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