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