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/frontswap.h> 39 #include <linux/swapfile.h> 40 #include <linux/export.h> 41 #include <linux/swap_slots.h> 42 #include <linux/sort.h> 43 #include <linux/completion.h> 44 45 #include <asm/tlbflush.h> 46 #include <linux/swapops.h> 47 #include <linux/swap_cgroup.h> 48 #include "swap.h" 49 50 static bool swap_count_continued(struct swap_info_struct *, pgoff_t, 51 unsigned char); 52 static void free_swap_count_continuations(struct swap_info_struct *); 53 54 static DEFINE_SPINLOCK(swap_lock); 55 static unsigned int nr_swapfiles; 56 atomic_long_t nr_swap_pages; 57 /* 58 * Some modules use swappable objects and may try to swap them out under 59 * memory pressure (via the shrinker). Before doing so, they may wish to 60 * check to see if any swap space is available. 61 */ 62 EXPORT_SYMBOL_GPL(nr_swap_pages); 63 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */ 64 long total_swap_pages; 65 static int least_priority = -1; 66 unsigned long swapfile_maximum_size; 67 #ifdef CONFIG_MIGRATION 68 bool swap_migration_ad_supported; 69 #endif /* CONFIG_MIGRATION */ 70 71 static const char Bad_file[] = "Bad swap file entry "; 72 static const char Unused_file[] = "Unused swap file entry "; 73 static const char Bad_offset[] = "Bad swap offset entry "; 74 static const char Unused_offset[] = "Unused swap offset entry "; 75 76 /* 77 * all active swap_info_structs 78 * protected with swap_lock, and ordered by priority. 79 */ 80 static PLIST_HEAD(swap_active_head); 81 82 /* 83 * all available (active, not full) swap_info_structs 84 * protected with swap_avail_lock, ordered by priority. 85 * This is used by folio_alloc_swap() instead of swap_active_head 86 * because swap_active_head includes all swap_info_structs, 87 * but folio_alloc_swap() doesn't need to look at full ones. 88 * This uses its own lock instead of swap_lock because when a 89 * swap_info_struct changes between not-full/full, it needs to 90 * add/remove itself to/from this list, but the swap_info_struct->lock 91 * is held and the locking order requires swap_lock to be taken 92 * before any swap_info_struct->lock. 93 */ 94 static struct plist_head *swap_avail_heads; 95 static DEFINE_SPINLOCK(swap_avail_lock); 96 97 struct swap_info_struct *swap_info[MAX_SWAPFILES]; 98 99 static DEFINE_MUTEX(swapon_mutex); 100 101 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait); 102 /* Activity counter to indicate that a swapon or swapoff has occurred */ 103 static atomic_t proc_poll_event = ATOMIC_INIT(0); 104 105 atomic_t nr_rotate_swap = ATOMIC_INIT(0); 106 107 static struct swap_info_struct *swap_type_to_swap_info(int type) 108 { 109 if (type >= MAX_SWAPFILES) 110 return NULL; 111 112 return READ_ONCE(swap_info[type]); /* rcu_dereference() */ 113 } 114 115 static inline unsigned char swap_count(unsigned char ent) 116 { 117 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */ 118 } 119 120 /* Reclaim the swap entry anyway if possible */ 121 #define TTRS_ANYWAY 0x1 122 /* 123 * Reclaim the swap entry if there are no more mappings of the 124 * corresponding page 125 */ 126 #define TTRS_UNMAPPED 0x2 127 /* Reclaim the swap entry if swap is getting full*/ 128 #define TTRS_FULL 0x4 129 130 /* returns 1 if swap entry is freed */ 131 static int __try_to_reclaim_swap(struct swap_info_struct *si, 132 unsigned long offset, unsigned long flags) 133 { 134 swp_entry_t entry = swp_entry(si->type, offset); 135 struct page *page; 136 int ret = 0; 137 138 page = find_get_page(swap_address_space(entry), offset); 139 if (!page) 140 return 0; 141 /* 142 * When this function is called from scan_swap_map_slots() and it's 143 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page, 144 * here. We have to use trylock for avoiding deadlock. This is a special 145 * case and you should use try_to_free_swap() with explicit lock_page() 146 * in usual operations. 147 */ 148 if (trylock_page(page)) { 149 if ((flags & TTRS_ANYWAY) || 150 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) || 151 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page))) 152 ret = try_to_free_swap(page); 153 unlock_page(page); 154 } 155 put_page(page); 156 return ret; 157 } 158 159 static inline struct swap_extent *first_se(struct swap_info_struct *sis) 160 { 161 struct rb_node *rb = rb_first(&sis->swap_extent_root); 162 return rb_entry(rb, struct swap_extent, rb_node); 163 } 164 165 static inline struct swap_extent *next_se(struct swap_extent *se) 166 { 167 struct rb_node *rb = rb_next(&se->rb_node); 168 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL; 169 } 170 171 /* 172 * swapon tell device that all the old swap contents can be discarded, 173 * to allow the swap device to optimize its wear-levelling. 174 */ 175 static int discard_swap(struct swap_info_struct *si) 176 { 177 struct swap_extent *se; 178 sector_t start_block; 179 sector_t nr_blocks; 180 int err = 0; 181 182 /* Do not discard the swap header page! */ 183 se = first_se(si); 184 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9); 185 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9); 186 if (nr_blocks) { 187 err = blkdev_issue_discard(si->bdev, start_block, 188 nr_blocks, GFP_KERNEL); 189 if (err) 190 return err; 191 cond_resched(); 192 } 193 194 for (se = next_se(se); se; se = next_se(se)) { 195 start_block = se->start_block << (PAGE_SHIFT - 9); 196 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9); 197 198 err = blkdev_issue_discard(si->bdev, start_block, 199 nr_blocks, GFP_KERNEL); 200 if (err) 201 break; 202 203 cond_resched(); 204 } 205 return err; /* That will often be -EOPNOTSUPP */ 206 } 207 208 static struct swap_extent * 209 offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset) 210 { 211 struct swap_extent *se; 212 struct rb_node *rb; 213 214 rb = sis->swap_extent_root.rb_node; 215 while (rb) { 216 se = rb_entry(rb, struct swap_extent, rb_node); 217 if (offset < se->start_page) 218 rb = rb->rb_left; 219 else if (offset >= se->start_page + se->nr_pages) 220 rb = rb->rb_right; 221 else 222 return se; 223 } 224 /* It *must* be present */ 225 BUG(); 226 } 227 228 sector_t swap_page_sector(struct page *page) 229 { 230 struct swap_info_struct *sis = page_swap_info(page); 231 struct swap_extent *se; 232 sector_t sector; 233 pgoff_t offset; 234 235 offset = __page_file_index(page); 236 se = offset_to_swap_extent(sis, offset); 237 sector = se->start_block + (offset - se->start_page); 238 return sector << (PAGE_SHIFT - 9); 239 } 240 241 /* 242 * swap allocation tell device that a cluster of swap can now be discarded, 243 * to allow the swap device to optimize its wear-levelling. 244 */ 245 static void discard_swap_cluster(struct swap_info_struct *si, 246 pgoff_t start_page, pgoff_t nr_pages) 247 { 248 struct swap_extent *se = offset_to_swap_extent(si, start_page); 249 250 while (nr_pages) { 251 pgoff_t offset = start_page - se->start_page; 252 sector_t start_block = se->start_block + offset; 253 sector_t nr_blocks = se->nr_pages - offset; 254 255 if (nr_blocks > nr_pages) 256 nr_blocks = nr_pages; 257 start_page += nr_blocks; 258 nr_pages -= nr_blocks; 259 260 start_block <<= PAGE_SHIFT - 9; 261 nr_blocks <<= PAGE_SHIFT - 9; 262 if (blkdev_issue_discard(si->bdev, start_block, 263 nr_blocks, GFP_NOIO)) 264 break; 265 266 se = next_se(se); 267 } 268 } 269 270 #ifdef CONFIG_THP_SWAP 271 #define SWAPFILE_CLUSTER HPAGE_PMD_NR 272 273 #define swap_entry_size(size) (size) 274 #else 275 #define SWAPFILE_CLUSTER 256 276 277 /* 278 * Define swap_entry_size() as constant to let compiler to optimize 279 * out some code if !CONFIG_THP_SWAP 280 */ 281 #define swap_entry_size(size) 1 282 #endif 283 #define LATENCY_LIMIT 256 284 285 static inline void cluster_set_flag(struct swap_cluster_info *info, 286 unsigned int flag) 287 { 288 info->flags = flag; 289 } 290 291 static inline unsigned int cluster_count(struct swap_cluster_info *info) 292 { 293 return info->data; 294 } 295 296 static inline void cluster_set_count(struct swap_cluster_info *info, 297 unsigned int c) 298 { 299 info->data = c; 300 } 301 302 static inline void cluster_set_count_flag(struct swap_cluster_info *info, 303 unsigned int c, unsigned int f) 304 { 305 info->flags = f; 306 info->data = c; 307 } 308 309 static inline unsigned int cluster_next(struct swap_cluster_info *info) 310 { 311 return info->data; 312 } 313 314 static inline void cluster_set_next(struct swap_cluster_info *info, 315 unsigned int n) 316 { 317 info->data = n; 318 } 319 320 static inline void cluster_set_next_flag(struct swap_cluster_info *info, 321 unsigned int n, unsigned int f) 322 { 323 info->flags = f; 324 info->data = n; 325 } 326 327 static inline bool cluster_is_free(struct swap_cluster_info *info) 328 { 329 return info->flags & CLUSTER_FLAG_FREE; 330 } 331 332 static inline bool cluster_is_null(struct swap_cluster_info *info) 333 { 334 return info->flags & CLUSTER_FLAG_NEXT_NULL; 335 } 336 337 static inline void cluster_set_null(struct swap_cluster_info *info) 338 { 339 info->flags = CLUSTER_FLAG_NEXT_NULL; 340 info->data = 0; 341 } 342 343 static inline bool cluster_is_huge(struct swap_cluster_info *info) 344 { 345 if (IS_ENABLED(CONFIG_THP_SWAP)) 346 return info->flags & CLUSTER_FLAG_HUGE; 347 return false; 348 } 349 350 static inline void cluster_clear_huge(struct swap_cluster_info *info) 351 { 352 info->flags &= ~CLUSTER_FLAG_HUGE; 353 } 354 355 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si, 356 unsigned long offset) 357 { 358 struct swap_cluster_info *ci; 359 360 ci = si->cluster_info; 361 if (ci) { 362 ci += offset / SWAPFILE_CLUSTER; 363 spin_lock(&ci->lock); 364 } 365 return ci; 366 } 367 368 static inline void unlock_cluster(struct swap_cluster_info *ci) 369 { 370 if (ci) 371 spin_unlock(&ci->lock); 372 } 373 374 /* 375 * Determine the locking method in use for this device. Return 376 * swap_cluster_info if SSD-style cluster-based locking is in place. 377 */ 378 static inline struct swap_cluster_info *lock_cluster_or_swap_info( 379 struct swap_info_struct *si, unsigned long offset) 380 { 381 struct swap_cluster_info *ci; 382 383 /* Try to use fine-grained SSD-style locking if available: */ 384 ci = lock_cluster(si, offset); 385 /* Otherwise, fall back to traditional, coarse locking: */ 386 if (!ci) 387 spin_lock(&si->lock); 388 389 return ci; 390 } 391 392 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si, 393 struct swap_cluster_info *ci) 394 { 395 if (ci) 396 unlock_cluster(ci); 397 else 398 spin_unlock(&si->lock); 399 } 400 401 static inline bool cluster_list_empty(struct swap_cluster_list *list) 402 { 403 return cluster_is_null(&list->head); 404 } 405 406 static inline unsigned int cluster_list_first(struct swap_cluster_list *list) 407 { 408 return cluster_next(&list->head); 409 } 410 411 static void cluster_list_init(struct swap_cluster_list *list) 412 { 413 cluster_set_null(&list->head); 414 cluster_set_null(&list->tail); 415 } 416 417 static void cluster_list_add_tail(struct swap_cluster_list *list, 418 struct swap_cluster_info *ci, 419 unsigned int idx) 420 { 421 if (cluster_list_empty(list)) { 422 cluster_set_next_flag(&list->head, idx, 0); 423 cluster_set_next_flag(&list->tail, idx, 0); 424 } else { 425 struct swap_cluster_info *ci_tail; 426 unsigned int tail = cluster_next(&list->tail); 427 428 /* 429 * Nested cluster lock, but both cluster locks are 430 * only acquired when we held swap_info_struct->lock 431 */ 432 ci_tail = ci + tail; 433 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING); 434 cluster_set_next(ci_tail, idx); 435 spin_unlock(&ci_tail->lock); 436 cluster_set_next_flag(&list->tail, idx, 0); 437 } 438 } 439 440 static unsigned int cluster_list_del_first(struct swap_cluster_list *list, 441 struct swap_cluster_info *ci) 442 { 443 unsigned int idx; 444 445 idx = cluster_next(&list->head); 446 if (cluster_next(&list->tail) == idx) { 447 cluster_set_null(&list->head); 448 cluster_set_null(&list->tail); 449 } else 450 cluster_set_next_flag(&list->head, 451 cluster_next(&ci[idx]), 0); 452 453 return idx; 454 } 455 456 /* Add a cluster to discard list and schedule it to do discard */ 457 static void swap_cluster_schedule_discard(struct swap_info_struct *si, 458 unsigned int idx) 459 { 460 /* 461 * If scan_swap_map_slots() can't find a free cluster, it will check 462 * si->swap_map directly. To make sure the discarding cluster isn't 463 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied). 464 * It will be cleared after discard 465 */ 466 memset(si->swap_map + idx * SWAPFILE_CLUSTER, 467 SWAP_MAP_BAD, SWAPFILE_CLUSTER); 468 469 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx); 470 471 schedule_work(&si->discard_work); 472 } 473 474 static void __free_cluster(struct swap_info_struct *si, unsigned long idx) 475 { 476 struct swap_cluster_info *ci = si->cluster_info; 477 478 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE); 479 cluster_list_add_tail(&si->free_clusters, ci, idx); 480 } 481 482 /* 483 * Doing discard actually. After a cluster discard is finished, the cluster 484 * will be added to free cluster list. caller should hold si->lock. 485 */ 486 static void swap_do_scheduled_discard(struct swap_info_struct *si) 487 { 488 struct swap_cluster_info *info, *ci; 489 unsigned int idx; 490 491 info = si->cluster_info; 492 493 while (!cluster_list_empty(&si->discard_clusters)) { 494 idx = cluster_list_del_first(&si->discard_clusters, info); 495 spin_unlock(&si->lock); 496 497 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER, 498 SWAPFILE_CLUSTER); 499 500 spin_lock(&si->lock); 501 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER); 502 __free_cluster(si, idx); 503 memset(si->swap_map + idx * SWAPFILE_CLUSTER, 504 0, SWAPFILE_CLUSTER); 505 unlock_cluster(ci); 506 } 507 } 508 509 static void swap_discard_work(struct work_struct *work) 510 { 511 struct swap_info_struct *si; 512 513 si = container_of(work, struct swap_info_struct, discard_work); 514 515 spin_lock(&si->lock); 516 swap_do_scheduled_discard(si); 517 spin_unlock(&si->lock); 518 } 519 520 static void swap_users_ref_free(struct percpu_ref *ref) 521 { 522 struct swap_info_struct *si; 523 524 si = container_of(ref, struct swap_info_struct, users); 525 complete(&si->comp); 526 } 527 528 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx) 529 { 530 struct swap_cluster_info *ci = si->cluster_info; 531 532 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx); 533 cluster_list_del_first(&si->free_clusters, ci); 534 cluster_set_count_flag(ci + idx, 0, 0); 535 } 536 537 static void free_cluster(struct swap_info_struct *si, unsigned long idx) 538 { 539 struct swap_cluster_info *ci = si->cluster_info + idx; 540 541 VM_BUG_ON(cluster_count(ci) != 0); 542 /* 543 * If the swap is discardable, prepare discard the cluster 544 * instead of free it immediately. The cluster will be freed 545 * after discard. 546 */ 547 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) == 548 (SWP_WRITEOK | SWP_PAGE_DISCARD)) { 549 swap_cluster_schedule_discard(si, idx); 550 return; 551 } 552 553 __free_cluster(si, idx); 554 } 555 556 /* 557 * The cluster corresponding to page_nr will be used. The cluster will be 558 * removed from free cluster list and its usage counter will be increased. 559 */ 560 static void inc_cluster_info_page(struct swap_info_struct *p, 561 struct swap_cluster_info *cluster_info, unsigned long page_nr) 562 { 563 unsigned long idx = page_nr / SWAPFILE_CLUSTER; 564 565 if (!cluster_info) 566 return; 567 if (cluster_is_free(&cluster_info[idx])) 568 alloc_cluster(p, idx); 569 570 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER); 571 cluster_set_count(&cluster_info[idx], 572 cluster_count(&cluster_info[idx]) + 1); 573 } 574 575 /* 576 * The cluster corresponding to page_nr decreases one usage. If the usage 577 * counter becomes 0, which means no page in the cluster is in using, we can 578 * optionally discard the cluster and add it to free cluster list. 579 */ 580 static void dec_cluster_info_page(struct swap_info_struct *p, 581 struct swap_cluster_info *cluster_info, unsigned long page_nr) 582 { 583 unsigned long idx = page_nr / SWAPFILE_CLUSTER; 584 585 if (!cluster_info) 586 return; 587 588 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0); 589 cluster_set_count(&cluster_info[idx], 590 cluster_count(&cluster_info[idx]) - 1); 591 592 if (cluster_count(&cluster_info[idx]) == 0) 593 free_cluster(p, idx); 594 } 595 596 /* 597 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free 598 * cluster list. Avoiding such abuse to avoid list corruption. 599 */ 600 static bool 601 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si, 602 unsigned long offset) 603 { 604 struct percpu_cluster *percpu_cluster; 605 bool conflict; 606 607 offset /= SWAPFILE_CLUSTER; 608 conflict = !cluster_list_empty(&si->free_clusters) && 609 offset != cluster_list_first(&si->free_clusters) && 610 cluster_is_free(&si->cluster_info[offset]); 611 612 if (!conflict) 613 return false; 614 615 percpu_cluster = this_cpu_ptr(si->percpu_cluster); 616 cluster_set_null(&percpu_cluster->index); 617 return true; 618 } 619 620 /* 621 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This 622 * might involve allocating a new cluster for current CPU too. 623 */ 624 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si, 625 unsigned long *offset, unsigned long *scan_base) 626 { 627 struct percpu_cluster *cluster; 628 struct swap_cluster_info *ci; 629 unsigned long tmp, max; 630 631 new_cluster: 632 cluster = this_cpu_ptr(si->percpu_cluster); 633 if (cluster_is_null(&cluster->index)) { 634 if (!cluster_list_empty(&si->free_clusters)) { 635 cluster->index = si->free_clusters.head; 636 cluster->next = cluster_next(&cluster->index) * 637 SWAPFILE_CLUSTER; 638 } else if (!cluster_list_empty(&si->discard_clusters)) { 639 /* 640 * we don't have free cluster but have some clusters in 641 * discarding, do discard now and reclaim them, then 642 * reread cluster_next_cpu since we dropped si->lock 643 */ 644 swap_do_scheduled_discard(si); 645 *scan_base = this_cpu_read(*si->cluster_next_cpu); 646 *offset = *scan_base; 647 goto new_cluster; 648 } else 649 return false; 650 } 651 652 /* 653 * Other CPUs can use our cluster if they can't find a free cluster, 654 * check if there is still free entry in the cluster 655 */ 656 tmp = cluster->next; 657 max = min_t(unsigned long, si->max, 658 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER); 659 if (tmp < max) { 660 ci = lock_cluster(si, tmp); 661 while (tmp < max) { 662 if (!si->swap_map[tmp]) 663 break; 664 tmp++; 665 } 666 unlock_cluster(ci); 667 } 668 if (tmp >= max) { 669 cluster_set_null(&cluster->index); 670 goto new_cluster; 671 } 672 cluster->next = tmp + 1; 673 *offset = tmp; 674 *scan_base = tmp; 675 return true; 676 } 677 678 static void __del_from_avail_list(struct swap_info_struct *p) 679 { 680 int nid; 681 682 for_each_node(nid) 683 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]); 684 } 685 686 static void del_from_avail_list(struct swap_info_struct *p) 687 { 688 spin_lock(&swap_avail_lock); 689 __del_from_avail_list(p); 690 spin_unlock(&swap_avail_lock); 691 } 692 693 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset, 694 unsigned int nr_entries) 695 { 696 unsigned int end = offset + nr_entries - 1; 697 698 if (offset == si->lowest_bit) 699 si->lowest_bit += nr_entries; 700 if (end == si->highest_bit) 701 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries); 702 WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries); 703 if (si->inuse_pages == si->pages) { 704 si->lowest_bit = si->max; 705 si->highest_bit = 0; 706 del_from_avail_list(si); 707 } 708 } 709 710 static void add_to_avail_list(struct swap_info_struct *p) 711 { 712 int nid; 713 714 spin_lock(&swap_avail_lock); 715 for_each_node(nid) { 716 WARN_ON(!plist_node_empty(&p->avail_lists[nid])); 717 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]); 718 } 719 spin_unlock(&swap_avail_lock); 720 } 721 722 static void swap_range_free(struct swap_info_struct *si, unsigned long offset, 723 unsigned int nr_entries) 724 { 725 unsigned long begin = offset; 726 unsigned long end = offset + nr_entries - 1; 727 void (*swap_slot_free_notify)(struct block_device *, unsigned long); 728 729 if (offset < si->lowest_bit) 730 si->lowest_bit = offset; 731 if (end > si->highest_bit) { 732 bool was_full = !si->highest_bit; 733 734 WRITE_ONCE(si->highest_bit, end); 735 if (was_full && (si->flags & SWP_WRITEOK)) 736 add_to_avail_list(si); 737 } 738 atomic_long_add(nr_entries, &nr_swap_pages); 739 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries); 740 if (si->flags & SWP_BLKDEV) 741 swap_slot_free_notify = 742 si->bdev->bd_disk->fops->swap_slot_free_notify; 743 else 744 swap_slot_free_notify = NULL; 745 while (offset <= end) { 746 arch_swap_invalidate_page(si->type, offset); 747 frontswap_invalidate_page(si->type, offset); 748 if (swap_slot_free_notify) 749 swap_slot_free_notify(si->bdev, offset); 750 offset++; 751 } 752 clear_shadow_from_swap_cache(si->type, begin, end); 753 } 754 755 static void set_cluster_next(struct swap_info_struct *si, unsigned long next) 756 { 757 unsigned long prev; 758 759 if (!(si->flags & SWP_SOLIDSTATE)) { 760 si->cluster_next = next; 761 return; 762 } 763 764 prev = this_cpu_read(*si->cluster_next_cpu); 765 /* 766 * Cross the swap address space size aligned trunk, choose 767 * another trunk randomly to avoid lock contention on swap 768 * address space if possible. 769 */ 770 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) != 771 (next >> SWAP_ADDRESS_SPACE_SHIFT)) { 772 /* No free swap slots available */ 773 if (si->highest_bit <= si->lowest_bit) 774 return; 775 next = si->lowest_bit + 776 prandom_u32_max(si->highest_bit - si->lowest_bit + 1); 777 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES); 778 next = max_t(unsigned int, next, si->lowest_bit); 779 } 780 this_cpu_write(*si->cluster_next_cpu, next); 781 } 782 783 static bool swap_offset_available_and_locked(struct swap_info_struct *si, 784 unsigned long offset) 785 { 786 if (data_race(!si->swap_map[offset])) { 787 spin_lock(&si->lock); 788 return true; 789 } 790 791 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) { 792 spin_lock(&si->lock); 793 return true; 794 } 795 796 return false; 797 } 798 799 static int scan_swap_map_slots(struct swap_info_struct *si, 800 unsigned char usage, int nr, 801 swp_entry_t slots[]) 802 { 803 struct swap_cluster_info *ci; 804 unsigned long offset; 805 unsigned long scan_base; 806 unsigned long last_in_cluster = 0; 807 int latency_ration = LATENCY_LIMIT; 808 int n_ret = 0; 809 bool scanned_many = false; 810 811 /* 812 * We try to cluster swap pages by allocating them sequentially 813 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this 814 * way, however, we resort to first-free allocation, starting 815 * a new cluster. This prevents us from scattering swap pages 816 * all over the entire swap partition, so that we reduce 817 * overall disk seek times between swap pages. -- sct 818 * But we do now try to find an empty cluster. -Andrea 819 * And we let swap pages go all over an SSD partition. Hugh 820 */ 821 822 si->flags += SWP_SCANNING; 823 /* 824 * Use percpu scan base for SSD to reduce lock contention on 825 * cluster and swap cache. For HDD, sequential access is more 826 * important. 827 */ 828 if (si->flags & SWP_SOLIDSTATE) 829 scan_base = this_cpu_read(*si->cluster_next_cpu); 830 else 831 scan_base = si->cluster_next; 832 offset = scan_base; 833 834 /* SSD algorithm */ 835 if (si->cluster_info) { 836 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base)) 837 goto scan; 838 } else if (unlikely(!si->cluster_nr--)) { 839 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) { 840 si->cluster_nr = SWAPFILE_CLUSTER - 1; 841 goto checks; 842 } 843 844 spin_unlock(&si->lock); 845 846 /* 847 * If seek is expensive, start searching for new cluster from 848 * start of partition, to minimize the span of allocated swap. 849 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info 850 * case, just handled by scan_swap_map_try_ssd_cluster() above. 851 */ 852 scan_base = offset = si->lowest_bit; 853 last_in_cluster = offset + SWAPFILE_CLUSTER - 1; 854 855 /* Locate the first empty (unaligned) cluster */ 856 for (; last_in_cluster <= si->highest_bit; offset++) { 857 if (si->swap_map[offset]) 858 last_in_cluster = offset + SWAPFILE_CLUSTER; 859 else if (offset == last_in_cluster) { 860 spin_lock(&si->lock); 861 offset -= SWAPFILE_CLUSTER - 1; 862 si->cluster_next = offset; 863 si->cluster_nr = SWAPFILE_CLUSTER - 1; 864 goto checks; 865 } 866 if (unlikely(--latency_ration < 0)) { 867 cond_resched(); 868 latency_ration = LATENCY_LIMIT; 869 } 870 } 871 872 offset = scan_base; 873 spin_lock(&si->lock); 874 si->cluster_nr = SWAPFILE_CLUSTER - 1; 875 } 876 877 checks: 878 if (si->cluster_info) { 879 while (scan_swap_map_ssd_cluster_conflict(si, offset)) { 880 /* take a break if we already got some slots */ 881 if (n_ret) 882 goto done; 883 if (!scan_swap_map_try_ssd_cluster(si, &offset, 884 &scan_base)) 885 goto scan; 886 } 887 } 888 if (!(si->flags & SWP_WRITEOK)) 889 goto no_page; 890 if (!si->highest_bit) 891 goto no_page; 892 if (offset > si->highest_bit) 893 scan_base = offset = si->lowest_bit; 894 895 ci = lock_cluster(si, offset); 896 /* reuse swap entry of cache-only swap if not busy. */ 897 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) { 898 int swap_was_freed; 899 unlock_cluster(ci); 900 spin_unlock(&si->lock); 901 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY); 902 spin_lock(&si->lock); 903 /* entry was freed successfully, try to use this again */ 904 if (swap_was_freed) 905 goto checks; 906 goto scan; /* check next one */ 907 } 908 909 if (si->swap_map[offset]) { 910 unlock_cluster(ci); 911 if (!n_ret) 912 goto scan; 913 else 914 goto done; 915 } 916 WRITE_ONCE(si->swap_map[offset], usage); 917 inc_cluster_info_page(si, si->cluster_info, offset); 918 unlock_cluster(ci); 919 920 swap_range_alloc(si, offset, 1); 921 slots[n_ret++] = swp_entry(si->type, offset); 922 923 /* got enough slots or reach max slots? */ 924 if ((n_ret == nr) || (offset >= si->highest_bit)) 925 goto done; 926 927 /* search for next available slot */ 928 929 /* time to take a break? */ 930 if (unlikely(--latency_ration < 0)) { 931 if (n_ret) 932 goto done; 933 spin_unlock(&si->lock); 934 cond_resched(); 935 spin_lock(&si->lock); 936 latency_ration = LATENCY_LIMIT; 937 } 938 939 /* try to get more slots in cluster */ 940 if (si->cluster_info) { 941 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base)) 942 goto checks; 943 } else if (si->cluster_nr && !si->swap_map[++offset]) { 944 /* non-ssd case, still more slots in cluster? */ 945 --si->cluster_nr; 946 goto checks; 947 } 948 949 /* 950 * Even if there's no free clusters available (fragmented), 951 * try to scan a little more quickly with lock held unless we 952 * have scanned too many slots already. 953 */ 954 if (!scanned_many) { 955 unsigned long scan_limit; 956 957 if (offset < scan_base) 958 scan_limit = scan_base; 959 else 960 scan_limit = si->highest_bit; 961 for (; offset <= scan_limit && --latency_ration > 0; 962 offset++) { 963 if (!si->swap_map[offset]) 964 goto checks; 965 } 966 } 967 968 done: 969 set_cluster_next(si, offset + 1); 970 si->flags -= SWP_SCANNING; 971 return n_ret; 972 973 scan: 974 spin_unlock(&si->lock); 975 while (++offset <= READ_ONCE(si->highest_bit)) { 976 if (swap_offset_available_and_locked(si, offset)) 977 goto checks; 978 if (unlikely(--latency_ration < 0)) { 979 cond_resched(); 980 latency_ration = LATENCY_LIMIT; 981 scanned_many = true; 982 } 983 } 984 offset = si->lowest_bit; 985 while (offset < scan_base) { 986 if (swap_offset_available_and_locked(si, offset)) 987 goto checks; 988 if (unlikely(--latency_ration < 0)) { 989 cond_resched(); 990 latency_ration = LATENCY_LIMIT; 991 scanned_many = true; 992 } 993 offset++; 994 } 995 spin_lock(&si->lock); 996 997 no_page: 998 si->flags -= SWP_SCANNING; 999 return n_ret; 1000 } 1001 1002 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot) 1003 { 1004 unsigned long idx; 1005 struct swap_cluster_info *ci; 1006 unsigned long offset; 1007 1008 /* 1009 * Should not even be attempting cluster allocations when huge 1010 * page swap is disabled. Warn and fail the allocation. 1011 */ 1012 if (!IS_ENABLED(CONFIG_THP_SWAP)) { 1013 VM_WARN_ON_ONCE(1); 1014 return 0; 1015 } 1016 1017 if (cluster_list_empty(&si->free_clusters)) 1018 return 0; 1019 1020 idx = cluster_list_first(&si->free_clusters); 1021 offset = idx * SWAPFILE_CLUSTER; 1022 ci = lock_cluster(si, offset); 1023 alloc_cluster(si, idx); 1024 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE); 1025 1026 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER); 1027 unlock_cluster(ci); 1028 swap_range_alloc(si, offset, SWAPFILE_CLUSTER); 1029 *slot = swp_entry(si->type, offset); 1030 1031 return 1; 1032 } 1033 1034 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx) 1035 { 1036 unsigned long offset = idx * SWAPFILE_CLUSTER; 1037 struct swap_cluster_info *ci; 1038 1039 ci = lock_cluster(si, offset); 1040 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER); 1041 cluster_set_count_flag(ci, 0, 0); 1042 free_cluster(si, idx); 1043 unlock_cluster(ci); 1044 swap_range_free(si, offset, SWAPFILE_CLUSTER); 1045 } 1046 1047 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size) 1048 { 1049 unsigned long size = swap_entry_size(entry_size); 1050 struct swap_info_struct *si, *next; 1051 long avail_pgs; 1052 int n_ret = 0; 1053 int node; 1054 1055 /* Only single cluster request supported */ 1056 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER); 1057 1058 spin_lock(&swap_avail_lock); 1059 1060 avail_pgs = atomic_long_read(&nr_swap_pages) / size; 1061 if (avail_pgs <= 0) { 1062 spin_unlock(&swap_avail_lock); 1063 goto noswap; 1064 } 1065 1066 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs); 1067 1068 atomic_long_sub(n_goal * size, &nr_swap_pages); 1069 1070 start_over: 1071 node = numa_node_id(); 1072 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) { 1073 /* requeue si to after same-priority siblings */ 1074 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]); 1075 spin_unlock(&swap_avail_lock); 1076 spin_lock(&si->lock); 1077 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) { 1078 spin_lock(&swap_avail_lock); 1079 if (plist_node_empty(&si->avail_lists[node])) { 1080 spin_unlock(&si->lock); 1081 goto nextsi; 1082 } 1083 WARN(!si->highest_bit, 1084 "swap_info %d in list but !highest_bit\n", 1085 si->type); 1086 WARN(!(si->flags & SWP_WRITEOK), 1087 "swap_info %d in list but !SWP_WRITEOK\n", 1088 si->type); 1089 __del_from_avail_list(si); 1090 spin_unlock(&si->lock); 1091 goto nextsi; 1092 } 1093 if (size == SWAPFILE_CLUSTER) { 1094 if (si->flags & SWP_BLKDEV) 1095 n_ret = swap_alloc_cluster(si, swp_entries); 1096 } else 1097 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE, 1098 n_goal, swp_entries); 1099 spin_unlock(&si->lock); 1100 if (n_ret || size == SWAPFILE_CLUSTER) 1101 goto check_out; 1102 pr_debug("scan_swap_map of si %d failed to find offset\n", 1103 si->type); 1104 1105 spin_lock(&swap_avail_lock); 1106 nextsi: 1107 /* 1108 * if we got here, it's likely that si was almost full before, 1109 * and since scan_swap_map_slots() can drop the si->lock, 1110 * multiple callers probably all tried to get a page from the 1111 * same si and it filled up before we could get one; or, the si 1112 * filled up between us dropping swap_avail_lock and taking 1113 * si->lock. Since we dropped the swap_avail_lock, the 1114 * swap_avail_head list may have been modified; so if next is 1115 * still in the swap_avail_head list then try it, otherwise 1116 * start over if we have not gotten any slots. 1117 */ 1118 if (plist_node_empty(&next->avail_lists[node])) 1119 goto start_over; 1120 } 1121 1122 spin_unlock(&swap_avail_lock); 1123 1124 check_out: 1125 if (n_ret < n_goal) 1126 atomic_long_add((long)(n_goal - n_ret) * size, 1127 &nr_swap_pages); 1128 noswap: 1129 return n_ret; 1130 } 1131 1132 static struct swap_info_struct *_swap_info_get(swp_entry_t entry) 1133 { 1134 struct swap_info_struct *p; 1135 unsigned long offset; 1136 1137 if (!entry.val) 1138 goto out; 1139 p = swp_swap_info(entry); 1140 if (!p) 1141 goto bad_nofile; 1142 if (data_race(!(p->flags & SWP_USED))) 1143 goto bad_device; 1144 offset = swp_offset(entry); 1145 if (offset >= p->max) 1146 goto bad_offset; 1147 if (data_race(!p->swap_map[swp_offset(entry)])) 1148 goto bad_free; 1149 return p; 1150 1151 bad_free: 1152 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val); 1153 goto out; 1154 bad_offset: 1155 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val); 1156 goto out; 1157 bad_device: 1158 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val); 1159 goto out; 1160 bad_nofile: 1161 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val); 1162 out: 1163 return NULL; 1164 } 1165 1166 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry, 1167 struct swap_info_struct *q) 1168 { 1169 struct swap_info_struct *p; 1170 1171 p = _swap_info_get(entry); 1172 1173 if (p != q) { 1174 if (q != NULL) 1175 spin_unlock(&q->lock); 1176 if (p != NULL) 1177 spin_lock(&p->lock); 1178 } 1179 return p; 1180 } 1181 1182 static unsigned char __swap_entry_free_locked(struct swap_info_struct *p, 1183 unsigned long offset, 1184 unsigned char usage) 1185 { 1186 unsigned char count; 1187 unsigned char has_cache; 1188 1189 count = p->swap_map[offset]; 1190 1191 has_cache = count & SWAP_HAS_CACHE; 1192 count &= ~SWAP_HAS_CACHE; 1193 1194 if (usage == SWAP_HAS_CACHE) { 1195 VM_BUG_ON(!has_cache); 1196 has_cache = 0; 1197 } else if (count == SWAP_MAP_SHMEM) { 1198 /* 1199 * Or we could insist on shmem.c using a special 1200 * swap_shmem_free() and free_shmem_swap_and_cache()... 1201 */ 1202 count = 0; 1203 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) { 1204 if (count == COUNT_CONTINUED) { 1205 if (swap_count_continued(p, offset, count)) 1206 count = SWAP_MAP_MAX | COUNT_CONTINUED; 1207 else 1208 count = SWAP_MAP_MAX; 1209 } else 1210 count--; 1211 } 1212 1213 usage = count | has_cache; 1214 if (usage) 1215 WRITE_ONCE(p->swap_map[offset], usage); 1216 else 1217 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE); 1218 1219 return usage; 1220 } 1221 1222 /* 1223 * Check whether swap entry is valid in the swap device. If so, 1224 * return pointer to swap_info_struct, and keep the swap entry valid 1225 * via preventing the swap device from being swapoff, until 1226 * put_swap_device() is called. Otherwise return NULL. 1227 * 1228 * Notice that swapoff or swapoff+swapon can still happen before the 1229 * percpu_ref_tryget_live() in get_swap_device() or after the 1230 * percpu_ref_put() in put_swap_device() if there isn't any other way 1231 * to prevent swapoff, such as page lock, page table lock, etc. The 1232 * caller must be prepared for that. For example, the following 1233 * situation is possible. 1234 * 1235 * CPU1 CPU2 1236 * do_swap_page() 1237 * ... swapoff+swapon 1238 * __read_swap_cache_async() 1239 * swapcache_prepare() 1240 * __swap_duplicate() 1241 * // check swap_map 1242 * // verify PTE not changed 1243 * 1244 * In __swap_duplicate(), the swap_map need to be checked before 1245 * changing partly because the specified swap entry may be for another 1246 * swap device which has been swapoff. And in do_swap_page(), after 1247 * the page is read from the swap device, the PTE is verified not 1248 * changed with the page table locked to check whether the swap device 1249 * has been swapoff or swapoff+swapon. 1250 */ 1251 struct swap_info_struct *get_swap_device(swp_entry_t entry) 1252 { 1253 struct swap_info_struct *si; 1254 unsigned long offset; 1255 1256 if (!entry.val) 1257 goto out; 1258 si = swp_swap_info(entry); 1259 if (!si) 1260 goto bad_nofile; 1261 if (!percpu_ref_tryget_live(&si->users)) 1262 goto out; 1263 /* 1264 * Guarantee the si->users are checked before accessing other 1265 * fields of swap_info_struct. 1266 * 1267 * Paired with the spin_unlock() after setup_swap_info() in 1268 * enable_swap_info(). 1269 */ 1270 smp_rmb(); 1271 offset = swp_offset(entry); 1272 if (offset >= si->max) 1273 goto put_out; 1274 1275 return si; 1276 bad_nofile: 1277 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val); 1278 out: 1279 return NULL; 1280 put_out: 1281 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val); 1282 percpu_ref_put(&si->users); 1283 return NULL; 1284 } 1285 1286 static unsigned char __swap_entry_free(struct swap_info_struct *p, 1287 swp_entry_t entry) 1288 { 1289 struct swap_cluster_info *ci; 1290 unsigned long offset = swp_offset(entry); 1291 unsigned char usage; 1292 1293 ci = lock_cluster_or_swap_info(p, offset); 1294 usage = __swap_entry_free_locked(p, offset, 1); 1295 unlock_cluster_or_swap_info(p, ci); 1296 if (!usage) 1297 free_swap_slot(entry); 1298 1299 return usage; 1300 } 1301 1302 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry) 1303 { 1304 struct swap_cluster_info *ci; 1305 unsigned long offset = swp_offset(entry); 1306 unsigned char count; 1307 1308 ci = lock_cluster(p, offset); 1309 count = p->swap_map[offset]; 1310 VM_BUG_ON(count != SWAP_HAS_CACHE); 1311 p->swap_map[offset] = 0; 1312 dec_cluster_info_page(p, p->cluster_info, offset); 1313 unlock_cluster(ci); 1314 1315 mem_cgroup_uncharge_swap(entry, 1); 1316 swap_range_free(p, offset, 1); 1317 } 1318 1319 /* 1320 * Caller has made sure that the swap device corresponding to entry 1321 * is still around or has not been recycled. 1322 */ 1323 void swap_free(swp_entry_t entry) 1324 { 1325 struct swap_info_struct *p; 1326 1327 p = _swap_info_get(entry); 1328 if (p) 1329 __swap_entry_free(p, entry); 1330 } 1331 1332 /* 1333 * Called after dropping swapcache to decrease refcnt to swap entries. 1334 */ 1335 void put_swap_page(struct page *page, swp_entry_t entry) 1336 { 1337 unsigned long offset = swp_offset(entry); 1338 unsigned long idx = offset / SWAPFILE_CLUSTER; 1339 struct swap_cluster_info *ci; 1340 struct swap_info_struct *si; 1341 unsigned char *map; 1342 unsigned int i, free_entries = 0; 1343 unsigned char val; 1344 int size = swap_entry_size(thp_nr_pages(page)); 1345 1346 si = _swap_info_get(entry); 1347 if (!si) 1348 return; 1349 1350 ci = lock_cluster_or_swap_info(si, offset); 1351 if (size == SWAPFILE_CLUSTER) { 1352 VM_BUG_ON(!cluster_is_huge(ci)); 1353 map = si->swap_map + offset; 1354 for (i = 0; i < SWAPFILE_CLUSTER; i++) { 1355 val = map[i]; 1356 VM_BUG_ON(!(val & SWAP_HAS_CACHE)); 1357 if (val == SWAP_HAS_CACHE) 1358 free_entries++; 1359 } 1360 cluster_clear_huge(ci); 1361 if (free_entries == SWAPFILE_CLUSTER) { 1362 unlock_cluster_or_swap_info(si, ci); 1363 spin_lock(&si->lock); 1364 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER); 1365 swap_free_cluster(si, idx); 1366 spin_unlock(&si->lock); 1367 return; 1368 } 1369 } 1370 for (i = 0; i < size; i++, entry.val++) { 1371 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) { 1372 unlock_cluster_or_swap_info(si, ci); 1373 free_swap_slot(entry); 1374 if (i == size - 1) 1375 return; 1376 lock_cluster_or_swap_info(si, offset); 1377 } 1378 } 1379 unlock_cluster_or_swap_info(si, ci); 1380 } 1381 1382 #ifdef CONFIG_THP_SWAP 1383 int split_swap_cluster(swp_entry_t entry) 1384 { 1385 struct swap_info_struct *si; 1386 struct swap_cluster_info *ci; 1387 unsigned long offset = swp_offset(entry); 1388 1389 si = _swap_info_get(entry); 1390 if (!si) 1391 return -EBUSY; 1392 ci = lock_cluster(si, offset); 1393 cluster_clear_huge(ci); 1394 unlock_cluster(ci); 1395 return 0; 1396 } 1397 #endif 1398 1399 static int swp_entry_cmp(const void *ent1, const void *ent2) 1400 { 1401 const swp_entry_t *e1 = ent1, *e2 = ent2; 1402 1403 return (int)swp_type(*e1) - (int)swp_type(*e2); 1404 } 1405 1406 void swapcache_free_entries(swp_entry_t *entries, int n) 1407 { 1408 struct swap_info_struct *p, *prev; 1409 int i; 1410 1411 if (n <= 0) 1412 return; 1413 1414 prev = NULL; 1415 p = NULL; 1416 1417 /* 1418 * Sort swap entries by swap device, so each lock is only taken once. 1419 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is 1420 * so low that it isn't necessary to optimize further. 1421 */ 1422 if (nr_swapfiles > 1) 1423 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL); 1424 for (i = 0; i < n; ++i) { 1425 p = swap_info_get_cont(entries[i], prev); 1426 if (p) 1427 swap_entry_free(p, entries[i]); 1428 prev = p; 1429 } 1430 if (p) 1431 spin_unlock(&p->lock); 1432 } 1433 1434 /* 1435 * How many references to page are currently swapped out? 1436 * This does not give an exact answer when swap count is continued, 1437 * but does include the high COUNT_CONTINUED flag to allow for that. 1438 */ 1439 static int page_swapcount(struct page *page) 1440 { 1441 int count = 0; 1442 struct swap_info_struct *p; 1443 struct swap_cluster_info *ci; 1444 swp_entry_t entry; 1445 unsigned long offset; 1446 1447 entry.val = page_private(page); 1448 p = _swap_info_get(entry); 1449 if (p) { 1450 offset = swp_offset(entry); 1451 ci = lock_cluster_or_swap_info(p, offset); 1452 count = swap_count(p->swap_map[offset]); 1453 unlock_cluster_or_swap_info(p, ci); 1454 } 1455 return count; 1456 } 1457 1458 int __swap_count(swp_entry_t entry) 1459 { 1460 struct swap_info_struct *si; 1461 pgoff_t offset = swp_offset(entry); 1462 int count = 0; 1463 1464 si = get_swap_device(entry); 1465 if (si) { 1466 count = swap_count(si->swap_map[offset]); 1467 put_swap_device(si); 1468 } 1469 return count; 1470 } 1471 1472 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry) 1473 { 1474 int count = 0; 1475 pgoff_t offset = swp_offset(entry); 1476 struct swap_cluster_info *ci; 1477 1478 ci = lock_cluster_or_swap_info(si, offset); 1479 count = swap_count(si->swap_map[offset]); 1480 unlock_cluster_or_swap_info(si, ci); 1481 return count; 1482 } 1483 1484 /* 1485 * How many references to @entry are currently swapped out? 1486 * This does not give an exact answer when swap count is continued, 1487 * but does include the high COUNT_CONTINUED flag to allow for that. 1488 */ 1489 int __swp_swapcount(swp_entry_t entry) 1490 { 1491 int count = 0; 1492 struct swap_info_struct *si; 1493 1494 si = get_swap_device(entry); 1495 if (si) { 1496 count = swap_swapcount(si, entry); 1497 put_swap_device(si); 1498 } 1499 return count; 1500 } 1501 1502 /* 1503 * How many references to @entry are currently swapped out? 1504 * This considers COUNT_CONTINUED so it returns exact answer. 1505 */ 1506 int swp_swapcount(swp_entry_t entry) 1507 { 1508 int count, tmp_count, n; 1509 struct swap_info_struct *p; 1510 struct swap_cluster_info *ci; 1511 struct page *page; 1512 pgoff_t offset; 1513 unsigned char *map; 1514 1515 p = _swap_info_get(entry); 1516 if (!p) 1517 return 0; 1518 1519 offset = swp_offset(entry); 1520 1521 ci = lock_cluster_or_swap_info(p, offset); 1522 1523 count = swap_count(p->swap_map[offset]); 1524 if (!(count & COUNT_CONTINUED)) 1525 goto out; 1526 1527 count &= ~COUNT_CONTINUED; 1528 n = SWAP_MAP_MAX + 1; 1529 1530 page = vmalloc_to_page(p->swap_map + offset); 1531 offset &= ~PAGE_MASK; 1532 VM_BUG_ON(page_private(page) != SWP_CONTINUED); 1533 1534 do { 1535 page = list_next_entry(page, lru); 1536 map = kmap_atomic(page); 1537 tmp_count = map[offset]; 1538 kunmap_atomic(map); 1539 1540 count += (tmp_count & ~COUNT_CONTINUED) * n; 1541 n *= (SWAP_CONT_MAX + 1); 1542 } while (tmp_count & COUNT_CONTINUED); 1543 out: 1544 unlock_cluster_or_swap_info(p, ci); 1545 return count; 1546 } 1547 1548 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si, 1549 swp_entry_t entry) 1550 { 1551 struct swap_cluster_info *ci; 1552 unsigned char *map = si->swap_map; 1553 unsigned long roffset = swp_offset(entry); 1554 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER); 1555 int i; 1556 bool ret = false; 1557 1558 ci = lock_cluster_or_swap_info(si, offset); 1559 if (!ci || !cluster_is_huge(ci)) { 1560 if (swap_count(map[roffset])) 1561 ret = true; 1562 goto unlock_out; 1563 } 1564 for (i = 0; i < SWAPFILE_CLUSTER; i++) { 1565 if (swap_count(map[offset + i])) { 1566 ret = true; 1567 break; 1568 } 1569 } 1570 unlock_out: 1571 unlock_cluster_or_swap_info(si, ci); 1572 return ret; 1573 } 1574 1575 static bool folio_swapped(struct folio *folio) 1576 { 1577 swp_entry_t entry; 1578 struct swap_info_struct *si; 1579 1580 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio))) 1581 return page_swapcount(&folio->page) != 0; 1582 1583 entry = folio_swap_entry(folio); 1584 si = _swap_info_get(entry); 1585 if (si) 1586 return swap_page_trans_huge_swapped(si, entry); 1587 return false; 1588 } 1589 1590 /* 1591 * If swap is getting full, or if there are no more mappings of this page, 1592 * then try_to_free_swap is called to free its swap space. 1593 */ 1594 int try_to_free_swap(struct page *page) 1595 { 1596 struct folio *folio = page_folio(page); 1597 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 1598 1599 if (!folio_test_swapcache(folio)) 1600 return 0; 1601 if (folio_test_writeback(folio)) 1602 return 0; 1603 if (folio_swapped(folio)) 1604 return 0; 1605 1606 /* 1607 * Once hibernation has begun to create its image of memory, 1608 * there's a danger that one of the calls to try_to_free_swap() 1609 * - most probably a call from __try_to_reclaim_swap() while 1610 * hibernation is allocating its own swap pages for the image, 1611 * but conceivably even a call from memory reclaim - will free 1612 * the swap from a page which has already been recorded in the 1613 * image as a clean swapcache page, and then reuse its swap for 1614 * another page of the image. On waking from hibernation, the 1615 * original page might be freed under memory pressure, then 1616 * later read back in from swap, now with the wrong data. 1617 * 1618 * Hibernation suspends storage while it is writing the image 1619 * to disk so check that here. 1620 */ 1621 if (pm_suspended_storage()) 1622 return 0; 1623 1624 delete_from_swap_cache(folio); 1625 folio_set_dirty(folio); 1626 return 1; 1627 } 1628 1629 /* 1630 * Free the swap entry like above, but also try to 1631 * free the page cache entry if it is the last user. 1632 */ 1633 int free_swap_and_cache(swp_entry_t entry) 1634 { 1635 struct swap_info_struct *p; 1636 unsigned char count; 1637 1638 if (non_swap_entry(entry)) 1639 return 1; 1640 1641 p = _swap_info_get(entry); 1642 if (p) { 1643 count = __swap_entry_free(p, entry); 1644 if (count == SWAP_HAS_CACHE && 1645 !swap_page_trans_huge_swapped(p, entry)) 1646 __try_to_reclaim_swap(p, swp_offset(entry), 1647 TTRS_UNMAPPED | TTRS_FULL); 1648 } 1649 return p != NULL; 1650 } 1651 1652 #ifdef CONFIG_HIBERNATION 1653 1654 swp_entry_t get_swap_page_of_type(int type) 1655 { 1656 struct swap_info_struct *si = swap_type_to_swap_info(type); 1657 swp_entry_t entry = {0}; 1658 1659 if (!si) 1660 goto fail; 1661 1662 /* This is called for allocating swap entry, not cache */ 1663 spin_lock(&si->lock); 1664 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry)) 1665 atomic_long_dec(&nr_swap_pages); 1666 spin_unlock(&si->lock); 1667 fail: 1668 return entry; 1669 } 1670 1671 /* 1672 * Find the swap type that corresponds to given device (if any). 1673 * 1674 * @offset - number of the PAGE_SIZE-sized block of the device, starting 1675 * from 0, in which the swap header is expected to be located. 1676 * 1677 * This is needed for the suspend to disk (aka swsusp). 1678 */ 1679 int swap_type_of(dev_t device, sector_t offset) 1680 { 1681 int type; 1682 1683 if (!device) 1684 return -1; 1685 1686 spin_lock(&swap_lock); 1687 for (type = 0; type < nr_swapfiles; type++) { 1688 struct swap_info_struct *sis = swap_info[type]; 1689 1690 if (!(sis->flags & SWP_WRITEOK)) 1691 continue; 1692 1693 if (device == sis->bdev->bd_dev) { 1694 struct swap_extent *se = first_se(sis); 1695 1696 if (se->start_block == offset) { 1697 spin_unlock(&swap_lock); 1698 return type; 1699 } 1700 } 1701 } 1702 spin_unlock(&swap_lock); 1703 return -ENODEV; 1704 } 1705 1706 int find_first_swap(dev_t *device) 1707 { 1708 int type; 1709 1710 spin_lock(&swap_lock); 1711 for (type = 0; type < nr_swapfiles; type++) { 1712 struct swap_info_struct *sis = swap_info[type]; 1713 1714 if (!(sis->flags & SWP_WRITEOK)) 1715 continue; 1716 *device = sis->bdev->bd_dev; 1717 spin_unlock(&swap_lock); 1718 return type; 1719 } 1720 spin_unlock(&swap_lock); 1721 return -ENODEV; 1722 } 1723 1724 /* 1725 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev 1726 * corresponding to given index in swap_info (swap type). 1727 */ 1728 sector_t swapdev_block(int type, pgoff_t offset) 1729 { 1730 struct swap_info_struct *si = swap_type_to_swap_info(type); 1731 struct swap_extent *se; 1732 1733 if (!si || !(si->flags & SWP_WRITEOK)) 1734 return 0; 1735 se = offset_to_swap_extent(si, offset); 1736 return se->start_block + (offset - se->start_page); 1737 } 1738 1739 /* 1740 * Return either the total number of swap pages of given type, or the number 1741 * of free pages of that type (depending on @free) 1742 * 1743 * This is needed for software suspend 1744 */ 1745 unsigned int count_swap_pages(int type, int free) 1746 { 1747 unsigned int n = 0; 1748 1749 spin_lock(&swap_lock); 1750 if ((unsigned int)type < nr_swapfiles) { 1751 struct swap_info_struct *sis = swap_info[type]; 1752 1753 spin_lock(&sis->lock); 1754 if (sis->flags & SWP_WRITEOK) { 1755 n = sis->pages; 1756 if (free) 1757 n -= sis->inuse_pages; 1758 } 1759 spin_unlock(&sis->lock); 1760 } 1761 spin_unlock(&swap_lock); 1762 return n; 1763 } 1764 #endif /* CONFIG_HIBERNATION */ 1765 1766 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte) 1767 { 1768 return pte_same(pte_swp_clear_flags(pte), swp_pte); 1769 } 1770 1771 /* 1772 * No need to decide whether this PTE shares the swap entry with others, 1773 * just let do_wp_page work it out if a write is requested later - to 1774 * force COW, vm_page_prot omits write permission from any private vma. 1775 */ 1776 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd, 1777 unsigned long addr, swp_entry_t entry, struct page *page) 1778 { 1779 struct page *swapcache; 1780 spinlock_t *ptl; 1781 pte_t *pte, new_pte; 1782 int ret = 1; 1783 1784 swapcache = page; 1785 page = ksm_might_need_to_copy(page, vma, addr); 1786 if (unlikely(!page)) 1787 return -ENOMEM; 1788 1789 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 1790 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) { 1791 ret = 0; 1792 goto out; 1793 } 1794 1795 if (unlikely(!PageUptodate(page))) { 1796 pte_t pteval; 1797 1798 dec_mm_counter(vma->vm_mm, MM_SWAPENTS); 1799 pteval = swp_entry_to_pte(make_swapin_error_entry(page)); 1800 set_pte_at(vma->vm_mm, addr, pte, pteval); 1801 swap_free(entry); 1802 ret = 0; 1803 goto out; 1804 } 1805 1806 /* See do_swap_page() */ 1807 BUG_ON(!PageAnon(page) && PageMappedToDisk(page)); 1808 BUG_ON(PageAnon(page) && PageAnonExclusive(page)); 1809 1810 dec_mm_counter(vma->vm_mm, MM_SWAPENTS); 1811 inc_mm_counter(vma->vm_mm, MM_ANONPAGES); 1812 get_page(page); 1813 if (page == swapcache) { 1814 rmap_t rmap_flags = RMAP_NONE; 1815 1816 /* 1817 * See do_swap_page(): PageWriteback() would be problematic. 1818 * However, we do a wait_on_page_writeback() just before this 1819 * call and have the page locked. 1820 */ 1821 VM_BUG_ON_PAGE(PageWriteback(page), page); 1822 if (pte_swp_exclusive(*pte)) 1823 rmap_flags |= RMAP_EXCLUSIVE; 1824 1825 page_add_anon_rmap(page, vma, addr, rmap_flags); 1826 } else { /* ksm created a completely new copy */ 1827 page_add_new_anon_rmap(page, vma, addr); 1828 lru_cache_add_inactive_or_unevictable(page, vma); 1829 } 1830 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot)); 1831 if (pte_swp_soft_dirty(*pte)) 1832 new_pte = pte_mksoft_dirty(new_pte); 1833 if (pte_swp_uffd_wp(*pte)) 1834 new_pte = pte_mkuffd_wp(new_pte); 1835 set_pte_at(vma->vm_mm, addr, pte, new_pte); 1836 swap_free(entry); 1837 out: 1838 pte_unmap_unlock(pte, ptl); 1839 if (page != swapcache) { 1840 unlock_page(page); 1841 put_page(page); 1842 } 1843 return ret; 1844 } 1845 1846 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd, 1847 unsigned long addr, unsigned long end, 1848 unsigned int type) 1849 { 1850 struct page *page; 1851 swp_entry_t entry; 1852 pte_t *pte; 1853 struct swap_info_struct *si; 1854 unsigned long offset; 1855 int ret = 0; 1856 volatile unsigned char *swap_map; 1857 1858 si = swap_info[type]; 1859 pte = pte_offset_map(pmd, addr); 1860 do { 1861 if (!is_swap_pte(*pte)) 1862 continue; 1863 1864 entry = pte_to_swp_entry(*pte); 1865 if (swp_type(entry) != type) 1866 continue; 1867 1868 offset = swp_offset(entry); 1869 pte_unmap(pte); 1870 swap_map = &si->swap_map[offset]; 1871 page = lookup_swap_cache(entry, vma, addr); 1872 if (!page) { 1873 struct vm_fault vmf = { 1874 .vma = vma, 1875 .address = addr, 1876 .real_address = addr, 1877 .pmd = pmd, 1878 }; 1879 1880 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, 1881 &vmf); 1882 } 1883 if (!page) { 1884 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD) 1885 goto try_next; 1886 return -ENOMEM; 1887 } 1888 1889 lock_page(page); 1890 wait_on_page_writeback(page); 1891 ret = unuse_pte(vma, pmd, addr, entry, page); 1892 if (ret < 0) { 1893 unlock_page(page); 1894 put_page(page); 1895 goto out; 1896 } 1897 1898 try_to_free_swap(page); 1899 unlock_page(page); 1900 put_page(page); 1901 try_next: 1902 pte = pte_offset_map(pmd, addr); 1903 } while (pte++, addr += PAGE_SIZE, addr != end); 1904 pte_unmap(pte - 1); 1905 1906 ret = 0; 1907 out: 1908 return ret; 1909 } 1910 1911 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud, 1912 unsigned long addr, unsigned long end, 1913 unsigned int type) 1914 { 1915 pmd_t *pmd; 1916 unsigned long next; 1917 int ret; 1918 1919 pmd = pmd_offset(pud, addr); 1920 do { 1921 cond_resched(); 1922 next = pmd_addr_end(addr, end); 1923 if (pmd_none_or_trans_huge_or_clear_bad(pmd)) 1924 continue; 1925 ret = unuse_pte_range(vma, pmd, addr, next, type); 1926 if (ret) 1927 return ret; 1928 } while (pmd++, addr = next, addr != end); 1929 return 0; 1930 } 1931 1932 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d, 1933 unsigned long addr, unsigned long end, 1934 unsigned int type) 1935 { 1936 pud_t *pud; 1937 unsigned long next; 1938 int ret; 1939 1940 pud = pud_offset(p4d, addr); 1941 do { 1942 next = pud_addr_end(addr, end); 1943 if (pud_none_or_clear_bad(pud)) 1944 continue; 1945 ret = unuse_pmd_range(vma, pud, addr, next, type); 1946 if (ret) 1947 return ret; 1948 } while (pud++, addr = next, addr != end); 1949 return 0; 1950 } 1951 1952 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd, 1953 unsigned long addr, unsigned long end, 1954 unsigned int type) 1955 { 1956 p4d_t *p4d; 1957 unsigned long next; 1958 int ret; 1959 1960 p4d = p4d_offset(pgd, addr); 1961 do { 1962 next = p4d_addr_end(addr, end); 1963 if (p4d_none_or_clear_bad(p4d)) 1964 continue; 1965 ret = unuse_pud_range(vma, p4d, addr, next, type); 1966 if (ret) 1967 return ret; 1968 } while (p4d++, addr = next, addr != end); 1969 return 0; 1970 } 1971 1972 static int unuse_vma(struct vm_area_struct *vma, unsigned int type) 1973 { 1974 pgd_t *pgd; 1975 unsigned long addr, end, next; 1976 int ret; 1977 1978 addr = vma->vm_start; 1979 end = vma->vm_end; 1980 1981 pgd = pgd_offset(vma->vm_mm, addr); 1982 do { 1983 next = pgd_addr_end(addr, end); 1984 if (pgd_none_or_clear_bad(pgd)) 1985 continue; 1986 ret = unuse_p4d_range(vma, pgd, addr, next, type); 1987 if (ret) 1988 return ret; 1989 } while (pgd++, addr = next, addr != end); 1990 return 0; 1991 } 1992 1993 static int unuse_mm(struct mm_struct *mm, unsigned int type) 1994 { 1995 struct vm_area_struct *vma; 1996 int ret = 0; 1997 1998 mmap_read_lock(mm); 1999 for (vma = mm->mmap; vma; vma = vma->vm_next) { 2000 if (vma->anon_vma) { 2001 ret = unuse_vma(vma, type); 2002 if (ret) 2003 break; 2004 } 2005 cond_resched(); 2006 } 2007 mmap_read_unlock(mm); 2008 return ret; 2009 } 2010 2011 /* 2012 * Scan swap_map from current position to next entry still in use. 2013 * Return 0 if there are no inuse entries after prev till end of 2014 * the map. 2015 */ 2016 static unsigned int find_next_to_unuse(struct swap_info_struct *si, 2017 unsigned int prev) 2018 { 2019 unsigned int i; 2020 unsigned char count; 2021 2022 /* 2023 * No need for swap_lock here: we're just looking 2024 * for whether an entry is in use, not modifying it; false 2025 * hits are okay, and sys_swapoff() has already prevented new 2026 * allocations from this area (while holding swap_lock). 2027 */ 2028 for (i = prev + 1; i < si->max; i++) { 2029 count = READ_ONCE(si->swap_map[i]); 2030 if (count && swap_count(count) != SWAP_MAP_BAD) 2031 break; 2032 if ((i % LATENCY_LIMIT) == 0) 2033 cond_resched(); 2034 } 2035 2036 if (i == si->max) 2037 i = 0; 2038 2039 return i; 2040 } 2041 2042 static int try_to_unuse(unsigned int type) 2043 { 2044 struct mm_struct *prev_mm; 2045 struct mm_struct *mm; 2046 struct list_head *p; 2047 int retval = 0; 2048 struct swap_info_struct *si = swap_info[type]; 2049 struct page *page; 2050 swp_entry_t entry; 2051 unsigned int i; 2052 2053 if (!READ_ONCE(si->inuse_pages)) 2054 return 0; 2055 2056 retry: 2057 retval = shmem_unuse(type); 2058 if (retval) 2059 return retval; 2060 2061 prev_mm = &init_mm; 2062 mmget(prev_mm); 2063 2064 spin_lock(&mmlist_lock); 2065 p = &init_mm.mmlist; 2066 while (READ_ONCE(si->inuse_pages) && 2067 !signal_pending(current) && 2068 (p = p->next) != &init_mm.mmlist) { 2069 2070 mm = list_entry(p, struct mm_struct, mmlist); 2071 if (!mmget_not_zero(mm)) 2072 continue; 2073 spin_unlock(&mmlist_lock); 2074 mmput(prev_mm); 2075 prev_mm = mm; 2076 retval = unuse_mm(mm, type); 2077 if (retval) { 2078 mmput(prev_mm); 2079 return retval; 2080 } 2081 2082 /* 2083 * Make sure that we aren't completely killing 2084 * interactive performance. 2085 */ 2086 cond_resched(); 2087 spin_lock(&mmlist_lock); 2088 } 2089 spin_unlock(&mmlist_lock); 2090 2091 mmput(prev_mm); 2092 2093 i = 0; 2094 while (READ_ONCE(si->inuse_pages) && 2095 !signal_pending(current) && 2096 (i = find_next_to_unuse(si, i)) != 0) { 2097 2098 entry = swp_entry(type, i); 2099 page = find_get_page(swap_address_space(entry), i); 2100 if (!page) 2101 continue; 2102 2103 /* 2104 * It is conceivable that a racing task removed this page from 2105 * swap cache just before we acquired the page lock. The page 2106 * might even be back in swap cache on another swap area. But 2107 * that is okay, try_to_free_swap() only removes stale pages. 2108 */ 2109 lock_page(page); 2110 wait_on_page_writeback(page); 2111 try_to_free_swap(page); 2112 unlock_page(page); 2113 put_page(page); 2114 } 2115 2116 /* 2117 * Lets check again to see if there are still swap entries in the map. 2118 * If yes, we would need to do retry the unuse logic again. 2119 * Under global memory pressure, swap entries can be reinserted back 2120 * into process space after the mmlist loop above passes over them. 2121 * 2122 * Limit the number of retries? No: when mmget_not_zero() 2123 * above fails, that mm is likely to be freeing swap from 2124 * exit_mmap(), which proceeds at its own independent pace; 2125 * and even shmem_writepage() could have been preempted after 2126 * folio_alloc_swap(), temporarily hiding that swap. It's easy 2127 * and robust (though cpu-intensive) just to keep retrying. 2128 */ 2129 if (READ_ONCE(si->inuse_pages)) { 2130 if (!signal_pending(current)) 2131 goto retry; 2132 return -EINTR; 2133 } 2134 2135 return 0; 2136 } 2137 2138 /* 2139 * After a successful try_to_unuse, if no swap is now in use, we know 2140 * we can empty the mmlist. swap_lock must be held on entry and exit. 2141 * Note that mmlist_lock nests inside swap_lock, and an mm must be 2142 * added to the mmlist just after page_duplicate - before would be racy. 2143 */ 2144 static void drain_mmlist(void) 2145 { 2146 struct list_head *p, *next; 2147 unsigned int type; 2148 2149 for (type = 0; type < nr_swapfiles; type++) 2150 if (swap_info[type]->inuse_pages) 2151 return; 2152 spin_lock(&mmlist_lock); 2153 list_for_each_safe(p, next, &init_mm.mmlist) 2154 list_del_init(p); 2155 spin_unlock(&mmlist_lock); 2156 } 2157 2158 /* 2159 * Free all of a swapdev's extent information 2160 */ 2161 static void destroy_swap_extents(struct swap_info_struct *sis) 2162 { 2163 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) { 2164 struct rb_node *rb = sis->swap_extent_root.rb_node; 2165 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node); 2166 2167 rb_erase(rb, &sis->swap_extent_root); 2168 kfree(se); 2169 } 2170 2171 if (sis->flags & SWP_ACTIVATED) { 2172 struct file *swap_file = sis->swap_file; 2173 struct address_space *mapping = swap_file->f_mapping; 2174 2175 sis->flags &= ~SWP_ACTIVATED; 2176 if (mapping->a_ops->swap_deactivate) 2177 mapping->a_ops->swap_deactivate(swap_file); 2178 } 2179 } 2180 2181 /* 2182 * Add a block range (and the corresponding page range) into this swapdev's 2183 * extent tree. 2184 * 2185 * This function rather assumes that it is called in ascending page order. 2186 */ 2187 int 2188 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page, 2189 unsigned long nr_pages, sector_t start_block) 2190 { 2191 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL; 2192 struct swap_extent *se; 2193 struct swap_extent *new_se; 2194 2195 /* 2196 * place the new node at the right most since the 2197 * function is called in ascending page order. 2198 */ 2199 while (*link) { 2200 parent = *link; 2201 link = &parent->rb_right; 2202 } 2203 2204 if (parent) { 2205 se = rb_entry(parent, struct swap_extent, rb_node); 2206 BUG_ON(se->start_page + se->nr_pages != start_page); 2207 if (se->start_block + se->nr_pages == start_block) { 2208 /* Merge it */ 2209 se->nr_pages += nr_pages; 2210 return 0; 2211 } 2212 } 2213 2214 /* No merge, insert a new extent. */ 2215 new_se = kmalloc(sizeof(*se), GFP_KERNEL); 2216 if (new_se == NULL) 2217 return -ENOMEM; 2218 new_se->start_page = start_page; 2219 new_se->nr_pages = nr_pages; 2220 new_se->start_block = start_block; 2221 2222 rb_link_node(&new_se->rb_node, parent, link); 2223 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root); 2224 return 1; 2225 } 2226 EXPORT_SYMBOL_GPL(add_swap_extent); 2227 2228 /* 2229 * A `swap extent' is a simple thing which maps a contiguous range of pages 2230 * onto a contiguous range of disk blocks. A rbtree of swap extents is 2231 * built at swapon time and is then used at swap_writepage/swap_readpage 2232 * time for locating where on disk a page belongs. 2233 * 2234 * If the swapfile is an S_ISBLK block device, a single extent is installed. 2235 * This is done so that the main operating code can treat S_ISBLK and S_ISREG 2236 * swap files identically. 2237 * 2238 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap 2239 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK 2240 * swapfiles are handled *identically* after swapon time. 2241 * 2242 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks 2243 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray 2244 * blocks are found which do not fall within the PAGE_SIZE alignment 2245 * requirements, they are simply tossed out - we will never use those blocks 2246 * for swapping. 2247 * 2248 * For all swap devices we set S_SWAPFILE across the life of the swapon. This 2249 * prevents users from writing to the swap device, which will corrupt memory. 2250 * 2251 * The amount of disk space which a single swap extent represents varies. 2252 * Typically it is in the 1-4 megabyte range. So we can have hundreds of 2253 * extents in the rbtree. - akpm. 2254 */ 2255 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span) 2256 { 2257 struct file *swap_file = sis->swap_file; 2258 struct address_space *mapping = swap_file->f_mapping; 2259 struct inode *inode = mapping->host; 2260 int ret; 2261 2262 if (S_ISBLK(inode->i_mode)) { 2263 ret = add_swap_extent(sis, 0, sis->max, 0); 2264 *span = sis->pages; 2265 return ret; 2266 } 2267 2268 if (mapping->a_ops->swap_activate) { 2269 ret = mapping->a_ops->swap_activate(sis, swap_file, span); 2270 if (ret < 0) 2271 return ret; 2272 sis->flags |= SWP_ACTIVATED; 2273 if ((sis->flags & SWP_FS_OPS) && 2274 sio_pool_init() != 0) { 2275 destroy_swap_extents(sis); 2276 return -ENOMEM; 2277 } 2278 return ret; 2279 } 2280 2281 return generic_swapfile_activate(sis, swap_file, span); 2282 } 2283 2284 static int swap_node(struct swap_info_struct *p) 2285 { 2286 struct block_device *bdev; 2287 2288 if (p->bdev) 2289 bdev = p->bdev; 2290 else 2291 bdev = p->swap_file->f_inode->i_sb->s_bdev; 2292 2293 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE; 2294 } 2295 2296 static void setup_swap_info(struct swap_info_struct *p, int prio, 2297 unsigned char *swap_map, 2298 struct swap_cluster_info *cluster_info) 2299 { 2300 int i; 2301 2302 if (prio >= 0) 2303 p->prio = prio; 2304 else 2305 p->prio = --least_priority; 2306 /* 2307 * the plist prio is negated because plist ordering is 2308 * low-to-high, while swap ordering is high-to-low 2309 */ 2310 p->list.prio = -p->prio; 2311 for_each_node(i) { 2312 if (p->prio >= 0) 2313 p->avail_lists[i].prio = -p->prio; 2314 else { 2315 if (swap_node(p) == i) 2316 p->avail_lists[i].prio = 1; 2317 else 2318 p->avail_lists[i].prio = -p->prio; 2319 } 2320 } 2321 p->swap_map = swap_map; 2322 p->cluster_info = cluster_info; 2323 } 2324 2325 static void _enable_swap_info(struct swap_info_struct *p) 2326 { 2327 p->flags |= SWP_WRITEOK; 2328 atomic_long_add(p->pages, &nr_swap_pages); 2329 total_swap_pages += p->pages; 2330 2331 assert_spin_locked(&swap_lock); 2332 /* 2333 * both lists are plists, and thus priority ordered. 2334 * swap_active_head needs to be priority ordered for swapoff(), 2335 * which on removal of any swap_info_struct with an auto-assigned 2336 * (i.e. negative) priority increments the auto-assigned priority 2337 * of any lower-priority swap_info_structs. 2338 * swap_avail_head needs to be priority ordered for folio_alloc_swap(), 2339 * which allocates swap pages from the highest available priority 2340 * swap_info_struct. 2341 */ 2342 plist_add(&p->list, &swap_active_head); 2343 add_to_avail_list(p); 2344 } 2345 2346 static void enable_swap_info(struct swap_info_struct *p, int prio, 2347 unsigned char *swap_map, 2348 struct swap_cluster_info *cluster_info, 2349 unsigned long *frontswap_map) 2350 { 2351 if (IS_ENABLED(CONFIG_FRONTSWAP)) 2352 frontswap_init(p->type, frontswap_map); 2353 spin_lock(&swap_lock); 2354 spin_lock(&p->lock); 2355 setup_swap_info(p, prio, swap_map, cluster_info); 2356 spin_unlock(&p->lock); 2357 spin_unlock(&swap_lock); 2358 /* 2359 * Finished initializing swap device, now it's safe to reference it. 2360 */ 2361 percpu_ref_resurrect(&p->users); 2362 spin_lock(&swap_lock); 2363 spin_lock(&p->lock); 2364 _enable_swap_info(p); 2365 spin_unlock(&p->lock); 2366 spin_unlock(&swap_lock); 2367 } 2368 2369 static void reinsert_swap_info(struct swap_info_struct *p) 2370 { 2371 spin_lock(&swap_lock); 2372 spin_lock(&p->lock); 2373 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info); 2374 _enable_swap_info(p); 2375 spin_unlock(&p->lock); 2376 spin_unlock(&swap_lock); 2377 } 2378 2379 bool has_usable_swap(void) 2380 { 2381 bool ret = true; 2382 2383 spin_lock(&swap_lock); 2384 if (plist_head_empty(&swap_active_head)) 2385 ret = false; 2386 spin_unlock(&swap_lock); 2387 return ret; 2388 } 2389 2390 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile) 2391 { 2392 struct swap_info_struct *p = NULL; 2393 unsigned char *swap_map; 2394 struct swap_cluster_info *cluster_info; 2395 unsigned long *frontswap_map; 2396 struct file *swap_file, *victim; 2397 struct address_space *mapping; 2398 struct inode *inode; 2399 struct filename *pathname; 2400 int err, found = 0; 2401 unsigned int old_block_size; 2402 2403 if (!capable(CAP_SYS_ADMIN)) 2404 return -EPERM; 2405 2406 BUG_ON(!current->mm); 2407 2408 pathname = getname(specialfile); 2409 if (IS_ERR(pathname)) 2410 return PTR_ERR(pathname); 2411 2412 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0); 2413 err = PTR_ERR(victim); 2414 if (IS_ERR(victim)) 2415 goto out; 2416 2417 mapping = victim->f_mapping; 2418 spin_lock(&swap_lock); 2419 plist_for_each_entry(p, &swap_active_head, list) { 2420 if (p->flags & SWP_WRITEOK) { 2421 if (p->swap_file->f_mapping == mapping) { 2422 found = 1; 2423 break; 2424 } 2425 } 2426 } 2427 if (!found) { 2428 err = -EINVAL; 2429 spin_unlock(&swap_lock); 2430 goto out_dput; 2431 } 2432 if (!security_vm_enough_memory_mm(current->mm, p->pages)) 2433 vm_unacct_memory(p->pages); 2434 else { 2435 err = -ENOMEM; 2436 spin_unlock(&swap_lock); 2437 goto out_dput; 2438 } 2439 del_from_avail_list(p); 2440 spin_lock(&p->lock); 2441 if (p->prio < 0) { 2442 struct swap_info_struct *si = p; 2443 int nid; 2444 2445 plist_for_each_entry_continue(si, &swap_active_head, list) { 2446 si->prio++; 2447 si->list.prio--; 2448 for_each_node(nid) { 2449 if (si->avail_lists[nid].prio != 1) 2450 si->avail_lists[nid].prio--; 2451 } 2452 } 2453 least_priority++; 2454 } 2455 plist_del(&p->list, &swap_active_head); 2456 atomic_long_sub(p->pages, &nr_swap_pages); 2457 total_swap_pages -= p->pages; 2458 p->flags &= ~SWP_WRITEOK; 2459 spin_unlock(&p->lock); 2460 spin_unlock(&swap_lock); 2461 2462 disable_swap_slots_cache_lock(); 2463 2464 set_current_oom_origin(); 2465 err = try_to_unuse(p->type); 2466 clear_current_oom_origin(); 2467 2468 if (err) { 2469 /* re-insert swap space back into swap_list */ 2470 reinsert_swap_info(p); 2471 reenable_swap_slots_cache_unlock(); 2472 goto out_dput; 2473 } 2474 2475 reenable_swap_slots_cache_unlock(); 2476 2477 /* 2478 * Wait for swap operations protected by get/put_swap_device() 2479 * to complete. 2480 * 2481 * We need synchronize_rcu() here to protect the accessing to 2482 * the swap cache data structure. 2483 */ 2484 percpu_ref_kill(&p->users); 2485 synchronize_rcu(); 2486 wait_for_completion(&p->comp); 2487 2488 flush_work(&p->discard_work); 2489 2490 destroy_swap_extents(p); 2491 if (p->flags & SWP_CONTINUED) 2492 free_swap_count_continuations(p); 2493 2494 if (!p->bdev || !bdev_nonrot(p->bdev)) 2495 atomic_dec(&nr_rotate_swap); 2496 2497 mutex_lock(&swapon_mutex); 2498 spin_lock(&swap_lock); 2499 spin_lock(&p->lock); 2500 drain_mmlist(); 2501 2502 /* wait for anyone still in scan_swap_map_slots */ 2503 p->highest_bit = 0; /* cuts scans short */ 2504 while (p->flags >= SWP_SCANNING) { 2505 spin_unlock(&p->lock); 2506 spin_unlock(&swap_lock); 2507 schedule_timeout_uninterruptible(1); 2508 spin_lock(&swap_lock); 2509 spin_lock(&p->lock); 2510 } 2511 2512 swap_file = p->swap_file; 2513 old_block_size = p->old_block_size; 2514 p->swap_file = NULL; 2515 p->max = 0; 2516 swap_map = p->swap_map; 2517 p->swap_map = NULL; 2518 cluster_info = p->cluster_info; 2519 p->cluster_info = NULL; 2520 frontswap_map = frontswap_map_get(p); 2521 spin_unlock(&p->lock); 2522 spin_unlock(&swap_lock); 2523 arch_swap_invalidate_area(p->type); 2524 frontswap_invalidate_area(p->type); 2525 frontswap_map_set(p, NULL); 2526 mutex_unlock(&swapon_mutex); 2527 free_percpu(p->percpu_cluster); 2528 p->percpu_cluster = NULL; 2529 free_percpu(p->cluster_next_cpu); 2530 p->cluster_next_cpu = NULL; 2531 vfree(swap_map); 2532 kvfree(cluster_info); 2533 kvfree(frontswap_map); 2534 /* Destroy swap account information */ 2535 swap_cgroup_swapoff(p->type); 2536 exit_swap_address_space(p->type); 2537 2538 inode = mapping->host; 2539 if (S_ISBLK(inode->i_mode)) { 2540 struct block_device *bdev = I_BDEV(inode); 2541 2542 set_blocksize(bdev, old_block_size); 2543 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL); 2544 } 2545 2546 inode_lock(inode); 2547 inode->i_flags &= ~S_SWAPFILE; 2548 inode_unlock(inode); 2549 filp_close(swap_file, NULL); 2550 2551 /* 2552 * Clear the SWP_USED flag after all resources are freed so that swapon 2553 * can reuse this swap_info in alloc_swap_info() safely. It is ok to 2554 * not hold p->lock after we cleared its SWP_WRITEOK. 2555 */ 2556 spin_lock(&swap_lock); 2557 p->flags = 0; 2558 spin_unlock(&swap_lock); 2559 2560 err = 0; 2561 atomic_inc(&proc_poll_event); 2562 wake_up_interruptible(&proc_poll_wait); 2563 2564 out_dput: 2565 filp_close(victim, NULL); 2566 out: 2567 putname(pathname); 2568 return err; 2569 } 2570 2571 #ifdef CONFIG_PROC_FS 2572 static __poll_t swaps_poll(struct file *file, poll_table *wait) 2573 { 2574 struct seq_file *seq = file->private_data; 2575 2576 poll_wait(file, &proc_poll_wait, wait); 2577 2578 if (seq->poll_event != atomic_read(&proc_poll_event)) { 2579 seq->poll_event = atomic_read(&proc_poll_event); 2580 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI; 2581 } 2582 2583 return EPOLLIN | EPOLLRDNORM; 2584 } 2585 2586 /* iterator */ 2587 static void *swap_start(struct seq_file *swap, loff_t *pos) 2588 { 2589 struct swap_info_struct *si; 2590 int type; 2591 loff_t l = *pos; 2592 2593 mutex_lock(&swapon_mutex); 2594 2595 if (!l) 2596 return SEQ_START_TOKEN; 2597 2598 for (type = 0; (si = swap_type_to_swap_info(type)); type++) { 2599 if (!(si->flags & SWP_USED) || !si->swap_map) 2600 continue; 2601 if (!--l) 2602 return si; 2603 } 2604 2605 return NULL; 2606 } 2607 2608 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos) 2609 { 2610 struct swap_info_struct *si = v; 2611 int type; 2612 2613 if (v == SEQ_START_TOKEN) 2614 type = 0; 2615 else 2616 type = si->type + 1; 2617 2618 ++(*pos); 2619 for (; (si = swap_type_to_swap_info(type)); type++) { 2620 if (!(si->flags & SWP_USED) || !si->swap_map) 2621 continue; 2622 return si; 2623 } 2624 2625 return NULL; 2626 } 2627 2628 static void swap_stop(struct seq_file *swap, void *v) 2629 { 2630 mutex_unlock(&swapon_mutex); 2631 } 2632 2633 static int swap_show(struct seq_file *swap, void *v) 2634 { 2635 struct swap_info_struct *si = v; 2636 struct file *file; 2637 int len; 2638 unsigned long bytes, inuse; 2639 2640 if (si == SEQ_START_TOKEN) { 2641 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n"); 2642 return 0; 2643 } 2644 2645 bytes = si->pages << (PAGE_SHIFT - 10); 2646 inuse = READ_ONCE(si->inuse_pages) << (PAGE_SHIFT - 10); 2647 2648 file = si->swap_file; 2649 len = seq_file_path(swap, file, " \t\n\\"); 2650 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n", 2651 len < 40 ? 40 - len : 1, " ", 2652 S_ISBLK(file_inode(file)->i_mode) ? 2653 "partition" : "file\t", 2654 bytes, bytes < 10000000 ? "\t" : "", 2655 inuse, inuse < 10000000 ? "\t" : "", 2656 si->prio); 2657 return 0; 2658 } 2659 2660 static const struct seq_operations swaps_op = { 2661 .start = swap_start, 2662 .next = swap_next, 2663 .stop = swap_stop, 2664 .show = swap_show 2665 }; 2666 2667 static int swaps_open(struct inode *inode, struct file *file) 2668 { 2669 struct seq_file *seq; 2670 int ret; 2671 2672 ret = seq_open(file, &swaps_op); 2673 if (ret) 2674 return ret; 2675 2676 seq = file->private_data; 2677 seq->poll_event = atomic_read(&proc_poll_event); 2678 return 0; 2679 } 2680 2681 static const struct proc_ops swaps_proc_ops = { 2682 .proc_flags = PROC_ENTRY_PERMANENT, 2683 .proc_open = swaps_open, 2684 .proc_read = seq_read, 2685 .proc_lseek = seq_lseek, 2686 .proc_release = seq_release, 2687 .proc_poll = swaps_poll, 2688 }; 2689 2690 static int __init procswaps_init(void) 2691 { 2692 proc_create("swaps", 0, NULL, &swaps_proc_ops); 2693 return 0; 2694 } 2695 __initcall(procswaps_init); 2696 #endif /* CONFIG_PROC_FS */ 2697 2698 #ifdef MAX_SWAPFILES_CHECK 2699 static int __init max_swapfiles_check(void) 2700 { 2701 MAX_SWAPFILES_CHECK(); 2702 return 0; 2703 } 2704 late_initcall(max_swapfiles_check); 2705 #endif 2706 2707 static struct swap_info_struct *alloc_swap_info(void) 2708 { 2709 struct swap_info_struct *p; 2710 struct swap_info_struct *defer = NULL; 2711 unsigned int type; 2712 int i; 2713 2714 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL); 2715 if (!p) 2716 return ERR_PTR(-ENOMEM); 2717 2718 if (percpu_ref_init(&p->users, swap_users_ref_free, 2719 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) { 2720 kvfree(p); 2721 return ERR_PTR(-ENOMEM); 2722 } 2723 2724 spin_lock(&swap_lock); 2725 for (type = 0; type < nr_swapfiles; type++) { 2726 if (!(swap_info[type]->flags & SWP_USED)) 2727 break; 2728 } 2729 if (type >= MAX_SWAPFILES) { 2730 spin_unlock(&swap_lock); 2731 percpu_ref_exit(&p->users); 2732 kvfree(p); 2733 return ERR_PTR(-EPERM); 2734 } 2735 if (type >= nr_swapfiles) { 2736 p->type = type; 2737 /* 2738 * Publish the swap_info_struct after initializing it. 2739 * Note that kvzalloc() above zeroes all its fields. 2740 */ 2741 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */ 2742 nr_swapfiles++; 2743 } else { 2744 defer = p; 2745 p = swap_info[type]; 2746 /* 2747 * Do not memset this entry: a racing procfs swap_next() 2748 * would be relying on p->type to remain valid. 2749 */ 2750 } 2751 p->swap_extent_root = RB_ROOT; 2752 plist_node_init(&p->list, 0); 2753 for_each_node(i) 2754 plist_node_init(&p->avail_lists[i], 0); 2755 p->flags = SWP_USED; 2756 spin_unlock(&swap_lock); 2757 if (defer) { 2758 percpu_ref_exit(&defer->users); 2759 kvfree(defer); 2760 } 2761 spin_lock_init(&p->lock); 2762 spin_lock_init(&p->cont_lock); 2763 init_completion(&p->comp); 2764 2765 return p; 2766 } 2767 2768 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode) 2769 { 2770 int error; 2771 2772 if (S_ISBLK(inode->i_mode)) { 2773 p->bdev = blkdev_get_by_dev(inode->i_rdev, 2774 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p); 2775 if (IS_ERR(p->bdev)) { 2776 error = PTR_ERR(p->bdev); 2777 p->bdev = NULL; 2778 return error; 2779 } 2780 p->old_block_size = block_size(p->bdev); 2781 error = set_blocksize(p->bdev, PAGE_SIZE); 2782 if (error < 0) 2783 return error; 2784 /* 2785 * Zoned block devices contain zones that have a sequential 2786 * write only restriction. Hence zoned block devices are not 2787 * suitable for swapping. Disallow them here. 2788 */ 2789 if (bdev_is_zoned(p->bdev)) 2790 return -EINVAL; 2791 p->flags |= SWP_BLKDEV; 2792 } else if (S_ISREG(inode->i_mode)) { 2793 p->bdev = inode->i_sb->s_bdev; 2794 } 2795 2796 return 0; 2797 } 2798 2799 2800 /* 2801 * Find out how many pages are allowed for a single swap device. There 2802 * are two limiting factors: 2803 * 1) the number of bits for the swap offset in the swp_entry_t type, and 2804 * 2) the number of bits in the swap pte, as defined by the different 2805 * architectures. 2806 * 2807 * In order to find the largest possible bit mask, a swap entry with 2808 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte, 2809 * decoded to a swp_entry_t again, and finally the swap offset is 2810 * extracted. 2811 * 2812 * This will mask all the bits from the initial ~0UL mask that can't 2813 * be encoded in either the swp_entry_t or the architecture definition 2814 * of a swap pte. 2815 */ 2816 unsigned long generic_max_swapfile_size(void) 2817 { 2818 return swp_offset(pte_to_swp_entry( 2819 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1; 2820 } 2821 2822 /* Can be overridden by an architecture for additional checks. */ 2823 __weak unsigned long arch_max_swapfile_size(void) 2824 { 2825 return generic_max_swapfile_size(); 2826 } 2827 2828 static unsigned long read_swap_header(struct swap_info_struct *p, 2829 union swap_header *swap_header, 2830 struct inode *inode) 2831 { 2832 int i; 2833 unsigned long maxpages; 2834 unsigned long swapfilepages; 2835 unsigned long last_page; 2836 2837 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) { 2838 pr_err("Unable to find swap-space signature\n"); 2839 return 0; 2840 } 2841 2842 /* swap partition endianness hack... */ 2843 if (swab32(swap_header->info.version) == 1) { 2844 swab32s(&swap_header->info.version); 2845 swab32s(&swap_header->info.last_page); 2846 swab32s(&swap_header->info.nr_badpages); 2847 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES) 2848 return 0; 2849 for (i = 0; i < swap_header->info.nr_badpages; i++) 2850 swab32s(&swap_header->info.badpages[i]); 2851 } 2852 /* Check the swap header's sub-version */ 2853 if (swap_header->info.version != 1) { 2854 pr_warn("Unable to handle swap header version %d\n", 2855 swap_header->info.version); 2856 return 0; 2857 } 2858 2859 p->lowest_bit = 1; 2860 p->cluster_next = 1; 2861 p->cluster_nr = 0; 2862 2863 maxpages = swapfile_maximum_size; 2864 last_page = swap_header->info.last_page; 2865 if (!last_page) { 2866 pr_warn("Empty swap-file\n"); 2867 return 0; 2868 } 2869 if (last_page > maxpages) { 2870 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n", 2871 maxpages << (PAGE_SHIFT - 10), 2872 last_page << (PAGE_SHIFT - 10)); 2873 } 2874 if (maxpages > last_page) { 2875 maxpages = last_page + 1; 2876 /* p->max is an unsigned int: don't overflow it */ 2877 if ((unsigned int)maxpages == 0) 2878 maxpages = UINT_MAX; 2879 } 2880 p->highest_bit = maxpages - 1; 2881 2882 if (!maxpages) 2883 return 0; 2884 swapfilepages = i_size_read(inode) >> PAGE_SHIFT; 2885 if (swapfilepages && maxpages > swapfilepages) { 2886 pr_warn("Swap area shorter than signature indicates\n"); 2887 return 0; 2888 } 2889 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode)) 2890 return 0; 2891 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES) 2892 return 0; 2893 2894 return maxpages; 2895 } 2896 2897 #define SWAP_CLUSTER_INFO_COLS \ 2898 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info)) 2899 #define SWAP_CLUSTER_SPACE_COLS \ 2900 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER) 2901 #define SWAP_CLUSTER_COLS \ 2902 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS) 2903 2904 static int setup_swap_map_and_extents(struct swap_info_struct *p, 2905 union swap_header *swap_header, 2906 unsigned char *swap_map, 2907 struct swap_cluster_info *cluster_info, 2908 unsigned long maxpages, 2909 sector_t *span) 2910 { 2911 unsigned int j, k; 2912 unsigned int nr_good_pages; 2913 int nr_extents; 2914 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER); 2915 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS; 2916 unsigned long i, idx; 2917 2918 nr_good_pages = maxpages - 1; /* omit header page */ 2919 2920 cluster_list_init(&p->free_clusters); 2921 cluster_list_init(&p->discard_clusters); 2922 2923 for (i = 0; i < swap_header->info.nr_badpages; i++) { 2924 unsigned int page_nr = swap_header->info.badpages[i]; 2925 if (page_nr == 0 || page_nr > swap_header->info.last_page) 2926 return -EINVAL; 2927 if (page_nr < maxpages) { 2928 swap_map[page_nr] = SWAP_MAP_BAD; 2929 nr_good_pages--; 2930 /* 2931 * Haven't marked the cluster free yet, no list 2932 * operation involved 2933 */ 2934 inc_cluster_info_page(p, cluster_info, page_nr); 2935 } 2936 } 2937 2938 /* Haven't marked the cluster free yet, no list operation involved */ 2939 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++) 2940 inc_cluster_info_page(p, cluster_info, i); 2941 2942 if (nr_good_pages) { 2943 swap_map[0] = SWAP_MAP_BAD; 2944 /* 2945 * Not mark the cluster free yet, no list 2946 * operation involved 2947 */ 2948 inc_cluster_info_page(p, cluster_info, 0); 2949 p->max = maxpages; 2950 p->pages = nr_good_pages; 2951 nr_extents = setup_swap_extents(p, span); 2952 if (nr_extents < 0) 2953 return nr_extents; 2954 nr_good_pages = p->pages; 2955 } 2956 if (!nr_good_pages) { 2957 pr_warn("Empty swap-file\n"); 2958 return -EINVAL; 2959 } 2960 2961 if (!cluster_info) 2962 return nr_extents; 2963 2964 2965 /* 2966 * Reduce false cache line sharing between cluster_info and 2967 * sharing same address space. 2968 */ 2969 for (k = 0; k < SWAP_CLUSTER_COLS; k++) { 2970 j = (k + col) % SWAP_CLUSTER_COLS; 2971 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) { 2972 idx = i * SWAP_CLUSTER_COLS + j; 2973 if (idx >= nr_clusters) 2974 continue; 2975 if (cluster_count(&cluster_info[idx])) 2976 continue; 2977 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE); 2978 cluster_list_add_tail(&p->free_clusters, cluster_info, 2979 idx); 2980 } 2981 } 2982 return nr_extents; 2983 } 2984 2985 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags) 2986 { 2987 struct swap_info_struct *p; 2988 struct filename *name; 2989 struct file *swap_file = NULL; 2990 struct address_space *mapping; 2991 struct dentry *dentry; 2992 int prio; 2993 int error; 2994 union swap_header *swap_header; 2995 int nr_extents; 2996 sector_t span; 2997 unsigned long maxpages; 2998 unsigned char *swap_map = NULL; 2999 struct swap_cluster_info *cluster_info = NULL; 3000 unsigned long *frontswap_map = NULL; 3001 struct page *page = NULL; 3002 struct inode *inode = NULL; 3003 bool inced_nr_rotate_swap = false; 3004 3005 if (swap_flags & ~SWAP_FLAGS_VALID) 3006 return -EINVAL; 3007 3008 if (!capable(CAP_SYS_ADMIN)) 3009 return -EPERM; 3010 3011 if (!swap_avail_heads) 3012 return -ENOMEM; 3013 3014 p = alloc_swap_info(); 3015 if (IS_ERR(p)) 3016 return PTR_ERR(p); 3017 3018 INIT_WORK(&p->discard_work, swap_discard_work); 3019 3020 name = getname(specialfile); 3021 if (IS_ERR(name)) { 3022 error = PTR_ERR(name); 3023 name = NULL; 3024 goto bad_swap; 3025 } 3026 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0); 3027 if (IS_ERR(swap_file)) { 3028 error = PTR_ERR(swap_file); 3029 swap_file = NULL; 3030 goto bad_swap; 3031 } 3032 3033 p->swap_file = swap_file; 3034 mapping = swap_file->f_mapping; 3035 dentry = swap_file->f_path.dentry; 3036 inode = mapping->host; 3037 3038 error = claim_swapfile(p, inode); 3039 if (unlikely(error)) 3040 goto bad_swap; 3041 3042 inode_lock(inode); 3043 if (d_unlinked(dentry) || cant_mount(dentry)) { 3044 error = -ENOENT; 3045 goto bad_swap_unlock_inode; 3046 } 3047 if (IS_SWAPFILE(inode)) { 3048 error = -EBUSY; 3049 goto bad_swap_unlock_inode; 3050 } 3051 3052 /* 3053 * Read the swap header. 3054 */ 3055 if (!mapping->a_ops->read_folio) { 3056 error = -EINVAL; 3057 goto bad_swap_unlock_inode; 3058 } 3059 page = read_mapping_page(mapping, 0, swap_file); 3060 if (IS_ERR(page)) { 3061 error = PTR_ERR(page); 3062 goto bad_swap_unlock_inode; 3063 } 3064 swap_header = kmap(page); 3065 3066 maxpages = read_swap_header(p, swap_header, inode); 3067 if (unlikely(!maxpages)) { 3068 error = -EINVAL; 3069 goto bad_swap_unlock_inode; 3070 } 3071 3072 /* OK, set up the swap map and apply the bad block list */ 3073 swap_map = vzalloc(maxpages); 3074 if (!swap_map) { 3075 error = -ENOMEM; 3076 goto bad_swap_unlock_inode; 3077 } 3078 3079 if (p->bdev && bdev_stable_writes(p->bdev)) 3080 p->flags |= SWP_STABLE_WRITES; 3081 3082 if (p->bdev && p->bdev->bd_disk->fops->rw_page) 3083 p->flags |= SWP_SYNCHRONOUS_IO; 3084 3085 if (p->bdev && bdev_nonrot(p->bdev)) { 3086 int cpu; 3087 unsigned long ci, nr_cluster; 3088 3089 p->flags |= SWP_SOLIDSTATE; 3090 p->cluster_next_cpu = alloc_percpu(unsigned int); 3091 if (!p->cluster_next_cpu) { 3092 error = -ENOMEM; 3093 goto bad_swap_unlock_inode; 3094 } 3095 /* 3096 * select a random position to start with to help wear leveling 3097 * SSD 3098 */ 3099 for_each_possible_cpu(cpu) { 3100 per_cpu(*p->cluster_next_cpu, cpu) = 3101 1 + prandom_u32_max(p->highest_bit); 3102 } 3103 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER); 3104 3105 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info), 3106 GFP_KERNEL); 3107 if (!cluster_info) { 3108 error = -ENOMEM; 3109 goto bad_swap_unlock_inode; 3110 } 3111 3112 for (ci = 0; ci < nr_cluster; ci++) 3113 spin_lock_init(&((cluster_info + ci)->lock)); 3114 3115 p->percpu_cluster = alloc_percpu(struct percpu_cluster); 3116 if (!p->percpu_cluster) { 3117 error = -ENOMEM; 3118 goto bad_swap_unlock_inode; 3119 } 3120 for_each_possible_cpu(cpu) { 3121 struct percpu_cluster *cluster; 3122 cluster = per_cpu_ptr(p->percpu_cluster, cpu); 3123 cluster_set_null(&cluster->index); 3124 } 3125 } else { 3126 atomic_inc(&nr_rotate_swap); 3127 inced_nr_rotate_swap = true; 3128 } 3129 3130 error = swap_cgroup_swapon(p->type, maxpages); 3131 if (error) 3132 goto bad_swap_unlock_inode; 3133 3134 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map, 3135 cluster_info, maxpages, &span); 3136 if (unlikely(nr_extents < 0)) { 3137 error = nr_extents; 3138 goto bad_swap_unlock_inode; 3139 } 3140 /* frontswap enabled? set up bit-per-page map for frontswap */ 3141 if (IS_ENABLED(CONFIG_FRONTSWAP)) 3142 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages), 3143 sizeof(long), 3144 GFP_KERNEL); 3145 3146 if ((swap_flags & SWAP_FLAG_DISCARD) && 3147 p->bdev && bdev_max_discard_sectors(p->bdev)) { 3148 /* 3149 * When discard is enabled for swap with no particular 3150 * policy flagged, we set all swap discard flags here in 3151 * order to sustain backward compatibility with older 3152 * swapon(8) releases. 3153 */ 3154 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD | 3155 SWP_PAGE_DISCARD); 3156 3157 /* 3158 * By flagging sys_swapon, a sysadmin can tell us to 3159 * either do single-time area discards only, or to just 3160 * perform discards for released swap page-clusters. 3161 * Now it's time to adjust the p->flags accordingly. 3162 */ 3163 if (swap_flags & SWAP_FLAG_DISCARD_ONCE) 3164 p->flags &= ~SWP_PAGE_DISCARD; 3165 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES) 3166 p->flags &= ~SWP_AREA_DISCARD; 3167 3168 /* issue a swapon-time discard if it's still required */ 3169 if (p->flags & SWP_AREA_DISCARD) { 3170 int err = discard_swap(p); 3171 if (unlikely(err)) 3172 pr_err("swapon: discard_swap(%p): %d\n", 3173 p, err); 3174 } 3175 } 3176 3177 error = init_swap_address_space(p->type, maxpages); 3178 if (error) 3179 goto bad_swap_unlock_inode; 3180 3181 /* 3182 * Flush any pending IO and dirty mappings before we start using this 3183 * swap device. 3184 */ 3185 inode->i_flags |= S_SWAPFILE; 3186 error = inode_drain_writes(inode); 3187 if (error) { 3188 inode->i_flags &= ~S_SWAPFILE; 3189 goto free_swap_address_space; 3190 } 3191 3192 mutex_lock(&swapon_mutex); 3193 prio = -1; 3194 if (swap_flags & SWAP_FLAG_PREFER) 3195 prio = 3196 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT; 3197 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map); 3198 3199 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n", 3200 p->pages<<(PAGE_SHIFT-10), name->name, p->prio, 3201 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10), 3202 (p->flags & SWP_SOLIDSTATE) ? "SS" : "", 3203 (p->flags & SWP_DISCARDABLE) ? "D" : "", 3204 (p->flags & SWP_AREA_DISCARD) ? "s" : "", 3205 (p->flags & SWP_PAGE_DISCARD) ? "c" : "", 3206 (frontswap_map) ? "FS" : ""); 3207 3208 mutex_unlock(&swapon_mutex); 3209 atomic_inc(&proc_poll_event); 3210 wake_up_interruptible(&proc_poll_wait); 3211 3212 error = 0; 3213 goto out; 3214 free_swap_address_space: 3215 exit_swap_address_space(p->type); 3216 bad_swap_unlock_inode: 3217 inode_unlock(inode); 3218 bad_swap: 3219 free_percpu(p->percpu_cluster); 3220 p->percpu_cluster = NULL; 3221 free_percpu(p->cluster_next_cpu); 3222 p->cluster_next_cpu = NULL; 3223 if (inode && S_ISBLK(inode->i_mode) && p->bdev) { 3224 set_blocksize(p->bdev, p->old_block_size); 3225 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL); 3226 } 3227 inode = NULL; 3228 destroy_swap_extents(p); 3229 swap_cgroup_swapoff(p->type); 3230 spin_lock(&swap_lock); 3231 p->swap_file = NULL; 3232 p->flags = 0; 3233 spin_unlock(&swap_lock); 3234 vfree(swap_map); 3235 kvfree(cluster_info); 3236 kvfree(frontswap_map); 3237 if (inced_nr_rotate_swap) 3238 atomic_dec(&nr_rotate_swap); 3239 if (swap_file) 3240 filp_close(swap_file, NULL); 3241 out: 3242 if (page && !IS_ERR(page)) { 3243 kunmap(page); 3244 put_page(page); 3245 } 3246 if (name) 3247 putname(name); 3248 if (inode) 3249 inode_unlock(inode); 3250 if (!error) 3251 enable_swap_slots_cache(); 3252 return error; 3253 } 3254 3255 void si_swapinfo(struct sysinfo *val) 3256 { 3257 unsigned int type; 3258 unsigned long nr_to_be_unused = 0; 3259 3260 spin_lock(&swap_lock); 3261 for (type = 0; type < nr_swapfiles; type++) { 3262 struct swap_info_struct *si = swap_info[type]; 3263 3264 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK)) 3265 nr_to_be_unused += READ_ONCE(si->inuse_pages); 3266 } 3267 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused; 3268 val->totalswap = total_swap_pages + nr_to_be_unused; 3269 spin_unlock(&swap_lock); 3270 } 3271 3272 /* 3273 * Verify that a swap entry is valid and increment its swap map count. 3274 * 3275 * Returns error code in following case. 3276 * - success -> 0 3277 * - swp_entry is invalid -> EINVAL 3278 * - swp_entry is migration entry -> EINVAL 3279 * - swap-cache reference is requested but there is already one. -> EEXIST 3280 * - swap-cache reference is requested but the entry is not used. -> ENOENT 3281 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM 3282 */ 3283 static int __swap_duplicate(swp_entry_t entry, unsigned char usage) 3284 { 3285 struct swap_info_struct *p; 3286 struct swap_cluster_info *ci; 3287 unsigned long offset; 3288 unsigned char count; 3289 unsigned char has_cache; 3290 int err; 3291 3292 p = get_swap_device(entry); 3293 if (!p) 3294 return -EINVAL; 3295 3296 offset = swp_offset(entry); 3297 ci = lock_cluster_or_swap_info(p, offset); 3298 3299 count = p->swap_map[offset]; 3300 3301 /* 3302 * swapin_readahead() doesn't check if a swap entry is valid, so the 3303 * swap entry could be SWAP_MAP_BAD. Check here with lock held. 3304 */ 3305 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) { 3306 err = -ENOENT; 3307 goto unlock_out; 3308 } 3309 3310 has_cache = count & SWAP_HAS_CACHE; 3311 count &= ~SWAP_HAS_CACHE; 3312 err = 0; 3313 3314 if (usage == SWAP_HAS_CACHE) { 3315 3316 /* set SWAP_HAS_CACHE if there is no cache and entry is used */ 3317 if (!has_cache && count) 3318 has_cache = SWAP_HAS_CACHE; 3319 else if (has_cache) /* someone else added cache */ 3320 err = -EEXIST; 3321 else /* no users remaining */ 3322 err = -ENOENT; 3323 3324 } else if (count || has_cache) { 3325 3326 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX) 3327 count += usage; 3328 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX) 3329 err = -EINVAL; 3330 else if (swap_count_continued(p, offset, count)) 3331 count = COUNT_CONTINUED; 3332 else 3333 err = -ENOMEM; 3334 } else 3335 err = -ENOENT; /* unused swap entry */ 3336 3337 WRITE_ONCE(p->swap_map[offset], count | has_cache); 3338 3339 unlock_out: 3340 unlock_cluster_or_swap_info(p, ci); 3341 put_swap_device(p); 3342 return err; 3343 } 3344 3345 /* 3346 * Help swapoff by noting that swap entry belongs to shmem/tmpfs 3347 * (in which case its reference count is never incremented). 3348 */ 3349 void swap_shmem_alloc(swp_entry_t entry) 3350 { 3351 __swap_duplicate(entry, SWAP_MAP_SHMEM); 3352 } 3353 3354 /* 3355 * Increase reference count of swap entry by 1. 3356 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required 3357 * but could not be atomically allocated. Returns 0, just as if it succeeded, 3358 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which 3359 * might occur if a page table entry has got corrupted. 3360 */ 3361 int swap_duplicate(swp_entry_t entry) 3362 { 3363 int err = 0; 3364 3365 while (!err && __swap_duplicate(entry, 1) == -ENOMEM) 3366 err = add_swap_count_continuation(entry, GFP_ATOMIC); 3367 return err; 3368 } 3369 3370 /* 3371 * @entry: swap entry for which we allocate swap cache. 3372 * 3373 * Called when allocating swap cache for existing swap entry, 3374 * This can return error codes. Returns 0 at success. 3375 * -EEXIST means there is a swap cache. 3376 * Note: return code is different from swap_duplicate(). 3377 */ 3378 int swapcache_prepare(swp_entry_t entry) 3379 { 3380 return __swap_duplicate(entry, SWAP_HAS_CACHE); 3381 } 3382 3383 struct swap_info_struct *swp_swap_info(swp_entry_t entry) 3384 { 3385 return swap_type_to_swap_info(swp_type(entry)); 3386 } 3387 3388 struct swap_info_struct *page_swap_info(struct page *page) 3389 { 3390 swp_entry_t entry = { .val = page_private(page) }; 3391 return swp_swap_info(entry); 3392 } 3393 3394 /* 3395 * out-of-line methods to avoid include hell. 3396 */ 3397 struct address_space *swapcache_mapping(struct folio *folio) 3398 { 3399 return page_swap_info(&folio->page)->swap_file->f_mapping; 3400 } 3401 EXPORT_SYMBOL_GPL(swapcache_mapping); 3402 3403 pgoff_t __page_file_index(struct page *page) 3404 { 3405 swp_entry_t swap = { .val = page_private(page) }; 3406 return swp_offset(swap); 3407 } 3408 EXPORT_SYMBOL_GPL(__page_file_index); 3409 3410 /* 3411 * add_swap_count_continuation - called when a swap count is duplicated 3412 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's 3413 * page of the original vmalloc'ed swap_map, to hold the continuation count 3414 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called 3415 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc. 3416 * 3417 * These continuation pages are seldom referenced: the common paths all work 3418 * on the original swap_map, only referring to a continuation page when the 3419 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. 3420 * 3421 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding 3422 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL) 3423 * can be called after dropping locks. 3424 */ 3425 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask) 3426 { 3427 struct swap_info_struct *si; 3428 struct swap_cluster_info *ci; 3429 struct page *head; 3430 struct page *page; 3431 struct page *list_page; 3432 pgoff_t offset; 3433 unsigned char count; 3434 int ret = 0; 3435 3436 /* 3437 * When debugging, it's easier to use __GFP_ZERO here; but it's better 3438 * for latency not to zero a page while GFP_ATOMIC and holding locks. 3439 */ 3440 page = alloc_page(gfp_mask | __GFP_HIGHMEM); 3441 3442 si = get_swap_device(entry); 3443 if (!si) { 3444 /* 3445 * An acceptable race has occurred since the failing 3446 * __swap_duplicate(): the swap device may be swapoff 3447 */ 3448 goto outer; 3449 } 3450 spin_lock(&si->lock); 3451 3452 offset = swp_offset(entry); 3453 3454 ci = lock_cluster(si, offset); 3455 3456 count = swap_count(si->swap_map[offset]); 3457 3458 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) { 3459 /* 3460 * The higher the swap count, the more likely it is that tasks 3461 * will race to add swap count continuation: we need to avoid 3462 * over-provisioning. 3463 */ 3464 goto out; 3465 } 3466 3467 if (!page) { 3468 ret = -ENOMEM; 3469 goto out; 3470 } 3471 3472 /* 3473 * We are fortunate that although vmalloc_to_page uses pte_offset_map, 3474 * no architecture is using highmem pages for kernel page tables: so it 3475 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps. 3476 */ 3477 head = vmalloc_to_page(si->swap_map + offset); 3478 offset &= ~PAGE_MASK; 3479 3480 spin_lock(&si->cont_lock); 3481 /* 3482 * Page allocation does not initialize the page's lru field, 3483 * but it does always reset its private field. 3484 */ 3485 if (!page_private(head)) { 3486 BUG_ON(count & COUNT_CONTINUED); 3487 INIT_LIST_HEAD(&head->lru); 3488 set_page_private(head, SWP_CONTINUED); 3489 si->flags |= SWP_CONTINUED; 3490 } 3491 3492 list_for_each_entry(list_page, &head->lru, lru) { 3493 unsigned char *map; 3494 3495 /* 3496 * If the previous map said no continuation, but we've found 3497 * a continuation page, free our allocation and use this one. 3498 */ 3499 if (!(count & COUNT_CONTINUED)) 3500 goto out_unlock_cont; 3501 3502 map = kmap_atomic(list_page) + offset; 3503 count = *map; 3504 kunmap_atomic(map); 3505 3506 /* 3507 * If this continuation count now has some space in it, 3508 * free our allocation and use this one. 3509 */ 3510 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX) 3511 goto out_unlock_cont; 3512 } 3513 3514 list_add_tail(&page->lru, &head->lru); 3515 page = NULL; /* now it's attached, don't free it */ 3516 out_unlock_cont: 3517 spin_unlock(&si->cont_lock); 3518 out: 3519 unlock_cluster(ci); 3520 spin_unlock(&si->lock); 3521 put_swap_device(si); 3522 outer: 3523 if (page) 3524 __free_page(page); 3525 return ret; 3526 } 3527 3528 /* 3529 * swap_count_continued - when the original swap_map count is incremented 3530 * from SWAP_MAP_MAX, check if there is already a continuation page to carry 3531 * into, carry if so, or else fail until a new continuation page is allocated; 3532 * when the original swap_map count is decremented from 0 with continuation, 3533 * borrow from the continuation and report whether it still holds more. 3534 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster 3535 * lock. 3536 */ 3537 static bool swap_count_continued(struct swap_info_struct *si, 3538 pgoff_t offset, unsigned char count) 3539 { 3540 struct page *head; 3541 struct page *page; 3542 unsigned char *map; 3543 bool ret; 3544 3545 head = vmalloc_to_page(si->swap_map + offset); 3546 if (page_private(head) != SWP_CONTINUED) { 3547 BUG_ON(count & COUNT_CONTINUED); 3548 return false; /* need to add count continuation */ 3549 } 3550 3551 spin_lock(&si->cont_lock); 3552 offset &= ~PAGE_MASK; 3553 page = list_next_entry(head, lru); 3554 map = kmap_atomic(page) + offset; 3555 3556 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */ 3557 goto init_map; /* jump over SWAP_CONT_MAX checks */ 3558 3559 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */ 3560 /* 3561 * Think of how you add 1 to 999 3562 */ 3563 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) { 3564 kunmap_atomic(map); 3565 page = list_next_entry(page, lru); 3566 BUG_ON(page == head); 3567 map = kmap_atomic(page) + offset; 3568 } 3569 if (*map == SWAP_CONT_MAX) { 3570 kunmap_atomic(map); 3571 page = list_next_entry(page, lru); 3572 if (page == head) { 3573 ret = false; /* add count continuation */ 3574 goto out; 3575 } 3576 map = kmap_atomic(page) + offset; 3577 init_map: *map = 0; /* we didn't zero the page */ 3578 } 3579 *map += 1; 3580 kunmap_atomic(map); 3581 while ((page = list_prev_entry(page, lru)) != head) { 3582 map = kmap_atomic(page) + offset; 3583 *map = COUNT_CONTINUED; 3584 kunmap_atomic(map); 3585 } 3586 ret = true; /* incremented */ 3587 3588 } else { /* decrementing */ 3589 /* 3590 * Think of how you subtract 1 from 1000 3591 */ 3592 BUG_ON(count != COUNT_CONTINUED); 3593 while (*map == COUNT_CONTINUED) { 3594 kunmap_atomic(map); 3595 page = list_next_entry(page, lru); 3596 BUG_ON(page == head); 3597 map = kmap_atomic(page) + offset; 3598 } 3599 BUG_ON(*map == 0); 3600 *map -= 1; 3601 if (*map == 0) 3602 count = 0; 3603 kunmap_atomic(map); 3604 while ((page = list_prev_entry(page, lru)) != head) { 3605 map = kmap_atomic(page) + offset; 3606 *map = SWAP_CONT_MAX | count; 3607 count = COUNT_CONTINUED; 3608 kunmap_atomic(map); 3609 } 3610 ret = count == COUNT_CONTINUED; 3611 } 3612 out: 3613 spin_unlock(&si->cont_lock); 3614 return ret; 3615 } 3616 3617 /* 3618 * free_swap_count_continuations - swapoff free all the continuation pages 3619 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it. 3620 */ 3621 static void free_swap_count_continuations(struct swap_info_struct *si) 3622 { 3623 pgoff_t offset; 3624 3625 for (offset = 0; offset < si->max; offset += PAGE_SIZE) { 3626 struct page *head; 3627 head = vmalloc_to_page(si->swap_map + offset); 3628 if (page_private(head)) { 3629 struct page *page, *next; 3630 3631 list_for_each_entry_safe(page, next, &head->lru, lru) { 3632 list_del(&page->lru); 3633 __free_page(page); 3634 } 3635 } 3636 } 3637 } 3638 3639 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP) 3640 void __cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask) 3641 { 3642 struct swap_info_struct *si, *next; 3643 int nid = page_to_nid(page); 3644 3645 if (!(gfp_mask & __GFP_IO)) 3646 return; 3647 3648 if (!blk_cgroup_congested()) 3649 return; 3650 3651 /* 3652 * We've already scheduled a throttle, avoid taking the global swap 3653 * lock. 3654 */ 3655 if (current->throttle_queue) 3656 return; 3657 3658 spin_lock(&swap_avail_lock); 3659 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid], 3660 avail_lists[nid]) { 3661 if (si->bdev) { 3662 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true); 3663 break; 3664 } 3665 } 3666 spin_unlock(&swap_avail_lock); 3667 } 3668 #endif 3669 3670 static int __init swapfile_init(void) 3671 { 3672 int nid; 3673 3674 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head), 3675 GFP_KERNEL); 3676 if (!swap_avail_heads) { 3677 pr_emerg("Not enough memory for swap heads, swap is disabled\n"); 3678 return -ENOMEM; 3679 } 3680 3681 for_each_node(nid) 3682 plist_head_init(&swap_avail_heads[nid]); 3683 3684 swapfile_maximum_size = arch_max_swapfile_size(); 3685 3686 #ifdef CONFIG_MIGRATION 3687 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS)) 3688 swap_migration_ad_supported = true; 3689 #endif /* CONFIG_MIGRATION */ 3690 3691 return 0; 3692 } 3693 subsys_initcall(swapfile_init); 3694