) { 3052 spin_unlock(&swap_lock); 3053 percpu_ref_exit(&p->users); 3054 kvfree(p); 3055 return ERR_PTR(-EPERM); 3056 } 3057 if (type >= nr_swapfiles) { 3058 p->type = type; 3059 /* 3060 * Publish the swap_info_struct after initializing it. 3061 * Note that kvzalloc() above zeroes all its fields. 3062 */ 3063 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */ 3064 nr_swapfiles++; 3065 } else { 3066 defer = p; 3067 p = swap_info[type]; 3068 /* 3069 * Do not memset this entry: a racing procfs swap_next() 3070 * would be relying on p->type to remain valid. 3071 */ 3072 } 3073 p->swap_extent_root = RB_ROOT; 3074 plist_node_init(&p->list, 0); 3075 for_each_node(i) 3076 plist_node_init(&p->avail_lists[i], 0); 3077 p->flags = SWP_USED; 3078 spin_unlock(&swap_lock); 3079 if (defer) { 3080 percpu_ref_exit(&defer->users); 3081 kvfree(defer); 3082 } 3083 spin_lock_init(&p->lock); 3084 spin_lock_init(&p->cont_lock); 3085 init_completion(&p->comp); 3086 3087 return p; 3088 } 3089 3090 static int claim_swapfile(struct swap_info_struct *si, struct inode *inode) 3091 { 3092 if (S_ISBLK(inode->i_mode)) { 3093 si->bdev = I_BDEV(inode); 3094 /* 3095 * Zoned block devices contain zones that have a sequential 3096 * write only restriction. Hence zoned block devices are not 3097 * suitable for swapping. Disallow them here. 3098 */ 3099 if (bdev_is_zoned(si->bdev)) 3100 return -EINVAL; 3101 si->flags |= SWP_BLKDEV; 3102 } else if (S_ISREG(inode->i_mode)) { 3103 si->bdev = inode->i_sb->s_bdev; 3104 } 3105 3106 return 0; 3107 } 3108 3109 3110 /* 3111 * Find out how many pages are allowed for a single swap device. There 3112 * are two limiting factors: 3113 * 1) the number of bits for the swap offset in the swp_entry_t type, and 3114 * 2) the number of bits in the swap pte, as defined by the different 3115 * architectures. 3116 * 3117 * In order to find the largest possible bit mask, a swap entry with 3118 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte, 3119 * decoded to a swp_entry_t again, and finally the swap offset is 3120 * extracted. 3121 * 3122 * This will mask all the bits from the initial ~0UL mask that can't 3123 * be encoded in either the swp_entry_t or the architecture definition 3124 * of a swap pte. 3125 */ 3126 unsigned long generic_max_swapfile_size(void) 3127 { 3128 return swp_offset(pte_to_swp_entry( 3129 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1; 3130 } 3131 3132 /* Can be overridden by an architecture for additional checks. */ 3133 __weak unsigned long arch_max_swapfile_size(void) 3134 { 3135 return generic_max_swapfile_size(); 3136 } 3137 3138 static unsigned long read_swap_header(struct swap_info_struct *si, 3139 union swap_header *swap_header, 3140 struct inode *inode) 3141 { 3142 int i; 3143 unsigned long maxpages; 3144 unsigned long swapfilepages; 3145 unsigned long last_page; 3146 3147 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) { 3148 pr_err("Unable to find swap-space signature\n"); 3149 return 0; 3150 } 3151 3152 /* swap partition endianness hack... */ 3153 if (swab32(swap_header->info.version) == 1) { 3154 swab32s(&swap_header->info.version); 3155 swab32s(&swap_header->info.last_page); 3156 swab32s(&swap_header->info.nr_badpages); 3157 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES) 3158 return 0; 3159 for (i = 0; i < swap_header->info.nr_badpages; i++) 3160 swab32s(&swap_header->info.badpages[i]); 3161 } 3162 /* Check the swap header's sub-version */ 3163 if (swap_header->info.version != 1) { 3164 pr_warn("Unable to handle swap header version %d\n", 3165 swap_header->info.version); 3166 return 0; 3167 } 3168 3169 si->lowest_bit = 1; 3170 si->cluster_next = 1; 3171 si->cluster_nr = 0; 3172 3173 maxpages = swapfile_maximum_size; 3174 last_page = swap_header->info.last_page; 3175 if (!last_page) { 3176 pr_warn("Empty swap-file\n"); 3177 return 0; 3178 } 3179 if (last_page > maxpages) { 3180 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n", 3181 K(maxpages), K(last_page)); 3182 } 3183 if (maxpages > last_page) { 3184 maxpages = last_page + 1; 3185 /* p->max is an unsigned int: don't overflow it */ 3186 if ((unsigned int)maxpages == 0) 3187 maxpages = UINT_MAX; 3188 } 3189 si->highest_bit = maxpages - 1; 3190 3191 if (!maxpages) 3192 return 0; 3193 swapfilepages = i_size_read(inode) >> PAGE_SHIFT; 3194 if (swapfilepages && maxpages > swapfilepages) { 3195 pr_warn("Swap area shorter than signature indicates\n"); 3196 return 0; 3197 } 3198 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode)) 3199 return 0; 3200 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES) 3201 return 0; 3202 3203 return maxpages; 3204 } 3205 3206 #define SWAP_CLUSTER_INFO_COLS \ 3207 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info)) 3208 #define SWAP_CLUSTER_SPACE_COLS \ 3209 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER) 3210 #define SWAP_CLUSTER_COLS \ 3211 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS) 3212 3213 static int setup_swap_map_and_extents(struct swap_info_struct *si, 3214 union swap_header *swap_header, 3215 unsigned char *swap_map, 3216 unsigned long maxpages, 3217 sector_t *span) 3218 { 3219 unsigned int nr_good_pages; 3220 unsigned long i; 3221 int nr_extents; 3222 3223 nr_good_pages = maxpages - 1; /* omit header page */ 3224 3225 for (i = 0; i < swap_header->info.nr_badpages; i++) { 3226 unsigned int page_nr = swap_header->info.badpages[i]; 3227 if (page_nr == 0 || page_nr > swap_header->info.last_page) 3228 return -EINVAL; 3229 if (page_nr < maxpages) { 3230 swap_map[page_nr] = SWAP_MAP_BAD; 3231 nr_good_pages--; 3232 } 3233 } 3234 3235 if (nr_good_pages) { 3236 swap_map[0] = SWAP_MAP_BAD; 3237 si->max = maxpages; 3238 si->pages = nr_good_pages; 3239 nr_extents = setup_swap_extents(si, span); 3240 if (nr_extents < 0) 3241 return nr_extents; 3242 nr_good_pages = si->pages; 3243 } 3244 if (!nr_good_pages) { 3245 pr_warn("Empty swap-file\n"); 3246 return -EINVAL; 3247 } 3248 3249 return nr_extents; 3250 } 3251 3252 static struct swap_cluster_info *setup_clusters(struct swap_info_struct *si, 3253 union swap_header *swap_header, 3254 unsigned long maxpages) 3255 { 3256 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER); 3257 unsigned long col = si->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS; 3258 struct swap_cluster_info *cluster_info; 3259 unsigned long i, j, k, idx; 3260 int cpu, err = -ENOMEM; 3261 3262 cluster_info = kvcalloc(nr_clusters, sizeof(*cluster_info), GFP_KERNEL); 3263 if (!cluster_info) 3264 goto err; 3265 3266 for (i = 0; i < nr_clusters; i++) 3267 spin_lock_init(&cluster_info[i].lock); 3268 3269 si->cluster_next_cpu = alloc_percpu(unsigned int); 3270 if (!si->cluster_next_cpu) 3271 goto err_free; 3272 3273 /* Random start position to help with wear leveling */ 3274 for_each_possible_cpu(cpu) 3275 per_cpu(*si->cluster_next_cpu, cpu) = 3276 get_random_u32_inclusive(1, si->highest_bit); 3277 3278 si->percpu_cluster = alloc_percpu(struct percpu_cluster); 3279 if (!si->percpu_cluster) 3280 goto err_free; 3281 3282 for_each_possible_cpu(cpu) { 3283 struct percpu_cluster *cluster; 3284 3285 cluster = per_cpu_ptr(si->percpu_cluster, cpu); 3286 for (i = 0; i < SWAP_NR_ORDERS; i++) 3287 cluster->next[i] = SWAP_NEXT_INVALID; 3288 } 3289 3290 /* 3291 * Mark unusable pages as unavailable. The clusters aren't 3292 * marked free yet, so no list operations are involved yet. 3293 * 3294 * See setup_swap_map_and_extents(): header page, bad pages, 3295 * and the EOF part of the last cluster. 3296 */ 3297 inc_cluster_info_page(si, cluster_info, 0); 3298 for (i = 0; i < swap_header->info.nr_badpages; i++) 3299 inc_cluster_info_page(si, cluster_info, 3300 swap_header->info.badpages[i]); 3301 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++) 3302 inc_cluster_info_page(si, cluster_info, i); 3303 3304 INIT_LIST_HEAD(&si->free_clusters); 3305 INIT_LIST_HEAD(&si->full_clusters); 3306 INIT_LIST_HEAD(&si->discard_clusters); 3307 3308 for (i = 0; i < SWAP_NR_ORDERS; i++) { 3309 INIT_LIST_HEAD(&si->nonfull_clusters[i]); 3310 INIT_LIST_HEAD(&si->frag_clusters[i]); 3311 si->frag_cluster_nr[i] = 0; 3312 } 3313 3314 /* 3315 * Reduce false cache line sharing between cluster_info and 3316 * sharing same address space. 3317 */ 3318 for (k = 0; k < SWAP_CLUSTER_COLS; k++) { 3319 j = (k + col) % SWAP_CLUSTER_COLS; 3320 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) { 3321 struct swap_cluster_info *ci; 3322 idx = i * SWAP_CLUSTER_COLS + j; 3323 ci = cluster_info + idx; 3324 if (idx >= nr_clusters) 3325 continue; 3326 if (ci->count) { 3327 ci->flags = CLUSTER_FLAG_NONFULL; 3328 list_add_tail(&ci->list, &si->nonfull_clusters[0]); 3329 continue; 3330 } 3331 ci->flags = CLUSTER_FLAG_FREE; 3332 list_add_tail(&ci->list, &si->free_clusters); 3333 } 3334 } 3335 3336 return cluster_info; 3337 3338 err_free: 3339 kvfree(cluster_info); 3340 err: 3341 return ERR_PTR(err); 3342 } 3343 3344 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags) 3345 { 3346 struct swap_info_struct *si; 3347 struct filename *name; 3348 struct file *swap_file = NULL; 3349 struct address_space *mapping; 3350 struct dentry *dentry; 3351 int prio; 3352 int error; 3353 union swap_header *swap_header; 3354 int nr_extents; 3355 sector_t span; 3356 unsigned long maxpages; 3357 unsigned char *swap_map = NULL; 3358 unsigned long *zeromap = NULL; 3359 struct swap_cluster_info *cluster_info = NULL; 3360 struct folio *folio = NULL; 3361 struct inode *inode = NULL; 3362 bool inced_nr_rotate_swap = false; 3363 3364 if (swap_flags & ~SWAP_FLAGS_VALID) 3365 return -EINVAL; 3366 3367 if (!capable(CAP_SYS_ADMIN)) 3368 return -EPERM; 3369 3370 if (!swap_avail_heads) 3371 return -ENOMEM; 3372 3373 si = alloc_swap_info(); 3374 if (IS_ERR(si)) 3375 return PTR_ERR(si); 3376 3377 INIT_WORK(&si->discard_work, swap_discard_work); 3378 3379 name = getname(specialfile); 3380 if (IS_ERR(name)) { 3381 error = PTR_ERR(name); 3382 name = NULL; 3383 goto bad_swap; 3384 } 3385 swap_file = file_open_name(name, O_RDWR | O_LARGEFILE | O_EXCL, 0); 3386 if (IS_ERR(swap_file)) { 3387 error = PTR_ERR(swap_file); 3388 swap_file = NULL; 3389 goto bad_swap; 3390 } 3391 3392 si->swap_file = swap_file; 3393 mapping = swap_file->f_mapping; 3394 dentry = swap_file->f_path.dentry; 3395 inode = mapping->host; 3396 3397 error = claim_swapfile(si, inode); 3398 if (unlikely(error)) 3399 goto bad_swap; 3400 3401 inode_lock(inode); 3402 if (d_unlinked(dentry) || cant_mount(dentry)) { 3403 error = -ENOENT; 3404 goto bad_swap_unlock_inode; 3405 } 3406 if (IS_SWAPFILE(inode)) { 3407 error = -EBUSY; 3408 goto bad_swap_unlock_inode; 3409 } 3410 3411 /* 3412 * Read the swap header. 3413 */ 3414 if (!mapping->a_ops->read_folio) { 3415 error = -EINVAL; 3416 goto bad_swap_unlock_inode; 3417 } 3418 folio = read_mapping_folio(mapping, 0, swap_file); 3419 if (IS_ERR(folio)) { 3420 error = PTR_ERR(folio); 3421 goto bad_swap_unlock_inode; 3422 } 3423 swap_header = kmap_local_folio(folio, 0); 3424 3425 maxpages = read_swap_header(si, swap_header, inode); 3426 if (unlikely(!maxpages)) { 3427 error = -EINVAL; 3428 goto bad_swap_unlock_inode; 3429 } 3430 3431 /* OK, set up the swap map and apply the bad block list */ 3432 swap_map = vzalloc(maxpages); 3433 if (!swap_map) { 3434 error = -ENOMEM; 3435 goto bad_swap_unlock_inode; 3436 } 3437 3438 error = swap_cgroup_swapon(si->type, maxpages); 3439 if (error) 3440 goto bad_swap_unlock_inode; 3441 3442 nr_extents = setup_swap_map_and_extents(si, swap_header, swap_map, 3443 maxpages, &span); 3444 if (unlikely(nr_extents < 0)) { 3445 error = nr_extents; 3446 goto bad_swap_unlock_inode; 3447 } 3448 3449 /* 3450 * Use kvmalloc_array instead of bitmap_zalloc as the allocation order might 3451 * be above MAX_PAGE_ORDER incase of a large swap file. 3452 */ 3453 zeromap = kvmalloc_array(BITS_TO_LONGS(maxpages), sizeof(long), 3454 GFP_KERNEL | __GFP_ZERO); 3455 if (!zeromap) { 3456 error = -ENOMEM; 3457 goto bad_swap_unlock_inode; 3458 } 3459 3460 if (si->bdev && bdev_stable_writes(si->bdev)) 3461 si->flags |= SWP_STABLE_WRITES; 3462 3463 if (si->bdev && bdev_synchronous(si->bdev)) 3464 si->flags |= SWP_SYNCHRONOUS_IO; 3465 3466 if (si->bdev && bdev_nonrot(si->bdev)) { 3467 si->flags |= SWP_SOLIDSTATE; 3468 3469 cluster_info = setup_clusters(si, swap_header, maxpages); 3470 if (IS_ERR(cluster_info)) { 3471 error = PTR_ERR(cluster_info); 3472 cluster_info = NULL; 3473 goto bad_swap_unlock_inode; 3474 } 3475 } else { 3476 atomic_inc(&nr_rotate_swap); 3477 inced_nr_rotate_swap = true; 3478 } 3479 3480 if ((swap_flags & SWAP_FLAG_DISCARD) && 3481 si->bdev && bdev_max_discard_sectors(si->bdev)) { 3482 /* 3483 * When discard is enabled for swap with no particular 3484 * policy flagged, we set all swap discard flags here in 3485 * order to sustain backward compatibility with older 3486 * swapon(8) releases. 3487 */ 3488 si->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD | 3489 SWP_PAGE_DISCARD); 3490 3491 /* 3492 * By flagging sys_swapon, a sysadmin can tell us to 3493 * either do single-time area discards only, or to just 3494 * perform discards for released swap page-clusters. 3495 * Now it's time to adjust the p->flags accordingly. 3496 */ 3497 if (swap_flags & SWAP_FLAG_DISCARD_ONCE) 3498 si->flags &= ~SWP_PAGE_DISCARD; 3499 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES) 3500 si->flags &= ~SWP_AREA_DISCARD; 3501 3502 /* issue a swapon-time discard if it's still required */ 3503 if (si->flags & SWP_AREA_DISCARD) { 3504 int err = discard_swap(si); 3505 if (unlikely(err)) 3506 pr_err("swapon: discard_swap(%p): %d\n", 3507 si, err); 3508 } 3509 } 3510 3511 error = init_swap_address_space(si->type, maxpages); 3512 if (error) 3513 goto bad_swap_unlock_inode; 3514 3515 error = zswap_swapon(si->type, maxpages); 3516 if (error) 3517 goto free_swap_address_space; 3518 3519 /* 3520 * Flush any pending IO and dirty mappings before we start using this 3521 * swap device. 3522 */ 3523 inode->i_flags |= S_SWAPFILE; 3524 error = inode_drain_writes(inode); 3525 if (error) { 3526 inode->i_flags &= ~S_SWAPFILE; 3527 goto free_swap_zswap; 3528 } 3529 3530 mutex_lock(&swapon_mutex); 3531 prio = -1; 3532 if (swap_flags & SWAP_FLAG_PREFER) 3533 prio = 3534 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT; 3535 enable_swap_info(si, prio, swap_map, cluster_info, zeromap); 3536 3537 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s\n", 3538 K(si->pages), name->name, si->prio, nr_extents, 3539 K((unsigned long long)span), 3540 (si->flags & SWP_SOLIDSTATE) ? "SS" : "", 3541 (si->flags & SWP_DISCARDABLE) ? "D" : "", 3542 (si->flags & SWP_AREA_DISCARD) ? "s" : "", 3543 (si->flags & SWP_PAGE_DISCARD) ? "c" : ""); 3544 3545 mutex_unlock(&swapon_mutex); 3546 atomic_inc(&proc_poll_event); 3547 wake_up_interruptible(&proc_poll_wait); 3548 3549 error = 0; 3550 goto out; 3551 free_swap_zswap: 3552 zswap_swapoff(si->type); 3553 free_swap_address_space: 3554 exit_swap_address_space(si->type); 3555 bad_swap_unlock_inode: 3556 inode_unlock(inode); 3557 bad_swap: 3558 free_percpu(si->percpu_cluster); 3559 si->percpu_cluster = NULL; 3560 free_percpu(si->cluster_next_cpu); 3561 si->cluster_next_cpu = NULL; 3562 inode = NULL; 3563 destroy_swap_extents(si); 3564 swap_cgroup_swapoff(si->type); 3565 spin_lock(&swap_lock); 3566 si->swap_file = NULL; 3567 si->flags = 0; 3568 spin_unlock(&swap_lock); 3569 vfree(swap_map); 3570 kvfree(zeromap); 3571 kvfree(cluster_info); 3572 if (inced_nr_rotate_swap) 3573 atomic_dec(&nr_rotate_swap); 3574 if (swap_file) 3575 filp_close(swap_file, NULL); 3576 out: 3577 if (!IS_ERR_OR_NULL(folio)) 3578 folio_release_kmap(folio, swap_header); 3579 if (name) 3580 putname(name); 3581 if (inode) 3582 inode_unlock(inode); 3583 if (!error) 3584 enable_swap_slots_cache(); 3585 return error; 3586 } 3587 3588 void si_swapinfo(struct sysinfo *val) 3589 { 3590 unsigned int type; 3591 unsigned long nr_to_be_unused = 0; 3592 3593 spin_lock(&swap_lock); 3594 for (type = 0; type < nr_swapfiles; type++) { 3595 struct swap_info_struct *si = swap_info[type]; 3596 3597 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK)) 3598 nr_to_be_unused += READ_ONCE(si->inuse_pages); 3599 } 3600 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused; 3601 val->totalswap = total_swap_pages + nr_to_be_unused; 3602 spin_unlock(&swap_lock); 3603 } 3604 3605 /* 3606 * Verify that nr swap entries are valid and increment their swap map counts. 3607 * 3608 * Returns error code in following case. 3609 * - success -> 0 3610 * - swp_entry is invalid -> EINVAL 3611 * - swp_entry is migration entry -> EINVAL 3612 * - swap-cache reference is requested but there is already one. -> EEXIST 3613 * - swap-cache reference is requested but the entry is not used. -> ENOENT 3614 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM 3615 */ 3616 static int __swap_duplicate(swp_entry_t entry, unsigned char usage, int nr) 3617 { 3618 struct swap_info_struct *si; 3619 struct swap_cluster_info *ci; 3620 unsigned long offset; 3621 unsigned char count; 3622 unsigned char has_cache; 3623 int err, i; 3624 3625 si = swp_swap_info(entry); 3626 3627 offset = swp_offset(entry); 3628 VM_WARN_ON(nr > SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER); 3629 VM_WARN_ON(usage == 1 && nr > 1); 3630 ci = lock_cluster_or_swap_info(si, offset); 3631 3632 err = 0; 3633 for (i = 0; i < nr; i++) { 3634 count = si->swap_map[offset + i]; 3635 3636 /* 3637 * swapin_readahead() doesn't check if a swap entry is valid, so the 3638 * swap entry could be SWAP_MAP_BAD. Check here with lock held. 3639 */ 3640 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) { 3641 err = -ENOENT; 3642 goto unlock_out; 3643 } 3644 3645 has_cache = count & SWAP_HAS_CACHE; 3646 count &= ~SWAP_HAS_CACHE; 3647 3648 if (!count && !has_cache) { 3649 err = -ENOENT; 3650 } else if (usage == SWAP_HAS_CACHE) { 3651 if (has_cache) 3652 err = -EEXIST; 3653 } else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX) { 3654 err = -EINVAL; 3655 } 3656 3657 if (err) 3658 goto unlock_out; 3659 } 3660 3661 for (i = 0; i < nr; i++) { 3662 count = si->swap_map[offset + i]; 3663 has_cache = count & SWAP_HAS_CACHE; 3664 count &= ~SWAP_HAS_CACHE; 3665 3666 if (usage == SWAP_HAS_CACHE) 3667 has_cache = SWAP_HAS_CACHE; 3668 else if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX) 3669 count += usage; 3670 else if (swap_count_continued(si, offset + i, count)) 3671 count = COUNT_CONTINUED; 3672 else { 3673 /* 3674 * Don't need to rollback changes, because if 3675 * usage == 1, there must be nr == 1. 3676 */ 3677 err = -ENOMEM; 3678 goto unlock_out; 3679 } 3680 3681 WRITE_ONCE(si->swap_map[offset + i], count | has_cache); 3682 } 3683 3684 unlock_out: 3685 unlock_cluster_or_swap_info(si, ci); 3686 return err; 3687 } 3688 3689 /* 3690 * Help swapoff by noting that swap entry belongs to shmem/tmpfs 3691 * (in which case its reference count is never incremented). 3692 */ 3693 void swap_shmem_alloc(swp_entry_t entry, int nr) 3694 { 3695 __swap_duplicate(entry, SWAP_MAP_SHMEM, nr); 3696 } 3697 3698 /* 3699 * Increase reference count of swap entry by 1. 3700 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required 3701 * but could not be atomically allocated. Returns 0, just as if it succeeded, 3702 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which 3703 * might occur if a page table entry has got corrupted. 3704 */ 3705 int swap_duplicate(swp_entry_t entry) 3706 { 3707 int err = 0; 3708 3709 while (!err && __swap_duplicate(entry, 1, 1) == -ENOMEM) 3710 err = add_swap_count_continuation(entry, GFP_ATOMIC); 3711 return err; 3712 } 3713 3714 /* 3715 * @entry: first swap entry from which we allocate nr swap cache. 3716 * 3717 * Called when allocating swap cache for existing swap entries, 3718 * This can return error codes. Returns 0 at success. 3719 * -EEXIST means there is a swap cache. 3720 * Note: return code is different from swap_duplicate(). 3721 */ 3722 int swapcache_prepare(swp_entry_t entry, int nr) 3723 { 3724 return __swap_duplicate(entry, SWAP_HAS_CACHE, nr); 3725 } 3726 3727 void swapcache_clear(struct swap_info_struct *si, swp_entry_t entry, int nr) 3728 { 3729 unsigned long offset = swp_offset(entry); 3730 3731 cluster_swap_free_nr(si, offset, nr, SWAP_HAS_CACHE); 3732 } 3733 3734 struct swap_info_struct *swp_swap_info(swp_entry_t entry) 3735 { 3736 return swap_type_to_swap_info(swp_type(entry)); 3737 } 3738 3739 /* 3740 * out-of-line methods to avoid include hell. 3741 */ 3742 struct address_space *swapcache_mapping(struct folio *folio) 3743 { 3744 return swp_swap_info(folio->swap)->swap_file->f_mapping; 3745 } 3746 EXPORT_SYMBOL_GPL(swapcache_mapping); 3747 3748 pgoff_t __folio_swap_cache_index(struct folio *folio) 3749 { 3750 return swap_cache_index(folio->swap); 3751 } 3752 EXPORT_SYMBOL_GPL(__folio_swap_cache_index); 3753 3754 /* 3755 * add_swap_count_continuation - called when a swap count is duplicated 3756 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's 3757 * page of the original vmalloc'ed swap_map, to hold the continuation count 3758 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called 3759 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc. 3760 * 3761 * These continuation pages are seldom referenced: the common paths all work 3762 * on the original swap_map, only referring to a continuation page when the 3763 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. 3764 * 3765 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding 3766 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL) 3767 * can be called after dropping locks. 3768 */ 3769 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask) 3770 { 3771 struct swap_info_struct *si; 3772 struct swap_cluster_info *ci; 3773 struct page *head; 3774 struct page *page; 3775 struct page *list_page; 3776 pgoff_t offset; 3777 unsigned char count; 3778 int ret = 0; 3779 3780 /* 3781 * When debugging, it's easier to use __GFP_ZERO here; but it's better 3782 * for latency not to zero a page while GFP_ATOMIC and holding locks. 3783 */ 3784 page = alloc_page(gfp_mask | __GFP_HIGHMEM); 3785 3786 si = get_swap_device(entry); 3787 if (!si) { 3788 /* 3789 * An acceptable race has occurred since the failing 3790 * __swap_duplicate(): the swap device may be swapoff 3791 */ 3792 goto outer; 3793 } 3794 spin_lock(&si->lock); 3795 3796 offset = swp_offset(entry); 3797 3798 ci = lock_cluster(si, offset); 3799 3800 count = swap_count(si->swap_map[offset]); 3801 3802 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) { 3803 /* 3804 * The higher the swap count, the more likely it is that tasks 3805 * will race to add swap count continuation: we need to avoid 3806 * over-provisioning. 3807 */ 3808 goto out; 3809 } 3810 3811 if (!page) { 3812 ret = -ENOMEM; 3813 goto out; 3814 } 3815 3816 head = vmalloc_to_page(si->swap_map + offset); 3817 offset &= ~PAGE_MASK; 3818 3819 spin_lock(&si->cont_lock); 3820 /* 3821 * Page allocation does not initialize the page's lru field, 3822 * but it does always reset its private field. 3823 */ 3824 if (!page_private(head)) { 3825 BUG_ON(count & COUNT_CONTINUED); 3826 INIT_LIST_HEAD(&head->lru); 3827 set_page_private(head, SWP_CONTINUED); 3828 si->flags |= SWP_CONTINUED; 3829 } 3830 3831 list_for_each_entry(list_page, &head->lru, lru) { 3832 unsigned char *map; 3833 3834 /* 3835 * If the previous map said no continuation, but we've found 3836 * a continuation page, free our allocation and use this one. 3837 */ 3838 if (!(count & COUNT_CONTINUED)) 3839 goto out_unlock_cont; 3840 3841 map = kmap_local_page(list_page) + offset; 3842 count = *map; 3843 kunmap_local(map); 3844 3845 /* 3846 * If this continuation count now has some space in it, 3847 * free our allocation and use this one. 3848 */ 3849 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX) 3850 goto out_unlock_cont; 3851 } 3852 3853 list_add_tail(&page->lru, &head->lru); 3854 page = NULL; /* now it's attached, don't free it */ 3855 out_unlock_cont: 3856 spin_unlock(&si->cont_lock); 3857 out: 3858 unlock_cluster(ci); 3859 spin_unlock(&si->lock); 3860 put_swap_device(si); 3861 outer: 3862 if (page) 3863 __free_page(page); 3864 return ret; 3865 } 3866 3867 /* 3868 * swap_count_continued - when the original swap_map count is incremented 3869 * from SWAP_MAP_MAX, check if there is already a continuation page to carry 3870 * into, carry if so, or else fail until a new continuation page is allocated; 3871 * when the original swap_map count is decremented from 0 with continuation, 3872 * borrow from the continuation and report whether it still holds more. 3873 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster 3874 * lock. 3875 */ 3876 static bool swap_count_continued(struct swap_info_struct *si, 3877 pgoff_t offset, unsigned char count) 3878 { 3879 struct page *head; 3880 struct page *page; 3881 unsigned char *map; 3882 bool ret; 3883 3884 head = vmalloc_to_page(si->swap_map + offset); 3885 if (page_private(head) != SWP_CONTINUED) { 3886 BUG_ON(count & COUNT_CONTINUED); 3887 return false; /* need to add count continuation */ 3888 } 3889 3890 spin_lock(&si->cont_lock); 3891 offset &= ~PAGE_MASK; 3892 page = list_next_entry(head, lru); 3893 map = kmap_local_page(page) + offset; 3894 3895 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */ 3896 goto init_map; /* jump over SWAP_CONT_MAX checks */ 3897 3898 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */ 3899 /* 3900 * Think of how you add 1 to 999 3901 */ 3902 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) { 3903 kunmap_local(map); 3904 page = list_next_entry(page, lru); 3905 BUG_ON(page == head); 3906 map = kmap_local_page(page) + offset; 3907 } 3908 if (*map == SWAP_CONT_MAX) { 3909 kunmap_local(map); 3910 page = list_next_entry(page, lru); 3911 if (page == head) { 3912 ret = false; /* add count continuation */ 3913 goto out; 3914 } 3915 map = kmap_local_page(page) + offset; 3916 init_map: *map = 0; /* we didn't zero the page */ 3917 } 3918 *map += 1; 3919 kunmap_local(map); 3920 while ((page = list_prev_entry(page, lru)) != head) { 3921 map = kmap_local_page(page) + offset; 3922 *map = COUNT_CONTINUED; 3923 kunmap_local(map); 3924 } 3925 ret = true; /* incremented */ 3926 3927 } else { /* decrementing */ 3928 /* 3929 * Think of how you subtract 1 from 1000 3930 */ 3931 BUG_ON(count != COUNT_CONTINUED); 3932 while (*map == COUNT_CONTINUED) { 3933 kunmap_local(map); 3934 page = list_next_entry(page, lru); 3935 BUG_ON(page == head); 3936 map = kmap_local_page(page) + offset; 3937 } 3938 BUG_ON(*map == 0); 3939 *map -= 1; 3940 if (*map == 0) 3941 count = 0; 3942 kunmap_local(map); 3943 while ((page = list_prev_entry(page, lru)) != head) { 3944 map = kmap_local_page(page) + offset; 3945 *map = SWAP_CONT_MAX | count; 3946 count = COUNT_CONTINUED; 3947 kunmap_local(map); 3948 } 3949 ret = count == COUNT_CONTINUED; 3950 } 3951 out: 3952 spin_unlock(&si->cont_lock); 3953 return ret; 3954 } 3955 3956 /* 3957 * free_swap_count_continuations - swapoff free all the continuation pages 3958 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it. 3959 */ 3960 static void free_swap_count_continuations(struct swap_info_struct *si) 3961 { 3962 pgoff_t offset; 3963 3964 for (offset = 0; offset < si->max; offset += PAGE_SIZE) { 3965 struct page *head; 3966 head = vmalloc_to_page(si->swap_map + offset); 3967 if (page_private(head)) { 3968 struct page *page, *next; 3969 3970 list_for_each_entry_safe(page, next, &head->lru, lru) { 3971 list_del(&page->lru); 3972 __free_page(page); 3973 } 3974 } 3975 } 3976 } 3977 3978 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP) 3979 void __folio_throttle_swaprate(struct folio *folio, gfp_t gfp) 3980 { 3981 struct swap_info_struct *si, *next; 3982 int nid = folio_nid(folio); 3983 3984 if (!(gfp & __GFP_IO)) 3985 return; 3986 3987 if (!__has_usable_swap()) 3988 return; 3989 3990 if (!blk_cgroup_congested()) 3991 return; 3992 3993 /* 3994 * We've already scheduled a throttle, avoid taking the global swap 3995 * lock. 3996 */ 3997 if (current->throttle_disk) 3998 return; 3999 4000 spin_lock(&swap_avail_lock); 4001 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid], 4002 avail_lists[nid]) { 4003 if (si->bdev) { 4004 blkcg_schedule_throttle(si->bdev->bd_disk, true); 4005 break; 4006 } 4007 } 4008 spin_unlock(&swap_avail_lock); 4009 } 4010 #endif 4011 4012 static int __init swapfile_init(void) 4013 { 4014 int nid; 4015 4016 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head), 4017 GFP_KERNEL); 4018 if (!swap_avail_heads) { 4019 pr_emerg("Not enough memory for swap heads, swap is disabled\n"); 4020 return -ENOMEM; 4021 } 4022 4023 for_each_node(nid) 4024 plist_head_init(&swap_avail_heads[nid]); 4025 4026 swapfile_maximum_size = arch_max_swapfile_size(); 4027 4028 #ifdef CONFIG_MIGRATION 4029 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS)) 4030 swap_migration_ad_supported = true; 4031 #endif /* CONFIG_MIGRATION */ 4032 4033 return 0; 4034 } 4035 subsys_initcall(swapfile_init); 4036