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